Explicit training and evaluation loops with correct metrics

1. Optimizing Model Parameters: `train_loop` and `test_loop`

   PyTorch - Learn the Basics

   Hyperparameters (epochs, batch size, learning rate); Loss Function (CrossEntropyLoss over logits); Optimizer (zero_grad, backward, step); Full Implementation (train_loop with model.train, test_loop with model.eval and no_grad, accuracy accumulation)

   Done
2. HOML: mini-batch loops and metric aggregation

   Hands-On Machine Learning with Scikit-Learn and PyTorch

   Chapter 10, Implementing Mini-Batch Gradient Descent Using DataLoaders and Model Evaluation (printed pp. 335-339): the train() and evaluate() functions, the batch-mean RMSE (0.5668) against the whole-set RMSE (0.6388), and the streaming TorchMetrics version

   Depth · Footnote 10 calls the short last batch a negligible effect on a batch-mean MSE. Your reproduce measures that gap instead of assuming it.

   Done

Focus on

• train_loop sets model.train() and test_loop sets model.eval() with no_grad. The comments call both best practice even where the tiny net needs neither.
• Zero before backward each iteration (or right after step, every iteration). step() applies whatever sits in .grad, including stale values.
• The tutorial test_loss divides by batch count, which weights batches equally. With an uneven tail batch that differs from the sample mean; correct over size is exact.
• CrossEntropyLoss takes raw logits and reduces by mean over the batch. Changing batch size changes the noise of that mean, not its expectation.

Reproduce closed-book

Traps

• Averaging per-batch losses with equal weights and calling it the epoch loss. Short tail batches then count as much as full ones.
• Evaluating without model.eval(). Dropout keeps dropping units, so evaluation becomes random, and BatchNorm normalizes with test-batch statistics while updating its running buffers on test data.
• Logging loss.item() but stepping on a different loss. The printed number and the backward graph must come from the same tensor.

Practice

Self-test

1. Test set 10000, batch 64. The tutorial reports test_loss /= num_batches. What does that number weight equally?

   Show answer Got it Missed

2. Where must zero_grad sit relative to backward and step, and what breaks if it is missing?

   Show answer Got it Missed

3. test_loop wraps evaluation in no_grad but the tutorial also sets model.eval(). What does each contribute?

   Show answer Got it Missed

4. Your evaluation loop uses CrossEntropyLoss with reduction="mean" per batch. How do you report the exact per-sample mean loss over the test set?

   Show answer Got it Missed

Leads to

• 2.8 Checkpoints, configuration and reproducibility Persist the loop state you can now run.
• 3.7 Objective checks, tiny-set overfit and shuffled-label controls Add overfit and shuffled-label controls to the loop.

Comes after

• 2.3 Modules, parameters, buffers and initializationTrain the module you can now inspect.
• 2.4 Backward, `zero_grad`, `gradcheck` and evaluation modesPut zeroing, backward and eval switching into a real loop.
• 2.6 Datasets, data loaders and samplersFeed the loop from a loader you trust.
• 3.8 Diagnosing failed runs: divergence, stale state, slices and seedsFix stale state with correct train and eval loops and metrics.

Resources for this unit

• PyTorch - Learn the Basics
  PyTorchprimarycourse
  Complete Tensors, Datasets and DataLoaders, Transforms, Build Model, Autograd, Optimization, and Save and Load Model. Rebuild each essential example in a small src package instead of retaining a tutorial notebook.
• Hands-On Machine Learning with Scikit-Learn and PyTorch
  Aurelien Geron (O'Reilly, 2025)supporttheory
  Read Chapter 10, Implementing Mini-Batch Gradient Descent Using DataLoaders and Model Evaluation, for the train and evaluate loops and how batch metrics must be aggregated.