การส่งต่อ ค่าความสูญเสีย ย้อนกลับ และก้าวปรับค่า

Every training iteration is just four moves repeated: forward, loss, backward, step. Learn this rhythm and the whole training loop stops feeling mysterious. 🥁

The forward pass sends your input through the model so it produces a prediction. You just call the model on a batch of data and read the output.

The forward output is usually raw scores called logits, not clean labels. The loss function knows how to read these scores directly, so you leave them as is.

Next you measure how wrong the prediction was. The loss compares your output to the true labels and boils the mistake down to one number.

A big loss means a poor prediction and a small one means a good prediction. Training has a single mission: drive this number steadily downward.

Now you ask which way to nudge each weight. The backward pass uses autograd to compute every gradient and fill each parameter's .grad for you.

Each gradient tells you how the loss would change if you tweaked that one weight. They are the map the optimizer follows toward a better model.

Finally the optimizer applies those gradients, nudging every weight a little. This step is the moment the model actually learns from its mistake.

Gradients pile up by default, so before the next backward you call zero_grad. Skip it and old gradients corrupt your next update.

Put the four moves together and you get one clean iteration. You will type this exact pattern in nearly every PyTorch project you build.

Run these moves on batch after batch and the loss falls a little each time. Thousands of tiny updates are what turn random weights into real skill.

Which call actually computes the gradients for every weight?

One iteration is four moves: forward for a prediction, loss to measure error, backward for gradients, and step to update weights. Repeat to learn. 🎉