Memory consumption U-Net

[ptrblck]

@Alex07

Why does my model use almost 5 times as much memory even though it has only ~0.5 as many parameters (14mio vs 7mio)? Where is all that memory used?

Depending on the model architecture the number of parameters and buffers might use less memory than the computed forward activations needed for the gradient calculation. This post explains it with an example.

@Alex07

Or is the parameter count just not reliable but looking at the printed model architecture also supports the fact that the pre-trained model is a lot bigger (it basically uses two of the layeredConv modules in a row instead of one like my model does).

Yes, using more layers, will create more intermediate activations, which need to be stored for the backward pass assuming you want to train the model.