Neural networks, deep learning, NLP · insight-0438

Compute CNN feature-map sizes on paper before writing the classifier head

Situation
You stack convolution and pooling layers for an image task and hard-code the input size of the first linear layer.
Action
Write out $H_{out} = \lfloor (H_{in} + 2p - d(k-1) - 1)/s \rfloor + 1$ per layer, assert it with a dummy forward, and end the feature extractor with adaptive pooling so the head does not depend on input resolution.
Reason
The output size is fixed arithmetic of input size, kernel, padding, stride and dilation, so it can be checked before training.
Boundary
CNNs, pooling layers and 1D convolutions over sequences.

Symptom or risk

A size mismatch crashes at the first forward, or worse, a different input resolution later breaks serving or forces silent resizing.

Pretrained backbones already document their output shapes; still verify with a dummy input.

Adaptive pooling discards spatial detail that some tasks need; segmentation keeps spatial maps.

Forgetting that floor division drops a pixel at odd sizes, so a decoder no longer aligns with the encoder.

224 x 224 input, kernel 7, stride 2, padding 3: floor((224 + 6 - 6 - 1)/2) + 1 = 112.

PyTorch Conv2d documents output height as floor((H_in + 2*padding - dilation*(kernel_size - 1) - 1)/stride + 1).
Conv2d (PyTorch documentation)
Dumoulin and Visin derive the relationships between input shape, kernel, padding, stride and output shape for convolution and pooling.
A guide to convolution arithmetic for deep learning