app.py

'''
https://huggingface.co/spaces/merve/OWLSAM

text,letter,watermark

vim run_text_mask.py

from gradio_client import Client, handle_file
from datasets import load_dataset, Image as HfImage
from PIL import ImageOps, Image
import numpy as np
import os
from tqdm import tqdm

# 初始化客户端
client = Client("http://localhost:7860")

# 加载数据集
dataset_name = "svjack/InfiniteYou_PosterCraft_Wang_Leehom_Poster_FP8_WAV"
dataset = load_dataset(dataset_name)

# 创建保存 mask 的文件夹
os.makedirs("mask_images", exist_ok=True)

#### 832, 1216
#### (864, 1152)
def combine_non_white_regions(annotations):
    canvas = None
    for i, annotation in enumerate(annotations):
        img = Image.open(annotation["image"]).convert("RGBA")
        img_array = np.array(img)
        if canvas is None:
            height, width = img_array.shape[:2]
            canvas = np.zeros((height, width, 4), dtype=np.uint8)
        rgb = img_array[..., :3]
        non_white_mask = np.any(rgb < 240, axis=-1, keepdims=True)
        alpha_layer = np.where(non_white_mask, img_array[..., 3:], 0)
        processed_img = np.concatenate([rgb, alpha_layer], axis=-1)
        canvas = np.where(processed_img[..., 3:] > 0, processed_img, canvas)
    if canvas is None:
        height = 1152
        width = 864
        result_array = np.zeros((height, width, 4), dtype=np.uint8)
        result_array[..., :3] = 255
        result_array[..., 3] = 255
        return Image.fromarray(result_array.astype(np.uint8))
        
    result_array = np.zeros((height, width, 4), dtype=np.uint8)
    result_array[..., :3] = 255
    result_array[..., 3] = 255
    result_array = np.where(canvas[..., 3:] > 0, canvas, result_array)

    non_white_mask = np.any(result_array[..., :3] < 255, axis=-1)
    result_array[non_white_mask] = [0, 0, 0, 255]

    return Image.fromarray(result_array.astype(np.uint8))

def generate_mask(image, idx):
    try:
        # 保存原始图片为临时文件
        temp_input_path = f"mask_images/temp_{idx:04d}.jpg"
        image.save(temp_input_path)

        # 调用 Gradio API
        result = client.predict(
            image=handle_file(temp_input_path),
            texts="text,letter,watermark",
            threshold=0.05,
            sam_threshold=0.88,
            api_name="/predict"
        )

        # 生成 mask 图像
        mask_image = combine_non_white_regions(result["annotations"])
        mask_image = ImageOps.invert(mask_image.convert("RGB"))

        # 保存 mask 图像
        output_mask_path = f"mask_images/mask_{idx:04d}.jpg"
        mask_image.save(output_mask_path)

        return {"mask_image": output_mask_path}

    except Exception as e:
        print(f"生成 mask 时出错 (index={idx}): {e}")
        return {"mask_image": None}

# 使用 map 处理整个数据集
updated_dataset = dataset["train"].map(
    lambda example, idx: generate_mask(example["Wang_Leehom_poster_image"], idx),
    with_indices=True,
    num_proc=1,
    batched=False
)

# 转换列类型为 Image
updated_dataset = updated_dataset.cast_column("mask_image", HfImage())

# 保存更新后的数据集
output_path = "Wang_Leehom_PosterCraft_with_Mask"
updated_dataset.save_to_disk(output_path)

print(f"✅ 已生成包含 mask 的数据集并保存至: {output_path}")
'''


from transformers import pipeline, SamModel, SamProcessor
import torch
import numpy as np
import gradio as gr
import spaces

checkpoint = "google/owlv2-base-patch16-ensemble"
detector = pipeline(model=checkpoint, task="zero-shot-object-detection", device="cuda")
sam_model = SamModel.from_pretrained("facebook/sam-vit-base").to("cuda")
sam_processor = SamProcessor.from_pretrained("facebook/sam-vit-base")


@spaces.GPU
def query(image, texts, threshold, sam_threshold):
  texts = texts.split(",")

  predictions = detector(
    image,
    candidate_labels=texts,
    threshold=threshold
  )

  result_labels = []
  for pred in predictions:

    box = pred["box"]
    score = pred["score"]
    label = pred["label"]
    box = [round(pred["box"]["xmin"], 2), round(pred["box"]["ymin"], 2),
        round(pred["box"]["xmax"], 2), round(pred["box"]["ymax"], 2)]
    inputs = sam_processor(
            image,
            input_boxes=[[box]],
            return_tensors="pt"
        ).to("cuda")
    with torch.no_grad():
      outputs = sam_model(**inputs)

    mask = sam_processor.image_processor.post_process_masks(
        outputs.pred_masks.cpu(),
        inputs["original_sizes"].cpu(),
        inputs["reshaped_input_sizes"].cpu()
    )
    iou_scores = outputs["iou_scores"]
    
    masks, iou_scores, boxes = sam_processor.image_processor.filter_masks(
        mask[0],
        iou_scores[0].cpu(),
        inputs["original_sizes"][0].cpu(),
        box,
        pred_iou_thresh=sam_threshold,
    )
    
    result_labels.append((mask[0][0][0].numpy(), label))
  return image, result_labels


description = "This Space combines OWLv2, the state-of-the-art zero-shot object detection model with SAM, the state-of-the-art mask generation model. SAM normally doesn't accept text input. Combining SAM with OWLv2 makes SAM text promptable. Try the example or input an image and comma separated candidate labels to segment."
demo = gr.Interface(
    query,
    inputs=[gr.Image(type="pil", label="Image Input"), gr.Textbox(label = "Candidate Labels"), gr.Slider(0, 1, value=0.05, label="Confidence Threshold for OWL"), gr.Slider(0, 1, value=0.88, label="IoU threshold for SAM")],
    outputs="annotatedimage",
    title="OWL 🤝 SAM",
    description=description,
    examples=[
        ["./cats.png", "cat", 0.1, 0.88],
    ],
    cache_examples=True
)
demo.launch(debug=True, share = True)