A better way to slice

1

Hi,

I use an RNN model to process video frame by frame such that the forward looks something like this:

for frame in range(x.shape[1]):
    out1, hidden = self.rnn_layer1(x[:,frame,:,:], hidden)
return out1

But that code returns error “one of the variables needed for gradient computation has been modified by an inplace operation”

After some trial and error i realized that the problem is that the sliced frame still holds gradient information of the whole video. I use an ugly hack that works just fine but i’m looking for a cleaner way to do this… here is the hack:

for frame in range(x.shape[1]):
    out1, hidden = self.rnn_layer1(torch.cat([x[:, frame, :, :]], dim=1).unsqueeze_(1), hidden)
return out1

Essentially concatenating the tensor to nothing…

2

Can you make sure your code is correct? It looks like your torch.cat call includes hidden as the second argument.

3

Despite the fact, that your actual code does not seem to work (@aplassard mentioned), this should work:

x_perm = x.transpose(0,1)
for frame in x_perm.split(1):
    out1, hidden = self.rnn_layer1(frame.squeeze(0), hidden)

Edit: what your hack actually does is allocating new storage and copying the tensor. Regarding this a simple x[:,frame,:,:].clone() or x[:,frame,:,:].contiguous() might do the trick as well