MCPcopy Create free account
hub / github.com/MotrixLab/ADHMR / mask_weights

Method mask_weights

ADHMR/lib/models/scorenet.py:177–186  ·  view source on GitHub ↗
(self, layer,mask,mshape)

Source from the content-addressed store, hash-verified

175 return torch.from_numpy(L)
176
177 def mask_weights(self, layer,mask,mshape):
178 assert isinstance(layer, nn.Linear), 'masked layer must be linear layer'
179
180 output_size, input_size = layer.weight.shape # pytorch weights [output_channel, input_channel]
181 input_size, output_size = int(input_size), int(output_size)
182 assert input_size % mshape == 0 and output_size % mshape == 0
183 in_F = int(input_size / mshape)
184 out_F = int(output_size / mshape)
185 weights = layer.weight.data.view([mshape, out_F, mshape, in_F])
186 weights.mul_(mask.t().view(mshape, 1, mshape, 1).to(device=weights.get_device()))
187
188 def get_local_feature(self, xinj, fmap):
189 bs = fmap.shape[0]

Callers 1

forwardMethod · 0.95

Calls 1

toMethod · 0.45

Tested by

no test coverage detected