[Python] Torch: how to insert a tensor into another tensor at certain index

I have a padded tensor X with shape (B, T1, C) and a padded tensor Y with shape (B, T2, C), I also know the sample lengths L for X. I want to insert the samples of X into Y at certain index I and pad at the end.

For example, if I = 5, the goal of the transformation is something like:

inputs = []
for i in range(X.shape[0]):
input = torch.cat([Y[0:5],
X[:L],
Y[5:],
torch.zeros(max(L) - L, Y.shape[2])],
dim=0)
inputs.append(input)
outputs = torch.stack(inputs, dim=0)


I want to know how to do this in tensor version instead of the for loop, which is too slow for training.

Continue reading...