Pytorch calculate loss backward when divide tensor into subgroup and put into different network

autograd
1

I am new to PyTorch coding. I want to calculate keypoints X and Y coordinate prediction. I am thinking to divide keypoints into subgroups and put those keypoints that belong to same group to same network. For example, I have tensor with shape [bs, 14, 128] (batch size, number of keypoints, feature size). I use tensor slicing to divide into 3 groups:

I am new to PyTorch coding. I want to calculate keypoints X and Y coordinate prediction. I am thinking to divide keypoints into subgroups and put those keypoints that belong to same group to same network. For example, I have tensor with shape [bs, 14, 128] (batch size, number of keypoints, feature size). I use tensor slicing to divide into 3 groups:

group_one[:, :4, :]

group_two[:, 4:9, :]

group_three[:, 9:14, :]

and then put those 3 subgroups into different network and concatenate back into [bs, 14, 128]. Is the loss backward can be calculated as usual and the parameters are updated correctly?

2

Views should be differentiable yes, so the gradients computed for each view should flow back to the base.