init

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/**/* filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,29 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+assets/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST

README.md

@@ -1,12 +1,14 @@
 ---
-title: Teset Demo
-emoji: 🖼
-colorFrom: purple
-colorTo: red
+title: ByteMorph Demo
+emoji: 📊
+colorFrom: blue
+colorTo: purple
 sdk: gradio
-sdk_version: 5.25.2
+sdk_version: 5.31.0
 app_file: app.py
 pinned: false
+license: other
+short_description: Online Demo for ByteMorph
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -1,154 +1,173 @@
 import gradio as gr
-import numpy as np
-import random
-
-# import spaces #[uncomment to use ZeroGPU]
-from diffusers import DiffusionPipeline
 import torch
-
-device = "cuda" if torch.cuda.is_available() else "cpu"
-model_repo_id = "stabilityai/sdxl-turbo"  # Replace to the model you would like to use
-
-if torch.cuda.is_available():
-    torch_dtype = torch.float16
-else:
-    torch_dtype = torch.float32
-
-pipe = DiffusionPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)
-pipe = pipe.to(device)
-
-MAX_SEED = np.iinfo(np.int32).max
-MAX_IMAGE_SIZE = 1024
-
-
-# @spaces.GPU #[uncomment to use ZeroGPU]
-def infer(
-    prompt,
-    negative_prompt,
-    seed,
-    randomize_seed,
-    width,
-    height,
-    guidance_scale,
-    num_inference_steps,
-    progress=gr.Progress(track_tqdm=True),
-):
-    if randomize_seed:
-        seed = random.randint(0, MAX_SEED)
-
-    generator = torch.Generator().manual_seed(seed)
-
-    image = pipe(
-        prompt=prompt,
-        negative_prompt=negative_prompt,
-        guidance_scale=guidance_scale,
-        num_inference_steps=num_inference_steps,
-        width=width,
-        height=height,
-        generator=generator,
-    ).images[0]
-
-    return image, seed
-
-
-examples = [
-    "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
-    "An astronaut riding a green horse",
-    "A delicious ceviche cheesecake slice",
-]
-
-css = """
-#col-container {
-    margin: 0 auto;
-    max-width: 640px;
-}
+import spaces
+import os
+import numpy as np
+from PIL import Image
+
+from huggingface_hub import hf_hub_download
+from safetensors.torch import load_file
+from omegaconf import OmegaConf
+from src.flux.util import load_ae, load_clip, load_flow_model2, load_t5, tensor_to_pil_image
+from src.flux.xflux_pipeline import XFluxSampler
+from image_datasets.dataset import image_resize
+
+# ===== No CUDA/model initialization globally =====
+args = OmegaConf.load("inference_configs/inference.yaml")
+is_schnell = args.model_name == "flux-schnell"
+
+# sampler = None
+device = torch.device("cuda")
+dtype = torch.bfloat16
+dit = load_flow_model2(args.model_name, device="cpu").to(device, dtype=dtype)
+vae = load_ae(args.model_name, device="cpu").to(device, dtype=dtype)
+t5 = load_t5(device="cpu", max_length=256 if is_schnell else 512).to(device, dtype=dtype)
+clip = load_clip("cpu").to(device, dtype=dtype)
+
+vae.requires_grad_(False)
+t5.requires_grad_(False)
+clip.requires_grad_(False)
+
+model_path = hf_hub_download(
+    repo_id="Boese0601/ByteMorpher",
+    filename="dit.safetensors",
+    use_auth_token=os.getenv("HF_TOKEN")
+)
+state_dict = load_file(model_path)
+dit.load_state_dict(state_dict)
+dit.eval()
+dit.to(device, dtype=dtype)
+
+sampler = XFluxSampler(
+    clip=clip,
+    t5=t5,
+    ae=vae,
+    model=dit,
+    device=device,
+    ip_loaded=False,
+    spatial_condition=False,
+    clip_image_processor=None,
+    image_encoder=None,
+    improj=None
+)
+#test push
+@spaces.GPU
+def generate(image: Image.Image, edit_prompt: str):
+    # global sampler
+    # device = torch.device("cuda")
+    # dtype = torch.bfloat16
+
+    # if sampler is None:
+        # dit = load_flow_model2(args.model_name, device="cpu").to(device, dtype=dtype)
+        # vae = load_ae(args.model_name, device="cpu").to(device, dtype=dtype)
+        # t5 = load_t5(device="cpu", max_length=256 if is_schnell else 512).to(device, dtype=dtype)
+        # clip = load_clip("cpu").to(device, dtype=dtype)
+
+        # vae.requires_grad_(False)
+        # t5.requires_grad_(False)
+        # clip.requires_grad_(False)
+
+        # model_path = hf_hub_download(
+        #     repo_id="Boese0601/ByteMorpher",
+        #     filename="dit.safetensors",
+        #     use_auth_token=os.getenv("HF_TOKEN")
+        # )
+        # state_dict = load_file(model_path)
+        # dit.load_state_dict(state_dict)
+        # dit.eval()
+
+        # sampler = XFluxSampler(
+        #     clip=clip,
+        #     t5=t5,
+        #     ae=vae,
+        #     model=dit,
+        #     device=device,
+        #     ip_loaded=False,
+        #     spatial_condition=False,
+        #     clip_image_processor=None,
+        #     image_encoder=None,
+        #     improj=None
+        # )
+
+    img = image_resize(image, 512)
+    w, h = img.size
+    img = img.resize(((w // 32) * 32, (h // 32) * 32))
+    img = torch.from_numpy((np.array(img) / 127.5) - 1)
+    img = img.permute(2, 0, 1).unsqueeze(0).to(device, dtype=dtype)
+
+    with torch.no_grad():
+        result = sampler(
+            prompt=edit_prompt,
+            width=args.sample_width,
+            height=args.sample_height,
+            num_steps=args.sample_steps,
+            image_prompt=None,
+            true_gs=args.cfg_scale,
+            seed=args.seed,
+            ip_scale=args.ip_scale if args.use_ip else 1.0,
+            source_image=img if args.use_spatial_condition else None,
+        )
+    return tensor_to_pil_image(result)
+
+def get_samples():
+    sample_list = [
+        {
+            "image": "assets/0_camera_zoom/20486354.png",
+            "edit_prompt": "Zoom in on the coral and add a small blue fish in the background.",
+        },
+    ]
+    return [
+        [
+            Image.open(sample["image"]).resize((512, 512)),
+            sample["edit_prompt"],
+        ]
+        for sample in sample_list
+    ]
+
+header = """
+# ByteMorph
+
+<div style="text-align: center; display: flex; justify-content: left; gap: 5px;">
+<a href=""><img src="https://img.shields.io/badge/ariXv-Paper-A42C25.svg" alt="arXiv"></a>
+<a href="https://huggingface.co/datasets/Boese0601/ByteMorph-Bench"><img src="https://img.shields.io/badge/🤗-Model-ffbd45.svg" alt="HuggingFace"></a>
+<a href="https://github.com/Boese0601/ByteMorph"><img src="https://img.shields.io/badge/GitHub-Code-blue.svg?logo=github&" alt="GitHub"></a>
+</div>
 """
 
-with gr.Blocks(css=css) as demo:
-    with gr.Column(elem_id="col-container"):
-        gr.Markdown(" # Text-to-Image Gradio Template")
-
-        with gr.Row():
-            prompt = gr.Text(
-                label="Prompt",
-                show_label=False,
-                max_lines=1,
-                placeholder="Enter your prompt",
-                container=False,
-            )
-
-            run_button = gr.Button("Run", scale=0, variant="primary")
-
-        result = gr.Image(label="Result", show_label=False)
-
-        with gr.Accordion("Advanced Settings", open=False):
-            negative_prompt = gr.Text(
-                label="Negative prompt",
-                max_lines=1,
-                placeholder="Enter a negative prompt",
-                visible=False,
-            )
-
-            seed = gr.Slider(
-                label="Seed",
-                minimum=0,
-                maximum=MAX_SEED,
-                step=1,
-                value=0,
-            )
-
-            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
-
-            with gr.Row():
-                width = gr.Slider(
-                    label="Width",
-                    minimum=256,
-                    maximum=MAX_IMAGE_SIZE,
-                    step=32,
-                    value=1024,  # Replace with defaults that work for your model
-                )
-
-                height = gr.Slider(
-                    label="Height",
-                    minimum=256,
-                    maximum=MAX_IMAGE_SIZE,
-                    step=32,
-                    value=1024,  # Replace with defaults that work for your model
+def create_app():
+    with gr.Blocks() as app:
+        gr.Markdown(header, elem_id="header")
+        with gr.Row(equal_height=False):
+            with gr.Column(variant="panel", elem_classes="inputPanel"):
+                original_image = gr.Image(
+                    type="pil", label="Condition Image", width=300, elem_id="input"
                 )
+                edit_prompt = gr.Textbox(lines=2, label="Edit Prompt", elem_id="edit_prompt")
+                submit_btn = gr.Button("Run", elem_id="submit_btn")
 
-            with gr.Row():
-                guidance_scale = gr.Slider(
-                    label="Guidance scale",
-                    minimum=0.0,
-                    maximum=10.0,
-                    step=0.1,
-                    value=0.0,  # Replace with defaults that work for your model
-                )
+            with gr.Column(variant="panel", elem_classes="outputPanel"):
+                output_image = gr.Image(type="pil", elem_id="output")
 
-                num_inference_steps = gr.Slider(
-                    label="Number of inference steps",
-                    minimum=1,
-                    maximum=50,
-                    step=1,
-                    value=2,  # Replace with defaults that work for your model
-                )
+        with gr.Row():
+            examples = gr.Examples(
+                examples=get_samples(),
+                inputs=[original_image, edit_prompt],
+                label="Examples",
+            )
 
-        gr.Examples(examples=examples, inputs=[prompt])
-    gr.on(
-        triggers=[run_button.click, prompt.submit],
-        fn=infer,
-        inputs=[
-            prompt,
-            negative_prompt,
-            seed,
-            randomize_seed,
-            width,
-            height,
-            guidance_scale,
-            num_inference_steps,
-        ],
-        outputs=[result, seed],
-    )
+        submit_btn.click(
+            fn=generate,
+            inputs=[original_image, edit_prompt],
+            outputs=output_image,
+        )
+        gr.HTML(
+            """
+            <div style="text-align: center;">
+                * This demo's template was modified from <a href="https://arxiv.org/abs/2411.15098" target="_blank">OminiControl</a>.
+            </div>
+            """
+        )
+    return app
 
 if __name__ == "__main__":
-    demo.launch()
+    create_app().launch(debug=False, share=False, ssr_mode=False)

app_old.py

@@ -0,0 +1,374 @@
+import gradio as gr
+import torch
+import spaces
+from PIL import Image, ImageDraw, ImageFont
+# from src.condition import Condition
+from diffusers.pipelines import FluxPipeline
+import numpy as np
+import requests
+from huggingface_hub import hf_hub_download
+from safetensors.torch import load_file
+import torch.multiprocessing as mp
+###
+import argparse
+import logging
+import math
+import os
+import re
+import random
+import shutil
+from contextlib import nullcontext
+from pathlib import Path
+from PIL import Image
+import accelerate
+import datasets
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import Tensor, nn
+import torch.utils.checkpoint
+import transformers
+from accelerate import Accelerator
+from accelerate.logging import get_logger
+from accelerate.state import AcceleratorState
+from accelerate.utils import ProjectConfiguration, set_seed
+from huggingface_hub import create_repo, upload_folder
+from packaging import version
+from tqdm.auto import tqdm
+from transformers import CLIPTextModel, CLIPTokenizer
+from transformers import CLIPVisionModelWithProjection, CLIPImageProcessor
+from transformers.utils import ContextManagers
+from omegaconf import OmegaConf
+from copy import deepcopy
+import diffusers
+from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionPipeline
+from diffusers.optimization import get_scheduler
+from diffusers.training_utils import EMAModel, compute_dream_and_update_latents, compute_snr
+from diffusers.utils import check_min_version, deprecate, make_image_grid
+from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
+from diffusers.utils.import_utils import is_xformers_available
+from diffusers.utils.torch_utils import is_compiled_module
+from einops import rearrange
+from src.flux.sampling import denoise, get_noise, get_schedule, prepare, unpack
+from src.flux.util import (configs, load_ae, load_clip,
+                       load_flow_model2, load_t5, save_image, tensor_to_pil_image, load_checkpoint)
+from src.flux.modules.layers import DoubleStreamBlockLoraProcessor, SingleStreamBlockLoraProcessor, IPDoubleStreamBlockProcessor, IPSingleStreamBlockProcessor, ImageProjModel
+from src.flux.xflux_pipeline import XFluxSampler
+
+from image_datasets.dataset import loader, eval_image_pair_loader, image_resize
+
+from safetensors.torch import load_file
+import json
+
+
+# logger = get_logger(__name__, log_level="INFO")
+
+
+def get_models(name: str, device, offload: bool, is_schnell: bool):
+    t5 = load_t5(device, max_length=256 if is_schnell else 512)
+    clip = load_clip(device)
+    clip.requires_grad_(False)
+    model = load_flow_model2(name, device="cpu")
+    vae = load_ae(name, device="cpu" if offload else device)
+    return model, vae, t5, clip
+
+args = OmegaConf.load("inference_configs/inference.yaml") #OmegaConf.load(parse_args())
+is_schnell = args.model_name == "flux-schnell"
+set_seed(args.seed)
+# logging_dir = os.path.join(args.output_dir, args.logging_dir)
+device = "cuda"
+dit, vae, t5, clip = get_models(name=args.model_name, device=device, offload=False, is_schnell=is_schnell)
+
+# # load image encoder
+# ip_image_encoder = CLIPVisionModelWithProjection.from_pretrained(os.getenv("CLIP_VIT")).to(
+#     # accelerator.device, dtype=torch.bfloat16
+#     device, dtype=torch.bfloat16
+# )
+# ip_clip_image_processor = CLIPImageProcessor()
+
+if args.use_ip:
+    sampler = XFluxSampler(clip=clip, t5=t5, ae=vae, model=dit, device=device, ip_loaded=True,  spatial_condition=False, clip_image_processor=ip_clip_image_processor, image_encoder=ip_image_encoder, improj=ip_improj)
+elif args.use_spatial_condition:
+    sampler = XFluxSampler(clip=clip, t5=t5, ae=vae, model=dit, device=device, ip_loaded=False,  spatial_condition=True, clip_image_processor=None, image_encoder=None, improj=None,share_position_embedding=args.share_position_embedding)
+else:
+    sampler = XFluxSampler(clip=clip, t5=t5, ae=vae, model=dit, device=device, ip_loaded=False,  spatial_condition=False, clip_image_processor=None, image_encoder=None, improj=None)
+
+
+# @spaces.GPU
+def generate(image, edit_prompt):
+    print("hello?????????!!!!!")
+
+    # accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
+
+    # accelerator = Accelerator(
+    #     gradient_accumulation_steps=1,
+    #     mixed_precision=args.mixed_precision,
+    #     log_with=args.report_to,
+    #     project_config=accelerator_project_config,
+    # )
+
+    # Make one log on every process with the configuration for debugging.
+    # logging.basicConfig(
+    #     format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+    #     datefmt="%m/%d/%Y %H:%M:%S",
+    #     level=logging.INFO,
+    # )
+    # logger.info(accelerator.state, main_process_only=False)
+    # if accelerator.is_local_main_process:
+    #     datasets.utils.logging.set_verbosity_warning()
+    #     transformers.utils.logging.set_verbosity_warning()
+    #     diffusers.utils.logging.set_verbosity_info()
+    # else:
+    #     datasets.utils.logging.set_verbosity_error()
+    #     transformers.utils.logging.set_verbosity_error()
+    #     diffusers.utils.logging.set_verbosity_error()
+
+
+    # if accelerator.is_main_process:
+    #     if args.output_dir is not None:
+    #         os.makedirs(args.output_dir, exist_ok=True)
+    #         gpt_eval_path = os.path.join(args.output_dir,"Eval")
+    #         os.makedirs(gpt_eval_path, exist_ok=True)
+
+    # dit, vae, t5, clip = get_models(name=args.model_name, device=accelerator.device, offload=False, is_schnell=is_schnell)
+    # dit, vae, t5, clip = get_models(name=args.model_name, device=device, offload=False, is_schnell=is_schnell)
+    
+    if args.use_lora:
+        lora_attn_procs = {}
+    if args.use_ip:
+        ip_attn_procs = {}
+    if args.double_blocks is None:
+        double_blocks_idx = list(range(19))
+    else:
+        double_blocks_idx = [int(idx) for idx in args.double_blocks.split(",")]
+
+    if args.single_blocks is None:
+        single_blocks_idx = list(range(38))
+    elif args.single_blocks is not None:
+        single_blocks_idx = [int(idx) for idx in args.single_blocks.split(",")]
+
+    if args.use_lora:
+        for name, attn_processor in dit.attn_processors.items():
+            match = re.search(r'\.(\d+)\.', name)
+            if match:
+                layer_index = int(match.group(1))
+
+            if name.startswith("double_blocks") and layer_index in double_blocks_idx:
+                # if accelerator.is_main_process:
+                #     print("setting LoRA Processor for", name)
+                lora_attn_procs[name] = DoubleStreamBlockLoraProcessor(
+                dim=3072, rank=args.rank
+                )
+            elif name.startswith("single_blocks") and layer_index in single_blocks_idx:
+                # if accelerator.is_main_process:
+                #     print("setting LoRA Processor for", name)
+                lora_attn_procs[name] = SingleStreamBlockLoraProcessor(
+                dim=3072, rank=args.rank
+                )
+            else:
+                lora_attn_procs[name] = attn_processor
+
+        dit.set_attn_processor(lora_attn_procs)
+    
+    # if args.use_ip:
+    #     # unpack checkpoint
+    #     checkpoint = load_checkpoint(args.ip_local_path, args.ip_repo_id, args.ip_name)
+    #     prefix = "double_blocks."
+    #     # blocks = {}
+    #     proj = {}
+
+    #     for key, value in checkpoint.items():
+    #         # if key.startswith(prefix):
+    #         #     blocks[key[len(prefix):].replace('.processor.', '.')] = value
+    #         if key.startswith("ip_adapter_proj_model"):
+    #             proj[key[len("ip_adapter_proj_model."):]] = value
+
+    #     # # load image encoder
+    #     # ip_image_encoder = CLIPVisionModelWithProjection.from_pretrained(os.getenv("CLIP_VIT")).to(
+    #     #     # accelerator.device, dtype=torch.bfloat16
+    #     #     device, dtype=torch.bfloat16
+    #     # )
+    #     # ip_clip_image_processor = CLIPImageProcessor()
+
+    #     # setup image embedding projection model
+    #     ip_improj = ImageProjModel(4096, 768, 4)
+    #     ip_improj.load_state_dict(proj)
+    #     # ip_improj = ip_improj.to(accelerator.device, dtype=torch.bfloat16)
+    #     ip_improj = ip_improj.to(device, dtype=torch.bfloat16)
+
+    #     ip_attn_procs = {}
+
+    #     for name, _ in dit.attn_processors.items():
+    #         ip_state_dict = {}
+    #         for k in checkpoint.keys():
+    #             if name in k:
+    #                 ip_state_dict[k.replace(f'{name}.', '')] = checkpoint[k]
+    #         if ip_state_dict:
+    #             ip_attn_procs[name] = IPDoubleStreamBlockProcessor(4096, 3072)
+    #             ip_attn_procs[name].load_state_dict(ip_state_dict)
+    #             ip_attn_procs[name].to(accelerator.device, dtype=torch.bfloat16)
+    #         else:
+    #             ip_attn_procs[name] = dit.attn_processors[name]
+    #     dit.set_attn_processor(ip_attn_procs)
+
+
+    vae.requires_grad_(False)
+    t5.requires_grad_(False)
+    clip.requires_grad_(False)
+
+
+
+    # weight_dtype = torch.float32
+    # if accelerator.mixed_precision == "fp16":
+    #     weight_dtype = torch.float16
+    #     args.mixed_precision = accelerator.mixed_precision
+    # elif accelerator.mixed_precision == "bf16":
+    #     weight_dtype = torch.bfloat16
+    #     args.mixed_precision = accelerator.mixed_precision
+
+
+    # print(f"Resuming from checkpoint {args.ckpt_dir}")
+    # dit_stat_dict = load_file(args.ckpt_dir)
+    # Get path from Hub
+    model_path = hf_hub_download(
+        repo_id="Boese0601/ByteMorpher",
+        filename="dit.safetensors"
+    )
+    state_dict = load_file(model_path)
+    dit.load_state_dict(state_dict)
+    dit = dit.to(weight_dtype)
+    dit.eval()
+
+    # test_dataloader = loader(**args.data_config)
+    test_dataloader = eval_image_pair_loader(**args.data_config)
+
+
+
+    # from deepspeed import initialize
+    dit = accelerator.prepare(dit)
+
+    # if accelerator.is_main_process:
+    #     accelerator.init_trackers(args.tracker_project_name, {"test": None})
+
+    # logger.info("***** Running Evaluation *****")
+    # logger.info(f"  Instantaneous batch size = {args.eval_batch_size}")
+
+
+
+    # progress_bar = tqdm(
+    #     range(0, len(test_dataloader)),
+    #     initial=0,
+    #     desc="Steps",
+    #     disable=not accelerator.is_local_main_process,
+    # )
+
+    # for step, batch in enumerate(test_dataloader):
+    # with accelerator.accumulate(dit):
+    # img, tgt_image, prompt, edit_prompt, img_name, edit_name = batch
+    img = image_resize(image, 512)
+    w, h = img.size
+    new_w = (w // 32) * 32
+    new_h = (h // 32) * 32
+    img = img.resize((new_w, new_h))
+    img = torch.from_numpy((np.array(img) / 127.5) - 1)
+    img = img.permute(2, 0, 1).unsqueeze(0)
+
+    edit_prompt = edit_prompt
+
+    # if args.use_ip:
+    #     sampler = XFluxSampler(clip=clip, t5=t5, ae=vae, model=dit, device=accelerator.device, ip_loaded=True,  spatial_condition=False, clip_image_processor=ip_clip_image_processor, image_encoder=ip_image_encoder, improj=ip_improj)
+    # elif args.use_spatial_condition:
+    #     sampler = XFluxSampler(clip=clip, t5=t5, ae=vae, model=dit, device=accelerator.device, ip_loaded=False,  spatial_condition=True, clip_image_processor=None, image_encoder=None, improj=None,share_position_embedding=args.share_position_embedding)
+    # else:
+    #     sampler = XFluxSampler(clip=clip, t5=t5, ae=vae, model=dit, device=accelerator.device, ip_loaded=False,  spatial_condition=False, clip_image_processor=None, image_encoder=None, improj=None)
+    with torch.no_grad():
+        result = sampler(prompt=edit_prompt,
+                            width=args.sample_width,
+                            height=args.sample_height,
+                            num_steps=args.sample_steps,
+                            image_prompt=None, # ip_adapter
+                            true_gs=args.cfg_scale,
+                            seed=args.seed,
+                            ip_scale=args.ip_scale if args.use_ip else 1.0,
+                            source_image=img if args.use_spatial_condition else None,
+                            )
+    gen_img = result
+
+
+
+    # progress_bar.update(1)
+
+    # accelerator.wait_for_everyone()
+    # accelerator.end_training()
+    return gen_img
+
+
+def get_samples():
+    sample_list = [
+        {
+            "image": "assets/0_camera_zoom/20486354.png",
+            "edit_prompt": "Zoom in on the coral and add a small blue fish in the background.",
+        },
+    ]
+    return [
+        [
+            Image.open(sample["image"]).resize((512, 512)),
+            sample["edit_prompt"],
+        ]
+        for sample in sample_list
+    ]
+
+
+header = """
+# ByteMoprh
+
+<div style="text-align: center; display: flex; justify-content: left; gap: 5px;">
+<a href=""><img src="https://img.shields.io/badge/ariXv-Paper-A42C25.svg" alt="arXiv"></a>
+<a href="https://huggingface.co/datasets/Boese0601/ByteMorph-Bench"><img src="https://img.shields.io/badge/🤗-Model-ffbd45.svg" alt="HuggingFace"></a>
+<a href="https://github.com/Boese0601/ByteMorph"><img src="https://img.shields.io/badge/GitHub-Code-blue.svg?logo=github&" alt="GitHub"></a>
+</div>
+"""
+
+
+def create_app():
+    with gr.Blocks() as app:
+        gr.Markdown(header, elem_id="header")
+        with gr.Row(equal_height=False):
+            with gr.Column(variant="panel", elem_classes="inputPanel"):
+                original_image = gr.Image(
+                    type="pil", label="Condition Image", width=300, elem_id="input"
+                )
+                edit_prompt = gr.Textbox(lines=2, label="Edit Prompt", elem_id="edit_prompt")
+                submit_btn = gr.Button("Run", elem_id="submit_btn")
+
+            with gr.Column(variant="panel", elem_classes="outputPanel"):
+                output_image = gr.Image(type="pil", elem_id="output")
+
+        with gr.Row():
+            examples = gr.Examples(
+                examples=get_samples(),
+                inputs=[original_image, edit_prompt],
+                label="Examples",
+            )
+
+        submit_btn.click(
+            fn=generate,
+            inputs=[original_image, edit_prompt],
+            outputs=output_image,
+        )
+        gr.HTML(
+            """
+            <div style="text-align: center;">
+                * This demo's template was modified from <a href="https://arxiv.org/abs/2411.15098" target="_blank">OminiControl</a>.
+            </div>
+            """
+        ) 
+    return app
+
+
+if __name__ == "__main__":
+    print("CUDA available:", torch.cuda.is_available())
+    print("CUDA version:", torch.version.cuda)
+    print("GPU device name:", torch.cuda.get_device_name(0) if torch.cuda.is_available() else "None")
+    # mp.set_start_method("spawn", force=True)
+    create_app().launch(debug=False, share=True, ssr_mode=False)

image_datasets/dataset.py

@@ -0,0 +1,231 @@
+import os
+import pandas as pd
+import numpy as np
+from PIL import Image
+import torch
+from torch.utils.data import Dataset, DataLoader
+import json
+import random
+import glob
+import torch
+import torchvision.transforms.functional as TF
+
+
+
+def image_resize(img, max_size=512):
+    w, h = img.size
+    if w >= h:
+        new_w = max_size
+        new_h = int((max_size / w) * h)
+    else:
+        new_h = max_size
+        new_w = int((max_size / h) * w)
+    return img.resize((new_w, new_h))
+
+def c_crop(image):
+    width, height = image.size
+    new_size = min(width, height)
+    left = (width - new_size) / 2
+    top = (height - new_size) / 2
+    right = (width + new_size) / 2
+    bottom = (height + new_size) / 2
+    return image.crop((left, top, right, bottom))
+
+def crop_to_aspect_ratio(image, ratio="16:9"):
+    width, height = image.size
+    ratio_map = {
+        "16:9": (16, 9),
+        "4:3": (4, 3),
+        "1:1": (1, 1)
+    }
+    target_w, target_h = ratio_map[ratio]
+    target_ratio_value = target_w / target_h
+
+    current_ratio = width / height
+
+    if current_ratio > target_ratio_value:
+        new_width = int(height * target_ratio_value)
+        offset = (width - new_width) // 2
+        crop_box = (offset, 0, offset + new_width, height)
+    else:
+        new_height = int(width / target_ratio_value)
+        offset = (height - new_height) // 2
+        crop_box = (0, offset, width, offset + new_height)
+
+    cropped_img = image.crop(crop_box)
+    return cropped_img
+
+
+class CustomImageDataset(Dataset):
+    def __init__(self, img_dir, img_size=512, caption_type='json', random_ratio=False):
+        self.images = [os.path.join(img_dir, i) for i in os.listdir(img_dir) if '.jpg' in i or '.png' in i]
+        # self.images = glob.glob(img_dir +'**/*.jpg', recursive=True) + glob.glob(img_dir +'**/*.png', recursive=True) + glob.glob(img_dir +'**/*.jpeg', recursive=True)
+        self.images.sort()
+        self.img_size = img_size
+        self.caption_type = caption_type
+        self.random_ratio = random_ratio
+
+    def __len__(self):
+        return len(self.images)
+
+    def __getitem__(self, idx):
+        try:
+            img = Image.open(self.images[idx]).convert('RGB')
+            
+            if self.random_ratio:
+                ratio = random.choice(["16:9", "default", "1:1", "4:3"])
+                if ratio != "default":
+                    img = crop_to_aspect_ratio(img, ratio)
+            img = image_resize(img, self.img_size)
+            w, h = img.size
+            new_w = (w // 32) * 32
+            new_h = (h // 32) * 32
+            img = img.resize((new_w, new_h))
+            img = torch.from_numpy((np.array(img) / 127.5) - 1)
+            img = img.permute(2, 0, 1)
+            json_path = self.images[idx].split('.')[0] + '.' + self.caption_type
+            if self.caption_type == "json":
+                prompt = json.load(open(json_path))['caption']
+            else:
+                prompt = open(json_path).read()
+            return img, prompt
+        except Exception as e:
+            print(e)
+            return self.__getitem__(random.randint(0, len(self.images) - 1))
+
+
+def loader(train_batch_size, num_workers, **args):
+    dataset = CustomImageDataset(**args)
+    return DataLoader(dataset, batch_size=train_batch_size, num_workers=num_workers, shuffle=True)
+
+
+
+class ImageEditPairDataset(Dataset):
+    def __init__(self, img_dir, img_size=512, caption_type='json', random_ratio=False, grayscale_editing=False, zoom_camera=False):
+        # self.images = [os.path.join(img_dir, i) for i in os.listdir(img_dir) if '.jpg' in i or '.png' in i]
+        self.images = glob.glob(img_dir +'**/*.jpg', recursive=True) + glob.glob(img_dir +'**/*.png', recursive=True) + glob.glob(img_dir +'**/*.jpeg', recursive=True)
+        self.images.sort()
+        self.img_size = img_size
+        self.caption_type = caption_type
+        self.random_ratio = random_ratio
+        self.grayscale_editing = grayscale_editing
+        self.zoom_camera = zoom_camera 
+        if "ByteMorph-Bench" or "InstructMove" in img_dir:
+            self.eval = True
+        else:
+            self.eval = False
+    def __len__(self):
+        return len(self.images)
+
+    def __getitem__(self, idx):
+        try:
+            img = Image.open(self.images[idx]).convert('RGB')
+            ori_width, ori_height = img.size
+            left_half = (0, 0, ori_width // 2, ori_height)
+            right_half = (ori_width // 2, 0, ori_width, ori_height)
+            src_image = img.crop(left_half)  # Left half
+            tgt_image = img.crop(right_half)  # Right half
+            # print("ori_width, ori_height: ",ori_width, ori_height)
+            if self.random_ratio:
+                ratio = random.choice(["16:9", "default", "1:1", "4:3"])
+                if ratio != "default":
+                    src_image = crop_to_aspect_ratio(src_image, ratio)
+                    tgt_image = crop_to_aspect_ratio(tgt_image, ratio)
+            src_image = image_resize(src_image, self.img_size)
+            tgt_image = image_resize(tgt_image, self.img_size)
+            w, h = src_image.size
+            new_w = (w // 32) * 32
+            new_h = (h // 32) * 32
+            # print("new_w, new_h: ",new_w, new_h)
+            src_image = src_image.resize((new_w, new_h))
+            src_image = torch.from_numpy((np.array(src_image) / 127.5) - 1)
+            src_image = src_image.permute(2, 0, 1)
+            tgt_image = tgt_image.resize((new_w, new_h))
+            tgt_image = torch.from_numpy((np.array(tgt_image) / 127.5) - 1)
+            tgt_image = tgt_image.permute(2, 0, 1)
+            json_path = self.images[idx].split('.')[0] + '.' + self.caption_type
+            if self.eval:
+                image_name = self.images[idx].split('.')[0].split("/")[-1]
+                edit_type = self.images[idx].split('.')[0].split("/")[-2]
+            if self.caption_type == "json":
+                if not self.eval:
+                    prompt = json.load(open(json_path))['caption']
+                    edit_prompt = json.load(open(json_path))['edit']
+                else:
+                    prompt = [] #json.load(open(json_path))['caption']
+                    edit_prompt = json.load(open(json_path))['edit']
+            else:
+                raise NotImplementedError
+                # prompt = open(json_path).read()
+            if (not self.grayscale_editing) and (not self.zoom_camera):
+                if not self.eval:
+                    return src_image, tgt_image, prompt, edit_prompt
+                else:
+                    return src_image, tgt_image, prompt, edit_prompt, image_name, edit_type
+            if self.grayscale_editing and (not self.zoom_camera):
+                # Grayscale = 0.2989 * R + 0.5870 * G + 0.1140 * B
+                grayscale_image = 0.2989 * src_image[0, :, :] + 0.5870 * src_image[1, :, :] + 0.1140 * src_image[2, :, :]
+                tgt_image = grayscale_image.unsqueeze(0).repeat(3, 1, 1)
+                edit_prompt = "Convert the input image to a black and white grayscale image while maintaining the original composition and details."
+                if not self.eval:
+                    return src_image, tgt_image, prompt, edit_prompt
+                else:
+                    return src_image, tgt_image, prompt, edit_prompt, image_name, edit_type
+            if (not self.grayscale_editing) and self.zoom_camera:
+                cropped = TF.center_crop(src_image, (256, 256))
+                tgt_image = TF.resize(cropped, (512, 512))
+                edit_prompt = "The central area of the input image is zoomed. The camera transitions from a wide shot to a closer position, narrowing its view."
+                if not self.eval:
+                    return src_image, tgt_image, prompt, edit_prompt
+                else:
+                    return src_image, tgt_image, prompt, edit_prompt, image_name, edit_type
+            if self.grayscale_editing and self.zoom_camera:
+                grayscale_image = 0.2989 * src_image[0, :, :] + 0.5870 * src_image[1, :, :] + 0.1140 * src_image[2, :, :]
+                tgt_image = grayscale_image.unsqueeze(0).repeat(3, 1, 1)
+                tgt_image = TF.center_crop(tgt_image, (256, 256))
+                tgt_image = TF.resize(tgt_image, (512, 512))
+                edit_prompt = "Convert the input image to a black and white grayscale image while maintaining the original composition and details. And the central area of the input image is zoomed, the camera transitions from a wide shot to a closer position, narrowing its view."
+                if not self.eval:
+                    return src_image, tgt_image, prompt, edit_prompt
+                else:
+                    return src_image, tgt_image, prompt, edit_prompt, image_name, edit_type
+        except Exception as e:
+            print(e)
+            return self.__getitem__(random.randint(0, len(self.images) - 1))
+
+
+def image_pair_loader(train_batch_size, num_workers, **args):
+    dataset = ImageEditPairDataset(**args)
+    return DataLoader(dataset, batch_size=train_batch_size, num_workers=num_workers, shuffle=True)
+
+def eval_image_pair_loader(eval_batch_size, num_workers, **args):
+    dataset = ImageEditPairDataset(**args)
+    return DataLoader(dataset, batch_size=eval_batch_size, num_workers=num_workers, shuffle=False)
+
+
+
+if __name__ == "__main__":
+    from src.flux.util import save_image
+    example_dataset = ImageEditPairDataset(
+        img_dir="", 
+        img_size=512, 
+        caption_type='json', 
+        random_ratio=False,
+        grayscale_editing=False,
+        zoom_camera=False,
+    )
+
+    train_dataloader = DataLoader(
+        example_dataset, 
+        batch_size=1, 
+        num_workers=4, 
+        shuffle=False,
+    )
+
+    for step, batch in enumerate(train_dataloader):
+        src_image, tgt_image, prompt, edit_prompt = batch
+        os.makedirs("./debug", exist_ok=True)
+        save_image(src_image, f"./debug/{step}-src_img.jpg")
+        save_image(tgt_image, f"./debug/{step}-tgt_img.jpg")
+        if step == 3:
+            breakpoint()

inference_configs/inference.yaml

@@ -0,0 +1,33 @@
+model_name: "flux-dev"
+use_spatial_condition: true
+share_position_embedding: true
+use_share_weight_referencenet: false
+use_ip: false
+ip_local_path: null
+ip_repo_id: null
+ip_name: null
+ip_scale: 1.0
+use_lora: false
+data_config:
+  eval_batch_size: 1
+  num_workers: 0
+  img_size: 512
+  img_dir: output_bench/ #./ByteMorph-Bench/
+  grayscale_editing: false
+  zoom_camera: false
+  random_ratio: false
+report_to: wandb
+eval_batch_size: 1
+ckpt_dir: ./pretrained_weights/ByteMorpher/dit.safetensors
+output_dir: ./test_log/seedmorpher/
+logging_dir: logs
+mixed_precision: "bf16"
+rank: 16
+single_blocks: null
+double_blocks: null
+disable_sampling: false
+sample_width: 512
+sample_height: 512
+sample_steps: 25
+seed: 42
+cfg_scale: 3.5

requirements.txt

@@ -1,6 +1,26 @@
-accelerate
+# --extra-index-url https://download.pytorch.org/whl/cu124
+# torch==2.6.0
+# torchvision==0.21.0
+# torchaudio==2.6.0
+
+gradio>=4.0
+accelerate==0.30.1
+deepspeed==0.14.4
+einops==0.8.0
+transformers==4.43.3
+huggingface-hub==0.24.5
+optimum-quanto
+datasets
+omegaconf
 diffusers
-invisible_watermark
-torch
-transformers
-xformers
+sentencepiece
+opencv-python
+matplotlib
+onnxruntime
+timm
+wandb
+
+setuptools
+wheel
+
+

src/flux/__init__.py

@@ -0,0 +1,11 @@
+try:
+    from ._version import version as __version__  # type: ignore
+    from ._version import version_tuple
+except ImportError:
+    __version__ = "unknown (no version information available)"
+    version_tuple = (0, 0, "unknown", "noinfo")
+
+from pathlib import Path
+
+PACKAGE = __package__.replace("_", "-")
+PACKAGE_ROOT = Path(__file__).parent

src/flux/__main__.py

@@ -0,0 +1,4 @@
+from .cli import app
+
+if __name__ == "__main__":
+    app()

src/flux/annotator/canny/__init__.py

@@ -0,0 +1,6 @@
+import cv2
+
+
+class CannyDetector:
+    def __call__(self, img, low_threshold, high_threshold):
+        return cv2.Canny(img, low_threshold, high_threshold)

src/flux/annotator/ckpts/ckpts.txt

@@ -0,0 +1 @@
+Weights here.

src/flux/annotator/dwpose/__init__.py

@@ -0,0 +1,68 @@
+# Openpose
+# Original from CMU https://github.com/CMU-Perceptual-Computing-Lab/openpose
+# 2nd Edited by https://github.com/Hzzone/pytorch-openpose
+# 3rd Edited by ControlNet
+# 4th Edited by ControlNet (added face and correct hands)
+
+import os
+os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE"
+
+import torch
+import numpy as np
+from . import util
+from .wholebody import Wholebody
+
+def draw_pose(pose, H, W):
+    bodies = pose['bodies']
+    faces = pose['faces']
+    hands = pose['hands']
+    candidate = bodies['candidate']
+    subset = bodies['subset']
+    canvas = np.zeros(shape=(H, W, 3), dtype=np.uint8)
+
+    canvas = util.draw_bodypose(canvas, candidate, subset)
+
+    canvas = util.draw_handpose(canvas, hands)
+
+    canvas = util.draw_facepose(canvas, faces)
+
+    return canvas
+
+
+class DWposeDetector:
+    def __init__(self, device):
+
+        self.pose_estimation = Wholebody(device)
+
+    def __call__(self, oriImg):
+        oriImg = oriImg.copy()
+        H, W, C = oriImg.shape
+        with torch.no_grad():
+            candidate, subset = self.pose_estimation(oriImg)
+            nums, keys, locs = candidate.shape
+            candidate[..., 0] /= float(W)
+            candidate[..., 1] /= float(H)
+            body = candidate[:,:18].copy()
+            body = body.reshape(nums*18, locs)
+            score = subset[:,:18]
+            for i in range(len(score)):
+                for j in range(len(score[i])):
+                    if score[i][j] > 0.3:
+                        score[i][j] = int(18*i+j)
+                    else:
+                        score[i][j] = -1
+
+            un_visible = subset<0.3
+            candidate[un_visible] = -1
+
+            foot = candidate[:,18:24]
+
+            faces = candidate[:,24:92]
+
+            hands = candidate[:,92:113]
+            hands = np.vstack([hands, candidate[:,113:]])
+
+            bodies = dict(candidate=body, subset=score)
+            pose = dict(bodies=bodies, hands=hands, faces=faces)
+
+            return draw_pose(pose, H, W)

src/flux/annotator/dwpose/onnxdet.py

@@ -0,0 +1,125 @@
+import cv2
+import numpy as np
+
+import onnxruntime
+
+def nms(boxes, scores, nms_thr):
+    """Single class NMS implemented in Numpy."""
+    x1 = boxes[:, 0]
+    y1 = boxes[:, 1]
+    x2 = boxes[:, 2]
+    y2 = boxes[:, 3]
+
+    areas = (x2 - x1 + 1) * (y2 - y1 + 1)
+    order = scores.argsort()[::-1]
+
+    keep = []
+    while order.size > 0:
+        i = order[0]
+        keep.append(i)
+        xx1 = np.maximum(x1[i], x1[order[1:]])
+        yy1 = np.maximum(y1[i], y1[order[1:]])
+        xx2 = np.minimum(x2[i], x2[order[1:]])
+        yy2 = np.minimum(y2[i], y2[order[1:]])
+
+        w = np.maximum(0.0, xx2 - xx1 + 1)
+        h = np.maximum(0.0, yy2 - yy1 + 1)
+        inter = w * h
+        ovr = inter / (areas[i] + areas[order[1:]] - inter)
+
+        inds = np.where(ovr <= nms_thr)[0]
+        order = order[inds + 1]
+
+    return keep
+
+def multiclass_nms(boxes, scores, nms_thr, score_thr):
+    """Multiclass NMS implemented in Numpy. Class-aware version."""
+    final_dets = []
+    num_classes = scores.shape[1]
+    for cls_ind in range(num_classes):
+        cls_scores = scores[:, cls_ind]
+        valid_score_mask = cls_scores > score_thr
+        if valid_score_mask.sum() == 0:
+            continue
+        else:
+            valid_scores = cls_scores[valid_score_mask]
+            valid_boxes = boxes[valid_score_mask]
+            keep = nms(valid_boxes, valid_scores, nms_thr)
+            if len(keep) > 0:
+                cls_inds = np.ones((len(keep), 1)) * cls_ind
+                dets = np.concatenate(
+                    [valid_boxes[keep], valid_scores[keep, None], cls_inds], 1
+                )
+                final_dets.append(dets)
+    if len(final_dets) == 0:
+        return None
+    return np.concatenate(final_dets, 0)
+
+def demo_postprocess(outputs, img_size, p6=False):
+    grids = []
+    expanded_strides = []
+    strides = [8, 16, 32] if not p6 else [8, 16, 32, 64]
+
+    hsizes = [img_size[0] // stride for stride in strides]
+    wsizes = [img_size[1] // stride for stride in strides]
+
+    for hsize, wsize, stride in zip(hsizes, wsizes, strides):
+        xv, yv = np.meshgrid(np.arange(wsize), np.arange(hsize))
+        grid = np.stack((xv, yv), 2).reshape(1, -1, 2)
+        grids.append(grid)
+        shape = grid.shape[:2]
+        expanded_strides.append(np.full((*shape, 1), stride))
+
+    grids = np.concatenate(grids, 1)
+    expanded_strides = np.concatenate(expanded_strides, 1)
+    outputs[..., :2] = (outputs[..., :2] + grids) * expanded_strides
+    outputs[..., 2:4] = np.exp(outputs[..., 2:4]) * expanded_strides
+
+    return outputs
+
+def preprocess(img, input_size, swap=(2, 0, 1)):
+    if len(img.shape) == 3:
+        padded_img = np.ones((input_size[0], input_size[1], 3), dtype=np.uint8) * 114
+    else:
+        padded_img = np.ones(input_size, dtype=np.uint8) * 114
+
+    r = min(input_size[0] / img.shape[0], input_size[1] / img.shape[1])
+    resized_img = cv2.resize(
+        img,
+        (int(img.shape[1] * r), int(img.shape[0] * r)),
+        interpolation=cv2.INTER_LINEAR,
+    ).astype(np.uint8)
+    padded_img[: int(img.shape[0] * r), : int(img.shape[1] * r)] = resized_img
+
+    padded_img = padded_img.transpose(swap)
+    padded_img = np.ascontiguousarray(padded_img, dtype=np.float32)
+    return padded_img, r
+
+def inference_detector(session, oriImg):
+    input_shape = (640,640)
+    img, ratio = preprocess(oriImg, input_shape)
+
+    ort_inputs = {session.get_inputs()[0].name: img[None, :, :, :]}
+    output = session.run(None, ort_inputs)
+    predictions = demo_postprocess(output[0], input_shape)[0]
+
+    boxes = predictions[:, :4]
+    scores = predictions[:, 4:5] * predictions[:, 5:]
+
+    boxes_xyxy = np.ones_like(boxes)
+    boxes_xyxy[:, 0] = boxes[:, 0] - boxes[:, 2]/2.
+    boxes_xyxy[:, 1] = boxes[:, 1] - boxes[:, 3]/2.
+    boxes_xyxy[:, 2] = boxes[:, 0] + boxes[:, 2]/2.
+    boxes_xyxy[:, 3] = boxes[:, 1] + boxes[:, 3]/2.
+    boxes_xyxy /= ratio
+    dets = multiclass_nms(boxes_xyxy, scores, nms_thr=0.45, score_thr=0.1)
+    if dets is not None:
+        final_boxes, final_scores, final_cls_inds = dets[:, :4], dets[:, 4], dets[:, 5]
+        isscore = final_scores>0.3
+        iscat = final_cls_inds == 0
+        isbbox = [ i and j for (i, j) in zip(isscore, iscat)]
+        final_boxes = final_boxes[isbbox]
+    else:
+        final_boxes = np.array([])
+
+    return final_boxes

src/flux/annotator/dwpose/onnxpose.py

@@ -0,0 +1,360 @@
+from typing import List, Tuple
+
+import cv2
+import numpy as np
+import onnxruntime as ort
+
+def preprocess(
+    img: np.ndarray, out_bbox, input_size: Tuple[int, int] = (192, 256)
+) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+    """Do preprocessing for RTMPose model inference.
+
+    Args:
+        img (np.ndarray): Input image in shape.
+        input_size (tuple): Input image size in shape (w, h).
+
+    Returns:
+        tuple:
+        - resized_img (np.ndarray): Preprocessed image.
+        - center (np.ndarray): Center of image.
+        - scale (np.ndarray): Scale of image.
+    """
+    # get shape of image
+    img_shape = img.shape[:2]
+    out_img, out_center, out_scale = [], [], []
+    if len(out_bbox) == 0:
+        out_bbox = [[0, 0, img_shape[1], img_shape[0]]]
+    for i in range(len(out_bbox)):
+        x0 = out_bbox[i][0]
+        y0 = out_bbox[i][1]
+        x1 = out_bbox[i][2]
+        y1 = out_bbox[i][3]
+        bbox = np.array([x0, y0, x1, y1])
+
+        # get center and scale
+        center, scale = bbox_xyxy2cs(bbox, padding=1.25)
+
+        # do affine transformation
+        resized_img, scale = top_down_affine(input_size, scale, center, img)
+
+        # normalize image
+        mean = np.array([123.675, 116.28, 103.53])
+        std = np.array([58.395, 57.12, 57.375])
+        resized_img = (resized_img - mean) / std
+
+        out_img.append(resized_img)
+        out_center.append(center)
+        out_scale.append(scale)
+
+    return out_img, out_center, out_scale
+
+
+def inference(sess: ort.InferenceSession, img: np.ndarray) -> np.ndarray:
+    """Inference RTMPose model.
+
+    Args:
+        sess (ort.InferenceSession): ONNXRuntime session.
+        img (np.ndarray): Input image in shape.
+
+    Returns:
+        outputs (np.ndarray): Output of RTMPose model.
+    """
+    all_out = []
+    # build input
+    for i in range(len(img)):
+        input = [img[i].transpose(2, 0, 1)]
+
+        # build output
+        sess_input = {sess.get_inputs()[0].name: input}
+        sess_output = []
+        for out in sess.get_outputs():
+            sess_output.append(out.name)
+
+        # run model
+        outputs = sess.run(sess_output, sess_input)
+        all_out.append(outputs)
+
+    return all_out
+
+
+def postprocess(outputs: List[np.ndarray],
+                model_input_size: Tuple[int, int],
+                center: Tuple[int, int],
+                scale: Tuple[int, int],
+                simcc_split_ratio: float = 2.0
+                ) -> Tuple[np.ndarray, np.ndarray]:
+    """Postprocess for RTMPose model output.
+
+    Args:
+        outputs (np.ndarray): Output of RTMPose model.
+        model_input_size (tuple): RTMPose model Input image size.
+        center (tuple): Center of bbox in shape (x, y).
+        scale (tuple): Scale of bbox in shape (w, h).
+        simcc_split_ratio (float): Split ratio of simcc.
+
+    Returns:
+        tuple:
+        - keypoints (np.ndarray): Rescaled keypoints.
+        - scores (np.ndarray): Model predict scores.
+    """
+    all_key = []
+    all_score = []
+    for i in range(len(outputs)):
+        # use simcc to decode
+        simcc_x, simcc_y = outputs[i]
+        keypoints, scores = decode(simcc_x, simcc_y, simcc_split_ratio)
+
+        # rescale keypoints
+        keypoints = keypoints / model_input_size * scale[i] + center[i] - scale[i] / 2
+        all_key.append(keypoints[0])
+        all_score.append(scores[0])
+
+    return np.array(all_key), np.array(all_score)
+
+
+def bbox_xyxy2cs(bbox: np.ndarray,
+                 padding: float = 1.) -> Tuple[np.ndarray, np.ndarray]:
+    """Transform the bbox format from (x,y,w,h) into (center, scale)
+
+    Args:
+        bbox (ndarray): Bounding box(es) in shape (4,) or (n, 4), formatted
+            as (left, top, right, bottom)
+        padding (float): BBox padding factor that will be multilied to scale.
+            Default: 1.0
+
+    Returns:
+        tuple: A tuple containing center and scale.
+        - np.ndarray[float32]: Center (x, y) of the bbox in shape (2,) or
+            (n, 2)
+        - np.ndarray[float32]: Scale (w, h) of the bbox in shape (2,) or
+            (n, 2)
+    """
+    # convert single bbox from (4, ) to (1, 4)
+    dim = bbox.ndim
+    if dim == 1:
+        bbox = bbox[None, :]
+
+    # get bbox center and scale
+    x1, y1, x2, y2 = np.hsplit(bbox, [1, 2, 3])
+    center = np.hstack([x1 + x2, y1 + y2]) * 0.5
+    scale = np.hstack([x2 - x1, y2 - y1]) * padding
+
+    if dim == 1:
+        center = center[0]
+        scale = scale[0]
+
+    return center, scale
+
+
+def _fix_aspect_ratio(bbox_scale: np.ndarray,
+                      aspect_ratio: float) -> np.ndarray:
+    """Extend the scale to match the given aspect ratio.
+
+    Args:
+        scale (np.ndarray): The image scale (w, h) in shape (2, )
+        aspect_ratio (float): The ratio of ``w/h``
+
+    Returns:
+        np.ndarray: The reshaped image scale in (2, )
+    """
+    w, h = np.hsplit(bbox_scale, [1])
+    bbox_scale = np.where(w > h * aspect_ratio,
+                          np.hstack([w, w / aspect_ratio]),
+                          np.hstack([h * aspect_ratio, h]))
+    return bbox_scale
+
+
+def _rotate_point(pt: np.ndarray, angle_rad: float) -> np.ndarray:
+    """Rotate a point by an angle.
+
+    Args:
+        pt (np.ndarray): 2D point coordinates (x, y) in shape (2, )
+        angle_rad (float): rotation angle in radian
+
+    Returns:
+        np.ndarray: Rotated point in shape (2, )
+    """
+    sn, cs = np.sin(angle_rad), np.cos(angle_rad)
+    rot_mat = np.array([[cs, -sn], [sn, cs]])
+    return rot_mat @ pt
+
+
+def _get_3rd_point(a: np.ndarray, b: np.ndarray) -> np.ndarray:
+    """To calculate the affine matrix, three pairs of points are required. This
+    function is used to get the 3rd point, given 2D points a & b.
+
+    The 3rd point is defined by rotating vector `a - b` by 90 degrees
+    anticlockwise, using b as the rotation center.
+
+    Args:
+        a (np.ndarray): The 1st point (x,y) in shape (2, )
+        b (np.ndarray): The 2nd point (x,y) in shape (2, )
+
+    Returns:
+        np.ndarray: The 3rd point.
+    """
+    direction = a - b
+    c = b + np.r_[-direction[1], direction[0]]
+    return c
+
+
+def get_warp_matrix(center: np.ndarray,
+                    scale: np.ndarray,
+                    rot: float,
+                    output_size: Tuple[int, int],
+                    shift: Tuple[float, float] = (0., 0.),
+                    inv: bool = False) -> np.ndarray:
+    """Calculate the affine transformation matrix that can warp the bbox area
+    in the input image to the output size.
+
+    Args:
+        center (np.ndarray[2, ]): Center of the bounding box (x, y).
+        scale (np.ndarray[2, ]): Scale of the bounding box
+            wrt [width, height].
+        rot (float): Rotation angle (degree).
+        output_size (np.ndarray[2, ] | list(2,)): Size of the
+            destination heatmaps.
+        shift (0-100%): Shift translation ratio wrt the width/height.
+            Default (0., 0.).
+        inv (bool): Option to inverse the affine transform direction.
+            (inv=False: src->dst or inv=True: dst->src)
+
+    Returns:
+        np.ndarray: A 2x3 transformation matrix
+    """
+    shift = np.array(shift)
+    src_w = scale[0]
+    dst_w = output_size[0]
+    dst_h = output_size[1]
+
+    # compute transformation matrix
+    rot_rad = np.deg2rad(rot)
+    src_dir = _rotate_point(np.array([0., src_w * -0.5]), rot_rad)
+    dst_dir = np.array([0., dst_w * -0.5])
+
+    # get four corners of the src rectangle in the original image
+    src = np.zeros((3, 2), dtype=np.float32)
+    src[0, :] = center + scale * shift
+    src[1, :] = center + src_dir + scale * shift
+    src[2, :] = _get_3rd_point(src[0, :], src[1, :])
+
+    # get four corners of the dst rectangle in the input image
+    dst = np.zeros((3, 2), dtype=np.float32)
+    dst[0, :] = [dst_w * 0.5, dst_h * 0.5]
+    dst[1, :] = np.array([dst_w * 0.5, dst_h * 0.5]) + dst_dir
+    dst[2, :] = _get_3rd_point(dst[0, :], dst[1, :])
+
+    if inv:
+        warp_mat = cv2.getAffineTransform(np.float32(dst), np.float32(src))
+    else:
+        warp_mat = cv2.getAffineTransform(np.float32(src), np.float32(dst))
+
+    return warp_mat
+
+
+def top_down_affine(input_size: dict, bbox_scale: dict, bbox_center: dict,
+                    img: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
+    """Get the bbox image as the model input by affine transform.
+
+    Args:
+        input_size (dict): The input size of the model.
+        bbox_scale (dict): The bbox scale of the img.
+        bbox_center (dict): The bbox center of the img.
+        img (np.ndarray): The original image.
+
+    Returns:
+        tuple: A tuple containing center and scale.
+        - np.ndarray[float32]: img after affine transform.
+        - np.ndarray[float32]: bbox scale after affine transform.
+    """
+    w, h = input_size
+    warp_size = (int(w), int(h))
+
+    # reshape bbox to fixed aspect ratio
+    bbox_scale = _fix_aspect_ratio(bbox_scale, aspect_ratio=w / h)
+
+    # get the affine matrix
+    center = bbox_center
+    scale = bbox_scale
+    rot = 0
+    warp_mat = get_warp_matrix(center, scale, rot, output_size=(w, h))
+
+    # do affine transform
+    img = cv2.warpAffine(img, warp_mat, warp_size, flags=cv2.INTER_LINEAR)
+
+    return img, bbox_scale
+
+
+def get_simcc_maximum(simcc_x: np.ndarray,
+                      simcc_y: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
+    """Get maximum response location and value from simcc representations.
+
+    Note:
+        instance number: N
+        num_keypoints: K
+        heatmap height: H
+        heatmap width: W
+
+    Args:
+        simcc_x (np.ndarray): x-axis SimCC in shape (K, Wx) or (N, K, Wx)
+        simcc_y (np.ndarray): y-axis SimCC in shape (K, Wy) or (N, K, Wy)
+
+    Returns:
+        tuple:
+        - locs (np.ndarray): locations of maximum heatmap responses in shape
+            (K, 2) or (N, K, 2)
+        - vals (np.ndarray): values of maximum heatmap responses in shape
+            (K,) or (N, K)
+    """
+    N, K, Wx = simcc_x.shape
+    simcc_x = simcc_x.reshape(N * K, -1)
+    simcc_y = simcc_y.reshape(N * K, -1)
+
+    # get maximum value locations
+    x_locs = np.argmax(simcc_x, axis=1)
+    y_locs = np.argmax(simcc_y, axis=1)
+    locs = np.stack((x_locs, y_locs), axis=-1).astype(np.float32)
+    max_val_x = np.amax(simcc_x, axis=1)
+    max_val_y = np.amax(simcc_y, axis=1)
+
+    # get maximum value across x and y axis
+    mask = max_val_x > max_val_y
+    max_val_x[mask] = max_val_y[mask]
+    vals = max_val_x
+    locs[vals <= 0.] = -1
+
+    # reshape
+    locs = locs.reshape(N, K, 2)
+    vals = vals.reshape(N, K)
+
+    return locs, vals
+
+
+def decode(simcc_x: np.ndarray, simcc_y: np.ndarray,
+           simcc_split_ratio) -> Tuple[np.ndarray, np.ndarray]:
+    """Modulate simcc distribution with Gaussian.
+
+    Args:
+        simcc_x (np.ndarray[K, Wx]): model predicted simcc in x.
+        simcc_y (np.ndarray[K, Wy]): model predicted simcc in y.
+        simcc_split_ratio (int): The split ratio of simcc.
+
+    Returns:
+        tuple: A tuple containing center and scale.
+        - np.ndarray[float32]: keypoints in shape (K, 2) or (n, K, 2)
+        - np.ndarray[float32]: scores in shape (K,) or (n, K)
+    """
+    keypoints, scores = get_simcc_maximum(simcc_x, simcc_y)
+    keypoints /= simcc_split_ratio
+
+    return keypoints, scores
+
+
+def inference_pose(session, out_bbox, oriImg):
+    h, w = session.get_inputs()[0].shape[2:]
+    model_input_size = (w, h)
+    resized_img, center, scale = preprocess(oriImg, out_bbox, model_input_size)
+    outputs = inference(session, resized_img)
+    keypoints, scores = postprocess(outputs, model_input_size, center, scale)
+
+    return keypoints, scores

src/flux/annotator/dwpose/util.py

@@ -0,0 +1,297 @@
+import math
+import numpy as np
+import matplotlib
+import cv2
+
+
+eps = 0.01
+
+
+def smart_resize(x, s):
+    Ht, Wt = s
+    if x.ndim == 2:
+        Ho, Wo = x.shape
+        Co = 1
+    else:
+        Ho, Wo, Co = x.shape
+    if Co == 3 or Co == 1:
+        k = float(Ht + Wt) / float(Ho + Wo)
+        return cv2.resize(x, (int(Wt), int(Ht)), interpolation=cv2.INTER_AREA if k < 1 else cv2.INTER_LANCZOS4)
+    else:
+        return np.stack([smart_resize(x[:, :, i], s) for i in range(Co)], axis=2)
+
+
+def smart_resize_k(x, fx, fy):
+    if x.ndim == 2:
+        Ho, Wo = x.shape
+        Co = 1
+    else:
+        Ho, Wo, Co = x.shape
+    Ht, Wt = Ho * fy, Wo * fx
+    if Co == 3 or Co == 1:
+        k = float(Ht + Wt) / float(Ho + Wo)
+        return cv2.resize(x, (int(Wt), int(Ht)), interpolation=cv2.INTER_AREA if k < 1 else cv2.INTER_LANCZOS4)
+    else:
+        return np.stack([smart_resize_k(x[:, :, i], fx, fy) for i in range(Co)], axis=2)
+
+
+def padRightDownCorner(img, stride, padValue):
+    h = img.shape[0]
+    w = img.shape[1]
+
+    pad = 4 * [None]
+    pad[0] = 0 # up
+    pad[1] = 0 # left
+    pad[2] = 0 if (h % stride == 0) else stride - (h % stride) # down
+    pad[3] = 0 if (w % stride == 0) else stride - (w % stride) # right
+
+    img_padded = img
+    pad_up = np.tile(img_padded[0:1, :, :]*0 + padValue, (pad[0], 1, 1))
+    img_padded = np.concatenate((pad_up, img_padded), axis=0)
+    pad_left = np.tile(img_padded[:, 0:1, :]*0 + padValue, (1, pad[1], 1))
+    img_padded = np.concatenate((pad_left, img_padded), axis=1)
+    pad_down = np.tile(img_padded[-2:-1, :, :]*0 + padValue, (pad[2], 1, 1))
+    img_padded = np.concatenate((img_padded, pad_down), axis=0)
+    pad_right = np.tile(img_padded[:, -2:-1, :]*0 + padValue, (1, pad[3], 1))
+    img_padded = np.concatenate((img_padded, pad_right), axis=1)
+
+    return img_padded, pad
+
+
+def transfer(model, model_weights):
+    transfered_model_weights = {}
+    for weights_name in model.state_dict().keys():
+        transfered_model_weights[weights_name] = model_weights['.'.join(weights_name.split('.')[1:])]
+    return transfered_model_weights
+
+
+def draw_bodypose(canvas, candidate, subset):
+    H, W, C = canvas.shape
+    candidate = np.array(candidate)
+    subset = np.array(subset)
+
+    stickwidth = 4
+
+    limbSeq = [[2, 3], [2, 6], [3, 4], [4, 5], [6, 7], [7, 8], [2, 9], [9, 10], \
+               [10, 11], [2, 12], [12, 13], [13, 14], [2, 1], [1, 15], [15, 17], \
+               [1, 16], [16, 18], [3, 17], [6, 18]]
+
+    colors = [[255, 0, 0], [255, 85, 0], [255, 170, 0], [255, 255, 0], [170, 255, 0], [85, 255, 0], [0, 255, 0], \
+              [0, 255, 85], [0, 255, 170], [0, 255, 255], [0, 170, 255], [0, 85, 255], [0, 0, 255], [85, 0, 255], \
+              [170, 0, 255], [255, 0, 255], [255, 0, 170], [255, 0, 85]]
+
+    for i in range(17):
+        for n in range(len(subset)):
+            index = subset[n][np.array(limbSeq[i]) - 1]
+            if -1 in index:
+                continue
+            Y = candidate[index.astype(int), 0] * float(W)
+            X = candidate[index.astype(int), 1] * float(H)
+            mX = np.mean(X)
+            mY = np.mean(Y)
+            length = ((X[0] - X[1]) ** 2 + (Y[0] - Y[1]) ** 2) ** 0.5
+            angle = math.degrees(math.atan2(X[0] - X[1], Y[0] - Y[1]))
+            polygon = cv2.ellipse2Poly((int(mY), int(mX)), (int(length / 2), stickwidth), int(angle), 0, 360, 1)
+            cv2.fillConvexPoly(canvas, polygon, colors[i])
+
+    canvas = (canvas * 0.6).astype(np.uint8)
+
+    for i in range(18):
+        for n in range(len(subset)):
+            index = int(subset[n][i])
+            if index == -1:
+                continue
+            x, y = candidate[index][0:2]
+            x = int(x * W)
+            y = int(y * H)
+            cv2.circle(canvas, (int(x), int(y)), 4, colors[i], thickness=-1)
+
+    return canvas
+
+
+def draw_handpose(canvas, all_hand_peaks):
+    H, W, C = canvas.shape
+
+    edges = [[0, 1], [1, 2], [2, 3], [3, 4], [0, 5], [5, 6], [6, 7], [7, 8], [0, 9], [9, 10], \
+             [10, 11], [11, 12], [0, 13], [13, 14], [14, 15], [15, 16], [0, 17], [17, 18], [18, 19], [19, 20]]
+
+    for peaks in all_hand_peaks:
+        peaks = np.array(peaks)
+
+        for ie, e in enumerate(edges):
+            x1, y1 = peaks[e[0]]
+            x2, y2 = peaks[e[1]]
+            x1 = int(x1 * W)
+            y1 = int(y1 * H)
+            x2 = int(x2 * W)
+            y2 = int(y2 * H)
+            if x1 > eps and y1 > eps and x2 > eps and y2 > eps:
+                cv2.line(canvas, (x1, y1), (x2, y2), matplotlib.colors.hsv_to_rgb([ie / float(len(edges)), 1.0, 1.0]) * 255, thickness=2)
+
+        for i, keyponit in enumerate(peaks):
+            x, y = keyponit
+            x = int(x * W)
+            y = int(y * H)
+            if x > eps and y > eps:
+                cv2.circle(canvas, (x, y), 4, (0, 0, 255), thickness=-1)
+    return canvas
+
+
+def draw_facepose(canvas, all_lmks):
+    H, W, C = canvas.shape
+    for lmks in all_lmks:
+        lmks = np.array(lmks)
+        for lmk in lmks:
+            x, y = lmk
+            x = int(x * W)
+            y = int(y * H)
+            if x > eps and y > eps:
+                cv2.circle(canvas, (x, y), 3, (255, 255, 255), thickness=-1)
+    return canvas
+
+
+# detect hand according to body pose keypoints
+# please refer to https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/src/openpose/hand/handDetector.cpp
+def handDetect(candidate, subset, oriImg):
+    # right hand: wrist 4, elbow 3, shoulder 2
+    # left hand: wrist 7, elbow 6, shoulder 5
+    ratioWristElbow = 0.33
+    detect_result = []
+    image_height, image_width = oriImg.shape[0:2]
+    for person in subset.astype(int):
+        # if any of three not detected
+        has_left = np.sum(person[[5, 6, 7]] == -1) == 0
+        has_right = np.sum(person[[2, 3, 4]] == -1) == 0
+        if not (has_left or has_right):
+            continue
+        hands = []
+        #left hand
+        if has_left:
+            left_shoulder_index, left_elbow_index, left_wrist_index = person[[5, 6, 7]]
+            x1, y1 = candidate[left_shoulder_index][:2]
+            x2, y2 = candidate[left_elbow_index][:2]
+            x3, y3 = candidate[left_wrist_index][:2]
+            hands.append([x1, y1, x2, y2, x3, y3, True])
+        # right hand
+        if has_right:
+            right_shoulder_index, right_elbow_index, right_wrist_index = person[[2, 3, 4]]
+            x1, y1 = candidate[right_shoulder_index][:2]
+            x2, y2 = candidate[right_elbow_index][:2]
+            x3, y3 = candidate[right_wrist_index][:2]
+            hands.append([x1, y1, x2, y2, x3, y3, False])
+
+        for x1, y1, x2, y2, x3, y3, is_left in hands:
+            # pos_hand = pos_wrist + ratio * (pos_wrist - pos_elbox) = (1 + ratio) * pos_wrist - ratio * pos_elbox
+            # handRectangle.x = posePtr[wrist*3] + ratioWristElbow * (posePtr[wrist*3] - posePtr[elbow*3]);
+            # handRectangle.y = posePtr[wrist*3+1] + ratioWristElbow * (posePtr[wrist*3+1] - posePtr[elbow*3+1]);
+            # const auto distanceWristElbow = getDistance(poseKeypoints, person, wrist, elbow);
+            # const auto distanceElbowShoulder = getDistance(poseKeypoints, person, elbow, shoulder);
+            # handRectangle.width = 1.5f * fastMax(distanceWristElbow, 0.9f * distanceElbowShoulder);
+            x = x3 + ratioWristElbow * (x3 - x2)
+            y = y3 + ratioWristElbow * (y3 - y2)
+            distanceWristElbow = math.sqrt((x3 - x2) ** 2 + (y3 - y2) ** 2)
+            distanceElbowShoulder = math.sqrt((x2 - x1) ** 2 + (y2 - y1) ** 2)
+            width = 1.5 * max(distanceWristElbow, 0.9 * distanceElbowShoulder)
+            # x-y refers to the center --> offset to topLeft point
+            # handRectangle.x -= handRectangle.width / 2.f;
+            # handRectangle.y -= handRectangle.height / 2.f;
+            x -= width / 2
+            y -= width / 2  # width = height
+            # overflow the image
+            if x < 0: x = 0
+            if y < 0: y = 0
+            width1 = width
+            width2 = width
+            if x + width > image_width: width1 = image_width - x
+            if y + width > image_height: width2 = image_height - y
+            width = min(width1, width2)
+            # the max hand box value is 20 pixels
+            if width >= 20:
+                detect_result.append([int(x), int(y), int(width), is_left])
+
+    '''
+    return value: [[x, y, w, True if left hand else False]].
+    width=height since the network require squared input.
+    x, y is the coordinate of top left 
+    '''
+    return detect_result
+
+
+# Written by Lvmin
+def faceDetect(candidate, subset, oriImg):
+    # left right eye ear 14 15 16 17
+    detect_result = []
+    image_height, image_width = oriImg.shape[0:2]
+    for person in subset.astype(int):
+        has_head = person[0] > -1
+        if not has_head:
+            continue
+
+        has_left_eye = person[14] > -1
+        has_right_eye = person[15] > -1
+        has_left_ear = person[16] > -1
+        has_right_ear = person[17] > -1
+
+        if not (has_left_eye or has_right_eye or has_left_ear or has_right_ear):
+            continue
+
+        head, left_eye, right_eye, left_ear, right_ear = person[[0, 14, 15, 16, 17]]
+
+        width = 0.0
+        x0, y0 = candidate[head][:2]
+
+        if has_left_eye:
+            x1, y1 = candidate[left_eye][:2]
+            d = max(abs(x0 - x1), abs(y0 - y1))
+            width = max(width, d * 3.0)
+
+        if has_right_eye:
+            x1, y1 = candidate[right_eye][:2]
+            d = max(abs(x0 - x1), abs(y0 - y1))
+            width = max(width, d * 3.0)
+
+        if has_left_ear:
+            x1, y1 = candidate[left_ear][:2]
+            d = max(abs(x0 - x1), abs(y0 - y1))
+            width = max(width, d * 1.5)
+
+        if has_right_ear:
+            x1, y1 = candidate[right_ear][:2]
+            d = max(abs(x0 - x1), abs(y0 - y1))
+            width = max(width, d * 1.5)
+
+        x, y = x0, y0
+
+        x -= width
+        y -= width
+
+        if x < 0:
+            x = 0
+
+        if y < 0:
+            y = 0
+
+        width1 = width * 2
+        width2 = width * 2
+
+        if x + width > image_width:
+            width1 = image_width - x
+
+        if y + width > image_height:
+            width2 = image_height - y
+
+        width = min(width1, width2)
+
+        if width >= 20:
+            detect_result.append([int(x), int(y), int(width)])
+
+    return detect_result
+
+
+# get max index of 2d array
+def npmax(array):
+    arrayindex = array.argmax(1)
+    arrayvalue = array.max(1)
+    i = arrayvalue.argmax()
+    j = arrayindex[i]
+    return i, j

src/flux/annotator/dwpose/wholebody.py

@@ -0,0 +1,48 @@
+import cv2
+import numpy as np
+
+import onnxruntime as ort
+from huggingface_hub import hf_hub_download
+from .onnxdet import inference_detector
+from .onnxpose import inference_pose
+
+
+class Wholebody:
+    def __init__(self, device="cuda:0"):
+        providers = ['CPUExecutionProvider'] if device == 'cpu' else ['CUDAExecutionProvider']
+        onnx_det = hf_hub_download("yzd-v/DWPose", "yolox_l.onnx")
+        onnx_pose = hf_hub_download("yzd-v/DWPose", "dw-ll_ucoco_384.onnx")
+
+        self.session_det = ort.InferenceSession(path_or_bytes=onnx_det, providers=providers)
+        self.session_pose = ort.InferenceSession(path_or_bytes=onnx_pose, providers=providers)
+
+    def __call__(self, oriImg):
+        det_result = inference_detector(self.session_det, oriImg)
+        keypoints, scores = inference_pose(self.session_pose, det_result, oriImg)
+
+        keypoints_info = np.concatenate(
+            (keypoints, scores[..., None]), axis=-1)
+        # compute neck joint
+        neck = np.mean(keypoints_info[:, [5, 6]], axis=1)
+        # neck score when visualizing pred
+        neck[:, 2:4] = np.logical_and(
+            keypoints_info[:, 5, 2:4] > 0.3,
+            keypoints_info[:, 6, 2:4] > 0.3).astype(int)
+        new_keypoints_info = np.insert(
+            keypoints_info, 17, neck, axis=1)
+        mmpose_idx = [
+            17, 6, 8, 10, 7, 9, 12, 14, 16, 13, 15, 2, 1, 4, 3
+        ]
+        openpose_idx = [
+            1, 2, 3, 4, 6, 7, 8, 9, 10, 12, 13, 14, 15, 16, 17
+        ]
+        new_keypoints_info[:, openpose_idx] = \
+            new_keypoints_info[:, mmpose_idx]
+        keypoints_info = new_keypoints_info
+
+        keypoints, scores = keypoints_info[
+            ..., :2], keypoints_info[..., 2]
+
+        return keypoints, scores
+
+

src/flux/annotator/hed/__init__.py

@@ -0,0 +1,95 @@
+# This is an improved version and model of HED edge detection with Apache License, Version 2.0.
+# Please use this implementation in your products
+# This implementation may produce slightly different results from Saining Xie's official implementations,
+# but it generates smoother edges and is more suitable for ControlNet as well as other image-to-image translations.
+# Different from official models and other implementations, this is an RGB-input model (rather than BGR)
+# and in this way it works better for gradio's RGB protocol
+
+import os
+import cv2
+import torch
+import numpy as np
+
+from huggingface_hub import hf_hub_download
+from einops import rearrange
+from ...annotator.util import annotator_ckpts_path
+
+
+class DoubleConvBlock(torch.nn.Module):
+    def __init__(self, input_channel, output_channel, layer_number):
+        super().__init__()
+        self.convs = torch.nn.Sequential()
+        self.convs.append(torch.nn.Conv2d(in_channels=input_channel, out_channels=output_channel, kernel_size=(3, 3), stride=(1, 1), padding=1))
+        for i in range(1, layer_number):
+            self.convs.append(torch.nn.Conv2d(in_channels=output_channel, out_channels=output_channel, kernel_size=(3, 3), stride=(1, 1), padding=1))
+        self.projection = torch.nn.Conv2d(in_channels=output_channel, out_channels=1, kernel_size=(1, 1), stride=(1, 1), padding=0)
+
+    def __call__(self, x, down_sampling=False):
+        h = x
+        if down_sampling:
+            h = torch.nn.functional.max_pool2d(h, kernel_size=(2, 2), stride=(2, 2))
+        for conv in self.convs:
+            h = conv(h)
+            h = torch.nn.functional.relu(h)
+        return h, self.projection(h)
+
+
+class ControlNetHED_Apache2(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.norm = torch.nn.Parameter(torch.zeros(size=(1, 3, 1, 1)))
+        self.block1 = DoubleConvBlock(input_channel=3, output_channel=64, layer_number=2)
+        self.block2 = DoubleConvBlock(input_channel=64, output_channel=128, layer_number=2)
+        self.block3 = DoubleConvBlock(input_channel=128, output_channel=256, layer_number=3)
+        self.block4 = DoubleConvBlock(input_channel=256, output_channel=512, layer_number=3)
+        self.block5 = DoubleConvBlock(input_channel=512, output_channel=512, layer_number=3)
+
+    def __call__(self, x):
+        h = x - self.norm
+        h, projection1 = self.block1(h)
+        h, projection2 = self.block2(h, down_sampling=True)
+        h, projection3 = self.block3(h, down_sampling=True)
+        h, projection4 = self.block4(h, down_sampling=True)
+        h, projection5 = self.block5(h, down_sampling=True)
+        return projection1, projection2, projection3, projection4, projection5
+
+
+class HEDdetector:
+    def __init__(self):
+        modelpath = os.path.join(annotator_ckpts_path, "ControlNetHED.pth")
+        if not os.path.exists(modelpath):
+            modelpath = hf_hub_download("lllyasviel/Annotators", "ControlNetHED.pth")
+        self.netNetwork = ControlNetHED_Apache2().float().cuda().eval()
+        self.netNetwork.load_state_dict(torch.load(modelpath))
+
+    def __call__(self, input_image):
+        assert input_image.ndim == 3
+        H, W, C = input_image.shape
+        with torch.no_grad():
+            image_hed = torch.from_numpy(input_image.copy()).float().cuda()
+            image_hed = rearrange(image_hed, 'h w c -> 1 c h w')
+            edges = self.netNetwork(image_hed)
+            edges = [e.detach().cpu().numpy().astype(np.float32)[0, 0] for e in edges]
+            edges = [cv2.resize(e, (W, H), interpolation=cv2.INTER_LINEAR) for e in edges]
+            edges = np.stack(edges, axis=2)
+            edge = 1 / (1 + np.exp(-np.mean(edges, axis=2).astype(np.float64)))
+            edge = (edge * 255.0).clip(0, 255).astype(np.uint8)
+            return edge
+
+
+def nms(x, t, s):
+    x = cv2.GaussianBlur(x.astype(np.float32), (0, 0), s)
+
+    f1 = np.array([[0, 0, 0], [1, 1, 1], [0, 0, 0]], dtype=np.uint8)
+    f2 = np.array([[0, 1, 0], [0, 1, 0], [0, 1, 0]], dtype=np.uint8)
+    f3 = np.array([[1, 0, 0], [0, 1, 0], [0, 0, 1]], dtype=np.uint8)
+    f4 = np.array([[0, 0, 1], [0, 1, 0], [1, 0, 0]], dtype=np.uint8)
+
+    y = np.zeros_like(x)
+
+    for f in [f1, f2, f3, f4]:
+        np.putmask(y, cv2.dilate(x, kernel=f) == x, x)
+
+    z = np.zeros_like(y, dtype=np.uint8)
+    z[y > t] = 255
+    return z

src/flux/annotator/midas/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2019 Intel ISL (Intel Intelligent Systems Lab)
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

src/flux/annotator/midas/__init__.py

@@ -0,0 +1,42 @@
+# Midas Depth Estimation
+# From https://github.com/isl-org/MiDaS
+# MIT LICENSE
+
+import cv2
+import numpy as np
+import torch
+
+from einops import rearrange
+from .api import MiDaSInference
+
+
+class MidasDetector:
+    def __init__(self):
+        self.model = MiDaSInference(model_type="dpt_hybrid").cuda()
+
+    def __call__(self, input_image, a=np.pi * 2.0, bg_th=0.1):
+        assert input_image.ndim == 3
+        image_depth = input_image
+        with torch.no_grad():
+            image_depth = torch.from_numpy(image_depth).float().cuda()
+            image_depth = image_depth / 127.5 - 1.0
+            image_depth = rearrange(image_depth, 'h w c -> 1 c h w')
+            depth = self.model(image_depth)[0]
+
+            depth_pt = depth.clone()
+            depth_pt -= torch.min(depth_pt)
+            depth_pt /= torch.max(depth_pt)
+            depth_pt = depth_pt.cpu().numpy()
+            depth_image = (depth_pt * 255.0).clip(0, 255).astype(np.uint8)
+
+            depth_np = depth.cpu().numpy()
+            x = cv2.Sobel(depth_np, cv2.CV_32F, 1, 0, ksize=3)
+            y = cv2.Sobel(depth_np, cv2.CV_32F, 0, 1, ksize=3)
+            z = np.ones_like(x) * a
+            x[depth_pt < bg_th] = 0
+            y[depth_pt < bg_th] = 0
+            normal = np.stack([x, y, z], axis=2)
+            normal /= np.sum(normal ** 2.0, axis=2, keepdims=True) ** 0.5
+            normal_image = (normal * 127.5 + 127.5).clip(0, 255).astype(np.uint8)
+
+            return depth_image, normal_image

src/flux/annotator/midas/api.py

@@ -0,0 +1,168 @@
+# based on https://github.com/isl-org/MiDaS
+
+import cv2
+import os
+import torch
+import torch.nn as nn
+from torchvision.transforms import Compose
+
+from huggingface_hub import hf_hub_download
+
+from .midas.dpt_depth import DPTDepthModel
+from .midas.midas_net import MidasNet
+from .midas.midas_net_custom import MidasNet_small
+from .midas.transforms import Resize, NormalizeImage, PrepareForNet
+from ...annotator.util import annotator_ckpts_path
+
+
+ISL_PATHS = {
+    "dpt_large": os.path.join(annotator_ckpts_path, "dpt_large-midas-2f21e586.pt"),
+    "dpt_hybrid": os.path.join(annotator_ckpts_path, "dpt_hybrid-midas-501f0c75.pt"),
+    "midas_v21": "",
+    "midas_v21_small": "",
+}
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+def load_midas_transform(model_type):
+    # https://github.com/isl-org/MiDaS/blob/master/run.py
+    # load transform only
+    if model_type == "dpt_large":  # DPT-Large
+        net_w, net_h = 384, 384
+        resize_mode = "minimal"
+        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+
+    elif model_type == "dpt_hybrid":  # DPT-Hybrid
+        net_w, net_h = 384, 384
+        resize_mode = "minimal"
+        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+
+    elif model_type == "midas_v21":
+        net_w, net_h = 384, 384
+        resize_mode = "upper_bound"
+        normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+
+    elif model_type == "midas_v21_small":
+        net_w, net_h = 256, 256
+        resize_mode = "upper_bound"
+        normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+
+    else:
+        assert False, f"model_type '{model_type}' not implemented, use: --model_type large"
+
+    transform = Compose(
+        [
+            Resize(
+                net_w,
+                net_h,
+                resize_target=None,
+                keep_aspect_ratio=True,
+                ensure_multiple_of=32,
+                resize_method=resize_mode,
+                image_interpolation_method=cv2.INTER_CUBIC,
+            ),
+            normalization,
+            PrepareForNet(),
+        ]
+    )
+
+    return transform
+
+
+def load_model(model_type):
+    # https://github.com/isl-org/MiDaS/blob/master/run.py
+    # load network
+    model_path = ISL_PATHS[model_type]
+    if model_type == "dpt_large":  # DPT-Large
+        model = DPTDepthModel(
+            path=model_path,
+            backbone="vitl16_384",
+            non_negative=True,
+        )
+        net_w, net_h = 384, 384
+        resize_mode = "minimal"
+        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+
+    elif model_type == "dpt_hybrid":  # DPT-Hybrid
+        if not os.path.exists(model_path):
+            model_path = hf_hub_download("lllyasviel/Annotators", "dpt_hybrid-midas-501f0c75.pt")
+
+        model = DPTDepthModel(
+            path=model_path,
+            backbone="vitb_rn50_384",
+            non_negative=True,
+        )
+        net_w, net_h = 384, 384
+        resize_mode = "minimal"
+        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+
+    elif model_type == "midas_v21":
+        model = MidasNet(model_path, non_negative=True)
+        net_w, net_h = 384, 384
+        resize_mode = "upper_bound"
+        normalization = NormalizeImage(
+            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
+        )
+
+    elif model_type == "midas_v21_small":
+        model = MidasNet_small(model_path, features=64, backbone="efficientnet_lite3", exportable=True,
+                               non_negative=True, blocks={'expand': True})
+        net_w, net_h = 256, 256
+        resize_mode = "upper_bound"
+        normalization = NormalizeImage(
+            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
+        )
+
+    else:
+        print(f"model_type '{model_type}' not implemented, use: --model_type large")
+        assert False
+
+    transform = Compose(
+        [
+            Resize(
+                net_w,
+                net_h,
+                resize_target=None,
+                keep_aspect_ratio=True,
+                ensure_multiple_of=32,
+                resize_method=resize_mode,
+                image_interpolation_method=cv2.INTER_CUBIC,
+            ),
+            normalization,
+            PrepareForNet(),
+        ]
+    )
+
+    return model.eval(), transform
+
+
+class MiDaSInference(nn.Module):
+    MODEL_TYPES_TORCH_HUB = [
+        "DPT_Large",
+        "DPT_Hybrid",
+        "MiDaS_small"
+    ]
+    MODEL_TYPES_ISL = [
+        "dpt_large",
+        "dpt_hybrid",
+        "midas_v21",
+        "midas_v21_small",
+    ]
+
+    def __init__(self, model_type):
+        super().__init__()
+        assert (model_type in self.MODEL_TYPES_ISL)
+        model, _ = load_model(model_type)
+        self.model = model
+        self.model.train = disabled_train
+
+    def forward(self, x):
+        with torch.no_grad():
+            prediction = self.model(x)
+        return prediction
+

src/flux/annotator/midas/midas/base_model.py

@@ -0,0 +1,16 @@
+import torch
+
+
+class BaseModel(torch.nn.Module):
+    def load(self, path):
+        """Load model from file.
+
+        Args:
+            path (str): file path
+        """
+        parameters = torch.load(path, map_location=torch.device('cpu'))
+
+        if "optimizer" in parameters:
+            parameters = parameters["model"]
+
+        self.load_state_dict(parameters)

src/flux/annotator/midas/midas/blocks.py

@@ -0,0 +1,342 @@
+import torch
+import torch.nn as nn
+
+from .vit import (
+    _make_pretrained_vitb_rn50_384,
+    _make_pretrained_vitl16_384,
+    _make_pretrained_vitb16_384,
+    forward_vit,
+)
+
+def _make_encoder(backbone, features, use_pretrained, groups=1, expand=False, exportable=True, hooks=None, use_vit_only=False, use_readout="ignore",):
+    if backbone == "vitl16_384":
+        pretrained = _make_pretrained_vitl16_384(
+            use_pretrained, hooks=hooks, use_readout=use_readout
+        )
+        scratch = _make_scratch(
+            [256, 512, 1024, 1024], features, groups=groups, expand=expand
+        )  # ViT-L/16 - 85.0% Top1 (backbone)
+    elif backbone == "vitb_rn50_384":
+        pretrained = _make_pretrained_vitb_rn50_384(
+            use_pretrained,
+            hooks=hooks,
+            use_vit_only=use_vit_only,
+            use_readout=use_readout,
+        )
+        scratch = _make_scratch(
+            [256, 512, 768, 768], features, groups=groups, expand=expand
+        )  # ViT-H/16 - 85.0% Top1 (backbone)
+    elif backbone == "vitb16_384":
+        pretrained = _make_pretrained_vitb16_384(
+            use_pretrained, hooks=hooks, use_readout=use_readout
+        )
+        scratch = _make_scratch(
+            [96, 192, 384, 768], features, groups=groups, expand=expand
+        )  # ViT-B/16 - 84.6% Top1 (backbone)
+    elif backbone == "resnext101_wsl":
+        pretrained = _make_pretrained_resnext101_wsl(use_pretrained)
+        scratch = _make_scratch([256, 512, 1024, 2048], features, groups=groups, expand=expand)     # efficientnet_lite3  
+    elif backbone == "efficientnet_lite3":
+        pretrained = _make_pretrained_efficientnet_lite3(use_pretrained, exportable=exportable)
+        scratch = _make_scratch([32, 48, 136, 384], features, groups=groups, expand=expand)  # efficientnet_lite3     
+    else:
+        print(f"Backbone '{backbone}' not implemented")
+        assert False
+        
+    return pretrained, scratch
+
+
+def _make_scratch(in_shape, out_shape, groups=1, expand=False):
+    scratch = nn.Module()
+
+    out_shape1 = out_shape
+    out_shape2 = out_shape
+    out_shape3 = out_shape
+    out_shape4 = out_shape
+    if expand==True:
+        out_shape1 = out_shape
+        out_shape2 = out_shape*2
+        out_shape3 = out_shape*4
+        out_shape4 = out_shape*8
+
+    scratch.layer1_rn = nn.Conv2d(
+        in_shape[0], out_shape1, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer2_rn = nn.Conv2d(
+        in_shape[1], out_shape2, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer3_rn = nn.Conv2d(
+        in_shape[2], out_shape3, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer4_rn = nn.Conv2d(
+        in_shape[3], out_shape4, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+
+    return scratch
+
+
+def _make_pretrained_efficientnet_lite3(use_pretrained, exportable=False):
+    efficientnet = torch.hub.load(
+        "rwightman/gen-efficientnet-pytorch",
+        "tf_efficientnet_lite3",
+        pretrained=use_pretrained,
+        exportable=exportable
+    )
+    return _make_efficientnet_backbone(efficientnet)
+
+
+def _make_efficientnet_backbone(effnet):
+    pretrained = nn.Module()
+
+    pretrained.layer1 = nn.Sequential(
+        effnet.conv_stem, effnet.bn1, effnet.act1, *effnet.blocks[0:2]
+    )
+    pretrained.layer2 = nn.Sequential(*effnet.blocks[2:3])
+    pretrained.layer3 = nn.Sequential(*effnet.blocks[3:5])
+    pretrained.layer4 = nn.Sequential(*effnet.blocks[5:9])
+
+    return pretrained
+    
+
+def _make_resnet_backbone(resnet):
+    pretrained = nn.Module()
+    pretrained.layer1 = nn.Sequential(
+        resnet.conv1, resnet.bn1, resnet.relu, resnet.maxpool, resnet.layer1
+    )
+
+    pretrained.layer2 = resnet.layer2
+    pretrained.layer3 = resnet.layer3
+    pretrained.layer4 = resnet.layer4
+
+    return pretrained
+
+
+def _make_pretrained_resnext101_wsl(use_pretrained):
+    resnet = torch.hub.load("facebookresearch/WSL-Images", "resnext101_32x8d_wsl")
+    return _make_resnet_backbone(resnet)
+
+
+
+class Interpolate(nn.Module):
+    """Interpolation module.
+    """
+
+    def __init__(self, scale_factor, mode, align_corners=False):
+        """Init.
+
+        Args:
+            scale_factor (float): scaling
+            mode (str): interpolation mode
+        """
+        super(Interpolate, self).__init__()
+
+        self.interp = nn.functional.interpolate
+        self.scale_factor = scale_factor
+        self.mode = mode
+        self.align_corners = align_corners
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input
+
+        Returns:
+            tensor: interpolated data
+        """
+
+        x = self.interp(
+            x, scale_factor=self.scale_factor, mode=self.mode, align_corners=self.align_corners
+        )
+
+        return x
+
+
+class ResidualConvUnit(nn.Module):
+    """Residual convolution module.
+    """
+
+    def __init__(self, features):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super().__init__()
+
+        self.conv1 = nn.Conv2d(
+            features, features, kernel_size=3, stride=1, padding=1, bias=True
+        )
+
+        self.conv2 = nn.Conv2d(
+            features, features, kernel_size=3, stride=1, padding=1, bias=True
+        )
+
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input
+
+        Returns:
+            tensor: output
+        """
+        out = self.relu(x)
+        out = self.conv1(out)
+        out = self.relu(out)
+        out = self.conv2(out)
+
+        return out + x
+
+
+class FeatureFusionBlock(nn.Module):
+    """Feature fusion block.
+    """
+
+    def __init__(self, features):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super(FeatureFusionBlock, self).__init__()
+
+        self.resConfUnit1 = ResidualConvUnit(features)
+        self.resConfUnit2 = ResidualConvUnit(features)
+
+    def forward(self, *xs):
+        """Forward pass.
+
+        Returns:
+            tensor: output
+        """
+        output = xs[0]
+
+        if len(xs) == 2:
+            output += self.resConfUnit1(xs[1])
+
+        output = self.resConfUnit2(output)
+
+        output = nn.functional.interpolate(
+            output, scale_factor=2, mode="bilinear", align_corners=True
+        )
+
+        return output
+
+
+
+
+class ResidualConvUnit_custom(nn.Module):
+    """Residual convolution module.
+    """
+
+    def __init__(self, features, activation, bn):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super().__init__()
+
+        self.bn = bn
+
+        self.groups=1
+
+        self.conv1 = nn.Conv2d(
+            features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups
+        )
+        
+        self.conv2 = nn.Conv2d(
+            features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups
+        )
+
+        if self.bn==True:
+            self.bn1 = nn.BatchNorm2d(features)
+            self.bn2 = nn.BatchNorm2d(features)
+
+        self.activation = activation
+
+        self.skip_add = nn.quantized.FloatFunctional()
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input
+
+        Returns:
+            tensor: output
+        """
+        
+        out = self.activation(x)
+        out = self.conv1(out)
+        if self.bn==True:
+            out = self.bn1(out)
+       
+        out = self.activation(out)
+        out = self.conv2(out)
+        if self.bn==True:
+            out = self.bn2(out)
+
+        if self.groups > 1:
+            out = self.conv_merge(out)
+
+        return self.skip_add.add(out, x)
+
+        # return out + x
+
+
+class FeatureFusionBlock_custom(nn.Module):
+    """Feature fusion block.
+    """
+
+    def __init__(self, features, activation, deconv=False, bn=False, expand=False, align_corners=True):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super(FeatureFusionBlock_custom, self).__init__()
+
+        self.deconv = deconv
+        self.align_corners = align_corners
+
+        self.groups=1
+
+        self.expand = expand
+        out_features = features
+        if self.expand==True:
+            out_features = features//2
+        
+        self.out_conv = nn.Conv2d(features, out_features, kernel_size=1, stride=1, padding=0, bias=True, groups=1)
+
+        self.resConfUnit1 = ResidualConvUnit_custom(features, activation, bn)
+        self.resConfUnit2 = ResidualConvUnit_custom(features, activation, bn)
+        
+        self.skip_add = nn.quantized.FloatFunctional()
+
+    def forward(self, *xs):
+        """Forward pass.
+
+        Returns:
+            tensor: output
+        """
+        output = xs[0]
+
+        if len(xs) == 2:
+            res = self.resConfUnit1(xs[1])
+            output = self.skip_add.add(output, res)
+            # output += res
+
+        output = self.resConfUnit2(output)
+
+        output = nn.functional.interpolate(
+            output, scale_factor=2, mode="bilinear", align_corners=self.align_corners
+        )
+
+        output = self.out_conv(output)
+
+        return output
+

src/flux/annotator/midas/midas/dpt_depth.py

@@ -0,0 +1,109 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .base_model import BaseModel
+from .blocks import (
+    FeatureFusionBlock,
+    FeatureFusionBlock_custom,
+    Interpolate,
+    _make_encoder,
+    forward_vit,
+)
+
+
+def _make_fusion_block(features, use_bn):
+    return FeatureFusionBlock_custom(
+        features,
+        nn.ReLU(False),
+        deconv=False,
+        bn=use_bn,
+        expand=False,
+        align_corners=True,
+    )
+
+
+class DPT(BaseModel):
+    def __init__(
+        self,
+        head,
+        features=256,
+        backbone="vitb_rn50_384",
+        readout="project",
+        channels_last=False,
+        use_bn=False,
+    ):
+
+        super(DPT, self).__init__()
+
+        self.channels_last = channels_last
+
+        hooks = {
+            "vitb_rn50_384": [0, 1, 8, 11],
+            "vitb16_384": [2, 5, 8, 11],
+            "vitl16_384": [5, 11, 17, 23],
+        }
+
+        # Instantiate backbone and reassemble blocks
+        self.pretrained, self.scratch = _make_encoder(
+            backbone,
+            features,
+            False, # Set to true of you want to train from scratch, uses ImageNet weights
+            groups=1,
+            expand=False,
+            exportable=False,
+            hooks=hooks[backbone],
+            use_readout=readout,
+        )
+
+        self.scratch.refinenet1 = _make_fusion_block(features, use_bn)
+        self.scratch.refinenet2 = _make_fusion_block(features, use_bn)
+        self.scratch.refinenet3 = _make_fusion_block(features, use_bn)
+        self.scratch.refinenet4 = _make_fusion_block(features, use_bn)
+
+        self.scratch.output_conv = head
+
+
+    def forward(self, x):
+        if self.channels_last == True:
+            x.contiguous(memory_format=torch.channels_last)
+
+        layer_1, layer_2, layer_3, layer_4 = forward_vit(self.pretrained, x)
+
+        layer_1_rn = self.scratch.layer1_rn(layer_1)
+        layer_2_rn = self.scratch.layer2_rn(layer_2)
+        layer_3_rn = self.scratch.layer3_rn(layer_3)
+        layer_4_rn = self.scratch.layer4_rn(layer_4)
+
+        path_4 = self.scratch.refinenet4(layer_4_rn)
+        path_3 = self.scratch.refinenet3(path_4, layer_3_rn)
+        path_2 = self.scratch.refinenet2(path_3, layer_2_rn)
+        path_1 = self.scratch.refinenet1(path_2, layer_1_rn)
+
+        out = self.scratch.output_conv(path_1)
+
+        return out
+
+
+class DPTDepthModel(DPT):
+    def __init__(self, path=None, non_negative=True, **kwargs):
+        features = kwargs["features"] if "features" in kwargs else 256
+
+        head = nn.Sequential(
+            nn.Conv2d(features, features // 2, kernel_size=3, stride=1, padding=1),
+            Interpolate(scale_factor=2, mode="bilinear", align_corners=True),
+            nn.Conv2d(features // 2, 32, kernel_size=3, stride=1, padding=1),
+            nn.ReLU(True),
+            nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0),
+            nn.ReLU(True) if non_negative else nn.Identity(),
+            nn.Identity(),
+        )
+
+        super().__init__(head, **kwargs)
+
+        if path is not None:
+           self.load(path)
+
+    def forward(self, x):
+        return super().forward(x).squeeze(dim=1)
+

src/flux/annotator/midas/midas/midas_net.py

@@ -0,0 +1,76 @@
+"""MidashNet: Network for monocular depth estimation trained by mixing several datasets.
+This file contains code that is adapted from
+https://github.com/thomasjpfan/pytorch_refinenet/blob/master/pytorch_refinenet/refinenet/refinenet_4cascade.py
+"""
+import torch
+import torch.nn as nn
+
+from .base_model import BaseModel
+from .blocks import FeatureFusionBlock, Interpolate, _make_encoder
+
+
+class MidasNet(BaseModel):
+    """Network for monocular depth estimation.
+    """
+
+    def __init__(self, path=None, features=256, non_negative=True):
+        """Init.
+
+        Args:
+            path (str, optional): Path to saved model. Defaults to None.
+            features (int, optional): Number of features. Defaults to 256.
+            backbone (str, optional): Backbone network for encoder. Defaults to resnet50
+        """
+        print("Loading weights: ", path)
+
+        super(MidasNet, self).__init__()
+
+        use_pretrained = False if path is None else True
+
+        self.pretrained, self.scratch = _make_encoder(backbone="resnext101_wsl", features=features, use_pretrained=use_pretrained)
+
+        self.scratch.refinenet4 = FeatureFusionBlock(features)
+        self.scratch.refinenet3 = FeatureFusionBlock(features)
+        self.scratch.refinenet2 = FeatureFusionBlock(features)
+        self.scratch.refinenet1 = FeatureFusionBlock(features)
+
+        self.scratch.output_conv = nn.Sequential(
+            nn.Conv2d(features, 128, kernel_size=3, stride=1, padding=1),
+            Interpolate(scale_factor=2, mode="bilinear"),
+            nn.Conv2d(128, 32, kernel_size=3, stride=1, padding=1),
+            nn.ReLU(True),
+            nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0),
+            nn.ReLU(True) if non_negative else nn.Identity(),
+        )
+
+        if path:
+            self.load(path)
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input data (image)
+
+        Returns:
+            tensor: depth
+        """
+
+        layer_1 = self.pretrained.layer1(x)
+        layer_2 = self.pretrained.layer2(layer_1)
+        layer_3 = self.pretrained.layer3(layer_2)
+        layer_4 = self.pretrained.layer4(layer_3)
+
+        layer_1_rn = self.scratch.layer1_rn(layer_1)
+        layer_2_rn = self.scratch.layer2_rn(layer_2)
+        layer_3_rn = self.scratch.layer3_rn(layer_3)
+        layer_4_rn = self.scratch.layer4_rn(layer_4)
+
+        path_4 = self.scratch.refinenet4(layer_4_rn)
+        path_3 = self.scratch.refinenet3(path_4, layer_3_rn)
+        path_2 = self.scratch.refinenet2(path_3, layer_2_rn)
+        path_1 = self.scratch.refinenet1(path_2, layer_1_rn)
+
+        out = self.scratch.output_conv(path_1)
+
+        return torch.squeeze(out, dim=1)

src/flux/annotator/midas/midas/midas_net_custom.py

@@ -0,0 +1,128 @@
+"""MidashNet: Network for monocular depth estimation trained by mixing several datasets.
+This file contains code that is adapted from
+https://github.com/thomasjpfan/pytorch_refinenet/blob/master/pytorch_refinenet/refinenet/refinenet_4cascade.py
+"""
+import torch
+import torch.nn as nn
+
+from .base_model import BaseModel
+from .blocks import FeatureFusionBlock, FeatureFusionBlock_custom, Interpolate, _make_encoder
+
+
+class MidasNet_small(BaseModel):
+    """Network for monocular depth estimation.
+    """
+
+    def __init__(self, path=None, features=64, backbone="efficientnet_lite3", non_negative=True, exportable=True, channels_last=False, align_corners=True,
+        blocks={'expand': True}):
+        """Init.
+
+        Args:
+            path (str, optional): Path to saved model. Defaults to None.
+            features (int, optional): Number of features. Defaults to 256.
+            backbone (str, optional): Backbone network for encoder. Defaults to resnet50
+        """
+        print("Loading weights: ", path)
+
+        super(MidasNet_small, self).__init__()
+
+        use_pretrained = False if path else True
+                
+        self.channels_last = channels_last
+        self.blocks = blocks
+        self.backbone = backbone
+
+        self.groups = 1
+
+        features1=features
+        features2=features
+        features3=features
+        features4=features
+        self.expand = False
+        if "expand" in self.blocks and self.blocks['expand'] == True:
+            self.expand = True
+            features1=features
+            features2=features*2
+            features3=features*4
+            features4=features*8
+
+        self.pretrained, self.scratch = _make_encoder(self.backbone, features, use_pretrained, groups=self.groups, expand=self.expand, exportable=exportable)
+  
+        self.scratch.activation = nn.ReLU(False)    
+
+        self.scratch.refinenet4 = FeatureFusionBlock_custom(features4, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners)
+        self.scratch.refinenet3 = FeatureFusionBlock_custom(features3, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners)
+        self.scratch.refinenet2 = FeatureFusionBlock_custom(features2, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners)
+        self.scratch.refinenet1 = FeatureFusionBlock_custom(features1, self.scratch.activation, deconv=False, bn=False, align_corners=align_corners)
+
+        
+        self.scratch.output_conv = nn.Sequential(
+            nn.Conv2d(features, features//2, kernel_size=3, stride=1, padding=1, groups=self.groups),
+            Interpolate(scale_factor=2, mode="bilinear"),
+            nn.Conv2d(features//2, 32, kernel_size=3, stride=1, padding=1),
+            self.scratch.activation,
+            nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0),
+            nn.ReLU(True) if non_negative else nn.Identity(),
+            nn.Identity(),
+        )
+        
+        if path:
+            self.load(path)
+
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input data (image)
+
+        Returns:
+            tensor: depth
+        """
+        if self.channels_last==True:
+            print("self.channels_last = ", self.channels_last)
+            x.contiguous(memory_format=torch.channels_last)
+
+
+        layer_1 = self.pretrained.layer1(x)
+        layer_2 = self.pretrained.layer2(layer_1)
+        layer_3 = self.pretrained.layer3(layer_2)
+        layer_4 = self.pretrained.layer4(layer_3)
+        
+        layer_1_rn = self.scratch.layer1_rn(layer_1)
+        layer_2_rn = self.scratch.layer2_rn(layer_2)
+        layer_3_rn = self.scratch.layer3_rn(layer_3)
+        layer_4_rn = self.scratch.layer4_rn(layer_4)
+
+
+        path_4 = self.scratch.refinenet4(layer_4_rn)
+        path_3 = self.scratch.refinenet3(path_4, layer_3_rn)
+        path_2 = self.scratch.refinenet2(path_3, layer_2_rn)
+        path_1 = self.scratch.refinenet1(path_2, layer_1_rn)
+        
+        out = self.scratch.output_conv(path_1)
+
+        return torch.squeeze(out, dim=1)
+
+
+
+def fuse_model(m):
+    prev_previous_type = nn.Identity()
+    prev_previous_name = ''
+    previous_type = nn.Identity()
+    previous_name = ''
+    for name, module in m.named_modules():
+        if prev_previous_type == nn.Conv2d and previous_type == nn.BatchNorm2d and type(module) == nn.ReLU:
+            # print("FUSED ", prev_previous_name, previous_name, name)
+            torch.quantization.fuse_modules(m, [prev_previous_name, previous_name, name], inplace=True)
+        elif prev_previous_type == nn.Conv2d and previous_type == nn.BatchNorm2d:
+            # print("FUSED ", prev_previous_name, previous_name)
+            torch.quantization.fuse_modules(m, [prev_previous_name, previous_name], inplace=True)
+        # elif previous_type == nn.Conv2d and type(module) == nn.ReLU:
+        #    print("FUSED ", previous_name, name)
+        #    torch.quantization.fuse_modules(m, [previous_name, name], inplace=True)
+
+        prev_previous_type = previous_type
+        prev_previous_name = previous_name
+        previous_type = type(module)
+        previous_name = name

src/flux/annotator/midas/midas/transforms.py

@@ -0,0 +1,234 @@
+import numpy as np
+import cv2
+import math
+
+
+def apply_min_size(sample, size, image_interpolation_method=cv2.INTER_AREA):
+    """Rezise the sample to ensure the given size. Keeps aspect ratio.
+
+    Args:
+        sample (dict): sample
+        size (tuple): image size
+
+    Returns:
+        tuple: new size
+    """
+    shape = list(sample["disparity"].shape)
+
+    if shape[0] >= size[0] and shape[1] >= size[1]:
+        return sample
+
+    scale = [0, 0]
+    scale[0] = size[0] / shape[0]
+    scale[1] = size[1] / shape[1]
+
+    scale = max(scale)
+
+    shape[0] = math.ceil(scale * shape[0])
+    shape[1] = math.ceil(scale * shape[1])
+
+    # resize
+    sample["image"] = cv2.resize(
+        sample["image"], tuple(shape[::-1]), interpolation=image_interpolation_method
+    )
+
+    sample["disparity"] = cv2.resize(
+        sample["disparity"], tuple(shape[::-1]), interpolation=cv2.INTER_NEAREST
+    )
+    sample["mask"] = cv2.resize(
+        sample["mask"].astype(np.float32),
+        tuple(shape[::-1]),
+        interpolation=cv2.INTER_NEAREST,
+    )
+    sample["mask"] = sample["mask"].astype(bool)
+
+    return tuple(shape)
+
+
+class Resize(object):
+    """Resize sample to given size (width, height).
+    """
+
+    def __init__(
+        self,
+        width,
+        height,
+        resize_target=True,
+        keep_aspect_ratio=False,
+        ensure_multiple_of=1,
+        resize_method="lower_bound",
+        image_interpolation_method=cv2.INTER_AREA,
+    ):
+        """Init.
+
+        Args:
+            width (int): desired output width
+            height (int): desired output height
+            resize_target (bool, optional):
+                True: Resize the full sample (image, mask, target).
+                False: Resize image only.
+                Defaults to True.
+            keep_aspect_ratio (bool, optional):
+                True: Keep the aspect ratio of the input sample.
+                Output sample might not have the given width and height, and
+                resize behaviour depends on the parameter 'resize_method'.
+                Defaults to False.
+            ensure_multiple_of (int, optional):
+                Output width and height is constrained to be multiple of this parameter.
+                Defaults to 1.
+            resize_method (str, optional):
+                "lower_bound": Output will be at least as large as the given size.
+                "upper_bound": Output will be at max as large as the given size. (Output size might be smaller than given size.)
+                "minimal": Scale as least as possible.  (Output size might be smaller than given size.)
+                Defaults to "lower_bound".
+        """
+        self.__width = width
+        self.__height = height
+
+        self.__resize_target = resize_target
+        self.__keep_aspect_ratio = keep_aspect_ratio
+        self.__multiple_of = ensure_multiple_of
+        self.__resize_method = resize_method
+        self.__image_interpolation_method = image_interpolation_method
+
+    def constrain_to_multiple_of(self, x, min_val=0, max_val=None):
+        y = (np.round(x / self.__multiple_of) * self.__multiple_of).astype(int)
+
+        if max_val is not None and y > max_val:
+            y = (np.floor(x / self.__multiple_of) * self.__multiple_of).astype(int)
+
+        if y < min_val:
+            y = (np.ceil(x / self.__multiple_of) * self.__multiple_of).astype(int)
+
+        return y
+
+    def get_size(self, width, height):
+        # determine new height and width
+        scale_height = self.__height / height
+        scale_width = self.__width / width
+
+        if self.__keep_aspect_ratio:
+            if self.__resize_method == "lower_bound":
+                # scale such that output size is lower bound
+                if scale_width > scale_height:
+                    # fit width
+                    scale_height = scale_width
+                else:
+                    # fit height
+                    scale_width = scale_height
+            elif self.__resize_method == "upper_bound":
+                # scale such that output size is upper bound
+                if scale_width < scale_height:
+                    # fit width
+                    scale_height = scale_width
+                else:
+                    # fit height
+                    scale_width = scale_height
+            elif self.__resize_method == "minimal":
+                # scale as least as possbile
+                if abs(1 - scale_width) < abs(1 - scale_height):
+                    # fit width
+                    scale_height = scale_width
+                else:
+                    # fit height
+                    scale_width = scale_height
+            else:
+                raise ValueError(
+                    f"resize_method {self.__resize_method} not implemented"
+                )
+
+        if self.__resize_method == "lower_bound":
+            new_height = self.constrain_to_multiple_of(
+                scale_height * height, min_val=self.__height
+            )
+            new_width = self.constrain_to_multiple_of(
+                scale_width * width, min_val=self.__width
+            )
+        elif self.__resize_method == "upper_bound":
+            new_height = self.constrain_to_multiple_of(
+                scale_height * height, max_val=self.__height
+            )
+            new_width = self.constrain_to_multiple_of(
+                scale_width * width, max_val=self.__width
+            )
+        elif self.__resize_method == "minimal":
+            new_height = self.constrain_to_multiple_of(scale_height * height)
+            new_width = self.constrain_to_multiple_of(scale_width * width)
+        else:
+            raise ValueError(f"resize_method {self.__resize_method} not implemented")
+
+        return (new_width, new_height)
+
+    def __call__(self, sample):
+        width, height = self.get_size(
+            sample["image"].shape[1], sample["image"].shape[0]
+        )
+
+        # resize sample
+        sample["image"] = cv2.resize(
+            sample["image"],
+            (width, height),
+            interpolation=self.__image_interpolation_method,
+        )
+
+        if self.__resize_target:
+            if "disparity" in sample:
+                sample["disparity"] = cv2.resize(
+                    sample["disparity"],
+                    (width, height),
+                    interpolation=cv2.INTER_NEAREST,
+                )
+
+            if "depth" in sample:
+                sample["depth"] = cv2.resize(
+                    sample["depth"], (width, height), interpolation=cv2.INTER_NEAREST
+                )
+
+            sample["mask"] = cv2.resize(
+                sample["mask"].astype(np.float32),
+                (width, height),
+                interpolation=cv2.INTER_NEAREST,
+            )
+            sample["mask"] = sample["mask"].astype(bool)
+
+        return sample
+
+
+class NormalizeImage(object):
+    """Normlize image by given mean and std.
+    """
+
+    def __init__(self, mean, std):
+        self.__mean = mean
+        self.__std = std
+
+    def __call__(self, sample):
+        sample["image"] = (sample["image"] - self.__mean) / self.__std
+
+        return sample
+
+
+class PrepareForNet(object):
+    """Prepare sample for usage as network input.
+    """
+
+    def __init__(self):
+        pass
+
+    def __call__(self, sample):
+        image = np.transpose(sample["image"], (2, 0, 1))
+        sample["image"] = np.ascontiguousarray(image).astype(np.float32)
+
+        if "mask" in sample:
+            sample["mask"] = sample["mask"].astype(np.float32)
+            sample["mask"] = np.ascontiguousarray(sample["mask"])
+
+        if "disparity" in sample:
+            disparity = sample["disparity"].astype(np.float32)
+            sample["disparity"] = np.ascontiguousarray(disparity)
+
+        if "depth" in sample:
+            depth = sample["depth"].astype(np.float32)
+            sample["depth"] = np.ascontiguousarray(depth)
+
+        return sample

src/flux/annotator/midas/midas/vit.py

@@ -0,0 +1,491 @@
+import torch
+import torch.nn as nn
+import timm
+import types
+import math
+import torch.nn.functional as F
+
+
+class Slice(nn.Module):
+    def __init__(self, start_index=1):
+        super(Slice, self).__init__()
+        self.start_index = start_index
+
+    def forward(self, x):
+        return x[:, self.start_index :]
+
+
+class AddReadout(nn.Module):
+    def __init__(self, start_index=1):
+        super(AddReadout, self).__init__()
+        self.start_index = start_index
+
+    def forward(self, x):
+        if self.start_index == 2:
+            readout = (x[:, 0] + x[:, 1]) / 2
+        else:
+            readout = x[:, 0]
+        return x[:, self.start_index :] + readout.unsqueeze(1)
+
+
+class ProjectReadout(nn.Module):
+    def __init__(self, in_features, start_index=1):
+        super(ProjectReadout, self).__init__()
+        self.start_index = start_index
+
+        self.project = nn.Sequential(nn.Linear(2 * in_features, in_features), nn.GELU())
+
+    def forward(self, x):
+        readout = x[:, 0].unsqueeze(1).expand_as(x[:, self.start_index :])
+        features = torch.cat((x[:, self.start_index :], readout), -1)
+
+        return self.project(features)
+
+
+class Transpose(nn.Module):
+    def __init__(self, dim0, dim1):
+        super(Transpose, self).__init__()
+        self.dim0 = dim0
+        self.dim1 = dim1
+
+    def forward(self, x):
+        x = x.transpose(self.dim0, self.dim1)
+        return x
+
+
+def forward_vit(pretrained, x):
+    b, c, h, w = x.shape
+
+    glob = pretrained.model.forward_flex(x)
+
+    layer_1 = pretrained.activations["1"]
+    layer_2 = pretrained.activations["2"]
+    layer_3 = pretrained.activations["3"]
+    layer_4 = pretrained.activations["4"]
+
+    layer_1 = pretrained.act_postprocess1[0:2](layer_1)
+    layer_2 = pretrained.act_postprocess2[0:2](layer_2)
+    layer_3 = pretrained.act_postprocess3[0:2](layer_3)
+    layer_4 = pretrained.act_postprocess4[0:2](layer_4)
+
+    unflatten = nn.Sequential(
+        nn.Unflatten(
+            2,
+            torch.Size(
+                [
+                    h // pretrained.model.patch_size[1],
+                    w // pretrained.model.patch_size[0],
+                ]
+            ),
+        )
+    )
+
+    if layer_1.ndim == 3:
+        layer_1 = unflatten(layer_1)
+    if layer_2.ndim == 3:
+        layer_2 = unflatten(layer_2)
+    if layer_3.ndim == 3:
+        layer_3 = unflatten(layer_3)
+    if layer_4.ndim == 3:
+        layer_4 = unflatten(layer_4)
+
+    layer_1 = pretrained.act_postprocess1[3 : len(pretrained.act_postprocess1)](layer_1)
+    layer_2 = pretrained.act_postprocess2[3 : len(pretrained.act_postprocess2)](layer_2)
+    layer_3 = pretrained.act_postprocess3[3 : len(pretrained.act_postprocess3)](layer_3)
+    layer_4 = pretrained.act_postprocess4[3 : len(pretrained.act_postprocess4)](layer_4)
+
+    return layer_1, layer_2, layer_3, layer_4
+
+
+def _resize_pos_embed(self, posemb, gs_h, gs_w):
+    posemb_tok, posemb_grid = (
+        posemb[:, : self.start_index],
+        posemb[0, self.start_index :],
+    )
+
+    gs_old = int(math.sqrt(len(posemb_grid)))
+
+    posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2)
+    posemb_grid = F.interpolate(posemb_grid, size=(gs_h, gs_w), mode="bilinear")
+    posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_h * gs_w, -1)
+
+    posemb = torch.cat([posemb_tok, posemb_grid], dim=1)
+
+    return posemb
+
+
+def forward_flex(self, x):
+    b, c, h, w = x.shape
+
+    pos_embed = self._resize_pos_embed(
+        self.pos_embed, h // self.patch_size[1], w // self.patch_size[0]
+    )
+
+    B = x.shape[0]
+
+    if hasattr(self.patch_embed, "backbone"):
+        x = self.patch_embed.backbone(x)
+        if isinstance(x, (list, tuple)):
+            x = x[-1]  # last feature if backbone outputs list/tuple of features
+
+    x = self.patch_embed.proj(x).flatten(2).transpose(1, 2)
+
+    if getattr(self, "dist_token", None) is not None:
+        cls_tokens = self.cls_token.expand(
+            B, -1, -1
+        )  # stole cls_tokens impl from Phil Wang, thanks
+        dist_token = self.dist_token.expand(B, -1, -1)
+        x = torch.cat((cls_tokens, dist_token, x), dim=1)
+    else:
+        cls_tokens = self.cls_token.expand(
+            B, -1, -1
+        )  # stole cls_tokens impl from Phil Wang, thanks
+        x = torch.cat((cls_tokens, x), dim=1)
+
+    x = x + pos_embed
+    x = self.pos_drop(x)
+
+    for blk in self.blocks:
+        x = blk(x)
+
+    x = self.norm(x)
+
+    return x
+
+
+activations = {}
+
+
+def get_activation(name):
+    def hook(model, input, output):
+        activations[name] = output
+
+    return hook
+
+
+def get_readout_oper(vit_features, features, use_readout, start_index=1):
+    if use_readout == "ignore":
+        readout_oper = [Slice(start_index)] * len(features)
+    elif use_readout == "add":
+        readout_oper = [AddReadout(start_index)] * len(features)
+    elif use_readout == "project":
+        readout_oper = [
+            ProjectReadout(vit_features, start_index) for out_feat in features
+        ]
+    else:
+        assert (
+            False
+        ), "wrong operation for readout token, use_readout can be 'ignore', 'add', or 'project'"
+
+    return readout_oper
+
+
+def _make_vit_b16_backbone(
+    model,
+    features=[96, 192, 384, 768],
+    size=[384, 384],
+    hooks=[2, 5, 8, 11],
+    vit_features=768,
+    use_readout="ignore",
+    start_index=1,
+):
+    pretrained = nn.Module()
+
+    pretrained.model = model
+    pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1"))
+    pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2"))
+    pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3"))
+    pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4"))
+
+    pretrained.activations = activations
+
+    readout_oper = get_readout_oper(vit_features, features, use_readout, start_index)
+
+    # 32, 48, 136, 384
+    pretrained.act_postprocess1 = nn.Sequential(
+        readout_oper[0],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[0],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+        nn.ConvTranspose2d(
+            in_channels=features[0],
+            out_channels=features[0],
+            kernel_size=4,
+            stride=4,
+            padding=0,
+            bias=True,
+            dilation=1,
+            groups=1,
+        ),
+    )
+
+    pretrained.act_postprocess2 = nn.Sequential(
+        readout_oper[1],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[1],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+        nn.ConvTranspose2d(
+            in_channels=features[1],
+            out_channels=features[1],
+            kernel_size=2,
+            stride=2,
+            padding=0,
+            bias=True,
+            dilation=1,
+            groups=1,
+        ),
+    )
+
+    pretrained.act_postprocess3 = nn.Sequential(
+        readout_oper[2],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[2],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+    )
+
+    pretrained.act_postprocess4 = nn.Sequential(
+        readout_oper[3],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[3],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+        nn.Conv2d(
+            in_channels=features[3],
+            out_channels=features[3],
+            kernel_size=3,
+            stride=2,
+            padding=1,
+        ),
+    )
+
+    pretrained.model.start_index = start_index
+    pretrained.model.patch_size = [16, 16]
+
+    # We inject this function into the VisionTransformer instances so that
+    # we can use it with interpolated position embeddings without modifying the library source.
+    pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model)
+    pretrained.model._resize_pos_embed = types.MethodType(
+        _resize_pos_embed, pretrained.model
+    )
+
+    return pretrained
+
+
+def _make_pretrained_vitl16_384(pretrained, use_readout="ignore", hooks=None):
+    model = timm.create_model("vit_large_patch16_384", pretrained=pretrained)
+
+    hooks = [5, 11, 17, 23] if hooks == None else hooks
+    return _make_vit_b16_backbone(
+        model,
+        features=[256, 512, 1024, 1024],
+        hooks=hooks,
+        vit_features=1024,
+        use_readout=use_readout,
+    )
+
+
+def _make_pretrained_vitb16_384(pretrained, use_readout="ignore", hooks=None):
+    model = timm.create_model("vit_base_patch16_384", pretrained=pretrained)
+
+    hooks = [2, 5, 8, 11] if hooks == None else hooks
+    return _make_vit_b16_backbone(
+        model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout
+    )
+
+
+def _make_pretrained_deitb16_384(pretrained, use_readout="ignore", hooks=None):
+    model = timm.create_model("vit_deit_base_patch16_384", pretrained=pretrained)
+
+    hooks = [2, 5, 8, 11] if hooks == None else hooks
+    return _make_vit_b16_backbone(
+        model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout
+    )
+
+
+def _make_pretrained_deitb16_distil_384(pretrained, use_readout="ignore", hooks=None):
+    model = timm.create_model(
+        "vit_deit_base_distilled_patch16_384", pretrained=pretrained
+    )
+
+    hooks = [2, 5, 8, 11] if hooks == None else hooks
+    return _make_vit_b16_backbone(
+        model,
+        features=[96, 192, 384, 768],
+        hooks=hooks,
+        use_readout=use_readout,
+        start_index=2,
+    )
+
+
+def _make_vit_b_rn50_backbone(
+    model,
+    features=[256, 512, 768, 768],
+    size=[384, 384],
+    hooks=[0, 1, 8, 11],
+    vit_features=768,
+    use_vit_only=False,
+    use_readout="ignore",
+    start_index=1,
+):
+    pretrained = nn.Module()
+
+    pretrained.model = model
+
+    if use_vit_only == True:
+        pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1"))
+        pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2"))
+    else:
+        pretrained.model.patch_embed.backbone.stages[0].register_forward_hook(
+            get_activation("1")
+        )
+        pretrained.model.patch_embed.backbone.stages[1].register_forward_hook(
+            get_activation("2")
+        )
+
+    pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3"))
+    pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4"))
+
+    pretrained.activations = activations
+
+    readout_oper = get_readout_oper(vit_features, features, use_readout, start_index)
+
+    if use_vit_only == True:
+        pretrained.act_postprocess1 = nn.Sequential(
+            readout_oper[0],
+            Transpose(1, 2),
+            nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+            nn.Conv2d(
+                in_channels=vit_features,
+                out_channels=features[0],
+                kernel_size=1,
+                stride=1,
+                padding=0,
+            ),
+            nn.ConvTranspose2d(
+                in_channels=features[0],
+                out_channels=features[0],
+                kernel_size=4,
+                stride=4,
+                padding=0,
+                bias=True,
+                dilation=1,
+                groups=1,
+            ),
+        )
+
+        pretrained.act_postprocess2 = nn.Sequential(
+            readout_oper[1],
+            Transpose(1, 2),
+            nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+            nn.Conv2d(
+                in_channels=vit_features,
+                out_channels=features[1],
+                kernel_size=1,
+                stride=1,
+                padding=0,
+            ),
+            nn.ConvTranspose2d(
+                in_channels=features[1],
+                out_channels=features[1],
+                kernel_size=2,
+                stride=2,
+                padding=0,
+                bias=True,
+                dilation=1,
+                groups=1,
+            ),
+        )
+    else:
+        pretrained.act_postprocess1 = nn.Sequential(
+            nn.Identity(), nn.Identity(), nn.Identity()
+        )
+        pretrained.act_postprocess2 = nn.Sequential(
+            nn.Identity(), nn.Identity(), nn.Identity()
+        )
+
+    pretrained.act_postprocess3 = nn.Sequential(
+        readout_oper[2],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[2],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+    )
+
+    pretrained.act_postprocess4 = nn.Sequential(
+        readout_oper[3],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[3],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+        nn.Conv2d(
+            in_channels=features[3],
+            out_channels=features[3],
+            kernel_size=3,
+            stride=2,
+            padding=1,
+        ),
+    )
+
+    pretrained.model.start_index = start_index
+    pretrained.model.patch_size = [16, 16]
+
+    # We inject this function into the VisionTransformer instances so that
+    # we can use it with interpolated position embeddings without modifying the library source.
+    pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model)
+
+    # We inject this function into the VisionTransformer instances so that
+    # we can use it with interpolated position embeddings without modifying the library source.
+    pretrained.model._resize_pos_embed = types.MethodType(
+        _resize_pos_embed, pretrained.model
+    )
+
+    return pretrained
+
+
+def _make_pretrained_vitb_rn50_384(
+    pretrained, use_readout="ignore", hooks=None, use_vit_only=False
+):
+    model = timm.create_model("vit_base_resnet50_384", pretrained=pretrained)
+
+    hooks = [0, 1, 8, 11] if hooks == None else hooks
+    return _make_vit_b_rn50_backbone(
+        model,
+        features=[256, 512, 768, 768],
+        size=[384, 384],
+        hooks=hooks,
+        use_vit_only=use_vit_only,
+        use_readout=use_readout,
+    )

src/flux/annotator/midas/utils.py

@@ -0,0 +1,189 @@
+"""Utils for monoDepth."""
+import sys
+import re
+import numpy as np
+import cv2
+import torch
+
+
+def read_pfm(path):
+    """Read pfm file.
+
+    Args:
+        path (str): path to file
+
+    Returns:
+        tuple: (data, scale)
+    """
+    with open(path, "rb") as file:
+
+        color = None
+        width = None
+        height = None
+        scale = None
+        endian = None
+
+        header = file.readline().rstrip()
+        if header.decode("ascii") == "PF":
+            color = True
+        elif header.decode("ascii") == "Pf":
+            color = False
+        else:
+            raise Exception("Not a PFM file: " + path)
+
+        dim_match = re.match(r"^(\d+)\s(\d+)\s$", file.readline().decode("ascii"))
+        if dim_match:
+            width, height = list(map(int, dim_match.groups()))
+        else:
+            raise Exception("Malformed PFM header.")
+
+        scale = float(file.readline().decode("ascii").rstrip())
+        if scale < 0:
+            # little-endian
+            endian = "<"
+            scale = -scale
+        else:
+            # big-endian
+            endian = ">"
+
+        data = np.fromfile(file, endian + "f")
+        shape = (height, width, 3) if color else (height, width)
+
+        data = np.reshape(data, shape)
+        data = np.flipud(data)
+
+        return data, scale
+
+
+def write_pfm(path, image, scale=1):
+    """Write pfm file.
+
+    Args:
+        path (str): pathto file
+        image (array): data
+        scale (int, optional): Scale. Defaults to 1.
+    """
+
+    with open(path, "wb") as file:
+        color = None
+
+        if image.dtype.name != "float32":
+            raise Exception("Image dtype must be float32.")
+
+        image = np.flipud(image)
+
+        if len(image.shape) == 3 and image.shape[2] == 3:  # color image
+            color = True
+        elif (
+            len(image.shape) == 2 or len(image.shape) == 3 and image.shape[2] == 1
+        ):  # greyscale
+            color = False
+        else:
+            raise Exception("Image must have H x W x 3, H x W x 1 or H x W dimensions.")
+
+        file.write("PF\n" if color else "Pf\n".encode())
+        file.write("%d %d\n".encode() % (image.shape[1], image.shape[0]))
+
+        endian = image.dtype.byteorder
+
+        if endian == "<" or endian == "=" and sys.byteorder == "little":
+            scale = -scale
+
+        file.write("%f\n".encode() % scale)
+
+        image.tofile(file)
+
+
+def read_image(path):
+    """Read image and output RGB image (0-1).
+
+    Args:
+        path (str): path to file
+
+    Returns:
+        array: RGB image (0-1)
+    """
+    img = cv2.imread(path)
+
+    if img.ndim == 2:
+        img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
+
+    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) / 255.0
+
+    return img
+
+
+def resize_image(img):
+    """Resize image and make it fit for network.
+
+    Args:
+        img (array): image
+
+    Returns:
+        tensor: data ready for network
+    """
+    height_orig = img.shape[0]
+    width_orig = img.shape[1]
+
+    if width_orig > height_orig:
+        scale = width_orig / 384
+    else:
+        scale = height_orig / 384
+
+    height = (np.ceil(height_orig / scale / 32) * 32).astype(int)
+    width = (np.ceil(width_orig / scale / 32) * 32).astype(int)
+
+    img_resized = cv2.resize(img, (width, height), interpolation=cv2.INTER_AREA)
+
+    img_resized = (
+        torch.from_numpy(np.transpose(img_resized, (2, 0, 1))).contiguous().float()
+    )
+    img_resized = img_resized.unsqueeze(0)
+
+    return img_resized
+
+
+def resize_depth(depth, width, height):
+    """Resize depth map and bring to CPU (numpy).
+
+    Args:
+        depth (tensor): depth
+        width (int): image width
+        height (int): image height
+
+    Returns:
+        array: processed depth
+    """
+    depth = torch.squeeze(depth[0, :, :, :]).to("cpu")
+
+    depth_resized = cv2.resize(
+        depth.numpy(), (width, height), interpolation=cv2.INTER_CUBIC
+    )
+
+    return depth_resized
+
+def write_depth(path, depth, bits=1):
+    """Write depth map to pfm and png file.
+
+    Args:
+        path (str): filepath without extension
+        depth (array): depth
+    """
+    write_pfm(path + ".pfm", depth.astype(np.float32))
+
+    depth_min = depth.min()
+    depth_max = depth.max()
+
+    max_val = (2**(8*bits))-1
+
+    if depth_max - depth_min > np.finfo("float").eps:
+        out = max_val * (depth - depth_min) / (depth_max - depth_min)
+    else:
+        out = np.zeros(depth.shape, dtype=depth.type)
+
+    if bits == 1:
+        cv2.imwrite(path + ".png", out.astype("uint8"))
+    elif bits == 2:
+        cv2.imwrite(path + ".png", out.astype("uint16"))
+
+    return

src/flux/annotator/mlsd/LICENSE

@@ -0,0 +1,201 @@
+                                Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "{}"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2021-present NAVER Corp.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

src/flux/annotator/mlsd/__init__.py

@@ -0,0 +1,40 @@
+# MLSD Line Detection
+# From https://github.com/navervision/mlsd
+# Apache-2.0 license
+
+import cv2
+import numpy as np
+import torch
+import os
+
+from einops import rearrange
+from huggingface_hub import hf_hub_download
+from .models.mbv2_mlsd_tiny import MobileV2_MLSD_Tiny
+from .models.mbv2_mlsd_large import MobileV2_MLSD_Large
+from .utils import pred_lines
+
+from ...annotator.util import annotator_ckpts_path
+
+
+class MLSDdetector:
+    def __init__(self):
+        model_path = os.path.join(annotator_ckpts_path, "mlsd_large_512_fp32.pth")
+        if not os.path.exists(model_path):
+            model_path = hf_hub_download("lllyasviel/Annotators", "mlsd_large_512_fp32.pth")
+        model = MobileV2_MLSD_Large()
+        model.load_state_dict(torch.load(model_path), strict=True)
+        self.model = model.cuda().eval()
+
+    def __call__(self, input_image, thr_v, thr_d):
+        assert input_image.ndim == 3
+        img = input_image
+        img_output = np.zeros_like(img)
+        try:
+            with torch.no_grad():
+                lines = pred_lines(img, self.model, [img.shape[0], img.shape[1]], thr_v, thr_d)
+                for line in lines:
+                    x_start, y_start, x_end, y_end = [int(val) for val in line]
+                    cv2.line(img_output, (x_start, y_start), (x_end, y_end), [255, 255, 255], 1)
+        except Exception as e:
+            pass
+        return img_output[:, :, 0]

src/flux/annotator/mlsd/models/mbv2_mlsd_large.py

@@ -0,0 +1,292 @@
+import os
+import sys
+import torch
+import torch.nn as nn
+import torch.utils.model_zoo as model_zoo
+from  torch.nn import  functional as F
+
+
+class BlockTypeA(nn.Module):
+    def __init__(self, in_c1, in_c2, out_c1, out_c2, upscale = True):
+        super(BlockTypeA, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.Conv2d(in_c2, out_c2, kernel_size=1),
+            nn.BatchNorm2d(out_c2),
+            nn.ReLU(inplace=True)
+        )
+        self.conv2 = nn.Sequential(
+            nn.Conv2d(in_c1, out_c1, kernel_size=1),
+            nn.BatchNorm2d(out_c1),
+            nn.ReLU(inplace=True)
+        )
+        self.upscale = upscale
+
+    def forward(self, a, b):
+        b = self.conv1(b)
+        a = self.conv2(a)
+        if self.upscale:
+             b = F.interpolate(b, scale_factor=2.0, mode='bilinear', align_corners=True)
+        return torch.cat((a, b), dim=1)
+
+
+class BlockTypeB(nn.Module):
+    def __init__(self, in_c, out_c):
+        super(BlockTypeB, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.Conv2d(in_c, in_c,  kernel_size=3, padding=1),
+            nn.BatchNorm2d(in_c),
+            nn.ReLU()
+        )
+        self.conv2 = nn.Sequential(
+            nn.Conv2d(in_c, out_c, kernel_size=3, padding=1),
+            nn.BatchNorm2d(out_c),
+            nn.ReLU()
+        )
+
+    def forward(self, x):
+        x = self.conv1(x) + x
+        x = self.conv2(x)
+        return x
+
+class BlockTypeC(nn.Module):
+    def __init__(self, in_c, out_c):
+        super(BlockTypeC, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.Conv2d(in_c, in_c,  kernel_size=3, padding=5, dilation=5),
+            nn.BatchNorm2d(in_c),
+            nn.ReLU()
+        )
+        self.conv2 = nn.Sequential(
+            nn.Conv2d(in_c, in_c,  kernel_size=3, padding=1),
+            nn.BatchNorm2d(in_c),
+            nn.ReLU()
+        )
+        self.conv3 = nn.Conv2d(in_c, out_c, kernel_size=1)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.conv2(x)
+        x = self.conv3(x)
+        return x
+
+def _make_divisible(v, divisor, min_value=None):
+    """
+    This function is taken from the original tf repo.
+    It ensures that all layers have a channel number that is divisible by 8
+    It can be seen here:
+    https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet.py
+    :param v:
+    :param divisor:
+    :param min_value:
+    :return:
+    """
+    if min_value is None:
+        min_value = divisor
+    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
+    # Make sure that round down does not go down by more than 10%.
+    if new_v < 0.9 * v:
+        new_v += divisor
+    return new_v
+
+
+class ConvBNReLU(nn.Sequential):
+    def __init__(self, in_planes, out_planes, kernel_size=3, stride=1, groups=1):
+        self.channel_pad = out_planes - in_planes
+        self.stride = stride
+        #padding = (kernel_size - 1) // 2
+
+        # TFLite uses slightly different padding than PyTorch
+        if stride == 2:
+            padding = 0
+        else:
+            padding = (kernel_size - 1) // 2
+
+        super(ConvBNReLU, self).__init__(
+            nn.Conv2d(in_planes, out_planes, kernel_size, stride, padding, groups=groups, bias=False),
+            nn.BatchNorm2d(out_planes),
+            nn.ReLU6(inplace=True)
+        )
+        self.max_pool = nn.MaxPool2d(kernel_size=stride, stride=stride)
+
+
+    def forward(self, x):
+        # TFLite uses  different padding
+        if self.stride == 2:
+            x = F.pad(x, (0, 1, 0, 1), "constant", 0)
+            #print(x.shape)
+
+        for module in self:
+            if not isinstance(module, nn.MaxPool2d):
+                x = module(x)
+        return x
+
+
+class InvertedResidual(nn.Module):
+    def __init__(self, inp, oup, stride, expand_ratio):
+        super(InvertedResidual, self).__init__()
+        self.stride = stride
+        assert stride in [1, 2]
+
+        hidden_dim = int(round(inp * expand_ratio))
+        self.use_res_connect = self.stride == 1 and inp == oup
+
+        layers = []
+        if expand_ratio != 1:
+            # pw
+            layers.append(ConvBNReLU(inp, hidden_dim, kernel_size=1))
+        layers.extend([
+            # dw
+            ConvBNReLU(hidden_dim, hidden_dim, stride=stride, groups=hidden_dim),
+            # pw-linear
+            nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
+            nn.BatchNorm2d(oup),
+        ])
+        self.conv = nn.Sequential(*layers)
+
+    def forward(self, x):
+        if self.use_res_connect:
+            return x + self.conv(x)
+        else:
+            return self.conv(x)
+
+
+class MobileNetV2(nn.Module):
+    def __init__(self, pretrained=True):
+        """
+        MobileNet V2 main class
+        Args:
+            num_classes (int): Number of classes
+            width_mult (float): Width multiplier - adjusts number of channels in each layer by this amount
+            inverted_residual_setting: Network structure
+            round_nearest (int): Round the number of channels in each layer to be a multiple of this number
+            Set to 1 to turn off rounding
+            block: Module specifying inverted residual building block for mobilenet
+        """
+        super(MobileNetV2, self).__init__()
+
+        block = InvertedResidual
+        input_channel = 32
+        last_channel = 1280
+        width_mult = 1.0
+        round_nearest = 8
+
+        inverted_residual_setting = [
+            # t, c, n, s
+            [1, 16, 1, 1],
+            [6, 24, 2, 2],
+            [6, 32, 3, 2],
+            [6, 64, 4, 2],
+            [6, 96, 3, 1],
+            #[6, 160, 3, 2],
+            #[6, 320, 1, 1],
+        ]
+
+        # only check the first element, assuming user knows t,c,n,s are required
+        if len(inverted_residual_setting) == 0 or len(inverted_residual_setting[0]) != 4:
+            raise ValueError("inverted_residual_setting should be non-empty "
+                             "or a 4-element list, got {}".format(inverted_residual_setting))
+
+        # building first layer
+        input_channel = _make_divisible(input_channel * width_mult, round_nearest)
+        self.last_channel = _make_divisible(last_channel * max(1.0, width_mult), round_nearest)
+        features = [ConvBNReLU(4, input_channel, stride=2)]
+        # building inverted residual blocks
+        for t, c, n, s in inverted_residual_setting:
+            output_channel = _make_divisible(c * width_mult, round_nearest)
+            for i in range(n):
+                stride = s if i == 0 else 1
+                features.append(block(input_channel, output_channel, stride, expand_ratio=t))
+                input_channel = output_channel
+
+        self.features = nn.Sequential(*features)
+        self.fpn_selected = [1, 3, 6, 10, 13]
+        # weight initialization
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out')
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+            elif isinstance(m, nn.BatchNorm2d):
+                nn.init.ones_(m.weight)
+                nn.init.zeros_(m.bias)
+            elif isinstance(m, nn.Linear):
+                nn.init.normal_(m.weight, 0, 0.01)
+                nn.init.zeros_(m.bias)
+        if pretrained:
+           self._load_pretrained_model()
+
+    def _forward_impl(self, x):
+        # This exists since TorchScript doesn't support inheritance, so the superclass method
+        # (this one) needs to have a name other than `forward` that can be accessed in a subclass
+        fpn_features = []
+        for i, f in enumerate(self.features):
+            if i > self.fpn_selected[-1]:
+                break
+            x = f(x)
+            if i in self.fpn_selected:
+                fpn_features.append(x)
+
+        c1, c2, c3, c4, c5 = fpn_features
+        return c1, c2, c3, c4, c5
+
+
+    def forward(self, x):
+        return self._forward_impl(x)
+
+    def _load_pretrained_model(self):
+        pretrain_dict = model_zoo.load_url('https://download.pytorch.org/models/mobilenet_v2-b0353104.pth')
+        model_dict = {}
+        state_dict = self.state_dict()
+        for k, v in pretrain_dict.items():
+            if k in state_dict:
+                model_dict[k] = v
+        state_dict.update(model_dict)
+        self.load_state_dict(state_dict)
+
+
+class MobileV2_MLSD_Large(nn.Module):
+    def __init__(self):
+        super(MobileV2_MLSD_Large, self).__init__()
+
+        self.backbone = MobileNetV2(pretrained=False)
+        ## A, B
+        self.block15 = BlockTypeA(in_c1= 64, in_c2= 96,
+                                  out_c1= 64, out_c2=64,
+                                  upscale=False)
+        self.block16 = BlockTypeB(128, 64)
+
+        ## A, B
+        self.block17 = BlockTypeA(in_c1 = 32,  in_c2 = 64,
+                                  out_c1= 64,  out_c2= 64)
+        self.block18 = BlockTypeB(128, 64)
+
+        ## A, B
+        self.block19 = BlockTypeA(in_c1=24, in_c2=64,
+                                  out_c1=64, out_c2=64)
+        self.block20 = BlockTypeB(128, 64)
+
+        ## A, B, C
+        self.block21 = BlockTypeA(in_c1=16, in_c2=64,
+                                  out_c1=64, out_c2=64)
+        self.block22 = BlockTypeB(128, 64)
+
+        self.block23 = BlockTypeC(64, 16)
+
+    def forward(self, x):
+        c1, c2, c3, c4, c5 = self.backbone(x)
+
+        x = self.block15(c4, c5)
+        x = self.block16(x)
+
+        x = self.block17(c3, x)
+        x = self.block18(x)
+
+        x = self.block19(c2, x)
+        x = self.block20(x)
+
+        x = self.block21(c1, x)
+        x = self.block22(x)
+        x = self.block23(x)
+        x = x[:, 7:, :, :]
+
+        return x

src/flux/annotator/mlsd/models/mbv2_mlsd_tiny.py

@@ -0,0 +1,275 @@
+import os
+import sys
+import torch
+import torch.nn as nn
+import torch.utils.model_zoo as model_zoo
+from  torch.nn import  functional as F
+
+
+class BlockTypeA(nn.Module):
+    def __init__(self, in_c1, in_c2, out_c1, out_c2, upscale = True):
+        super(BlockTypeA, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.Conv2d(in_c2, out_c2, kernel_size=1),
+            nn.BatchNorm2d(out_c2),
+            nn.ReLU(inplace=True)
+        )
+        self.conv2 = nn.Sequential(
+            nn.Conv2d(in_c1, out_c1, kernel_size=1),
+            nn.BatchNorm2d(out_c1),
+            nn.ReLU(inplace=True)
+        )
+        self.upscale = upscale
+
+    def forward(self, a, b):
+        b = self.conv1(b)
+        a = self.conv2(a)
+        b = F.interpolate(b, scale_factor=2.0, mode='bilinear', align_corners=True)
+        return torch.cat((a, b), dim=1)
+
+
+class BlockTypeB(nn.Module):
+    def __init__(self, in_c, out_c):
+        super(BlockTypeB, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.Conv2d(in_c, in_c,  kernel_size=3, padding=1),
+            nn.BatchNorm2d(in_c),
+            nn.ReLU()
+        )
+        self.conv2 = nn.Sequential(
+            nn.Conv2d(in_c, out_c, kernel_size=3, padding=1),
+            nn.BatchNorm2d(out_c),
+            nn.ReLU()
+        )
+
+    def forward(self, x):
+        x = self.conv1(x) + x
+        x = self.conv2(x)
+        return x
+
+class BlockTypeC(nn.Module):
+    def __init__(self, in_c, out_c):
+        super(BlockTypeC, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.Conv2d(in_c, in_c,  kernel_size=3, padding=5, dilation=5),
+            nn.BatchNorm2d(in_c),
+            nn.ReLU()
+        )
+        self.conv2 = nn.Sequential(
+            nn.Conv2d(in_c, in_c,  kernel_size=3, padding=1),
+            nn.BatchNorm2d(in_c),
+            nn.ReLU()
+        )
+        self.conv3 = nn.Conv2d(in_c, out_c, kernel_size=1)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.conv2(x)
+        x = self.conv3(x)
+        return x
+
+def _make_divisible(v, divisor, min_value=None):
+    """
+    This function is taken from the original tf repo.
+    It ensures that all layers have a channel number that is divisible by 8
+    It can be seen here:
+    https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet.py
+    :param v:
+    :param divisor:
+    :param min_value:
+    :return:
+    """
+    if min_value is None:
+        min_value = divisor
+    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
+    # Make sure that round down does not go down by more than 10%.
+    if new_v < 0.9 * v:
+        new_v += divisor
+    return new_v
+
+
+class ConvBNReLU(nn.Sequential):
+    def __init__(self, in_planes, out_planes, kernel_size=3, stride=1, groups=1):
+        self.channel_pad = out_planes - in_planes
+        self.stride = stride
+        #padding = (kernel_size - 1) // 2
+
+        # TFLite uses slightly different padding than PyTorch
+        if stride == 2:
+            padding = 0
+        else:
+            padding = (kernel_size - 1) // 2
+
+        super(ConvBNReLU, self).__init__(
+            nn.Conv2d(in_planes, out_planes, kernel_size, stride, padding, groups=groups, bias=False),
+            nn.BatchNorm2d(out_planes),
+            nn.ReLU6(inplace=True)
+        )
+        self.max_pool = nn.MaxPool2d(kernel_size=stride, stride=stride)
+
+
+    def forward(self, x):
+        # TFLite uses  different padding
+        if self.stride == 2:
+            x = F.pad(x, (0, 1, 0, 1), "constant", 0)
+            #print(x.shape)
+
+        for module in self:
+            if not isinstance(module, nn.MaxPool2d):
+                x = module(x)
+        return x
+
+
+class InvertedResidual(nn.Module):
+    def __init__(self, inp, oup, stride, expand_ratio):
+        super(InvertedResidual, self).__init__()
+        self.stride = stride
+        assert stride in [1, 2]
+
+        hidden_dim = int(round(inp * expand_ratio))
+        self.use_res_connect = self.stride == 1 and inp == oup
+
+        layers = []
+        if expand_ratio != 1:
+            # pw
+            layers.append(ConvBNReLU(inp, hidden_dim, kernel_size=1))
+        layers.extend([
+            # dw
+            ConvBNReLU(hidden_dim, hidden_dim, stride=stride, groups=hidden_dim),
+            # pw-linear
+            nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
+            nn.BatchNorm2d(oup),
+        ])
+        self.conv = nn.Sequential(*layers)
+
+    def forward(self, x):
+        if self.use_res_connect:
+            return x + self.conv(x)
+        else:
+            return self.conv(x)
+
+
+class MobileNetV2(nn.Module):
+    def __init__(self, pretrained=True):
+        """
+        MobileNet V2 main class
+        Args:
+            num_classes (int): Number of classes
+            width_mult (float): Width multiplier - adjusts number of channels in each layer by this amount
+            inverted_residual_setting: Network structure
+            round_nearest (int): Round the number of channels in each layer to be a multiple of this number
+            Set to 1 to turn off rounding
+            block: Module specifying inverted residual building block for mobilenet
+        """
+        super(MobileNetV2, self).__init__()
+
+        block = InvertedResidual
+        input_channel = 32
+        last_channel = 1280
+        width_mult = 1.0
+        round_nearest = 8
+
+        inverted_residual_setting = [
+            # t, c, n, s
+            [1, 16, 1, 1],
+            [6, 24, 2, 2],
+            [6, 32, 3, 2],
+            [6, 64, 4, 2],
+            #[6, 96, 3, 1],
+            #[6, 160, 3, 2],
+            #[6, 320, 1, 1],
+        ]
+
+        # only check the first element, assuming user knows t,c,n,s are required
+        if len(inverted_residual_setting) == 0 or len(inverted_residual_setting[0]) != 4:
+            raise ValueError("inverted_residual_setting should be non-empty "
+                             "or a 4-element list, got {}".format(inverted_residual_setting))
+
+        # building first layer
+        input_channel = _make_divisible(input_channel * width_mult, round_nearest)
+        self.last_channel = _make_divisible(last_channel * max(1.0, width_mult), round_nearest)
+        features = [ConvBNReLU(4, input_channel, stride=2)]
+        # building inverted residual blocks
+        for t, c, n, s in inverted_residual_setting:
+            output_channel = _make_divisible(c * width_mult, round_nearest)
+            for i in range(n):
+                stride = s if i == 0 else 1
+                features.append(block(input_channel, output_channel, stride, expand_ratio=t))
+                input_channel = output_channel
+        self.features = nn.Sequential(*features)
+
+        self.fpn_selected = [3, 6, 10]
+        # weight initialization
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out')
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+            elif isinstance(m, nn.BatchNorm2d):
+                nn.init.ones_(m.weight)
+                nn.init.zeros_(m.bias)
+            elif isinstance(m, nn.Linear):
+                nn.init.normal_(m.weight, 0, 0.01)
+                nn.init.zeros_(m.bias)
+
+        #if pretrained:
+        #    self._load_pretrained_model()
+
+    def _forward_impl(self, x):
+        # This exists since TorchScript doesn't support inheritance, so the superclass method
+        # (this one) needs to have a name other than `forward` that can be accessed in a subclass
+        fpn_features = []
+        for i, f in enumerate(self.features):
+            if i > self.fpn_selected[-1]:
+                break
+            x = f(x)
+            if i in self.fpn_selected:
+                fpn_features.append(x)
+
+        c2, c3, c4 = fpn_features
+        return c2, c3, c4
+
+
+    def forward(self, x):
+        return self._forward_impl(x)
+
+    def _load_pretrained_model(self):
+        pretrain_dict = model_zoo.load_url('https://download.pytorch.org/models/mobilenet_v2-b0353104.pth')
+        model_dict = {}
+        state_dict = self.state_dict()
+        for k, v in pretrain_dict.items():
+            if k in state_dict:
+                model_dict[k] = v
+        state_dict.update(model_dict)
+        self.load_state_dict(state_dict)
+
+
+class MobileV2_MLSD_Tiny(nn.Module):
+    def __init__(self):
+        super(MobileV2_MLSD_Tiny, self).__init__()
+
+        self.backbone = MobileNetV2(pretrained=True)
+
+        self.block12 = BlockTypeA(in_c1= 32, in_c2= 64,
+                                  out_c1= 64, out_c2=64)
+        self.block13 = BlockTypeB(128, 64)
+
+        self.block14 = BlockTypeA(in_c1 = 24,  in_c2 = 64,
+                                  out_c1= 32,  out_c2= 32)
+        self.block15 = BlockTypeB(64, 64)
+
+        self.block16 = BlockTypeC(64, 16)
+
+    def forward(self, x):
+        c2, c3, c4 = self.backbone(x)
+
+        x = self.block12(c3, c4)
+        x = self.block13(x)
+        x = self.block14(c2, x)
+        x = self.block15(x)
+        x = self.block16(x)
+        x = x[:, 7:, :, :]
+        #print(x.shape)
+        x = F.interpolate(x, scale_factor=2.0, mode='bilinear', align_corners=True)
+
+        return x

src/flux/annotator/mlsd/utils.py

@@ -0,0 +1,580 @@
+'''
+modified by  lihaoweicv
+pytorch version
+'''
+
+'''
+M-LSD
+Copyright 2021-present NAVER Corp.
+Apache License v2.0
+'''
+
+import os
+import numpy as np
+import cv2
+import torch
+from  torch.nn import  functional as F
+
+
+def deccode_output_score_and_ptss(tpMap, topk_n = 200, ksize = 5):
+    '''
+    tpMap:
+    center: tpMap[1, 0, :, :]
+    displacement: tpMap[1, 1:5, :, :]
+    '''
+    b, c, h, w = tpMap.shape
+    assert  b==1, 'only support bsize==1'
+    displacement = tpMap[:, 1:5, :, :][0]
+    center = tpMap[:, 0, :, :]
+    heat = torch.sigmoid(center)
+    hmax = F.max_pool2d( heat, (ksize, ksize), stride=1, padding=(ksize-1)//2)
+    keep = (hmax == heat).float()
+    heat = heat * keep
+    heat = heat.reshape(-1, )
+
+    scores, indices = torch.topk(heat, topk_n, dim=-1, largest=True)
+    yy = torch.floor_divide(indices, w).unsqueeze(-1)
+    xx = torch.fmod(indices, w).unsqueeze(-1)
+    ptss = torch.cat((yy, xx),dim=-1)
+
+    ptss   = ptss.detach().cpu().numpy()
+    scores = scores.detach().cpu().numpy()
+    displacement = displacement.detach().cpu().numpy()
+    displacement = displacement.transpose((1,2,0))
+    return  ptss, scores, displacement
+
+
+def pred_lines(image, model,
+               input_shape=[512, 512],
+               score_thr=0.10,
+               dist_thr=20.0):
+    h, w, _ = image.shape
+    h_ratio, w_ratio = [h / input_shape[0], w / input_shape[1]]
+
+    resized_image = np.concatenate([cv2.resize(image, (input_shape[1], input_shape[0]), interpolation=cv2.INTER_AREA),
+                                    np.ones([input_shape[0], input_shape[1], 1])], axis=-1)
+
+    resized_image = resized_image.transpose((2,0,1))
+    batch_image = np.expand_dims(resized_image, axis=0).astype('float32')
+    batch_image = (batch_image / 127.5) - 1.0
+
+    batch_image = torch.from_numpy(batch_image).float().to("cuda:4")
+    outputs = model(batch_image)
+    pts, pts_score, vmap = deccode_output_score_and_ptss(outputs, 200, 3)
+    start = vmap[:, :, :2]
+    end = vmap[:, :, 2:]
+    dist_map = np.sqrt(np.sum((start - end) ** 2, axis=-1))
+
+    segments_list = []
+    for center, score in zip(pts, pts_score):
+        y, x = center
+        distance = dist_map[y, x]
+        if score > score_thr and distance > dist_thr:
+            disp_x_start, disp_y_start, disp_x_end, disp_y_end = vmap[y, x, :]
+            x_start = x + disp_x_start
+            y_start = y + disp_y_start
+            x_end = x + disp_x_end
+            y_end = y + disp_y_end
+            segments_list.append([x_start, y_start, x_end, y_end])
+
+    lines = 2 * np.array(segments_list)  # 256 > 512
+    lines[:, 0] = lines[:, 0] * w_ratio
+    lines[:, 1] = lines[:, 1] * h_ratio
+    lines[:, 2] = lines[:, 2] * w_ratio
+    lines[:, 3] = lines[:, 3] * h_ratio
+
+    return lines
+
+
+def pred_squares(image,
+                 model,
+                 input_shape=[512, 512],
+                 params={'score': 0.06,
+                         'outside_ratio': 0.28,
+                         'inside_ratio': 0.45,
+                         'w_overlap': 0.0,
+                         'w_degree': 1.95,
+                         'w_length': 0.0,
+                         'w_area': 1.86,
+                         'w_center': 0.14}):
+    '''
+    shape = [height, width]
+    '''
+    h, w, _ = image.shape
+    original_shape = [h, w]
+
+    resized_image = np.concatenate([cv2.resize(image, (input_shape[0], input_shape[1]), interpolation=cv2.INTER_AREA),
+                                    np.ones([input_shape[0], input_shape[1], 1])], axis=-1)
+    resized_image = resized_image.transpose((2, 0, 1))
+    batch_image = np.expand_dims(resized_image, axis=0).astype('float32')
+    batch_image = (batch_image / 127.5) - 1.0
+
+    batch_image = torch.from_numpy(batch_image).float().cuda()
+    outputs = model(batch_image)
+
+    pts, pts_score, vmap = deccode_output_score_and_ptss(outputs, 200, 3)
+    start = vmap[:, :, :2]  # (x, y)
+    end = vmap[:, :, 2:]  # (x, y)
+    dist_map = np.sqrt(np.sum((start - end) ** 2, axis=-1))
+
+    junc_list = []
+    segments_list = []
+    for junc, score in zip(pts, pts_score):
+        y, x = junc
+        distance = dist_map[y, x]
+        if score > params['score'] and distance > 20.0:
+            junc_list.append([x, y])
+            disp_x_start, disp_y_start, disp_x_end, disp_y_end = vmap[y, x, :]
+            d_arrow = 1.0
+            x_start = x + d_arrow * disp_x_start
+            y_start = y + d_arrow * disp_y_start
+            x_end = x + d_arrow * disp_x_end
+            y_end = y + d_arrow * disp_y_end
+            segments_list.append([x_start, y_start, x_end, y_end])
+
+    segments = np.array(segments_list)
+
+    ####### post processing for squares
+    # 1. get unique lines
+    point = np.array([[0, 0]])
+    point = point[0]
+    start = segments[:, :2]
+    end = segments[:, 2:]
+    diff = start - end
+    a = diff[:, 1]
+    b = -diff[:, 0]
+    c = a * start[:, 0] + b * start[:, 1]
+
+    d = np.abs(a * point[0] + b * point[1] - c) / np.sqrt(a ** 2 + b ** 2 + 1e-10)
+    theta = np.arctan2(diff[:, 0], diff[:, 1]) * 180 / np.pi
+    theta[theta < 0.0] += 180
+    hough = np.concatenate([d[:, None], theta[:, None]], axis=-1)
+
+    d_quant = 1
+    theta_quant = 2
+    hough[:, 0] //= d_quant
+    hough[:, 1] //= theta_quant
+    _, indices, counts = np.unique(hough, axis=0, return_index=True, return_counts=True)
+
+    acc_map = np.zeros([512 // d_quant + 1, 360 // theta_quant + 1], dtype='float32')
+    idx_map = np.zeros([512 // d_quant + 1, 360 // theta_quant + 1], dtype='int32') - 1
+    yx_indices = hough[indices, :].astype('int32')
+    acc_map[yx_indices[:, 0], yx_indices[:, 1]] = counts
+    idx_map[yx_indices[:, 0], yx_indices[:, 1]] = indices
+
+    acc_map_np = acc_map
+    # acc_map = acc_map[None, :, :, None]
+    #
+    # ### fast suppression using tensorflow op
+    # acc_map = tf.constant(acc_map, dtype=tf.float32)
+    # max_acc_map = tf.keras.layers.MaxPool2D(pool_size=(5, 5), strides=1, padding='same')(acc_map)
+    # acc_map = acc_map * tf.cast(tf.math.equal(acc_map, max_acc_map), tf.float32)
+    # flatten_acc_map = tf.reshape(acc_map, [1, -1])
+    # topk_values, topk_indices = tf.math.top_k(flatten_acc_map, k=len(pts))
+    # _, h, w, _ = acc_map.shape
+    # y = tf.expand_dims(topk_indices // w, axis=-1)
+    # x = tf.expand_dims(topk_indices % w, axis=-1)
+    # yx = tf.concat([y, x], axis=-1)
+
+    ### fast suppression using pytorch op
+    acc_map = torch.from_numpy(acc_map_np).unsqueeze(0).unsqueeze(0)
+    _,_, h, w = acc_map.shape
+    max_acc_map = F.max_pool2d(acc_map,kernel_size=5, stride=1, padding=2)
+    acc_map = acc_map * ( (acc_map == max_acc_map).float() )
+    flatten_acc_map = acc_map.reshape([-1, ])
+
+    scores, indices = torch.topk(flatten_acc_map, len(pts), dim=-1, largest=True)
+    yy = torch.div(indices, w, rounding_mode='floor').unsqueeze(-1)
+    xx = torch.fmod(indices, w).unsqueeze(-1)
+    yx = torch.cat((yy, xx), dim=-1)
+
+    yx = yx.detach().cpu().numpy()
+
+    topk_values = scores.detach().cpu().numpy()
+    indices = idx_map[yx[:, 0], yx[:, 1]]
+    basis = 5 // 2
+
+    merged_segments = []
+    for yx_pt, max_indice, value in zip(yx, indices, topk_values):
+        y, x = yx_pt
+        if max_indice == -1 or value == 0:
+            continue
+        segment_list = []
+        for y_offset in range(-basis, basis + 1):
+            for x_offset in range(-basis, basis + 1):
+                indice = idx_map[y + y_offset, x + x_offset]
+                cnt = int(acc_map_np[y + y_offset, x + x_offset])
+                if indice != -1:
+                    segment_list.append(segments[indice])
+                if cnt > 1:
+                    check_cnt = 1
+                    current_hough = hough[indice]
+                    for new_indice, new_hough in enumerate(hough):
+                        if (current_hough == new_hough).all() and indice != new_indice:
+                            segment_list.append(segments[new_indice])
+                            check_cnt += 1
+                        if check_cnt == cnt:
+                            break
+        group_segments = np.array(segment_list).reshape([-1, 2])
+        sorted_group_segments = np.sort(group_segments, axis=0)
+        x_min, y_min = sorted_group_segments[0, :]
+        x_max, y_max = sorted_group_segments[-1, :]
+
+        deg = theta[max_indice]
+        if deg >= 90:
+            merged_segments.append([x_min, y_max, x_max, y_min])
+        else:
+            merged_segments.append([x_min, y_min, x_max, y_max])
+
+    # 2. get intersections
+    new_segments = np.array(merged_segments)  # (x1, y1, x2, y2)
+    start = new_segments[:, :2]  # (x1, y1)
+    end = new_segments[:, 2:]  # (x2, y2)
+    new_centers = (start + end) / 2.0
+    diff = start - end
+    dist_segments = np.sqrt(np.sum(diff ** 2, axis=-1))
+
+    # ax + by = c
+    a = diff[:, 1]
+    b = -diff[:, 0]
+    c = a * start[:, 0] + b * start[:, 1]
+    pre_det = a[:, None] * b[None, :]
+    det = pre_det - np.transpose(pre_det)
+
+    pre_inter_y = a[:, None] * c[None, :]
+    inter_y = (pre_inter_y - np.transpose(pre_inter_y)) / (det + 1e-10)
+    pre_inter_x = c[:, None] * b[None, :]
+    inter_x = (pre_inter_x - np.transpose(pre_inter_x)) / (det + 1e-10)
+    inter_pts = np.concatenate([inter_x[:, :, None], inter_y[:, :, None]], axis=-1).astype('int32')
+
+    # 3. get corner information
+    # 3.1 get distance
+    '''
+    dist_segments:
+        | dist(0), dist(1), dist(2), ...|
+    dist_inter_to_segment1:
+        | dist(inter,0), dist(inter,0), dist(inter,0), ... |
+        | dist(inter,1), dist(inter,1), dist(inter,1), ... |
+        ...
+    dist_inter_to_semgnet2:
+        | dist(inter,0), dist(inter,1), dist(inter,2), ... |
+        | dist(inter,0), dist(inter,1), dist(inter,2), ... |
+        ...
+    '''
+
+    dist_inter_to_segment1_start = np.sqrt(
+        np.sum(((inter_pts - start[:, None, :]) ** 2), axis=-1, keepdims=True))  # [n_batch, n_batch, 1]
+    dist_inter_to_segment1_end = np.sqrt(
+        np.sum(((inter_pts - end[:, None, :]) ** 2), axis=-1, keepdims=True))  # [n_batch, n_batch, 1]
+    dist_inter_to_segment2_start = np.sqrt(
+        np.sum(((inter_pts - start[None, :, :]) ** 2), axis=-1, keepdims=True))  # [n_batch, n_batch, 1]
+    dist_inter_to_segment2_end = np.sqrt(
+        np.sum(((inter_pts - end[None, :, :]) ** 2), axis=-1, keepdims=True))  # [n_batch, n_batch, 1]
+
+    # sort ascending
+    dist_inter_to_segment1 = np.sort(
+        np.concatenate([dist_inter_to_segment1_start, dist_inter_to_segment1_end], axis=-1),
+        axis=-1)  # [n_batch, n_batch, 2]
+    dist_inter_to_segment2 = np.sort(
+        np.concatenate([dist_inter_to_segment2_start, dist_inter_to_segment2_end], axis=-1),
+        axis=-1)  # [n_batch, n_batch, 2]
+
+    # 3.2 get degree
+    inter_to_start = new_centers[:, None, :] - inter_pts
+    deg_inter_to_start = np.arctan2(inter_to_start[:, :, 1], inter_to_start[:, :, 0]) * 180 / np.pi
+    deg_inter_to_start[deg_inter_to_start < 0.0] += 360
+    inter_to_end = new_centers[None, :, :] - inter_pts
+    deg_inter_to_end = np.arctan2(inter_to_end[:, :, 1], inter_to_end[:, :, 0]) * 180 / np.pi
+    deg_inter_to_end[deg_inter_to_end < 0.0] += 360
+
+    '''
+    B -- G
+    |    |
+    C -- R
+    B : blue / G: green / C: cyan / R: red
+
+    0 -- 1
+    |    |
+    3 -- 2
+    '''
+    # rename variables
+    deg1_map, deg2_map = deg_inter_to_start, deg_inter_to_end
+    # sort deg ascending
+    deg_sort = np.sort(np.concatenate([deg1_map[:, :, None], deg2_map[:, :, None]], axis=-1), axis=-1)
+
+    deg_diff_map = np.abs(deg1_map - deg2_map)
+    # we only consider the smallest degree of intersect
+    deg_diff_map[deg_diff_map > 180] = 360 - deg_diff_map[deg_diff_map > 180]
+
+    # define available degree range
+    deg_range = [60, 120]
+
+    corner_dict = {corner_info: [] for corner_info in range(4)}
+    inter_points = []
+    for i in range(inter_pts.shape[0]):
+        for j in range(i + 1, inter_pts.shape[1]):
+            # i, j > line index, always i < j
+            x, y = inter_pts[i, j, :]
+            deg1, deg2 = deg_sort[i, j, :]
+            deg_diff = deg_diff_map[i, j]
+
+            check_degree = deg_diff > deg_range[0] and deg_diff < deg_range[1]
+
+            outside_ratio = params['outside_ratio']  # over ratio >>> drop it!
+            inside_ratio = params['inside_ratio']  # over ratio >>> drop it!
+            check_distance = ((dist_inter_to_segment1[i, j, 1] >= dist_segments[i] and \
+                               dist_inter_to_segment1[i, j, 0] <= dist_segments[i] * outside_ratio) or \
+                              (dist_inter_to_segment1[i, j, 1] <= dist_segments[i] and \
+                               dist_inter_to_segment1[i, j, 0] <= dist_segments[i] * inside_ratio)) and \
+                             ((dist_inter_to_segment2[i, j, 1] >= dist_segments[j] and \
+                               dist_inter_to_segment2[i, j, 0] <= dist_segments[j] * outside_ratio) or \
+                              (dist_inter_to_segment2[i, j, 1] <= dist_segments[j] and \
+                               dist_inter_to_segment2[i, j, 0] <= dist_segments[j] * inside_ratio))
+
+            if check_degree and check_distance:
+                corner_info = None
+
+                if (deg1 >= 0 and deg1 <= 45 and deg2 >= 45 and deg2 <= 120) or \
+                        (deg2 >= 315 and deg1 >= 45 and deg1 <= 120):
+                    corner_info, color_info = 0, 'blue'
+                elif (deg1 >= 45 and deg1 <= 125 and deg2 >= 125 and deg2 <= 225):
+                    corner_info, color_info = 1, 'green'
+                elif (deg1 >= 125 and deg1 <= 225 and deg2 >= 225 and deg2 <= 315):
+                    corner_info, color_info = 2, 'black'
+                elif (deg1 >= 0 and deg1 <= 45 and deg2 >= 225 and deg2 <= 315) or \
+                        (deg2 >= 315 and deg1 >= 225 and deg1 <= 315):
+                    corner_info, color_info = 3, 'cyan'
+                else:
+                    corner_info, color_info = 4, 'red'  # we don't use it
+                    continue
+
+                corner_dict[corner_info].append([x, y, i, j])
+                inter_points.append([x, y])
+
+    square_list = []
+    connect_list = []
+    segments_list = []
+    for corner0 in corner_dict[0]:
+        for corner1 in corner_dict[1]:
+            connect01 = False
+            for corner0_line in corner0[2:]:
+                if corner0_line in corner1[2:]:
+                    connect01 = True
+                    break
+            if connect01:
+                for corner2 in corner_dict[2]:
+                    connect12 = False
+                    for corner1_line in corner1[2:]:
+                        if corner1_line in corner2[2:]:
+                            connect12 = True
+                            break
+                    if connect12:
+                        for corner3 in corner_dict[3]:
+                            connect23 = False
+                            for corner2_line in corner2[2:]:
+                                if corner2_line in corner3[2:]:
+                                    connect23 = True
+                                    break
+                            if connect23:
+                                for corner3_line in corner3[2:]:
+                                    if corner3_line in corner0[2:]:
+                                        # SQUARE!!!
+                                        '''
+                                        0 -- 1
+                                        |    |
+                                        3 -- 2
+                                        square_list:
+                                            order: 0 > 1 > 2 > 3
+                                            | x0, y0, x1, y1, x2, y2, x3, y3 |
+                                            | x0, y0, x1, y1, x2, y2, x3, y3 |
+                                            ...
+                                        connect_list:
+                                            order: 01 > 12 > 23 > 30
+                                            | line_idx01, line_idx12, line_idx23, line_idx30 |
+                                            | line_idx01, line_idx12, line_idx23, line_idx30 |
+                                            ...
+                                        segments_list:
+                                            order: 0 > 1 > 2 > 3
+                                            | line_idx0_i, line_idx0_j, line_idx1_i, line_idx1_j, line_idx2_i, line_idx2_j, line_idx3_i, line_idx3_j |
+                                            | line_idx0_i, line_idx0_j, line_idx1_i, line_idx1_j, line_idx2_i, line_idx2_j, line_idx3_i, line_idx3_j |
+                                            ...
+                                        '''
+                                        square_list.append(corner0[:2] + corner1[:2] + corner2[:2] + corner3[:2])
+                                        connect_list.append([corner0_line, corner1_line, corner2_line, corner3_line])
+                                        segments_list.append(corner0[2:] + corner1[2:] + corner2[2:] + corner3[2:])
+
+    def check_outside_inside(segments_info, connect_idx):
+        # return 'outside or inside', min distance, cover_param, peri_param
+        if connect_idx == segments_info[0]:
+            check_dist_mat = dist_inter_to_segment1
+        else:
+            check_dist_mat = dist_inter_to_segment2
+
+        i, j = segments_info
+        min_dist, max_dist = check_dist_mat[i, j, :]
+        connect_dist = dist_segments[connect_idx]
+        if max_dist > connect_dist:
+            return 'outside', min_dist, 0, 1
+        else:
+            return 'inside', min_dist, -1, -1
+
+    top_square = None
+
+    try:
+        map_size = input_shape[0] / 2
+        squares = np.array(square_list).reshape([-1, 4, 2])
+        score_array = []
+        connect_array = np.array(connect_list)
+        segments_array = np.array(segments_list).reshape([-1, 4, 2])
+
+        # get degree of corners:
+        squares_rollup = np.roll(squares, 1, axis=1)
+        squares_rolldown = np.roll(squares, -1, axis=1)
+        vec1 = squares_rollup - squares
+        normalized_vec1 = vec1 / (np.linalg.norm(vec1, axis=-1, keepdims=True) + 1e-10)
+        vec2 = squares_rolldown - squares
+        normalized_vec2 = vec2 / (np.linalg.norm(vec2, axis=-1, keepdims=True) + 1e-10)
+        inner_products = np.sum(normalized_vec1 * normalized_vec2, axis=-1)  # [n_squares, 4]
+        squares_degree = np.arccos(inner_products) * 180 / np.pi  # [n_squares, 4]
+
+        # get square score
+        overlap_scores = []
+        degree_scores = []
+        length_scores = []
+
+        for connects, segments, square, degree in zip(connect_array, segments_array, squares, squares_degree):
+            '''
+            0 -- 1
+            |    |
+            3 -- 2
+
+            # segments: [4, 2]
+            # connects: [4]
+            '''
+
+            ###################################### OVERLAP SCORES
+            cover = 0
+            perimeter = 0
+            # check 0 > 1 > 2 > 3
+            square_length = []
+
+            for start_idx in range(4):
+                end_idx = (start_idx + 1) % 4
+
+                connect_idx = connects[start_idx]  # segment idx of segment01
+                start_segments = segments[start_idx]
+                end_segments = segments[end_idx]
+
+                start_point = square[start_idx]
+                end_point = square[end_idx]
+
+                # check whether outside or inside
+                start_position, start_min, start_cover_param, start_peri_param = check_outside_inside(start_segments,
+                                                                                                      connect_idx)
+                end_position, end_min, end_cover_param, end_peri_param = check_outside_inside(end_segments, connect_idx)
+
+                cover += dist_segments[connect_idx] + start_cover_param * start_min + end_cover_param * end_min
+                perimeter += dist_segments[connect_idx] + start_peri_param * start_min + end_peri_param * end_min
+
+                square_length.append(
+                    dist_segments[connect_idx] + start_peri_param * start_min + end_peri_param * end_min)
+
+            overlap_scores.append(cover / perimeter)
+            ######################################
+            ###################################### DEGREE SCORES
+            '''
+            deg0 vs deg2
+            deg1 vs deg3
+            '''
+            deg0, deg1, deg2, deg3 = degree
+            deg_ratio1 = deg0 / deg2
+            if deg_ratio1 > 1.0:
+                deg_ratio1 = 1 / deg_ratio1
+            deg_ratio2 = deg1 / deg3
+            if deg_ratio2 > 1.0:
+                deg_ratio2 = 1 / deg_ratio2
+            degree_scores.append((deg_ratio1 + deg_ratio2) / 2)
+            ######################################
+            ###################################### LENGTH SCORES
+            '''
+            len0 vs len2
+            len1 vs len3
+            '''
+            len0, len1, len2, len3 = square_length
+            len_ratio1 = len0 / len2 if len2 > len0 else len2 / len0
+            len_ratio2 = len1 / len3 if len3 > len1 else len3 / len1
+            length_scores.append((len_ratio1 + len_ratio2) / 2)
+
+            ######################################
+
+        overlap_scores = np.array(overlap_scores)
+        overlap_scores /= np.max(overlap_scores)
+
+        degree_scores = np.array(degree_scores)
+        # degree_scores /= np.max(degree_scores)
+
+        length_scores = np.array(length_scores)
+
+        ###################################### AREA SCORES
+        area_scores = np.reshape(squares, [-1, 4, 2])
+        area_x = area_scores[:, :, 0]
+        area_y = area_scores[:, :, 1]
+        correction = area_x[:, -1] * area_y[:, 0] - area_y[:, -1] * area_x[:, 0]
+        area_scores = np.sum(area_x[:, :-1] * area_y[:, 1:], axis=-1) - np.sum(area_y[:, :-1] * area_x[:, 1:], axis=-1)
+        area_scores = 0.5 * np.abs(area_scores + correction)
+        area_scores /= (map_size * map_size)  # np.max(area_scores)
+        ######################################
+
+        ###################################### CENTER SCORES
+        centers = np.array([[256 // 2, 256 // 2]], dtype='float32')  # [1, 2]
+        # squares: [n, 4, 2]
+        square_centers = np.mean(squares, axis=1)  # [n, 2]
+        center2center = np.sqrt(np.sum((centers - square_centers) ** 2))
+        center_scores = center2center / (map_size / np.sqrt(2.0))
+
+        '''
+        score_w = [overlap, degree, area, center, length]
+        '''
+        score_w = [0.0, 1.0, 10.0, 0.5, 1.0]
+        score_array = params['w_overlap'] * overlap_scores \
+                      + params['w_degree'] * degree_scores \
+                      + params['w_area'] * area_scores \
+                      - params['w_center'] * center_scores \
+                      + params['w_length'] * length_scores
+
+        best_square = []
+
+        sorted_idx = np.argsort(score_array)[::-1]
+        score_array = score_array[sorted_idx]
+        squares = squares[sorted_idx]
+
+    except Exception as e:
+        pass
+
+    '''return list
+    merged_lines, squares, scores
+    '''
+
+    try:
+        new_segments[:, 0] = new_segments[:, 0] * 2 / input_shape[1] * original_shape[1]
+        new_segments[:, 1] = new_segments[:, 1] * 2 / input_shape[0] * original_shape[0]
+        new_segments[:, 2] = new_segments[:, 2] * 2 / input_shape[1] * original_shape[1]
+        new_segments[:, 3] = new_segments[:, 3] * 2 / input_shape[0] * original_shape[0]
+    except:
+        new_segments = []
+
+    try:
+        squares[:, :, 0] = squares[:, :, 0] * 2 / input_shape[1] * original_shape[1]
+        squares[:, :, 1] = squares[:, :, 1] * 2 / input_shape[0] * original_shape[0]
+    except:
+        squares = []
+        score_array = []
+
+    try:
+        inter_points = np.array(inter_points)
+        inter_points[:, 0] = inter_points[:, 0] * 2 / input_shape[1] * original_shape[1]
+        inter_points[:, 1] = inter_points[:, 1] * 2 / input_shape[0] * original_shape[0]
+    except:
+        inter_points = []
+
+    return new_segments, squares, score_array, inter_points

src/flux/annotator/tile/__init__.py

@@ -0,0 +1,26 @@
+import random
+import cv2
+from .guided_filter import FastGuidedFilter
+
+
+class TileDetector:
+    # https://huggingface.co/xinsir/controlnet-tile-sdxl-1.0
+    def __init__(self):
+        pass
+
+    def __call__(self, image):
+        blur_strength = random.sample([i / 10. for i in range(10, 201, 2)], k=1)[0]
+        radius = random.sample([i for i in range(1, 40, 2)], k=1)[0]
+        eps = random.sample([i / 1000. for i in range(1, 101, 2)], k=1)[0]
+        scale_factor = random.sample([i / 10. for i in range(10, 181, 5)], k=1)[0]
+
+        ksize = int(blur_strength)
+        if ksize % 2 == 0:
+            ksize += 1
+
+        if random.random() > 0.5:
+            image = cv2.GaussianBlur(image, (ksize, ksize), blur_strength / 2)
+        if random.random() > 0.5:
+            filter = FastGuidedFilter(image, radius, eps, scale_factor)
+            image = filter.filter(image)
+        return image

src/flux/annotator/tile/guided_filter.py

@@ -0,0 +1,280 @@
+# -*- coding: utf-8 -*-
+## @package guided_filter.core.filters
+#
+#  Implementation of guided filter.
+#  * GuidedFilter: Original guided filter.
+#  * FastGuidedFilter: Fast version of the guided filter.
+#  @author      tody
+#  @date        2015/08/26
+
+import numpy as np
+import cv2
+
+## Convert image into float32 type.
+def to32F(img):
+    if img.dtype == np.float32:
+        return img
+    return (1.0 / 255.0) * np.float32(img)
+
+## Convert image into uint8 type.
+def to8U(img):
+    if img.dtype == np.uint8:
+        return img
+    return np.clip(np.uint8(255.0 * img), 0, 255)
+
+## Return if the input image is gray or not.
+def _isGray(I):
+    return len(I.shape) == 2
+
+
+## Return down sampled image.
+#  @param scale (w/s, h/s) image will be created.
+#  @param shape I.shape[:2]=(h, w). numpy friendly size parameter.
+def _downSample(I, scale=4, shape=None):
+    if shape is not None:
+        h, w = shape
+        return cv2.resize(I, (w, h), interpolation=cv2.INTER_NEAREST)
+
+    h, w = I.shape[:2]
+    return cv2.resize(I, (int(w / scale), int(h / scale)), interpolation=cv2.INTER_NEAREST)
+
+
+## Return up sampled image.
+#  @param scale (w*s, h*s) image will be created.
+#  @param shape I.shape[:2]=(h, w). numpy friendly size parameter.
+def _upSample(I, scale=2, shape=None):
+    if shape is not None:
+        h, w = shape
+        return cv2.resize(I, (w, h), interpolation=cv2.INTER_LINEAR)
+
+    h, w = I.shape[:2]
+    return cv2.resize(I, (int(w * scale), int(h * scale)), interpolation=cv2.INTER_LINEAR)
+
+## Fast guide filter.
+class FastGuidedFilter:
+    ## Constructor.
+    #  @param I Input guidance image. Color or gray.
+    #  @param radius Radius of Guided Filter.
+    #  @param epsilon Regularization term of Guided Filter.
+    #  @param scale Down sampled scale.
+    def __init__(self, I, radius=5, epsilon=0.4, scale=4):
+        I_32F = to32F(I)
+        self._I = I_32F
+        h, w = I.shape[:2]
+
+        I_sub = _downSample(I_32F, scale)
+
+        self._I_sub = I_sub
+        radius = int(radius / scale)
+
+        if _isGray(I):
+            self._guided_filter = GuidedFilterGray(I_sub, radius, epsilon)
+        else:
+            self._guided_filter = GuidedFilterColor(I_sub, radius, epsilon)
+
+    ## Apply filter for the input image.
+    #  @param p Input image for the filtering.
+    def filter(self, p):
+        p_32F = to32F(p)
+        shape_original = p.shape[:2]
+
+        p_sub = _downSample(p_32F, shape=self._I_sub.shape[:2])
+
+        if _isGray(p_sub):
+            return self._filterGray(p_sub, shape_original)
+
+        cs = p.shape[2]
+        q = np.array(p_32F)
+
+        for ci in range(cs):
+            q[:, :, ci] = self._filterGray(p_sub[:, :, ci], shape_original)
+        return to8U(q)
+
+    def _filterGray(self, p_sub, shape_original):
+        ab_sub = self._guided_filter._computeCoefficients(p_sub)
+        ab = [_upSample(abi, shape=shape_original) for abi in ab_sub]
+        return self._guided_filter._computeOutput(ab, self._I)
+
+
+## Guide filter.
+class GuidedFilter:
+    ## Constructor.
+    #  @param I Input guidance image. Color or gray.
+    #  @param radius Radius of Guided Filter.
+    #  @param epsilon Regularization term of Guided Filter.
+    def __init__(self, I, radius=5, epsilon=0.4):
+        I_32F = to32F(I)
+
+        if _isGray(I):
+            self._guided_filter = GuidedFilterGray(I_32F, radius, epsilon)
+        else:
+            self._guided_filter = GuidedFilterColor(I_32F, radius, epsilon)
+
+    ## Apply filter for the input image.
+    #  @param p Input image for the filtering.
+    def filter(self, p):
+        return to8U(self._guided_filter.filter(p))
+
+
+## Common parts of guided filter.
+#
+#  This class is used by guided_filter class. GuidedFilterGray and GuidedFilterColor.
+#  Based on guided_filter._computeCoefficients, guided_filter._computeOutput,
+#  GuidedFilterCommon.filter computes filtered image for color and gray.
+class GuidedFilterCommon:
+    def __init__(self, guided_filter):
+        self._guided_filter = guided_filter
+
+    ## Apply filter for the input image.
+    #  @param p Input image for the filtering.
+    def filter(self, p):
+        p_32F = to32F(p)
+        if _isGray(p_32F):
+            return self._filterGray(p_32F)
+
+        cs = p.shape[2]
+        q = np.array(p_32F)
+
+        for ci in range(cs):
+            q[:, :, ci] = self._filterGray(p_32F[:, :, ci])
+        return q
+
+    def _filterGray(self, p):
+        ab = self._guided_filter._computeCoefficients(p)
+        return self._guided_filter._computeOutput(ab, self._guided_filter._I)
+
+
+## Guided filter for gray guidance image.
+class GuidedFilterGray:
+    #  @param I Input gray guidance image.
+    #  @param radius Radius of Guided Filter.
+    #  @param epsilon Regularization term of Guided Filter.
+    def __init__(self, I, radius=5, epsilon=0.4):
+        self._radius = 2 * radius + 1
+        self._epsilon = epsilon
+        self._I = to32F(I)
+        self._initFilter()
+        self._filter_common = GuidedFilterCommon(self)
+
+    ## Apply filter for the input image.
+    #  @param p Input image for the filtering.
+    def filter(self, p):
+        return self._filter_common.filter(p)
+
+    def _initFilter(self):
+        I = self._I
+        r = self._radius
+        self._I_mean = cv2.blur(I, (r, r))
+        I_mean_sq = cv2.blur(I ** 2, (r, r))
+        self._I_var = I_mean_sq - self._I_mean ** 2
+
+    def _computeCoefficients(self, p):
+        r = self._radius
+        p_mean = cv2.blur(p, (r, r))
+        p_cov = p_mean - self._I_mean * p_mean
+        a = p_cov / (self._I_var + self._epsilon)
+        b = p_mean - a * self._I_mean
+        a_mean = cv2.blur(a, (r, r))
+        b_mean = cv2.blur(b, (r, r))
+        return a_mean, b_mean
+
+    def _computeOutput(self, ab, I):
+        a_mean, b_mean = ab
+        return a_mean * I + b_mean
+
+
+## Guided filter for color guidance image.
+class GuidedFilterColor:
+    #  @param I Input color guidance image.
+   #  @param radius Radius of Guided Filter.
+    #  @param epsilon Regularization term of Guided Filter.
+    def __init__(self, I, radius=5, epsilon=0.2):
+        self._radius = 2 * radius + 1
+        self._epsilon = epsilon
+        self._I = to32F(I)
+        self._initFilter()
+        self._filter_common = GuidedFilterCommon(self)
+
+    ## Apply filter for the input image.
+    #  @param p Input image for the filtering.
+    def filter(self, p):
+        return self._filter_common.filter(p)
+
+    def _initFilter(self):
+        I = self._I
+        r = self._radius
+        eps = self._epsilon
+
+        Ir, Ig, Ib = I[:, :, 0], I[:, :, 1], I[:, :, 2]
+
+        self._Ir_mean = cv2.blur(Ir, (r, r))
+        self._Ig_mean = cv2.blur(Ig, (r, r))
+        self._Ib_mean = cv2.blur(Ib, (r, r))
+
+        Irr_var = cv2.blur(Ir ** 2, (r, r)) - self._Ir_mean ** 2 + eps
+        Irg_var = cv2.blur(Ir * Ig, (r, r)) - self._Ir_mean * self._Ig_mean
+        Irb_var = cv2.blur(Ir * Ib, (r, r)) - self._Ir_mean * self._Ib_mean
+        Igg_var = cv2.blur(Ig * Ig, (r, r)) - self._Ig_mean * self._Ig_mean + eps
+        Igb_var = cv2.blur(Ig * Ib, (r, r)) - self._Ig_mean * self._Ib_mean
+        Ibb_var = cv2.blur(Ib * Ib, (r, r)) - self._Ib_mean * self._Ib_mean + eps
+
+        Irr_inv = Igg_var * Ibb_var - Igb_var * Igb_var
+        Irg_inv = Igb_var * Irb_var - Irg_var * Ibb_var
+        Irb_inv = Irg_var * Igb_var - Igg_var * Irb_var
+        Igg_inv = Irr_var * Ibb_var - Irb_var * Irb_var
+        Igb_inv = Irb_var * Irg_var - Irr_var * Igb_var
+        Ibb_inv = Irr_var * Igg_var - Irg_var * Irg_var
+
+        I_cov = Irr_inv * Irr_var + Irg_inv * Irg_var + Irb_inv * Irb_var
+        Irr_inv /= I_cov
+        Irg_inv /= I_cov
+        Irb_inv /= I_cov
+        Igg_inv /= I_cov
+        Igb_inv /= I_cov
+        Ibb_inv /= I_cov
+
+        self._Irr_inv = Irr_inv
+        self._Irg_inv = Irg_inv
+        self._Irb_inv = Irb_inv
+        self._Igg_inv = Igg_inv
+        self._Igb_inv = Igb_inv
+        self._Ibb_inv = Ibb_inv
+
+    def _computeCoefficients(self, p):
+        r = self._radius
+        I = self._I
+        Ir, Ig, Ib = I[:, :, 0], I[:, :, 1], I[:, :, 2]
+
+        p_mean = cv2.blur(p, (r, r))
+
+        Ipr_mean = cv2.blur(Ir * p, (r, r))
+        Ipg_mean = cv2.blur(Ig * p, (r, r))
+        Ipb_mean = cv2.blur(Ib * p, (r, r))
+
+        Ipr_cov = Ipr_mean - self._Ir_mean * p_mean
+        Ipg_cov = Ipg_mean - self._Ig_mean * p_mean
+        Ipb_cov = Ipb_mean - self._Ib_mean * p_mean
+
+        ar = self._Irr_inv * Ipr_cov + self._Irg_inv * Ipg_cov + self._Irb_inv * Ipb_cov
+        ag = self._Irg_inv * Ipr_cov + self._Igg_inv * Ipg_cov + self._Igb_inv * Ipb_cov
+        ab = self._Irb_inv * Ipr_cov + self._Igb_inv * Ipg_cov + self._Ibb_inv * Ipb_cov
+        b = p_mean - ar * self._Ir_mean - ag * self._Ig_mean - ab * self._Ib_mean
+
+        ar_mean = cv2.blur(ar, (r, r))
+        ag_mean = cv2.blur(ag, (r, r))
+        ab_mean = cv2.blur(ab, (r, r))
+        b_mean = cv2.blur(b, (r, r))
+
+        return ar_mean, ag_mean, ab_mean, b_mean
+
+    def _computeOutput(self, ab, I):
+        ar_mean, ag_mean, ab_mean, b_mean = ab
+
+        Ir, Ig, Ib = I[:, :, 0], I[:, :, 1], I[:, :, 2]
+
+        q = (ar_mean * Ir +
+             ag_mean * Ig +
+             ab_mean * Ib +
+             b_mean)
+
+        return q

src/flux/annotator/util.py

@@ -0,0 +1,38 @@
+import numpy as np
+import cv2
+import os
+
+
+annotator_ckpts_path = os.path.join(os.path.dirname(__file__), 'ckpts')
+
+
+def HWC3(x):
+    assert x.dtype == np.uint8
+    if x.ndim == 2:
+        x = x[:, :, None]
+    assert x.ndim == 3
+    H, W, C = x.shape
+    assert C == 1 or C == 3 or C == 4
+    if C == 3:
+        return x
+    if C == 1:
+        return np.concatenate([x, x, x], axis=2)
+    if C == 4:
+        color = x[:, :, 0:3].astype(np.float32)
+        alpha = x[:, :, 3:4].astype(np.float32) / 255.0
+        y = color * alpha + 255.0 * (1.0 - alpha)
+        y = y.clip(0, 255).astype(np.uint8)
+        return y
+
+
+def resize_image(input_image, resolution):
+    H, W, C = input_image.shape
+    H = float(H)
+    W = float(W)
+    k = float(resolution) / min(H, W)
+    H *= k
+    W *= k
+    H = int(np.round(H / 64.0)) * 64
+    W = int(np.round(W / 64.0)) * 64
+    img = cv2.resize(input_image, (W, H), interpolation=cv2.INTER_LANCZOS4 if k > 1 else cv2.INTER_AREA)
+    return img

src/flux/annotator/zoe/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2022 Intelligent Systems Lab Org
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

src/flux/annotator/zoe/__init__.py

@@ -0,0 +1,48 @@
+# ZoeDepth
+# https://github.com/isl-org/ZoeDepth
+
+import os
+import cv2
+import numpy as np
+import torch
+
+from einops import rearrange
+from .zoedepth.models.zoedepth.zoedepth_v1 import ZoeDepth
+from .zoedepth.utils.config import get_config
+from ...annotator.util import annotator_ckpts_path
+from huggingface_hub import hf_hub_download
+
+
+class ZoeDetector:
+    def __init__(self):
+        model_path = os.path.join(annotator_ckpts_path, "ZoeD_M12_N.pt")
+        if not os.path.exists(model_path):
+            model_path = hf_hub_download("lllyasviel/Annotators", "ZoeD_M12_N.pt")
+        conf = get_config("zoedepth", "infer")
+        model = ZoeDepth.build_from_config(conf)
+        model.load_state_dict(torch.load(model_path)['model'], strict=False)
+        model = model.cuda()
+        model.device = 'cuda'
+        model.eval()
+        self.model = model
+
+    def __call__(self, input_image):
+        assert input_image.ndim == 3
+        image_depth = input_image
+        with torch.no_grad():
+            image_depth = torch.from_numpy(image_depth).float().cuda()
+            image_depth = image_depth / 255.0
+            image_depth = rearrange(image_depth, 'h w c -> 1 c h w')
+            depth = self.model.infer(image_depth)
+
+            depth = depth[0, 0].cpu().numpy()
+
+            vmin = np.percentile(depth, 2)
+            vmax = np.percentile(depth, 85)
+
+            depth -= vmin
+            depth /= vmax - vmin
+            depth = 1.0 - depth
+            depth_image = (depth * 255.0).clip(0, 255).astype(np.uint8)
+
+            return depth_image

src/flux/annotator/zoe/zoedepth/data/__init__.py

@@ -0,0 +1,24 @@
+# MIT License
+
+# Copyright (c) 2022 Intelligent Systems Lab Org
+
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+# File author: Shariq Farooq Bhat
+

src/flux/annotator/zoe/zoedepth/data/data_mono.py

@@ -0,0 +1,573 @@
+# MIT License
+
+# Copyright (c) 2022 Intelligent Systems Lab Org
+
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+# File author: Shariq Farooq Bhat
+
+# This file is partly inspired from BTS (https://github.com/cleinc/bts/blob/master/pytorch/bts_dataloader.py); author: Jin Han Lee
+
+import itertools
+import os
+import random
+
+import numpy as np
+import cv2
+import torch
+import torch.nn as nn
+import torch.utils.data.distributed
+from zoedepth.utils.easydict import EasyDict as edict
+from PIL import Image, ImageOps
+from torch.utils.data import DataLoader, Dataset
+from torchvision import transforms
+
+from zoedepth.utils.config import change_dataset
+
+from .ddad import get_ddad_loader
+from .diml_indoor_test import get_diml_indoor_loader
+from .diml_outdoor_test import get_diml_outdoor_loader
+from .diode import get_diode_loader
+from .hypersim import get_hypersim_loader
+from .ibims import get_ibims_loader
+from .sun_rgbd_loader import get_sunrgbd_loader
+from .vkitti import get_vkitti_loader
+from .vkitti2 import get_vkitti2_loader
+
+from .preprocess import CropParams, get_white_border, get_black_border
+
+
+def _is_pil_image(img):
+    return isinstance(img, Image.Image)
+
+
+def _is_numpy_image(img):
+    return isinstance(img, np.ndarray) and (img.ndim in {2, 3})
+
+
+def preprocessing_transforms(mode, **kwargs):
+    return transforms.Compose([
+        ToTensor(mode=mode, **kwargs)
+    ])
+
+
+class DepthDataLoader(object):
+    def __init__(self, config, mode, device='cpu', transform=None, **kwargs):
+        """
+        Data loader for depth datasets
+
+        Args:
+            config (dict): Config dictionary. Refer to utils/config.py
+            mode (str): "train" or "online_eval"
+            device (str, optional): Device to load the data on. Defaults to 'cpu'.
+            transform (torchvision.transforms, optional): Transform to apply to the data. Defaults to None.
+        """
+
+        self.config = config
+
+        if config.dataset == 'ibims':
+            self.data = get_ibims_loader(config, batch_size=1, num_workers=1)
+            return
+
+        if config.dataset == 'sunrgbd':
+            self.data = get_sunrgbd_loader(
+                data_dir_root=config.sunrgbd_root, batch_size=1, num_workers=1)
+            return
+
+        if config.dataset == 'diml_indoor':
+            self.data = get_diml_indoor_loader(
+                data_dir_root=config.diml_indoor_root, batch_size=1, num_workers=1)
+            return
+
+        if config.dataset == 'diml_outdoor':
+            self.data = get_diml_outdoor_loader(
+                data_dir_root=config.diml_outdoor_root, batch_size=1, num_workers=1)
+            return
+
+        if "diode" in config.dataset:
+            self.data = get_diode_loader(
+                config[config.dataset+"_root"], batch_size=1, num_workers=1)
+            return
+
+        if config.dataset == 'hypersim_test':
+            self.data = get_hypersim_loader(
+                config.hypersim_test_root, batch_size=1, num_workers=1)
+            return
+
+        if config.dataset == 'vkitti':
+            self.data = get_vkitti_loader(
+                config.vkitti_root, batch_size=1, num_workers=1)
+            return
+
+        if config.dataset == 'vkitti2':
+            self.data = get_vkitti2_loader(
+                config.vkitti2_root, batch_size=1, num_workers=1)
+            return
+
+        if config.dataset == 'ddad':
+            self.data = get_ddad_loader(config.ddad_root, resize_shape=(
+                352, 1216), batch_size=1, num_workers=1)
+            return
+
+        img_size = self.config.get("img_size", None)
+        img_size = img_size if self.config.get(
+            "do_input_resize", False) else None
+
+        if transform is None:
+            transform = preprocessing_transforms(mode, size=img_size)
+
+        if mode == 'train':
+
+            Dataset = DataLoadPreprocess
+            self.training_samples = Dataset(
+                config, mode, transform=transform, device=device)
+
+            if config.distributed:
+                self.train_sampler = torch.utils.data.distributed.DistributedSampler(
+                    self.training_samples)
+            else:
+                self.train_sampler = None
+
+            self.data = DataLoader(self.training_samples,
+                                   batch_size=config.batch_size,
+                                   shuffle=(self.train_sampler is None),
+                                   num_workers=config.workers,
+                                   pin_memory=True,
+                                   persistent_workers=True,
+                                #    prefetch_factor=2,
+                                   sampler=self.train_sampler)
+
+        elif mode == 'online_eval':
+            self.testing_samples = DataLoadPreprocess(
+                config, mode, transform=transform)
+            if config.distributed:  # redundant. here only for readability and to be more explicit
+                # Give whole test set to all processes (and report evaluation only on one) regardless
+                self.eval_sampler = None
+            else:
+                self.eval_sampler = None
+            self.data = DataLoader(self.testing_samples, 1,
+                                   shuffle=kwargs.get("shuffle_test", False),
+                                   num_workers=1,
+                                   pin_memory=False,
+                                   sampler=self.eval_sampler)
+
+        elif mode == 'test':
+            self.testing_samples = DataLoadPreprocess(
+                config, mode, transform=transform)
+            self.data = DataLoader(self.testing_samples,
+                                   1, shuffle=False, num_workers=1)
+
+        else:
+            print(
+                'mode should be one of \'train, test, online_eval\'. Got {}'.format(mode))
+
+
+def repetitive_roundrobin(*iterables):
+    """
+    cycles through iterables but sample wise
+    first yield first sample from first iterable then first sample from second iterable and so on
+    then second sample from first iterable then second sample from second iterable and so on
+
+    If one iterable is shorter than the others, it is repeated until all iterables are exhausted
+    repetitive_roundrobin('ABC', 'D', 'EF') --> A D E B D F C D E
+    """
+    # Repetitive roundrobin
+    iterables_ = [iter(it) for it in iterables]
+    exhausted = [False] * len(iterables)
+    while not all(exhausted):
+        for i, it in enumerate(iterables_):
+            try:
+                yield next(it)
+            except StopIteration:
+                exhausted[i] = True
+                iterables_[i] = itertools.cycle(iterables[i])
+                # First elements may get repeated if one iterable is shorter than the others
+                yield next(iterables_[i])
+
+
+class RepetitiveRoundRobinDataLoader(object):
+    def __init__(self, *dataloaders):
+        self.dataloaders = dataloaders
+
+    def __iter__(self):
+        return repetitive_roundrobin(*self.dataloaders)
+
+    def __len__(self):
+        # First samples get repeated, thats why the plus one
+        return len(self.dataloaders) * (max(len(dl) for dl in self.dataloaders) + 1)
+
+
+class MixedNYUKITTI(object):
+    def __init__(self, config, mode, device='cpu', **kwargs):
+        config = edict(config)
+        config.workers = config.workers // 2
+        self.config = config
+        nyu_conf = change_dataset(edict(config), 'nyu')
+        kitti_conf = change_dataset(edict(config), 'kitti')
+
+        # make nyu default for testing
+        self.config = config = nyu_conf
+        img_size = self.config.get("img_size", None)
+        img_size = img_size if self.config.get(
+            "do_input_resize", False) else None
+        if mode == 'train':
+            nyu_loader = DepthDataLoader(
+                nyu_conf, mode, device=device, transform=preprocessing_transforms(mode, size=img_size)).data
+            kitti_loader = DepthDataLoader(
+                kitti_conf, mode, device=device, transform=preprocessing_transforms(mode, size=img_size)).data
+            # It has been changed to repetitive roundrobin
+            self.data = RepetitiveRoundRobinDataLoader(
+                nyu_loader, kitti_loader)
+        else:
+            self.data = DepthDataLoader(nyu_conf, mode, device=device).data
+
+
+def remove_leading_slash(s):
+    if s[0] == '/' or s[0] == '\\':
+        return s[1:]
+    return s
+
+
+class CachedReader:
+    def __init__(self, shared_dict=None):
+        if shared_dict:
+            self._cache = shared_dict
+        else:
+            self._cache = {}
+
+    def open(self, fpath):
+        im = self._cache.get(fpath, None)
+        if im is None:
+            im = self._cache[fpath] = Image.open(fpath)
+        return im
+
+
+class ImReader:
+    def __init__(self):
+        pass
+
+    # @cache
+    def open(self, fpath):
+        return Image.open(fpath)
+
+
+class DataLoadPreprocess(Dataset):
+    def __init__(self, config, mode, transform=None, is_for_online_eval=False, **kwargs):
+        self.config = config
+        if mode == 'online_eval':
+            with open(config.filenames_file_eval, 'r') as f:
+                self.filenames = f.readlines()
+        else:
+            with open(config.filenames_file, 'r') as f:
+                self.filenames = f.readlines()
+
+        self.mode = mode
+        self.transform = transform
+        self.to_tensor = ToTensor(mode)
+        self.is_for_online_eval = is_for_online_eval
+        if config.use_shared_dict:
+            self.reader = CachedReader(config.shared_dict)
+        else:
+            self.reader = ImReader()
+
+    def postprocess(self, sample):
+        return sample
+
+    def __getitem__(self, idx):
+        sample_path = self.filenames[idx]
+        focal = float(sample_path.split()[2])
+        sample = {}
+
+        if self.mode == 'train':
+            if self.config.dataset == 'kitti' and self.config.use_right and random.random() > 0.5:
+                image_path = os.path.join(
+                    self.config.data_path, remove_leading_slash(sample_path.split()[3]))
+                depth_path = os.path.join(
+                    self.config.gt_path, remove_leading_slash(sample_path.split()[4]))
+            else:
+                image_path = os.path.join(
+                    self.config.data_path, remove_leading_slash(sample_path.split()[0]))
+                depth_path = os.path.join(
+                    self.config.gt_path, remove_leading_slash(sample_path.split()[1]))
+
+            image = self.reader.open(image_path)
+            depth_gt = self.reader.open(depth_path)
+            w, h = image.size
+
+            if self.config.do_kb_crop:
+                height = image.height
+                width = image.width
+                top_margin = int(height - 352)
+                left_margin = int((width - 1216) / 2)
+                depth_gt = depth_gt.crop(
+                    (left_margin, top_margin, left_margin + 1216, top_margin + 352))
+                image = image.crop(
+                    (left_margin, top_margin, left_margin + 1216, top_margin + 352))
+
+            # Avoid blank boundaries due to pixel registration?
+            # Train images have white border. Test images have black border.
+            if self.config.dataset == 'nyu' and self.config.avoid_boundary:
+                # print("Avoiding Blank Boundaries!")
+                # We just crop and pad again with reflect padding to original size
+                # original_size = image.size
+                crop_params = get_white_border(np.array(image, dtype=np.uint8))
+                image = image.crop((crop_params.left, crop_params.top, crop_params.right, crop_params.bottom))
+                depth_gt = depth_gt.crop((crop_params.left, crop_params.top, crop_params.right, crop_params.bottom))
+
+                # Use reflect padding to fill the blank
+                image = np.array(image)
+                image = np.pad(image, ((crop_params.top, h - crop_params.bottom), (crop_params.left, w - crop_params.right), (0, 0)), mode='reflect')
+                image = Image.fromarray(image)
+
+                depth_gt = np.array(depth_gt)
+                depth_gt = np.pad(depth_gt, ((crop_params.top, h - crop_params.bottom), (crop_params.left, w - crop_params.right)), 'constant', constant_values=0)
+                depth_gt = Image.fromarray(depth_gt)
+
+
+            if self.config.do_random_rotate and (self.config.aug):
+                random_angle = (random.random() - 0.5) * 2 * self.config.degree
+                image = self.rotate_image(image, random_angle)
+                depth_gt = self.rotate_image(
+                    depth_gt, random_angle, flag=Image.NEAREST)
+
+            image = np.asarray(image, dtype=np.float32) / 255.0
+            depth_gt = np.asarray(depth_gt, dtype=np.float32)
+            depth_gt = np.expand_dims(depth_gt, axis=2)
+
+            if self.config.dataset == 'nyu':
+                depth_gt = depth_gt / 1000.0
+            else:
+                depth_gt = depth_gt / 256.0
+
+            if self.config.aug and (self.config.random_crop):
+                image, depth_gt = self.random_crop(
+                    image, depth_gt, self.config.input_height, self.config.input_width)
+            
+            if self.config.aug and self.config.random_translate:
+                # print("Random Translation!")
+                image, depth_gt = self.random_translate(image, depth_gt, self.config.max_translation)
+
+            image, depth_gt = self.train_preprocess(image, depth_gt)
+            mask = np.logical_and(depth_gt > self.config.min_depth,
+                                  depth_gt < self.config.max_depth).squeeze()[None, ...]
+            sample = {'image': image, 'depth': depth_gt, 'focal': focal,
+                      'mask': mask, **sample}
+
+        else:
+            if self.mode == 'online_eval':
+                data_path = self.config.data_path_eval
+            else:
+                data_path = self.config.data_path
+
+            image_path = os.path.join(
+                data_path, remove_leading_slash(sample_path.split()[0]))
+            image = np.asarray(self.reader.open(image_path),
+                               dtype=np.float32) / 255.0
+
+            if self.mode == 'online_eval':
+                gt_path = self.config.gt_path_eval
+                depth_path = os.path.join(
+                    gt_path, remove_leading_slash(sample_path.split()[1]))
+                has_valid_depth = False
+                try:
+                    depth_gt = self.reader.open(depth_path)
+                    has_valid_depth = True
+                except IOError:
+                    depth_gt = False
+                    # print('Missing gt for {}'.format(image_path))
+
+                if has_valid_depth:
+                    depth_gt = np.asarray(depth_gt, dtype=np.float32)
+                    depth_gt = np.expand_dims(depth_gt, axis=2)
+                    if self.config.dataset == 'nyu':
+                        depth_gt = depth_gt / 1000.0
+                    else:
+                        depth_gt = depth_gt / 256.0
+
+                    mask = np.logical_and(
+                        depth_gt >= self.config.min_depth, depth_gt <= self.config.max_depth).squeeze()[None, ...]
+                else:
+                    mask = False
+
+            if self.config.do_kb_crop:
+                height = image.shape[0]
+                width = image.shape[1]
+                top_margin = int(height - 352)
+                left_margin = int((width - 1216) / 2)
+                image = image[top_margin:top_margin + 352,
+                              left_margin:left_margin + 1216, :]
+                if self.mode == 'online_eval' and has_valid_depth:
+                    depth_gt = depth_gt[top_margin:top_margin +
+                                        352, left_margin:left_margin + 1216, :]
+
+            if self.mode == 'online_eval':
+                sample = {'image': image, 'depth': depth_gt, 'focal': focal, 'has_valid_depth': has_valid_depth,
+                          'image_path': sample_path.split()[0], 'depth_path': sample_path.split()[1],
+                          'mask': mask}
+            else:
+                sample = {'image': image, 'focal': focal}
+
+        if (self.mode == 'train') or ('has_valid_depth' in sample and sample['has_valid_depth']):
+            mask = np.logical_and(depth_gt > self.config.min_depth,
+                                  depth_gt < self.config.max_depth).squeeze()[None, ...]
+            sample['mask'] = mask
+
+        if self.transform:
+            sample = self.transform(sample)
+
+        sample = self.postprocess(sample)
+        sample['dataset'] = self.config.dataset
+        sample = {**sample, 'image_path': sample_path.split()[0], 'depth_path': sample_path.split()[1]}
+
+        return sample
+
+    def rotate_image(self, image, angle, flag=Image.BILINEAR):
+        result = image.rotate(angle, resample=flag)
+        return result
+
+    def random_crop(self, img, depth, height, width):
+        assert img.shape[0] >= height
+        assert img.shape[1] >= width
+        assert img.shape[0] == depth.shape[0]
+        assert img.shape[1] == depth.shape[1]
+        x = random.randint(0, img.shape[1] - width)
+        y = random.randint(0, img.shape[0] - height)
+        img = img[y:y + height, x:x + width, :]
+        depth = depth[y:y + height, x:x + width, :]
+
+        return img, depth
+    
+    def random_translate(self, img, depth, max_t=20):
+        assert img.shape[0] == depth.shape[0]
+        assert img.shape[1] == depth.shape[1]
+        p = self.config.translate_prob
+        do_translate = random.random()
+        if do_translate > p:
+            return img, depth
+        x = random.randint(-max_t, max_t)
+        y = random.randint(-max_t, max_t)
+        M = np.float32([[1, 0, x], [0, 1, y]])
+        # print(img.shape, depth.shape)
+        img = cv2.warpAffine(img, M, (img.shape[1], img.shape[0]))
+        depth = cv2.warpAffine(depth, M, (depth.shape[1], depth.shape[0]))
+        depth = depth.squeeze()[..., None]  # add channel dim back. Affine warp removes it
+        # print("after", img.shape, depth.shape)
+        return img, depth
+
+    def train_preprocess(self, image, depth_gt):
+        if self.config.aug:
+            # Random flipping
+            do_flip = random.random()
+            if do_flip > 0.5:
+                image = (image[:, ::-1, :]).copy()
+                depth_gt = (depth_gt[:, ::-1, :]).copy()
+
+            # Random gamma, brightness, color augmentation
+            do_augment = random.random()
+            if do_augment > 0.5:
+                image = self.augment_image(image)
+
+        return image, depth_gt
+
+    def augment_image(self, image):
+        # gamma augmentation
+        gamma = random.uniform(0.9, 1.1)
+        image_aug = image ** gamma
+
+        # brightness augmentation
+        if self.config.dataset == 'nyu':
+            brightness = random.uniform(0.75, 1.25)
+        else:
+            brightness = random.uniform(0.9, 1.1)
+        image_aug = image_aug * brightness
+
+        # color augmentation
+        colors = np.random.uniform(0.9, 1.1, size=3)
+        white = np.ones((image.shape[0], image.shape[1]))
+        color_image = np.stack([white * colors[i] for i in range(3)], axis=2)
+        image_aug *= color_image
+        image_aug = np.clip(image_aug, 0, 1)
+
+        return image_aug
+
+    def __len__(self):
+        return len(self.filenames)
+
+
+class ToTensor(object):
+    def __init__(self, mode, do_normalize=False, size=None):
+        self.mode = mode
+        self.normalize = transforms.Normalize(
+            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) if do_normalize else nn.Identity()
+        self.size = size
+        if size is not None:
+            self.resize = transforms.Resize(size=size)
+        else:
+            self.resize = nn.Identity()
+
+    def __call__(self, sample):
+        image, focal = sample['image'], sample['focal']
+        image = self.to_tensor(image)
+        image = self.normalize(image)
+        image = self.resize(image)
+
+        if self.mode == 'test':
+            return {'image': image, 'focal': focal}
+
+        depth = sample['depth']
+        if self.mode == 'train':
+            depth = self.to_tensor(depth)
+            return {**sample, 'image': image, 'depth': depth, 'focal': focal}
+        else:
+            has_valid_depth = sample['has_valid_depth']
+            image = self.resize(image)
+            return {**sample, 'image': image, 'depth': depth, 'focal': focal, 'has_valid_depth': has_valid_depth,
+                    'image_path': sample['image_path'], 'depth_path': sample['depth_path']}
+
+    def to_tensor(self, pic):
+        if not (_is_pil_image(pic) or _is_numpy_image(pic)):
+            raise TypeError(
+                'pic should be PIL Image or ndarray. Got {}'.format(type(pic)))
+
+        if isinstance(pic, np.ndarray):
+            img = torch.from_numpy(pic.transpose((2, 0, 1)))
+            return img
+
+        # handle PIL Image
+        if pic.mode == 'I':
+            img = torch.from_numpy(np.array(pic, np.int32, copy=False))
+        elif pic.mode == 'I;16':
+            img = torch.from_numpy(np.array(pic, np.int16, copy=False))
+        else:
+            img = torch.ByteTensor(
+                torch.ByteStorage.from_buffer(pic.tobytes()))
+        # PIL image mode: 1, L, P, I, F, RGB, YCbCr, RGBA, CMYK
+        if pic.mode == 'YCbCr':
+            nchannel = 3
+        elif pic.mode == 'I;16':
+            nchannel = 1
+        else:
+            nchannel = len(pic.mode)
+        img = img.view(pic.size[1], pic.size[0], nchannel)
+
+        img = img.transpose(0, 1).transpose(0, 2).contiguous()
+        if isinstance(img, torch.ByteTensor):
+            return img.float()
+        else:
+            return img

src/flux/annotator/zoe/zoedepth/data/ddad.py

@@ -0,0 +1,117 @@
+# MIT License
+
+# Copyright (c) 2022 Intelligent Systems Lab Org
+
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+# File author: Shariq Farooq Bhat
+
+import os
+
+import numpy as np
+import torch
+from PIL import Image
+from torch.utils.data import DataLoader, Dataset
+from torchvision import transforms
+
+
+class ToTensor(object):
+    def __init__(self, resize_shape):
+        # self.normalize = transforms.Normalize(
+        #     mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+        self.normalize = lambda x : x
+        self.resize = transforms.Resize(resize_shape)
+
+    def __call__(self, sample):
+        image, depth = sample['image'], sample['depth']
+        image = self.to_tensor(image)
+        image = self.normalize(image)
+        depth = self.to_tensor(depth)
+
+        image = self.resize(image)
+
+        return {'image': image, 'depth': depth, 'dataset': "ddad"}
+
+    def to_tensor(self, pic):
+
+        if isinstance(pic, np.ndarray):
+            img = torch.from_numpy(pic.transpose((2, 0, 1)))
+            return img
+
+        #         # handle PIL Image
+        if pic.mode == 'I':
+            img = torch.from_numpy(np.array(pic, np.int32, copy=False))
+        elif pic.mode == 'I;16':
+            img = torch.from_numpy(np.array(pic, np.int16, copy=False))
+        else:
+            img = torch.ByteTensor(
+                torch.ByteStorage.from_buffer(pic.tobytes()))
+        # PIL image mode: 1, L, P, I, F, RGB, YCbCr, RGBA, CMYK
+        if pic.mode == 'YCbCr':
+            nchannel = 3
+        elif pic.mode == 'I;16':
+            nchannel = 1
+        else:
+            nchannel = len(pic.mode)
+        img = img.view(pic.size[1], pic.size[0], nchannel)
+
+        img = img.transpose(0, 1).transpose(0, 2).contiguous()
+
+        if isinstance(img, torch.ByteTensor):
+            return img.float()
+        else:
+            return img
+
+
+class DDAD(Dataset):
+    def __init__(self, data_dir_root, resize_shape):
+        import glob
+
+        # image paths are of the form <data_dir_root>/{outleft, depthmap}/*.png
+        self.image_files = glob.glob(os.path.join(data_dir_root, '*.png'))
+        self.depth_files = [r.replace("_rgb.png", "_depth.npy")
+                            for r in self.image_files]
+        self.transform = ToTensor(resize_shape)
+
+    def __getitem__(self, idx):
+
+        image_path = self.image_files[idx]
+        depth_path = self.depth_files[idx]
+
+        image = np.asarray(Image.open(image_path), dtype=np.float32) / 255.0
+        depth = np.load(depth_path)  # meters
+
+        # depth[depth > 8] = -1
+        depth = depth[..., None]
+
+        sample = dict(image=image, depth=depth)
+        sample = self.transform(sample)
+
+        if idx == 0:
+            print(sample["image"].shape)
+
+        return sample
+
+    def __len__(self):
+        return len(self.image_files)
+
+
+def get_ddad_loader(data_dir_root, resize_shape, batch_size=1, **kwargs):
+    dataset = DDAD(data_dir_root, resize_shape)
+    return DataLoader(dataset, batch_size, **kwargs)

src/flux/annotator/zoe/zoedepth/data/diml_indoor_test.py

@@ -0,0 +1,125 @@
+# MIT License
+
+# Copyright (c) 2022 Intelligent Systems Lab Org
+
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+# File author: Shariq Farooq Bhat
+
+import os
+
+import numpy as np
+import torch
+from PIL import Image
+from torch.utils.data import DataLoader, Dataset
+from torchvision import transforms
+
+
+class ToTensor(object):
+    def __init__(self):
+        # self.normalize = transforms.Normalize(
+        #     mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+        self.normalize = lambda x : x
+        self.resize = transforms.Resize((480, 640))
+
+    def __call__(self, sample):
+        image, depth = sample['image'], sample['depth']
+        image = self.to_tensor(image)
+        image = self.normalize(image)
+        depth = self.to_tensor(depth)
+
+        image = self.resize(image)
+
+        return {'image': image, 'depth': depth, 'dataset': "diml_indoor"}
+
+    def to_tensor(self, pic):
+
+        if isinstance(pic, np.ndarray):
+            img = torch.from_numpy(pic.transpose((2, 0, 1)))
+            return img
+
+        #         # handle PIL Image
+        if pic.mode == 'I':
+            img = torch.from_numpy(np.array(pic, np.int32, copy=False))
+        elif pic.mode == 'I;16':
+            img = torch.from_numpy(np.array(pic, np.int16, copy=False))
+        else:
+            img = torch.ByteTensor(
+                torch.ByteStorage.from_buffer(pic.tobytes()))
+        # PIL image mode: 1, L, P, I, F, RGB, YCbCr, RGBA, CMYK
+        if pic.mode == 'YCbCr':
+            nchannel = 3
+        elif pic.mode == 'I;16':
+            nchannel = 1
+        else:
+            nchannel = len(pic.mode)
+        img = img.view(pic.size[1], pic.size[0], nchannel)
+
+        img = img.transpose(0, 1).transpose(0, 2).contiguous()
+        if isinstance(img, torch.ByteTensor):
+            return img.float()
+        else:
+            return img
+
+
+class DIML_Indoor(Dataset):
+    def __init__(self, data_dir_root):
+        import glob
+
+        # image paths are of the form <data_dir_root>/{HR, LR}/<scene>/{color, depth_filled}/*.png
+        self.image_files = glob.glob(os.path.join(
+            data_dir_root, "LR", '*', 'color', '*.png'))
+        self.depth_files = [r.replace("color", "depth_filled").replace(
+            "_c.png", "_depth_filled.png") for r in self.image_files]
+        self.transform = ToTensor()
+
+    def __getitem__(self, idx):
+        image_path = self.image_files[idx]
+        depth_path = self.depth_files[idx]
+
+        image = np.asarray(Image.open(image_path), dtype=np.float32) / 255.0
+        depth = np.asarray(Image.open(depth_path),
+                           dtype='uint16') / 1000.0  # mm to meters
+
+        # print(np.shape(image))
+        # print(np.shape(depth))
+
+        # depth[depth > 8] = -1
+        depth = depth[..., None]
+
+        sample = dict(image=image, depth=depth)
+
+        # return sample
+        sample = self.transform(sample)
+
+        if idx == 0:
+            print(sample["image"].shape)
+
+        return sample
+
+    def __len__(self):
+        return len(self.image_files)
+
+
+def get_diml_indoor_loader(data_dir_root, batch_size=1, **kwargs):
+    dataset = DIML_Indoor(data_dir_root)
+    return DataLoader(dataset, batch_size, **kwargs)
+
+# get_diml_indoor_loader(data_dir_root="datasets/diml/indoor/test/HR")
+# get_diml_indoor_loader(data_dir_root="datasets/diml/indoor/test/LR")

src/flux/annotator/zoe/zoedepth/data/diml_outdoor_test.py

@@ -0,0 +1,114 @@
+# MIT License
+
+# Copyright (c) 2022 Intelligent Systems Lab Org
+
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+# File author: Shariq Farooq Bhat
+
+import os
+
+import numpy as np
+import torch
+from PIL import Image
+from torch.utils.data import DataLoader, Dataset
+from torchvision import transforms
+
+
+class ToTensor(object):
+    def __init__(self):
+        # self.normalize = transforms.Normalize(
+        #     mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+        self.normalize = lambda x : x
+
+    def __call__(self, sample):
+        image, depth = sample['image'], sample['depth']
+        image = self.to_tensor(image)
+        image = self.normalize(image)
+        depth = self.to_tensor(depth)
+
+        return {'image': image, 'depth': depth, 'dataset': "diml_outdoor"}
+
+    def to_tensor(self, pic):
+
+        if isinstance(pic, np.ndarray):
+            img = torch.from_numpy(pic.transpose((2, 0, 1)))
+            return img
+
+        #         # handle PIL Image
+        if pic.mode == 'I':
+            img = torch.from_numpy(np.array(pic, np.int32, copy=False))
+        elif pic.mode == 'I;16':
+            img = torch.from_numpy(np.array(pic, np.int16, copy=False))
+        else:
+            img = torch.ByteTensor(
+                torch.ByteStorage.from_buffer(pic.tobytes()))
+        # PIL image mode: 1, L, P, I, F, RGB, YCbCr, RGBA, CMYK
+        if pic.mode == 'YCbCr':
+            nchannel = 3
+        elif pic.mode == 'I;16':
+            nchannel = 1
+        else:
+            nchannel = len(pic.mode)
+        img = img.view(pic.size[1], pic.size[0], nchannel)
+
+        img = img.transpose(0, 1).transpose(0, 2).contiguous()
+        if isinstance(img, torch.ByteTensor):
+            return img.float()
+        else:
+            return img
+
+
+class DIML_Outdoor(Dataset):
+    def __init__(self, data_dir_root):
+        import glob
+
+        # image paths are of the form <data_dir_root>/{outleft, depthmap}/*.png
+        self.image_files = glob.glob(os.path.join(
+            data_dir_root, "*", 'outleft', '*.png'))
+        self.depth_files = [r.replace("outleft", "depthmap")
+                            for r in self.image_files]
+        self.transform = ToTensor()
+
+    def __getitem__(self, idx):
+        image_path = self.image_files[idx]
+        depth_path = self.depth_files[idx]
+
+        image = np.asarray(Image.open(image_path), dtype=np.float32) / 255.0
+        depth = np.asarray(Image.open(depth_path),
+                           dtype='uint16') / 1000.0  # mm to meters
+
+        # depth[depth > 8] = -1
+        depth = depth[..., None]
+
+        sample = dict(image=image, depth=depth, dataset="diml_outdoor")
+
+        # return sample
+        return self.transform(sample)
+
+    def __len__(self):
+        return len(self.image_files)
+
+
+def get_diml_outdoor_loader(data_dir_root, batch_size=1, **kwargs):
+    dataset = DIML_Outdoor(data_dir_root)
+    return DataLoader(dataset, batch_size, **kwargs)
+
+# get_diml_outdoor_loader(data_dir_root="datasets/diml/outdoor/test/HR")
+# get_diml_outdoor_loader(data_dir_root="datasets/diml/outdoor/test/LR")

src/flux/annotator/zoe/zoedepth/data/diode.py

@@ -0,0 +1,125 @@
+# MIT License
+
+# Copyright (c) 2022 Intelligent Systems Lab Org
+
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+# File author: Shariq Farooq Bhat
+
+import os
+
+import numpy as np
+import torch
+from PIL import Image
+from torch.utils.data import DataLoader, Dataset
+from torchvision import transforms
+
+
+class ToTensor(object):
+    def __init__(self):
+        # self.normalize = transforms.Normalize(
+        #     mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+        self.normalize = lambda x : x
+        self.resize = transforms.Resize(480)
+
+    def __call__(self, sample):
+        image, depth = sample['image'], sample['depth']
+        image = self.to_tensor(image)
+        image = self.normalize(image)
+        depth = self.to_tensor(depth)
+
+        image = self.resize(image)
+
+        return {'image': image, 'depth': depth, 'dataset': "diode"}
+
+    def to_tensor(self, pic):
+
+        if isinstance(pic, np.ndarray):
+            img = torch.from_numpy(pic.transpose((2, 0, 1)))
+            return img
+
+        #         # handle PIL Image
+        if pic.mode == 'I':
+            img = torch.from_numpy(np.array(pic, np.int32, copy=False))
+        elif pic.mode == 'I;16':
+            img = torch.from_numpy(np.array(pic, np.int16, copy=False))
+        else:
+            img = torch.ByteTensor(
+                torch.ByteStorage.from_buffer(pic.tobytes()))
+        # PIL image mode: 1, L, P, I, F, RGB, YCbCr, RGBA, CMYK
+        if pic.mode == 'YCbCr':
+            nchannel = 3
+        elif pic.mode == 'I;16':
+            nchannel = 1
+        else:
+            nchannel = len(pic.mode)
+        img = img.view(pic.size[1], pic.size[0], nchannel)
+
+        img = img.transpose(0, 1).transpose(0, 2).contiguous()
+
+        if isinstance(img, torch.ByteTensor):
+            return img.float()
+        else:
+            return img
+
+
+class DIODE(Dataset):
+    def __init__(self, data_dir_root):
+        import glob
+
+        # image paths are of the form <data_dir_root>/scene_#/scan_#/*.png
+        self.image_files = glob.glob(
+            os.path.join(data_dir_root, '*', '*', '*.png'))
+        self.depth_files = [r.replace(".png", "_depth.npy")
+                            for r in self.image_files]
+        self.depth_mask_files = [
+            r.replace(".png", "_depth_mask.npy") for r in self.image_files]
+        self.transform = ToTensor()
+
+    def __getitem__(self, idx):
+        image_path = self.image_files[idx]
+        depth_path = self.depth_files[idx]
+        depth_mask_path = self.depth_mask_files[idx]
+
+        image = np.asarray(Image.open(image_path), dtype=np.float32) / 255.0
+        depth = np.load(depth_path)  # in meters
+        valid = np.load(depth_mask_path)  # binary
+
+        # depth[depth > 8] = -1
+        # depth = depth[..., None]
+
+        sample = dict(image=image, depth=depth, valid=valid)
+
+        # return sample
+        sample = self.transform(sample)
+
+        if idx == 0:
+            print(sample["image"].shape)
+
+        return sample
+
+    def __len__(self):
+        return len(self.image_files)
+
+
+def get_diode_loader(data_dir_root, batch_size=1, **kwargs):
+    dataset = DIODE(data_dir_root)
+    return DataLoader(dataset, batch_size, **kwargs)
+
+# get_diode_loader(data_dir_root="datasets/diode/val/outdoor")