Full CNN Classifier

The problem, from first principles

Every piece needed for an actual image classifier already exists across this curriculum: a convolutional backbone that extracts features (03-stack-multiple-blocks), a way to flatten those features into a vector (01-flatten), a Linear layer that turns a feature vector into class scores (01-classical-ml/01-linear-regression/01-hypothesis-function), and softmax plus cross-entropy to turn those scores into probabilities and a trainable loss (01-classical-ml/02-classification/06-softmax-cce). This question is the capstone: wire all of them into one real, working classifier's forward pass.

From theory to code

Theory says: run each image through the convolutional backbone, flatten the result, stack a whole batch of flattened vectors into one matrix, run that matrix through the final Linear layer, then softmax. If true labels are given, also score the result with cross-entropy — this is a genuine, if small, image classifier.

Implement full_cnn_classifier(images, kernels, fc_weight, fc_bias, labels=None, pool_size=2) against that reasoning, reusing every piece named above rather than reimplementing any of them.

Constraints

• images: shape (N, C, H, W) — a batch of N images.
• kernels: list of conv kernels, passed straight through to 03-stack-multiple-blocks.
• fc_weight: shape (num_classes, flattened_feature_dim). fc_bias: shape (num_classes,).
• labels: optional, shape (N,), integer class indices.
• Returns probs (shape (N, num_classes), rows summing to 1) if labels is None; otherwise returns (probs, loss) where loss is a plain Python float.
• images, every kernel, fc_weight and fc_bias are never modified in place.

Hints

Open one at a time. Each gives away a little more than the last.