Day15 :Mini Trainer

The model.fit() Trap

Here is how most beginners train a neural network:

python

Four lines. It works. Your accuracy prints. You feel productive.

But here is what you just gave up: you have no idea what happened during those 200 iterations. Did the loss decrease monotonically or oscillate? Did the gradients explode on epoch 12 and silently recover? Was early stopping triggered internally? Did any batch produce a NaN that got masked by a np.nan_to_num call buried three layers deep in sklearn's source?

You cannot answer any of these questions because fit() is a black box with a contract: give me data, I give you weights. That contract is fine for production pipelines where you trust the library. It is fatal for learning, because the exact mechanics hidden inside that function — the batch loop, the gradient accumulation, the loss plateau detection — are the entire subject matter of this lesson.

The cost of abstraction here is not performance. It is understanding. And understanding is the only thing that lets you debug a training run that silently converges to 51% accuracy on a problem that should hit 94%.

The Failure Mode

Let's say you decide to implement the training loop yourself, naively:

python