search

hassony2 / kinetics_i3d_pytorch

Inflated i3d network with inception backbone, weights transfered from tensorflow
MIT License
523 stars 114 forks source link

why use Ugly hack? #3

Closed rimchang closed 6 years ago

rimchang commented 6 years ago

Thanks for great job!!

Ugly hack because I didn't use tensorflow's exact padding function self.pad_5a = torch.nn.ConstantPad3d((0, 0, 0, 0, 0, 1), 0)

https://www.tensorflow.org/versions/r0.12/api_docs/python/nn/convolution pad_along_height = ((out_height - 1) strides[1] + filter_height - in_height) pad_along_width = ((out_width - 1) strides[2] + filter_width - in_width) pad_top = pad_along_height / 2 pad_left = pad_along_width / 2

Why use ugly hack?

I think pad_along_front = ((8-1)*2 + 2 - 16) = 0 pad_back = 0/2 = 0

hassony2 commented 6 years ago

First, I saw you did a very similar conversion in one of your repos almost at the same time as me ! :)

You are right, removing this padding does not create any problems. At some version of my code, I obtained an error if I did not pad, and I forgot to remove this padding after I resolved it ! I just pushed the removal of this line to the master branch.

Thank you for your help !