CNN shape calculator for PyTorch (2D & 3D)

11/1/2020

This is a small, practical tool to avoid off-by-one mistakes when designing CNNs. You set the input tensor once, add layers, tweak hyperparameters, and see the shapes and parameter counts update live.

Why this is useful

  • Quick shape checks while sketching an architecture
  • Early catch for invalid settings (e.g., output becomes ≤ 0)
  • Fast parameter estimates per layer

What it supports

  • 2D: nn.Conv2d, nn.ConvTranspose2d, *Pool2d
  • 3D: nn.Conv3d, nn.ConvTranspose3d, *Pool3d
  • Output padding for transpose convs
  • Per-layer input/output summary and learnable parameter counts
  • Add/remove layers, instant re-compute

Notes:

  • Dilation is assumed to be 1.
  • Parameter counts shown exclude bias. If you use bias=True, add out_channels.

Formulas

For a single spatial dimension:

  • Conv / Pool out = floor((in + 2*pad − (kernel − 1) − 1) / stride + 1)

  • ConvTranspose out = (in − 1) * stride − 2*pad + (kernel − 1) + output_padding + 1

These apply independently to each dimension (H and W for 2D; D, H, W for 3D).

CNN Dimensions Calculator

Live shape propagation for nn.Conv2d, nn.ConvTranspose2d, *Pool2d and nn.Conv3d, nn.ConvTranspose3d, *Pool3d.

(N, C, H, W)N=C=H×W=×
↓
Conv2d
nn.Conv2d (in_channels=,out_channels=,kernel_size=×,stride=×,padding=×)
IN: [1, 28, 28]
CONV: CIN=1, COUT=10, F=(3, 3), S=(1, 1), P=(0, 0), 90 params
OUT: [10, 26, 26]
↓
(N, C, H, W)N=C=H×W=×