train_from_scratch.py

import argparse
from pathlib import Path

import torch
from datasets import load_dataset
from torch import nn
from torch.utils.data import DataLoader
from tqdm import tqdm

from calority_scratch_model import CalorityFoodCNN, image_to_tensor, save_checkpoint


def parse_args():
    parser = argparse.ArgumentParser(description="Train Calority's food model from scratch on a Hugging Face dataset.")
    parser.add_argument("--dataset", default="food101", help="Hugging Face dataset name, for example food101")
    parser.add_argument("--image-column", default="image")
    parser.add_argument("--label-column", default="label")
    parser.add_argument("--train-split", default="train")
    parser.add_argument("--eval-split", default="validation")
    parser.add_argument("--output-dir", default="./calority-scratch-model")
    parser.add_argument("--epochs", type=int, default=12)
    parser.add_argument("--batch-size", type=int, default=32)
    parser.add_argument("--learning-rate", type=float, default=3e-4)
    parser.add_argument("--num-workers", type=int, default=0)
    parser.add_argument("--limit-train", type=int, default=0, help="Optional small limit for quick smoke tests")
    parser.add_argument("--limit-eval", type=int, default=0, help="Optional small limit for quick smoke tests")
    return parser.parse_args()


def get_labels(dataset, split: str, label_column: str) -> list[str]:
    feature = dataset[split].features[label_column]
    if hasattr(feature, "names") and feature.names:
        return list(feature.names)

    values = sorted(set(dataset[split][label_column]))
    return [str(value) for value in values]


def make_collate_fn(image_column: str, label_column: str):
    def collate(batch):
        images = torch.stack([image_to_tensor(item[image_column]) for item in batch])
        labels = torch.tensor([int(item[label_column]) for item in batch], dtype=torch.long)
        return images, labels

    return collate


def evaluate(model, loader, loss_fn, device):
    model.eval()
    total_loss = 0.0
    total_correct = 0
    total_seen = 0

    with torch.no_grad():
        for images, labels in loader:
            images = images.to(device)
            labels = labels.to(device)
            logits = model(images)
            loss = loss_fn(logits, labels)
            total_loss += loss.item() * labels.size(0)
            total_correct += (logits.argmax(dim=1) == labels).sum().item()
            total_seen += labels.size(0)

    return total_loss / max(total_seen, 1), total_correct / max(total_seen, 1)


def main():
    args = parse_args()
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    dataset = load_dataset(args.dataset)

    if args.limit_train:
        dataset[args.train_split] = dataset[args.train_split].shuffle(seed=42).select(range(args.limit_train))
    if args.limit_eval:
        dataset[args.eval_split] = dataset[args.eval_split].shuffle(seed=42).select(range(args.limit_eval))

    labels = get_labels(dataset, args.train_split, args.label_column)
    model = CalorityFoodCNN(num_labels=len(labels)).to(device)

    collate_fn = make_collate_fn(args.image_column, args.label_column)
    train_loader = DataLoader(
        dataset[args.train_split],
        batch_size=args.batch_size,
        shuffle=True,
        num_workers=args.num_workers,
        collate_fn=collate_fn,
    )
    eval_loader = DataLoader(
        dataset[args.eval_split],
        batch_size=args.batch_size,
        shuffle=False,
        num_workers=args.num_workers,
        collate_fn=collate_fn,
    )

    loss_fn = nn.CrossEntropyLoss()
    optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate, weight_decay=1e-4)
    scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=args.epochs)

    best_acc = 0.0
    output_dir = Path(args.output_dir)
    output_dir.mkdir(parents=True, exist_ok=True)

    for epoch in range(1, args.epochs + 1):
        model.train()
        running_loss = 0.0
        total_seen = 0
        total_correct = 0

        progress = tqdm(train_loader, desc=f"epoch {epoch}/{args.epochs}", leave=False)
        for images, labels_batch in progress:
            images = images.to(device)
            labels_batch = labels_batch.to(device)

            optimizer.zero_grad(set_to_none=True)
            logits = model(images)
            loss = loss_fn(logits, labels_batch)
            loss.backward()
            optimizer.step()

            running_loss += loss.item() * labels_batch.size(0)
            total_correct += (logits.argmax(dim=1) == labels_batch).sum().item()
            total_seen += labels_batch.size(0)
            progress.set_postfix(
                loss=round(running_loss / max(total_seen, 1), 4),
                acc=round(total_correct / max(total_seen, 1), 4),
            )

        scheduler.step()
        eval_loss, eval_acc = evaluate(model, eval_loader, loss_fn, device)
        print(f"epoch={epoch} eval_loss={eval_loss:.4f} eval_acc={eval_acc:.4f}")

        if eval_acc >= best_acc:
            best_acc = eval_acc
            save_checkpoint(model, labels, output_dir)
            print(f"saved best model to {output_dir} with eval_acc={best_acc:.4f}")

    print(f"done. best_eval_acc={best_acc:.4f}")


if __name__ == "__main__":
    main()