2026-09-11
Daily Tech Challenge
An image classifier trains fine, but its validation accuracy jumps around by more than a point between epochs and sits well below what the same checkpoint scores in production. The transform pipeline is below. What is actually wrong?
train_tf = transforms.Compose([
transforms.RandomResizedCrop(224, scale=(0.3, 1.0)),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(MEAN, STD),
])
# reused for validation "so the distributions match"
val_tf = transforms.Compose([
transforms.RandomResizedCrop(224, scale=(0.3, 1.0)),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(MEAN, STD),
])
val_loader = DataLoader(ImageFolder(VAL_DIR, val_tf), batch_size=64)
model.eval()
with torch.no_grad():
acc = evaluate(model, val_loader)Friday, September 11, 2026 · A new challenge drops every day