Torch sum a tensor along an axis

The simplest and best solution is to use torch.sum().

To sum all elements of a tensor:

torch.sum(outputs) # gives back a scalar

To sum over all rows (i.e. for each column):

torch.sum(outputs, dim=0) # size = [1, ncol]

To sum over all columns (i.e. for each row):

torch.sum(outputs, dim=1) # size = [nrow, 1]

Alternatively, you can use tensor.sum(axis) where axis indicates 0 and 1 for summing over rows and columns respectively, for a 2D tensor.

In [210]: XOut[210]: tensor([[  1,  -3,   0,  10],        [  9,   3,   2,  10],        [  0,   3, -12,  32]])In [211]: X.sum(1)Out[211]: tensor([ 8, 24, 23])In [212]: X.sum(0)Out[212]: tensor([ 10,   3, -10,  52])

As, we can see from the above outputs, in both cases, the output is a 1D tensor. If you, on the other hand, wish to retain the dimension of the original tensor in the output as well, then you've set the boolean kwarg keepdim to True as in:

In [217]: X.sum(0, keepdim=True)Out[217]: tensor([[ 10,   3, -10,  52]])In [218]: X.sum(1, keepdim=True)Out[218]: tensor([[ 8],        [24],        [23]])

If you have tensor my_tensor, and you wish to sum across the second array dimension (that is, the one with index 1, which is the column-dimension, if the tensor is 2-dimensional, as yours is), use torch.sum(my_tensor,1) or equivalently my_tensor.sum(1) see documentation here.

One thing that is not mentioned explicitly in the documentation is: you can sum across the last array-dimension by using -1 (or the second-to last dimension, with -2, etc.)

So, in your example, you could use: outputs.sum(1) or torch.sum(outputs,1), or, equivalently, outputs.sum(-1) or torch.sum(outputs,-1). All of these would give the same result, an output tensor of size torch.Size([10]), with each entry being the sum over the all rows in a given column of the tensor outputs.

To illustrate with a 3-dimensional tensor:

In [1]: my_tensor = torch.arange(24).view(2, 3, 4) Out[1]: tensor([[[ 0,  1,  2,  3],         [ 4,  5,  6,  7],         [ 8,  9, 10, 11]],        [[12, 13, 14, 15],         [16, 17, 18, 19],         [20, 21, 22, 23]]])In [2]: my_tensor.sum(2)Out[2]:tensor([[ 6, 22, 38],        [54, 70, 86]])In [3]: my_tensor.sum(-1)Out[3]:tensor([[ 6, 22, 38],        [54, 70, 86]])