src/training/datasets.py

"""
src/training/datasets.py

Shared dataset utilities used by scripts/ entrypoints.
"""
from __future__ import annotations

import csv
from pathlib import Path
from typing import Optional

import numpy as np
import torch
from PIL import Image
from torch.utils.data import Dataset
from torchvision import transforms

IMAGE_EXTS = {".jpg", ".jpeg", ".png", ".webp"}

IMAGENET_MEAN = [0.485, 0.456, 0.406]
IMAGENET_STD  = [0.229, 0.224, 0.225]


def get_train_transform(size: int = 224):
    return transforms.Compose([
        transforms.RandomResizedCrop(size, scale=(0.8, 1.0)),
        transforms.RandomHorizontalFlip(),
        transforms.ColorJitter(0.2, 0.2, 0.2, 0.1),
        transforms.ToTensor(),
        transforms.Normalize(IMAGENET_MEAN, IMAGENET_STD),
    ])


def get_val_transform(size: int = 224):
    return transforms.Compose([
        transforms.Resize(int(size * 256 / 224)),
        transforms.CenterCrop(size),
        transforms.ToTensor(),
        transforms.Normalize(IMAGENET_MEAN, IMAGENET_STD),
    ])


class ImageManifestDataset(Dataset):
    """
    Generic image dataset driven by a manifest CSV.

    Manifest format: filepath, label (0=real, 1=fake), [generator (int)]
    """

    def __init__(
        self,
        manifest_path: Path,
        transform=None,
        root_dir: Optional[Path] = None,
    ):
        self.transform = transform
        self.root_dir  = Path(root_dir) if root_dir else None
        self.samples   = []

        with open(manifest_path) as f:
            reader = csv.DictReader(f)
            for row in reader:
                filepath  = Path(row["filepath"])
                if self.root_dir and not filepath.is_absolute():
                    filepath = self.root_dir / filepath
                label     = int(row["label"])
                generator = int(row.get("generator", 0))
                self.samples.append((filepath, label, generator))

    def __len__(self) -> int:
        return len(self.samples)

    def __getitem__(self, idx: int) -> dict:
        path, label, generator = self.samples[idx]
        img = Image.open(path).convert("RGB")
        if self.transform:
            img = self.transform(img)
        return {
            "image":     img,
            "label":     label,
            "generator": generator,
            "filepath":  str(path),
        }

    def get_class_weights(self) -> torch.Tensor:
        labels = [s[1] for s in self.samples]
        n_real = labels.count(0)
        n_fake = labels.count(1)
        w_real = 1.0 / max(n_real, 1)
        w_fake = 1.0 / max(n_fake, 1)
        return torch.tensor([w_real, w_fake], dtype=torch.float32)