jhkim89 / pyramidnet Goto Github PK

in resnet-34, the block is basicBlock.

but in the code, you always use bottleneck (https://github.com/jhkim89/PyramidNet/blob/master/addpyramidnet.lua#L125), do we also need bottleneck when using resnet-34?

I encounter out-of-memory problem each time I train the net with depth=272, alpha = 200. My training configuration is as follows:
th main.lua -netType addpyramidnet -depth 272 -batchSize 128 -nGPU 4 -shareGradInput true
and the out-of-memory problem occurs when the 2nd epoch begins.
After I changed batchSize into 120 and nGPU into 6, the problems occurs when the 3rd epoch begins.
Are there any solutions?

in resnet basic block, there is no bn in the head (https://github.com/facebook/fb.resnet.torch/blob/master/models/resnet.lua#L53)

Is there any reason to add bn in the head of the basic block in PyramidNet (https://github.com/jhkim89/PyramidNet/blob/master/addpyramidnet.lua#L53)?