why use add operation in TimeBlock layer

FelixOpolka / STGCN-PyTorch

🚗 Implementation of spatio-temporal graph convolutional network with PyTorch

MIT License

335 stars 70 forks source link

why use add operation in TimeBlock layer #6

Open shirenyu opened 5 years ago

shirenyu commented 5 years ago

In the tensorflow implementation, the temporal_conv_layer result is the product of conv and sigmoid. “ return (x_conv[:, :, :, 0:c_out] + x_input) * tf.nn.sigmoid(x_conv[:, :, :, -c_out:]) ”

could you explain why the add operation is used here?

temp = self.conv1(X) + torch.sigmoid(self.conv2(X)) out = F.relu(temp + self.conv3(X))

NotEnded99 commented 3 years ago

Same question as you

lp2900 commented 2 years ago

Yeah, no GLU .

ManlioWu commented 1 year ago

I have the same question: Why addition rather than multiplication?