import os import random import numpy as np from PIL import Image from scipy import io import torch from torch.utils.data import Dataset, DataLoader, Subset from torch.utils.data.distributed import DistributedSampler import torchvision.transforms.functional as TF class CustomDataset(Dataset): def __init__(self, root_dir, crop_size=None, augment=False): """ Initialize the dataset. Args: root_dir: Dataset root directory. crop_size: Crop size as an integer or a (height, width) tuple. augment: Whether to enable data augmentation for the training set. """ self.root_dir = root_dir self.input_dir = os.path.join(root_dir, 'input') self.output_dir = os.path.join(root_dir, 'output') # Sort strictly for deterministic multi-GPU and repeated runs. self.samples = sorted(os.listdir(self.input_dir)) self.augment = augment # Process the crop dimensions. if crop_size is not None: if isinstance(crop_size, (int, float)): self.crop_height = self.crop_width = int(crop_size) elif isinstance(crop_size, (tuple, list)) and len(crop_size) == 2: self.crop_height, self.crop_width = crop_size else: raise ValueError("crop_size should be an integer or a tuple of (height, width)") else: self.crop_height = self.crop_width = None def __len__(self): """Return the dataset size.""" return len(self.samples) def __getitem__(self, idx): sample_name = self.samples[idx] # 1. Read each input image: (H, W) -> [1, H, W]. img_folder = os.path.join(self.input_dir, sample_name) img1 = torch.from_numpy(np.array(Image.open(os.path.join(img_folder, 'image1.png')), dtype=np.float32)).unsqueeze(0) img2 = torch.from_numpy(np.array(Image.open(os.path.join(img_folder, 'image2.png')), dtype=np.float32)).unsqueeze(0) img3 = torch.from_numpy(np.array(Image.open(os.path.join(img_folder, 'image3.png')), dtype=np.float32)).unsqueeze(0) # Concatenate three images into [3, H, W] so transforms remain synchronized. input_tensor = torch.cat([img1, img2, img3], dim=0) # 2. Read the .mat label and convert it to a [1, H, W] tensor. label_folder = os.path.join(self.output_dir, sample_name) mat_data = io.loadmat(os.path.join(label_folder, 'label.mat')) M_tensor = torch.from_numpy(mat_data['M'].astype(np.float32)).unsqueeze(0) D_tensor = torch.from_numpy(mat_data['D'].astype(np.float32)).unsqueeze(0) # 3. Apply synchronized geometric augmentation to training data with 75% probability. if self.augment and random.random() < 0.75: # Use small random transforms to favor physical patterns over memorization. angle = random.uniform(-5.0, 5.0) # Rotation: [-5, 5] degrees. tx = random.randint(-12, 12) # X translation: [-12, 12]. ty = random.randint(-12, 12) # Y translation: [-12, 12]. scale = random.uniform(0.95, 1.05) # Scale: [0.95, 1.05]. # Apply all geometric transforms in one affine operation with bilinear interpolation. input_tensor = TF.affine(input_tensor, angle=angle, translate=[tx, ty], scale=scale, shear=0, interpolation=TF.InterpolationMode.BILINEAR) M_tensor = TF.affine(M_tensor, angle=angle, translate=[tx, ty], scale=scale, shear=0, interpolation=TF.InterpolationMode.BILINEAR) D_tensor = TF.affine(D_tensor, angle=angle, translate=[tx, ty], scale=scale, shear=0, interpolation=TF.InterpolationMode.BILINEAR) # 4. Compute sina, cosa, and the mask from the transformed physical images. img1_t, img2_t, img3_t = input_tensor[0], input_tensor[1], input_tensor[2] sina = -(img1_t * np.sin(2 * np.pi * 0 / 3) + img2_t * np.sin(2 * np.pi * 1 / 3) + img3_t * np.sin(2 * np.pi * 2 / 3)) cosa = (img1_t * np.cos(2 * np.pi * 0 / 3) + img2_t * np.cos(2 * np.pi * 1 / 3) + img3_t * np.cos(2 * np.pi * 2 / 3)) # Create the mask in 2D first to avoid a dimension mismatch: [H, W]. mask_2d = torch.ones_like(img1_t, dtype=torch.float32) # Apply Boolean indexing while the mask is 2D. mask_2d[(sina == 0) & (cosa == 0)] = 0 # Add the channel dimension after assignment to obtain [1, H, W]. mask_tensor = mask_2d.unsqueeze(0) # 5. Center-crop the inputs, labels, and mask together. if self.crop_height is not None and self.crop_width is not None: _, H, W = input_tensor.shape start_h = (H - self.crop_height) // 2 start_w = (W - self.crop_width) // 2 input_tensor = input_tensor[:, start_h:start_h + self.crop_height, start_w:start_w + self.crop_width] M_tensor = M_tensor[:, start_h:start_h + self.crop_height, start_w:start_w + self.crop_width] D_tensor = D_tensor[:, start_h:start_h + self.crop_height, start_w:start_w + self.crop_width] mask_tensor = mask_tensor[:, start_h:start_h + self.crop_height, start_w:start_w + self.crop_width] # 6. Apply consistent value normalization by dividing by 1000. input_tensor = input_tensor / 1000.0 M_tensor = M_tensor / 1000.0 D_tensor = D_tensor / 1000.0 return input_tensor, {'M': M_tensor, 'D': D_tensor, 'mask': mask_tensor} def seed_everything(seed=42): """Set global seeds for reproducible results.""" random.seed(seed) os.environ['PYTHONHASHSEED'] = str(seed) np.random.seed(seed) torch.manual_seed(seed) torch.cuda.manual_seed(seed) torch.cuda.manual_seed_all(seed) torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = False def seed_worker(worker_id): """Set a distinct seed for each DataLoader worker.""" worker_seed = torch.initial_seed() % 2**32 np.random.seed(worker_seed) random.seed(worker_seed) def create_data_loaders( data_dir, batch_size=4, train_ratio=0.8, num_workers=0, crop_size=None, seed=42, distributed=False ): """ Create data loaders with a deterministic split and training-only augmentation. """ seed_everything(seed) # Enable augmentation for training while keeping validation images unchanged. train_dataset_full = CustomDataset(data_dir, crop_size=crop_size, augment=True) val_dataset_full = CustomDataset(data_dir, crop_size=crop_size, augment=False) # Split deterministic indices instead of sharing random_split dataset state. dataset_size = len(train_dataset_full) indices = list(range(dataset_size)) np_rng = np.random.RandomState(seed) np_rng.shuffle(indices) train_size = int(train_ratio * dataset_size) train_indices = indices[:train_size] val_indices = indices[train_size:] # Separate training and validation data with Subset. train_dataset = Subset(train_dataset_full, train_indices) val_dataset = Subset(val_dataset_full, val_indices) g = torch.Generator() g.manual_seed(seed) if distributed: train_sampler = DistributedSampler(train_dataset, shuffle=True, seed=seed, drop_last=True) val_sampler = DistributedSampler(val_dataset, shuffle=False, drop_last=False) else: train_sampler = None val_sampler = None train_loader = DataLoader( train_dataset, batch_size=batch_size, shuffle=(train_sampler is None), sampler=train_sampler, num_workers=num_workers, pin_memory=True, drop_last=True, persistent_workers=(num_workers > 0), worker_init_fn=seed_worker, generator=g, ) val_loader = DataLoader( val_dataset, batch_size=batch_size, shuffle=False, sampler=val_sampler, num_workers=num_workers, pin_memory=True, drop_last=False, # Keep every validation sample for complete evaluation. persistent_workers=(num_workers > 0), worker_init_fn=seed_worker, generator=g, ) return train_loader, val_loader if __name__ == "__main__": # Basic smoke test. print("Dataloader script is ready!")