import argparse import json import time from pathlib import Path import torch import torch.nn as nn from torch.utils.data import DataLoader from torchvision import datasets, transforms from model import CNN, count_params def get_loaders(data_dir, batch_size): transform = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,)), ]) train_set = datasets.MNIST(data_dir, train=True, download=True, transform=transform) val_set = datasets.MNIST(data_dir, train=False, download=True, transform=transform) train_loader = DataLoader(train_set, batch_size=batch_size, shuffle=True, num_workers=2) val_loader = DataLoader(val_set, batch_size=512, shuffle=False, num_workers=2) return train_loader, val_loader @torch.no_grad() def evaluate(model, loader, device): model.eval() correct = total = 0 for x, y in loader: x, y = x.to(device), y.to(device) correct += (model(x).argmax(1) == y).sum().item() total += y.size(0) model.train() return correct / total def main(): parser = argparse.ArgumentParser() parser.add_argument("--depth", type=int, required=True, help="number of blocks (each = 2 conv layers)") parser.add_argument("--residual", action="store_true") parser.add_argument("--no-bn", action="store_true") parser.add_argument("--channels", type=int, default=64) parser.add_argument("--epochs", type=int, default=15) parser.add_argument("--batch-size", type=int, default=128) parser.add_argument("--lr", type=float, default=1e-3) parser.add_argument("--data-dir", default="data") parser.add_argument("--out-dir", default=None) args = parser.parse_args() device = "cuda" if torch.cuda.is_available() else "cpu" suffix = ("_res" if args.residual else "") + ("_nobn" if args.no_bn else "") out_dir = Path(args.out_dir or f"runs/depth{args.depth}{suffix}") out_dir.mkdir(parents=True, exist_ok=True) model = CNN(args.depth, channels=args.channels, use_residual=args.residual, bn=not args.no_bn).to(device) train_loader, val_loader = get_loaders(args.data_dir, args.batch_size) opt = torch.optim.Adam(model.parameters(), lr=args.lr) sched = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=args.epochs) criterion = nn.CrossEntropyLoss() print(f"depth={args.depth} residual={args.residual} bn={not args.no_bn} params={count_params(model):,} device={device}") history = [] for epoch in range(1, args.epochs + 1): model.train() t0 = time.time() running_loss, running_correct, n = 0.0, 0, 0 for x, y in train_loader: x, y = x.to(device), y.to(device) opt.zero_grad(set_to_none=True) logits = model(x) loss = criterion(logits, y) loss.backward() opt.step() running_loss += loss.item() * y.size(0) running_correct += (logits.argmax(1) == y).sum().item() n += y.size(0) sched.step() train_acc = running_correct / n val_acc = evaluate(model, val_loader, device) train_loss = running_loss / n history.append({"epoch": epoch, "train_loss": train_loss, "train_acc": train_acc, "val_acc": val_acc}) print(f"epoch {epoch:2d}/{args.epochs} loss={train_loss:.4f} " f"train_acc={train_acc:.4f} val_acc={val_acc:.4f} ({time.time()-t0:.1f}s)") torch.save(model.state_dict(), out_dir / "model.pt") (out_dir / "history.json").write_text(json.dumps(history, indent=2)) print(f"saved to {out_dir}") if __name__ == "__main__": main()