Hi,

I think I found an issue with the accounting of MAC operations/# weights read in the memory (MEM) when using the dynamic channel pruning (DCP) feature. The number of MAC/MEM reported for gate.density values < 1 are more than they should be. This happens when a conv layer follows a non convolution layer like dropout. In case of density values of < 1, there is sparsity benefit from both input and output sides. However, when a conv follows a dropout layer, input side sparsity is not accounted for, even though the input to dropout is a sparse tensor from a conv layer.

Take the CIFARNET example with gate density values of 1.0 and 0.1. The results from the framework are pasted below. The errors are in the estimate of conv4 and conv7. For GD of 0.1, the estimates for conv4 and conv7 are 10x higher than they should be because of the above mentioned issue.

Gate Density = 1.0

And Gate Density =0.1