search

AliaksandrSiarohin / monkey-net

Animating Arbitrary Objects via Deep Motion Transfer
467 stars 81 forks source link

assigning zero weights to hourglass decoder while predicting mask #27

Closed saandeepa93 closed 3 years ago

saandeepa93 commented 3 years ago

Hello,

In dense motion module, while running the predictions through the hourglass to calculate the mask, the hourglass' decoder network's weights are assigned to 0 and then the bias is initialized to a particular value.

Can you please let me know the reason for this as I am trying to correlate this with the original paper? modules -> dense_motion_module.py self.hourglass.decoder.conv.weight.data.zero() self.hourglass.decoder.conv.bias.data.copy(torch.tensor(bias_init, dtype=torch.float))

AliaksandrSiarohin commented 3 years ago

This is just a way to have everything assigned to the background (so no motion initially).