Training & Evaluation

Every supervised network trains with the same loop, repeated over epochs (passes through the training set) in mini-batches:

1. Forward pass — compute predictions for a batch.
2. Loss — measure how wrong they are; for classification, cross-entropy = −log(probability of the true class).
3. Backward pass — backpropagation computes the gradient of the loss with respect to every weight.
4. Update — an optimiser moves the weights against the gradient: SGD with momentum, or AdamW (Adam with decoupled weight decay).

for x, y in loader:
    opt.zero_grad()
    loss = F.cross_entropy(model(x), y)
    loss.backward()
    opt.step()

The learning rate is the most important hyper-parameter: too high and the loss jumps around or diverges; too low and training crawls. Warm-up plus cosine decay is a robust default schedule.

• Training set — the model learns from it.
• Validation set — you choose hyper-parameters, architecture, and the epoch to stop at.
• Test set — touched once, at the end, to report performance.

Because you make many decisions using the validation set, it slowly becomes "seen"; the test set is the honest estimate.

Leakage in vision is often subtle: the same patient's scans in train and test, near-duplicate frames from one video split across sets, or augmenting before splitting. Split by the unit that must generalise — patient, video, location — not by image.

Plot training and validation loss per epoch:

Batch normalisation behaves differently in the two modes: in training it normalises with the current batch's statistics and updates running averages; in evaluation it uses the stored running averages. Dropout is active only in training. Forgetting model.eval() at test time gives noisy, wrong results.

Accuracy fails on imbalanced data: if 99% of scans are healthy, "always healthy" scores 99% and finds no disease.

From the confusion matrix:

• Precision = TP / (TP + FP) — of my positive predictions, how many were right?
• Recall = TP / (TP + FN) — of the real positives, how many did I find?
• F1 = 2PR / (P + R) — their harmonic mean (punishes imbalance between them).
• Specificity = TN / (TN + FP).

Which matters depends on the cost of each mistake: cancer screening wants high recall; a spam filter that deletes mail wants high precision. Curves (ROC, precision–recall) show the trade-off across thresholds.