Backprop through scatter_add_

autograd
1

Hi, I was wondering if there would be any issues with backprop while using scatter_add_, since it is an in-place operation. Specifically, in the code below:

x = conv(x)
p = x.new_zeros((n,c,h))
p.scatter_add_(2, idx, x)

Will gradients flow correctly to the weights of conv, if the loss is calculated using p?

2

Hi,

Yes the autograd works fine with inplace operations. If it cannot deal with things, it will raise an error.