AUTOMATIC DIFFERENTIATION WITH TORCH.AUTOGRAD - Jacobian Product Tutorial Shape Confusion

[nb8g13]

In the tutorial AUTOMATIC DIFFERENTIATION WITH TORCH.AUTOGRAD a working toy example is presented in Optional Reading: Tensor Gradients and Jacobian Products.

However, the shape of the input tensor inp is used when specifying the vector $v$ in the the backwards call instead of the shape of the output tensor out, as highlighted in the image below.

This code runs correctly because the input and output tensors happen to be the same size in this example. However, if the output tensor is a different shape then using the input shape in the backwards() call leads to an exception. I believe it would be more instructive to use out in place of inp in the positions highlighted. This would also better match the text above the code snippet.

cc @suraj813

[malfet]

Thank you for being a thorough reader. Indeed, it would be better to use out as argument to tensor.ones_like