Suppose the size of a is (128,) and the size of b is (128, 3, 64, 64). How to do broadcast multiplication with these two tensors?
I guess u need to convert tensor A to size (128,1) and then multiply with tensor B.
output = A. unsqueeze(1) * B
@kelam_goutam Thanks for your reply. It works if the size of b is (128,3), i.e, b is 2-dimensional tensor, but doesn’t work when the dimension of b is more than 2 dimensions.
the most generic way would be to do something like this (which is not broadcasting, but explicit expanding):
output = a.expand_as(b)*b
if you want to do it with broadcasting, you can simply do
output = a * b
if the shapes cannot be broadcasted, an error will be raised, but you will also receive this error if you use the explicit approach
Thanks for your apply. I tried the solution but it does not work as I expected.
A related question. Suppose I have a = torch.tensor([1,2,3]), how to expand it to
tensor([[[ 1, 1, 1],
[ 1, 1, 1]],
[[ 2, 2, 2], [ 2, 2, 2]], [[ 3, 3, 3], [ 3, 3, 3]]])
A crude way to solve your problem could be the extension of my previous solution.
output = A.unsqueeze(1).unsqueeze(2).unsqueeze(3) * B
This basically makes A a 4D tensor.
These also work:
output = a.reshape(128, 1, 1, 1) * b
output = a.reshape(-1, 1, 1, 1) * b
You can broadcast a vector to a higher dimensional tensor like so:
def row_mult(t,vec): extra_dims = (1,)*(t.dim()-1) return t * vec.view(-1, *extra_dims)