test_real_save_output.py

"""
对 real 测试集跑模型，并把输出图保存到指定文件夹。
输入：./dataset/test/real_input（或 --real_inp）
输出：保存到 --output_dir，文件名与输入一致。
"""
import os
import argparse

import torch
from PIL import Image
from torchvision.transforms import ToTensor

from restormer import ChannelShuffleWithGBPDeep
import utils as utils


def load_model(weights_path: str, device: torch.device) -> torch.nn.Module:
    model = ChannelShuffleWithGBPDeep().to(device)
    utils.load_checkpointG1(model, weights_path, strict=False)
    model.eval()
    return model


def list_image_files(folder: str):
    exts = (".png", ".jpg", ".jpeg", ".bmp", ".tif", ".tiff", ".gif")
    return sorted([f for f in os.listdir(folder) if f.lower().endswith(exts)])


def main():
    parser = argparse.ArgumentParser(description="Run model on real test images and save outputs.")
    parser.add_argument("--weights", type=str, required=True, help="Path to model checkpoint.")
    parser.add_argument(
        "--real_inp",
        type=str,
        default="./dataset/test/real_input",
        help="Path to real-world test input folder.",
    )
    parser.add_argument(
        "--output_dir",
        type=str,
        default="./dataset/test/real_output",
        help="Folder to save output images.",
    )
    parser.add_argument("--gpu", type=str, default="0")
    args = parser.parse_args()

    os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
    os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print(f"Using device: {device}")

    assert os.path.exists(args.real_inp), f"real_inp not found: {args.real_inp}"
    os.makedirs(args.output_dir, exist_ok=True)
    print(f"Output dir: {args.output_dir}")

    model = load_model(args.weights, device)
    to_tensor = ToTensor()
    img_names = list_image_files(args.real_inp)
    assert len(img_names) > 0, f"No images found in {args.real_inp}"

    for name in img_names:
        inp_path = os.path.join(args.real_inp, name)
        inp_img = Image.open(inp_path).convert("RGB")
        inp = to_tensor(inp_img).unsqueeze(0).to(device)

        with torch.no_grad():
            out = model(inp)

        # (1, 3, H, W) RGB 0~1 -> 保存为 PNG
        out_np = out.squeeze(0).cpu().clamp(0, 1).numpy()
        out_np = (out_np.transpose(1, 2, 0) * 255.0).clip(0, 255).astype("uint8")
        out_pil = Image.fromarray(out_np)

        # 输出文件名与输入一致，扩展名统一为 .png
        base, _ = os.path.splitext(name)
        out_name = base + ".png"
        out_path = os.path.join(args.output_dir, out_name)
        out_pil.save(out_path)
        print(f"Saved: {out_path}")

    print(f"Done. {len(img_names)} images saved to {args.output_dir}")


if __name__ == "__main__":
    main()