SeungjunNah/DeepDeblur-PyTorch

Initialization method

maradona20200204 opened this issue · 2 comments

My god, the code is a masterpiece!

I have one simple question: what is the initialization method of the convolutional weights, e.g., Xavier initialization or He initialization? Or the code just uses the default initialization method of PyTorch? I cannot find any initialization method in the code, therefore, I think that the code uses the default initialization method of PyTorch. Am I right?

I appreciate your help.

Hi @maradona20200204,

I used the PyTorch default initialization which is the He initialization.
The initialization function for conv layers, reset_parameters, can be found here: conv.py
The function, kaiming_uniform_ is defined here: init.py

Thanks!