up

parti_prompts.py

@@ -1,6 +1,7 @@
 #!/usr/bin/env python3
 from diffusers import DiffusionPipeline, DDIMScheduler
 import argparse
+from diffusers.pipelines.stable_diffusion import safety_checker
 import torch
 from datasets import load_dataset
 import PIL
@@ -12,12 +13,12 @@ def resize(image: PIL.Image):
     return image.resize(IMAGE_OUTPUT_SIZE, resample=PIL.Image.Resampling.LANCZOS)
 
 def get_sd_eval(ckpt, guidance_scale=7.5):
-    pipe = DiffusionPipeline.from_pretrained(ckpt, torch_dtype=torch.float16)
+    pipe = DiffusionPipeline.from_pretrained(ckpt, torch_dtype=torch.float16, safety_checker=None)
     pipe.to("cuda")
-    pipe.scheduler = DDIMScheduler.from_config(pipe.config)
+    pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
 
-    def sd_eval(prompt):
-        images = pipe(prompt, num_inference_steps=100, guidance_scale=guidance_scale).images
+    def sd_eval(prompt, generator=None):
+        images = pipe(prompt, generator=generator, num_inference_steps=NUM_INFERENCE_STEPS, guidance_scale=guidance_scale).images
         images = [resize(image) for image in images]
         return images
 
@@ -28,28 +29,28 @@ def get_karlo_eval(ckpt):
     pipe.to("cuda")
 
     def karlo_eval(prompt):
-        images = pipe(prompt, prior_num_inference_steps=50, decoder_num_inference_steps=100).images
+        images = pipe(prompt, prior_num_inference_steps=50, decoder_num_inference_steps=NUM_INFERENCE_STEPS).images
         return images
 
     return karlo_eval
 
 def get_if_eval(ckpt):
-    pipe_low = DiffusionPipeline.from_pretrained(ckpt, torch_dtype=torch.float16)
+    pipe_low = DiffusionPipeline.from_pretrained(ckpt, safety_checker=None, torch_dtype=torch.float16)
     pipe_low.enable_model_cpu_offload()
 
-    pipe_up = DiffusionPipeline.from_pretrained("DeepFloyd/IF-II-L-v1.0", text_encoder=pipe_low.text_encoder, torch_dtype=torch.float16)
+    pipe_up = DiffusionPipeline.from_pretrained("DeepFloyd/IF-II-L-v1.0", safety_checker=None, text_encoder=pipe_low.text_encoder, torch_dtype=torch.float16)
     pipe_up.enable_model_cpu_offload()
 
-    def sd_eval(prompt):
-        images = pipe_low(prompt, num_inference_steps=100, output_type="pt").images
-        images = pipe_up(promtp=prompt, images=images, num_inference_steps=100).images
+    def if_eval(prompt, generator=None):
+        images = pipe_low(prompt, num_inference_steps=NUM_INFERENCE_STEPS, generator=generator, output_type="pt").images
+        images = pipe_up(promtp=prompt, images=images, num_inference_steps=NUM_INFERENCE_STEPS).images
         return images
 
-    return sd_eval
+    return if_eval
 
 MODELS = {
     "runwayml/stable-diffusion-v1-5": get_sd_eval,
-    "stabilityai/stable-diffusion-v2-1": get_sd_eval,
+    "stabilityai/stable-diffusion-2-1": get_sd_eval,
     "kakaobrain/karlo-alpha": get_karlo_eval,
     "DeepFloyd/IF-I-XL-v1.0": get_if_eval,
 }
@@ -59,24 +60,28 @@ MODELS = {
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(description='Run Parti Prompt Evaluation')
-    parser.add_argument('model_repo_or_id', type=str, help='ID or URL of the model repository.', required=True)
+    parser.add_argument('model_repo_or_id', type=str, help='ID or URL of the model repository.')
     parser.add_argument('--dataset_repo_or_id', type=str, default='diffusers/prompt_generations', help='ID or URL of the dataset repository (default: "diffusers/prompt_generations")')
     parser.add_argument('--batch_size', type=int, default=8, help="Batch size for the eval function")
     parser.add_argument('--upload_to_hub', action='store_true', help='whether to upload the dataset to the Hugging Face dataset hub')
+    parser.add_argument('--seed', type=int, default=0, help='Random seed')
 
     args = parser.parse_args()
 
-    eval_fn = MODELS[args.model_repo_or_id](args.model_repo_or_id)
+    dataset = load_dataset("nateraw/parti-prompts")["train"]
 
-    dataset = load_dataset("nateraw/parti-prompts")
+    eval_fn = MODELS[args.model_repo_or_id](args.model_repo_or_id)
 
     def map_fn(batch):
-        batch["images"] = eval_fn(batch["prompt"])
+        generators = [torch.Generator(device="cuda").manual_seed(args.seed) for _ in range(args.batch_size)]
+        batch["images"] = eval_fn(batch["Prompt"], generator=generators)
+        batch["model_name"] = len(batch["images"]) * [args.model_repo_or_id]
+        batch["seed"] = len(batch["images"]) * [args.seed]
         return batch
 
-    dataset_images = dataset.map(map_fn, batched=True, batch_size=8)
+    dataset_images = dataset.map(map_fn, batched=True, batch_size=args.batch_size)
 
     if args.upload_to_hub:
-        dataset.push_to_hub(args.dataset_repo_or_id)
+        dataset_images.push_to_hub(args.dataset_repo_or_id)
     else:
-        dataset.save_to_disk(args.dataset_repo_or_id.split("/")[-1])
+        dataset_images.save_to_disk(args.dataset_repo_or_id.split("/")[-1])