app.py

# app.py
import os, json
import numpy as np
from PIL import Image
import torch
from transformers import (
    AutoConfig,
    AutoModelForSemanticSegmentation,
    SegformerImageProcessor,
)
import gradio as gr

# ===== Config =====
MODEL_ID = "Itbanque/fashion_segformer"
PROCESSOR_ID = MODEL_ID

# ===== Load processor =====
try:
    processor = SegformerImageProcessor.from_pretrained(PROCESSOR_ID)
except Exception:
    # 兜底：没有 preprocessor_config.json 时，手动构造
    processor = SegformerImageProcessor(
        size={"height": 512, "width": 512},
        do_resize=True,
        do_normalize=True
    )

# ===== Load model =====
try:
    cfg = AutoConfig.from_pretrained(MODEL_ID)
    model = AutoModelForSemanticSegmentation.from_pretrained(MODEL_ID, config=cfg)
except Exception:
    # 兼容老的只存了权重的目录
    model = AutoModelForSemanticSegmentation.from_pretrained(MODEL_ID)

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if (device == "cuda") else torch.float32
model.to(device=device, dtype=dtype)
model.eval()

# ===== id2label / palette =====
id2label = getattr(model.config, "id2label", None)
if isinstance(id2label, dict):
    id2label = {int(k): v for k, v in id2label.items()}
else:
    id2label = {i: str(i) for i in range(model.config.num_labels)}
NUM_CLASSES = len(id2label)

def make_palette(n: int) -> np.ndarray:
    # 固定随机种子，稳定配色；0类设为黑色
    rng = np.random.default_rng(0)
    colors = rng.integers(0, 255, size=(n, 3), dtype=np.uint8)
    colors[0] = np.array([0, 0, 0], dtype=np.uint8)
    return colors

PALETTE = make_palette(NUM_CLASSES)

# ===== Inference =====
@torch.no_grad()
def predict(pil_img: Image.Image, alpha: float = 0.5, show_overlay: bool = True):
    if pil_img is None:
        return None, None

    img = pil_img.convert("RGB")
    W, H = img.size

    inputs = processor(images=img, return_tensors="pt")
    pixel_values = inputs["pixel_values"].to(device, dtype=dtype)

    outputs = model(pixel_values=pixel_values)
    logits = outputs.logits  # (1, C, h, w)

    # 上采样到原图尺寸再 argmax
    up = torch.nn.functional.interpolate(
        logits, size=(H, W), mode="bilinear", align_corners=False
    )
    pred = up.argmax(dim=1)[0].to(torch.uint8).cpu().numpy()  # (H, W)

    # 生成彩色掩码 & 叠加
    color_mask = PALETTE[pred]  # (H, W, 3)
    mask_pil = Image.fromarray(color_mask, mode="RGB")

    if show_overlay:
        overlay = Image.blend(img, mask_pil, float(alpha))
        return overlay, mask_pil
    else:
        return None, mask_pil

# ===== Gradio UI =====
with gr.Blocks(title="Fashion Segmentation") as demo:
    gr.Markdown("## Fashion Segmentation\nUpload an image and run SegFormer inference.")
    with gr.Row():
        inp = gr.Image(type="pil", label="Upload image")
        with gr.Column():
            alpha = gr.Slider(0.0, 1.0, value=0.5, step=0.05, label="Overlay alpha")
            show_overlay = gr.Checkbox(value=True, label="Show overlay")
            btn = gr.Button("Predict", variant="primary")
    with gr.Row():
        out_overlay = gr.Image(label="Overlay", interactive=False)
        out_mask = gr.Image(label="Colored mask", interactive=False)

    def _run(image, a, show):
        return predict(image, alpha=a, show_overlay=show)

    btn.click(_run, inputs=[inp, alpha, show_overlay], outputs=[out_overlay, out_mask])

if __name__ == "__main__":
    # 本地运行：python app.py
    # HF Spaces：把本文件命名为 app.py 即可
    demo.launch()