48. Small CNN Image Project in PyTorch
48.4 Holdout Check
Why. Always check a fresh batch — Chapter 38 overfit lesson applies to nets too.
What. Build a new batch with the same rules; evaluate with model.eval() and torch.no_grad().
How.
Holdout check
- make_batch(64) as test.
- Predict argmax classes.
- Print accuracy.
model.eval()
X_test, y_test = make_batch(64)
with torch.no_grad():
pred = model(X_test).argmax(1)
print("test acc", float((pred == y_test).float().mean()))
Ravindra Bagale's Tip
Synthetic easy task = learning the API. Real photos need more data + augmentation later.