Thinking in Shapes

.shape is the first thing experienced practitioners print. It tells you what the tensor is:

• (1000, 20) — 1000 examples × 20 features: a dataset
• (224, 224, 3) — height × width × RGB channels: one image
• (32, 224, 224, 3) — a batch of 32 images
• (128, 50, 768) — 128 sequences of 50 tokens, each a 768-dim vector: transformer activations

Convention runs deep: rows are examples, columns are features. When a model says it expects (batch, features), it's describing a shape contract — and most shape bugs are delivering the columns where the rows belong.