วงรอบยุคพร้อมการตรวจสอบ

One epoch is a single full pass over your entire training set. Real training repeats this pass many times so the model keeps improving.

Wrap your whole routine in a loop over epochs. Each turn of this outer loop trains once and then checks progress on validation.

Inside an epoch you loop over batches from the DataLoader. Each batch runs the familiar forward, loss, backward, step rhythm.

Before training a pass, call model.train(). It turns on layers like dropout and batch norm that should behave differently while learning.

Before validating, call model.eval(). This freezes dropout and uses running batch-norm stats so your scores are steady and fair.

Validation only reads the model, so wrap it in torch.no_grad(). Skipping the graph saves memory and runs noticeably faster.

Add up each batch's loss across the epoch, then divide by the count. This average loss is one clean number to compare epoch to epoch.

Compute the same average on the validation set. This validation loss reveals how well the model generalizes beyond what it trained on.

Log both losses every epoch so you can watch the curves. Seeing the trend live makes overfitting easy to catch the moment it starts.

Healthy training shows both losses falling. When validation loss starts climbing while training keeps dropping, the model is beginning to overfit.

A frequent bug is calling backward() during validation. Never update weights from val data, or you contaminate your honest measurement.

What should you call right before running the validation pass?

Each epoch trains in train mode, then validates in eval mode under no_grad. Track both losses and watch the gap to catch overfitting early. 📈