binwang-shu / mypythoncnn

Writing some cnn layers ans the computation graph in python

MyPythonCNN

Writing some cnn layers ans the computation graph in python

This repository is mainly based on https://github.com/wuziheng/CNN-Numpy

There are some problems in https://github.com/wuziheng/CNN-Numpy. When the stride != 1, the forward and backward of the convolution will error. When the kernel size != the stride, the error will happen on pooling. So have improved the algorithm of the convolution and pool.

convolution:

forward(x):

x * conv = out
im2col(x) dot col(conv) ==> out.

backward(eta):

imcol(eta) dot col(conv.T) ==> input.eta
if stride != 1, I expand the eta based on the stride and backwards it like the stride == 1.

MaxPooling:

forward(x):

x * pool = out
save the index of the max ==> self.index

backward(eta):

assgin each pixel in eta to the related place of input based on the index.
if the kernelsize > stride, some pixels in input will be used over once.So in the backward, I use add to assign.

AvgPooling:

like MaxPooling.