Appendix A: Tensor Shapes Cheat Sheet

“Because debugging starts with dimensions.”

Task / Layer	Expected Shape
Single image (grayscale)	[1, H, W]
Single image (RGB)	[3, H, W]
Batch of grayscale images	[B, 1, H, W]
Batch of RGB images	[B, 3, H, W]
Fully connected input	[B, features]
LSTM input	[seq_len, batch, input_size] or [B, T, F] (batch_first=True)
Transformer input	[B, seq_len] or [B, seq_len, emb_dim]
nn.Embedding input	[B, T] (int64 tokens) → Output: [B, T, emb_dim]
CNN output to FC layer	[B, C, H, W] → x.view(B, -1)
Classification output	[B, num_classes]
Regression output	[B, 1] or [B]

Always print(.shape) at each step of your model to avoid dimensional disasters.