code/refuge2.py

# REFUGE2 Dataset for PixelGen Medical Image Generation
# Optic disc/cup segmentation: 1200 RGB fundus images with 3-class masks
# Mask values: 0=background, 128=optic cup, 255=optic disc

import os
import torch
import random
import numpy as np
from torch.utils.data import Dataset
from PIL import Image
import torchvision.transforms as transforms
import torchvision.transforms.functional as TF
from torchvision.transforms import Normalize


class REFUGE2Dataset(Dataset):
    """
    REFUGE2 dataset for mask-conditional image generation.

    Data format:
    - Images: RGB fundus photographs (varying sizes ~1600-2100px)
    - Masks: 3-class segmentation (0=background, 128=optic cup, 255=optic disc)
    - 400 images per split (train/val/test)
    - Mask files: train/test=.bmp, val=.png

    Returns format compatible with PixelGen:
    - normalized_image: [3, H, W] in range [-1, 1]
    - label: class label (0 for all)
    - metadata: dict with 'raw_image', 'mask', 'class'
    """

    def __init__(self, data_root, resolution=256, splits=('train', 'val'),
                 augment=True, seed=42, max_samples=None, random_flip=True,
                 val_ratio=0.0):
        super().__init__()
        self.data_root = data_root
        self.resolution = resolution
        self.augment = augment
        self.random_flip = random_flip

        # Collect files from specified splits
        all_pairs = []
        for split in splits:
            img_dir = os.path.join(data_root, split, 'images')
            mask_dir = os.path.join(data_root, split, 'mask')

            img_files = sorted([f for f in os.listdir(img_dir)
                               if f.endswith(('.jpg', '.png', '.jpeg'))])

            for img_f in img_files:
                img_path = os.path.join(img_dir, img_f)
                # Mask may have different extension (.bmp or .png)
                base_name = os.path.splitext(img_f)[0]
                mask_path = None
                for ext in ['.bmp', '.png', '.jpg']:
                    candidate = os.path.join(mask_dir, base_name + ext)
                    if os.path.exists(candidate):
                        mask_path = candidate
                        break
                if mask_path is not None:
                    all_pairs.append((img_path, mask_path))

        # Optional: hold out a portion for validation
        if val_ratio > 0:
            random.seed(seed)
            random.shuffle(all_pairs)
            split_idx = int(len(all_pairs) * (1 - val_ratio))
            self.pairs = all_pairs[:split_idx]
        else:
            self.pairs = all_pairs

        # Limit samples if specified
        if max_samples is not None and max_samples < len(self.pairs):
            random.seed(seed)
            self.pairs = random.sample(self.pairs, max_samples)

        # Normalization for images ([-1, 1] range)
        self.normalize = Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])

        print(f"[REFUGE2Dataset] splits={splits}: {len(self.pairs)} image pairs")

    def __len__(self):
        return len(self.pairs)

    def _load_and_process(self, idx):
        """Load and process a single sample."""
        img_path, mask_path = self.pairs[idx]

        image = Image.open(img_path).convert('RGB')
        mask = Image.open(mask_path).convert('L')

        # Resize to target size (square)
        image = TF.resize(image, (self.resolution, self.resolution),
                         interpolation=transforms.InterpolationMode.BILINEAR)
        mask = TF.resize(mask, (self.resolution, self.resolution),
                        interpolation=transforms.InterpolationMode.NEAREST)

        # Data augmentation
        if self.augment:
            if self.random_flip and random.random() > 0.5:
                image = TF.hflip(image)
                mask = TF.hflip(mask)

            if self.random_flip and random.random() > 0.5:
                image = TF.vflip(image)
                mask = TF.vflip(mask)

            # Random color jitter for image only
            if random.random() > 0.5:
                brightness_factor = random.uniform(0.85, 1.15)
                image = TF.adjust_brightness(image, brightness_factor)
                contrast_factor = random.uniform(0.85, 1.15)
                image = TF.adjust_contrast(image, contrast_factor)
                saturation_factor = random.uniform(0.85, 1.15)
                image = TF.adjust_saturation(image, saturation_factor)

        return image, mask

    def __getitem__(self, idx):
        max_retries = 10
        for retry in range(max_retries):
            try:
                actual_idx = (idx + retry) % len(self.pairs)
                image, mask = self._load_and_process(actual_idx)
                break
            except Exception as e:
                if retry == max_retries - 1:
                    raise RuntimeError(f"Failed to load image after {max_retries} retries: {e}")
                continue

        raw_image = TF.to_tensor(image)  # [3, H, W], range [0, 1]
        normalized_image = self.normalize(raw_image)

        # Normalize mask: 0->0.0, 128->0.5, 255->1.0
        mask_tensor = TF.to_tensor(mask)  # [1, H, W], range [0, 1]

        label = 0
        metadata = {
            "raw_image": raw_image,
            "mask": mask_tensor,
            "class": label,
        }

        return normalized_image, label, metadata


class REFUGE2ValDataset(Dataset):
    """Validation subset from REFUGE2 (held-out from training splits)."""

    def __init__(self, data_root, resolution=256, splits=('train', 'val'),
                 val_ratio=0.1, seed=42):
        super().__init__()
        self.resolution = resolution
        self.normalize = Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])

        all_pairs = []
        for split in splits:
            img_dir = os.path.join(data_root, split, 'images')
            mask_dir = os.path.join(data_root, split, 'mask')
            img_files = sorted([f for f in os.listdir(img_dir)
                               if f.endswith(('.jpg', '.png', '.jpeg'))])
            for img_f in img_files:
                base_name = os.path.splitext(img_f)[0]
                img_path = os.path.join(img_dir, img_f)
                mask_path = None
                for ext in ['.bmp', '.png', '.jpg']:
                    candidate = os.path.join(mask_dir, base_name + ext)
                    if os.path.exists(candidate):
                        mask_path = candidate
                        break
                if mask_path is not None:
                    all_pairs.append((img_path, mask_path))

        random.seed(seed)
        random.shuffle(all_pairs)
        split_idx = int(len(all_pairs) * (1 - val_ratio))
        self.pairs = all_pairs[split_idx:]

        print(f"[REFUGE2ValDataset] {len(self.pairs)} val samples")

    def __len__(self):
        return len(self.pairs)

    def __getitem__(self, idx):
        img_path, mask_path = self.pairs[idx]
        image = Image.open(img_path).convert('RGB')
        mask = Image.open(mask_path).convert('L')
        image = TF.resize(image, (self.resolution, self.resolution),
                         interpolation=transforms.InterpolationMode.BILINEAR)
        mask = TF.resize(mask, (self.resolution, self.resolution),
                        interpolation=transforms.InterpolationMode.NEAREST)

        raw_image = TF.to_tensor(image)
        normalized_image = self.normalize(raw_image)
        mask_tensor = TF.to_tensor(mask)

        label = 0
        metadata = {
            "raw_image": raw_image,
            "mask": mask_tensor,
            "class": label,
        }
        return normalized_image, label, metadata


class REFUGE2RandnDataset(Dataset):
    """Random noise dataset for evaluation/prediction with REFUGE2 masks."""

    def __init__(self, data_root, resolution=256, max_num_instances=1000,
                 noise_scale=1.0, seed=42, splits=('train', 'val', 'test')):
        super().__init__()
        self.resolution = resolution
        self.noise_scale = noise_scale

        # Collect all mask paths
        all_masks = []
        for split in splits:
            mask_dir = os.path.join(data_root, split, 'mask')
            if not os.path.exists(mask_dir):
                continue
            mask_files = sorted([f for f in os.listdir(mask_dir)
                                if f.endswith(('.bmp', '.png', '.jpg'))])
            for mf in mask_files:
                all_masks.append(os.path.join(mask_dir, mf))

        random.seed(seed)
        if max_num_instances <= len(all_masks):
            self.mask_paths = random.sample(all_masks, max_num_instances)
        else:
            self.mask_paths = all_masks * (max_num_instances // len(all_masks) + 1)
            self.mask_paths = self.mask_paths[:max_num_instances]

        print(f"[REFUGE2RandnDataset] {len(self.mask_paths)} samples for generation")

    def __len__(self):
        return len(self.mask_paths)

    def __getitem__(self, idx):
        xT = self.noise_scale * torch.randn(3, self.resolution, self.resolution)

        mask = Image.open(self.mask_paths[idx]).convert('L')
        mask = TF.resize(mask, (self.resolution, self.resolution),
                        interpolation=transforms.InterpolationMode.NEAREST)
        mask_tensor = TF.to_tensor(mask)

        label = 0
        metadata = {
            "mask": mask_tensor,
            "class": label,
        }
        return xT, label, metadata