Converting model from Keras to Pytorch (Problem with BatchNorm)

1

Hi, I am trying to convert pix2pix GAN network from keras to pytorch. I would like to convert Keras BatchNorm layer to Pytorch BatchNorm2d, because i think that the reason why my discriminator loss is falling faster than in keras.

This is original Keras implementation: BatchNormalization(momentum=0.8) and I also would like to set up init weight and bias as it is in Keras.

Thanks for help.

2

Hi,

The batchnorm in pytorch takes a momentum parameter as well so you can do the same thing (doc here).
For the initialization, you can initialize the .weight and .bias of the batchnorm as you want.

3

this one is clear for me, but i am not sure about the others, which are set as default in keras version