What is mean by res5c_relu

[eng100200]

Please tell me what is mean by res5c_relu? and why
Estimates for a single full pass of model at input size 224 x 224 takes 8GFLOPs
while, res5c_relu features with the network 224 x 224 input sizes using a batch size of 128 takes 974 GFLOPs?

[albanie]

Hi @eng100200, it refers to a certain feature map - the one produced by the relu following block 5c (this is in the fifth stage of the network using the resnet naming scheme)

[CMCDragonkai]

@albanie It would really useful if you wrote down your exact formula for calculating 974 GigaFLOPs. Actually if the README could have detail the command line invocations to reproduce your reports that would be really useful.

[albanie]

Hi @CMCDragonkai

The number of flops is computed here - the code comments also mention some of the approximations and assumptions. In practice, for most architectures, almost all of the flops will come from the convolutions/deconvolutions (so those are the key layers from a flop perspective). Writing down a formula would be a little tricky, since it would essentially be a big switch statement.

The tool is written in MATLAB, so the invocation to generate the reports is just to run the burden() function. If you prefer PyTorch, there is a tool which can be called from the command line here (although I haven’t used it personally): /~https://github.com/Swall0w/torchstat