search

fyu / drn

Dilated Residual Networks
https://www.vis.xyz/pub/drn
BSD 3-Clause "New" or "Revised" License
1.1k stars 219 forks source link

The mean of fill_up_weights #41

Open UpCoder opened 5 years ago

UpCoder commented 5 years ago

Hi, I am reading the code. And I am wondering what's the mean of the function----fill_up_weights. The code is here. It seems that use the fill_up_weight to init the parameter of ConvTranspose2d. However, the parameter of ConvTranspose2d maybe not update in the training process. So why do freeze the weight of ConvTranspose2d?

fyu commented 5 years ago

The code is for bilinear upsampling. In the early versions of pytorch, upsampling is aligned to the corners, which will actually cause pixel misalignment. Therefore, I made my own version of bilinear upsampling by using "fill_up_weights" and "ConvTranspose2d". It has been resolved in the current pytorch. PyTorch documentation also has some explanation: https://pytorch.org/docs/stable/nn.html?highlight=bilinear#torch.nn.Upsample.

UpCoder commented 5 years ago

So, the weight which named as up.weight is useless? In the inference state, could be the ConvTranspose2d operation replaced with the resize operation with the bilinear kernel?