wangch-g / lanet

MIT License
30 stars 12 forks source link

Rectification factor in location decoder #2

Closed ericzzj1989 closed 1 year ago

ericzzj1989 commented 1 year ago

Thanks for your great work.

The location decoder has a learnable rectification factor you mentioned in the original paper, however, this learning parameter is not reflected in the code. https://github.com/wangch-g/lanet/blob/74dd27cafa2d3d76ba9ddcb3a8804728f9998653/network_v0/modules.py#L58 self.loc_out = nn.Conv2d(256, 2, kernel_size=3, stride=1, padding=1) The 'loc_out' just has 2 channels in the output as above.

Can you give me some advice?

wangch-g commented 1 year ago

https://github.com/wangch-g/lanet/blob/74dd27cafa2d3d76ba9ddcb3a8804728f9998653/network_v0/modules.py#L65 self.shift_out = nn.Conv2d(256, 1, kernel_size=3, stride=1, padding=1) The 'shift_out' is the rectification factor.

ericzzj1989 commented 1 year ago

Got it. Many thanks for your reply.