InPeerReview/RemoteSensingChangeDetection-RSCD.HA2F

1import sys2from model.trainer import Trainer3 4sys.path.insert(0, '.')5 6import torch7import torch.nn.functional as F8import torch.backends.cudnn as cudnn9from torch.nn.parallel import gather10import torch.optim.lr_scheduler11 12import dataset.dataset as myDataLoader13import dataset.Transforms as myTransforms14from model.metric_tool import ConfuseMatrixMeter15from model.utils import BCEDiceLoss, init_seed16from PIL import Image17import os18import time19import numpy as np20from argparse import ArgumentParser21from tqdm import tqdm22 23 24@torch.no_grad()25def validate(args, val_loader, model, save_masks=False):26    model.eval()27 28    # 确保所有BatchNorm层使用全局统计量29    for m in model.modules():30        if isinstance(m, (torch.nn.BatchNorm2d, torch.nn.BatchNorm1d)):31            m.track_running_stats = True32            m.eval()33 34    salEvalVal = ConfuseMatrixMeter(n_class=2)35    epoch_loss = []36 37    if save_masks:38        mask_dir = f"{args.savedir}/pred_masks"39        os.makedirs(mask_dir, exist_ok=True)40        print(f"Saving prediction masks to: {mask_dir}")41 42    pbar = tqdm(enumerate(val_loader), total=len(val_loader), desc="Validating")43 44    for batch_idx, batched_inputs in pbar:45        img, target = batched_inputs46        # 获取当前batch的所有文件名47        batch_file_names = val_loader.sampler.data_source.file_list[48            batch_idx * args.batch_size : (batch_idx + 1) * args.batch_size49        ]50        51        pre_img = img[:, 0:3]52        post_img = img[:, 3:6]53 54        if args.onGPU:55            pre_img = pre_img.cuda()56            post_img = post_img.cuda()57            target = target.cuda()58 59        target = target.float()60        output = model(pre_img, post_img)61        loss = BCEDiceLoss(output, target)62        pred = (output > 0.5).long()63 64        if save_masks:65            pred_np = pred.cpu().numpy().astype(np.uint8)66            67            print(f"\nDebug - Batch {batch_idx}: {len(batch_file_names)} files, Mask shape: {pred_np.shape}")68            69            try:70                for i in range(pred_np.shape[0]):71                    if i >= len(batch_file_names):  # 防止文件名不足72                        print(f"Warning: Missing filename for mask {i}, using default")73                        base_name = f"batch_{batch_idx}_mask_{i}"74                    else:75                        base_name = os.path.splitext(os.path.basename(batch_file_names[i]))[0]76                    77                    single_mask = pred_np[i, 0]  # 获取(1, 256, 256)中的(256, 256)78                    79                    if single_mask.ndim != 2:80                        raise ValueError(f"Invalid mask shape: {single_mask.shape}")81                    82                    mask_path = f"{mask_dir}/{base_name}_pred.png"83                    Image.fromarray(single_mask * 255).save(mask_path)84                    print(f"Saved: {mask_path}")85 86            except Exception as e:87                print(f"\nError saving batch {batch_idx}: {str(e)}")88                print(f"Current mask shape: {single_mask.shape if 'single_mask' in locals() else 'N/A'}")89                print(f"Current file: {base_name if 'base_name' in locals() else 'N/A'}")90 91        if args.onGPU and torch.cuda.device_count() > 1:92            pred = gather(pred, 0, dim=0)93 94        f1 = salEvalVal.update_cm(pr=pred.cpu().numpy(), gt=target.cpu().numpy())95        epoch_loss.append(loss.item())96 97        pbar.set_postfix({'Loss': f"{loss.item():.4f}", 'F1': f"{f1:.4f}"})98 99    average_loss = sum(epoch_loss) / len(epoch_loss)100    scores = salEvalVal.get_scores()101    return average_loss, scores102    103def ValidateSegmentation(args):104    """完整的验证流程主函数"""105    # 初始化设置106    os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu_id)107    torch.backends.cudnn.benchmark = True108    init_seed(args.seed)  # 固定随机种子保证可重复性109 110    # 模型路径设置111    args.savedir = os.path.join(args.savedir,112                                f"{args.file_root}_iter_{args.max_steps}_lr_{args.lr}")113    os.makedirs(args.savedir, exist_ok=True)114 115    # 数据集路径配置116    dataset_mapping = {117        'LEVIR': './levir_cd_256',118        'WHU': './whu_cd_256',119        'CLCD': './clcd_256',120        'SYSU': './sysu_256',121        'OSCD': './oscd_256'122    }123    args.file_root = dataset_mapping.get(args.file_root, args.file_root)124 125    # 初始化模型126    model = Trainer(args.model_type).float()127    if args.onGPU:128        model = model.cuda()129 130    # 数据预处理 - 保持与训练时验证集相同的预处理131    mean = [0.406, 0.456, 0.485, 0.406, 0.456, 0.485]132    std = [0.225, 0.224, 0.229, 0.225, 0.224, 0.229]133 134    valDataset = myTransforms.Compose([135        myTransforms.Normalize(mean=mean, std=std),136        myTransforms.Scale(args.inWidth, args.inHeight),137        myTransforms.ToTensor()138    ])139 140    # 数据加载141    test_data = myDataLoader.Dataset(file_root=args.file_root, mode="test", transform=valDataset)142    testLoader = torch.utils.data.DataLoader(143        test_data,144        batch_size=args.batch_size,145        shuffle=False,146        num_workers=args.num_workers,147        pin_memory=True148    )149 150    # 日志设置151    logFileLoc = os.path.join(args.savedir, args.logFile)152    logger = open(logFileLoc, 'a' if os.path.exists(logFileLoc) else 'w')153    if not os.path.exists(logFileLoc):154        logger.write("\n%s\t%s\t%s\t%s\t%s\t%s\t%s" %155                     ('Epoch', 'Kappa', 'IoU', 'F1', 'Recall', 'Precision', 'OA'))156    logger.flush()157 158    # 加载最佳模型159    model_file_name = os.path.join(args.savedir, 'best_model.pth')160    if not os.path.exists(model_file_name):161        raise FileNotFoundError(f"Model file not found: {model_file_name}")162 163    state_dict = torch.load(model_file_name)164    model.load_state_dict(state_dict)165    print(f"Loaded model from {model_file_name}")166 167    # 执行验证168    loss_test, score_test = validate(args, testLoader, model, save_masks=args.save_masks)169 170    # 输出结果171    print("\nTest Results:")172    print(f"Loss: {loss_test:.4f}")173    print(f"Kappa: {score_test['Kappa']:.4f}")174    print(f"IoU: {score_test['IoU']:.4f}")175    print(f"F1: {score_test['F1']:.4f}")176    print(f"Recall: {score_test['recall']:.4f}")177    print(f"Precision: {score_test['precision']:.4f}")178    print(f"OA: {score_test['OA']:.4f}")179 180    # 记录日志181    logger.write("\n%s\t\t%.4f\t\t%.4f\t\t%.4f\t\t%.4f\t\t%.4f\t\t%.4f" %182                 ('Test', score_test['Kappa'], score_test['IoU'], score_test['F1'],183                  score_test['recall'], score_test['precision'], score_test['OA']))184    logger.close()185 186 187if __name__ == '__main__':188    parser = ArgumentParser()189    parser.add_argument('--file_root', default="LEVIR",190                        help='Data directory | LEVIR | WHU | CLCD | SYSU | OSCD')191    parser.add_argument('--inWidth', type=int, default=256, help='Width of input image')192    parser.add_argument('--inHeight', type=int, default=256, help='Height of input image')193    parser.add_argument('--max_steps', type=int, default=80000,194                        help='Max. number of iterations (for path naming)')195    parser.add_argument('--num_workers', type=int, default=4,196                        help='Number of data loading workers')197    parser.add_argument('--model_type', type=str, default='small',198                        help='Model type | tiny | small')199    parser.add_argument('--batch_size', type=int, default=16,200                        help='Batch size for validation')201    parser.add_argument('--lr', type=float, default=2e-4,202                        help='Learning rate (for path naming)')203    parser.add_argument('--seed', type=int, default=16,204                        help='Random seed for reproducibility')205    parser.add_argument('--savedir', default='./results',206                        help='Base directory to save results')207    parser.add_argument('--logFile', default='testLog.txt',208                        help='File to save validation logs')209    parser.add_argument('--onGPU', default=True,210                        type=lambda x: (str(x).lower() == 'true'),211                        help='Run on GPU if True')212    parser.add_argument('--gpu_id', type=int, default=0,213                        help='GPU device id')214    parser.add_argument('--save_masks', action='store_true',215                        help='Save predicted masks to disk')216 217    args = parser.parse_args()218    print('Validation with args:')219    print(args)220 221    ValidateSegmentation(args)222 223