Train SimCLR on your images

Start from the source installation in the quickstart. The runnable example accepts ImageFolder directories with matching class names:

images/
  train/
    class_a/001.jpg
    class_b/002.jpg
  val/
    class_a/101.jpg
    class_b/102.jpg

Keep validation images separate. Provide at least eight training images and two classes for this example’s classification probe. Labels train the probe only; SimCLR uses the paired views, not the class labels.

python -m stable_pretraining.quickstart --data ./images --epochs 2 --cache-dir ./image-runs
spt web ./image-runs/runs

The example intentionally uses 16x16 crops, batch size 8, and a small CPU encoder for a fast first run. Before a research experiment, copy the starter and choose the crop size, encoder, augmentation, batch size, and training duration for your data. The quickstart’s hyperparameters are not an accuracy baseline.

Data wiring uses the library’s existing adapters:

from pathlib import Path
from torch.utils.data import DataLoader
from torchvision.datasets import ImageFolder
import stable_pretraining as spt

root = Path("images")
t = spt.data.transforms
augment = t.Compose(
    t.RandomResizedCrop((16, 16)),
    t.RandomHorizontalFlip(),
    t.ToImage(scale=True),
)
dataset = spt.data.FromTorchDataset(
    ImageFolder(root / "train"),
    names=["image", "label"],
    transform=t.MultiViewTransform([augment, augment]),
)
loader = DataLoader(dataset, batch_size=8, drop_last=True)
batch = next(iter(loader))
assert len(batch["views"]) == 2
assert batch["views"][0]["image"].shape == (8, 3, 16, 16)

Compose takes transforms as positional arguments; MultiViewTransform takes a list (or a dict of named views). Default PyTorch collation preserves the {"views": [{"image": ..., "label": ...}, ...]} structure used by spt.forward.simclr. For your own objective, inspect an actual collated batch before writing the forward function.