Suppose I have a list of indices and wish to modify an existing array with this list. Currently the only way I can do this is by using a for loop as follows. Just wondering if there is a faster/ efficient way.
torch.manual_seed(0)
a = torch.randn(5,3)
idx = torch.Tensor([[1,2], [3,2]]).to(torch.long)
for i,j in idx:
a[i,j] = 1
I initially assumed that gather or index_select would go some way in answering this question, but looking at documentation this doesn’t seem to be the answer.
In my particular case, a is a 5 dimensional vector and idx is a Nx5 vector. So the output (after subscripting with something like a[idx]) I’d expect is a (N,) shaped vector.
@sachinruk Thanks for mentioning, and I would like to know how to extend this to higher dimension.
If we have a 4-D tensor of [batch, channel, height, width], and a set of [x, y] coordinates with shape [batch, num_points, 2] , how to select this 4-D tensor without loop?
Here is my implementation with loop:
Suppose the 4-D feature map [10, 256, 64, 64] in dimension, and coordinates is [10, 68, 2] in dimension(for each batch, there are 68 points that we wanted to select from feature map)
coord_features = torch.zeros(10, 68, 256)
feature_map = feature_map.transpose(1,2).transpose(2,3) #reshape to [10, 64, 64, 256]
for i in xrange(coords.shape[0]): #loop through each sample in batch
for j in xrange(coords.shape[1]): #loop through each points
#select coordinate on feature map
coord_features[i][j] = feature_map[i][coords[i][j][1].type(torch.int64)][coords[i][j][0].type(torch.int64)]
To do this with a batch dimension, you can just create index tensors for the other dimensions as needed.
For a simple example, let’s say we have a tensor data of grayscale images with size b, h, w and a set of random pixels coordinates coords with size b, 2 of pixel coordinates (one pixel per image in the batch) that we want to zero out. We can do this with:
Your example is a little more complicated, because you have multiple points per image, but you can just do something like bselect = torch.range(batch, dtype=torch.long)[:, None].expand(batch, num_points).view(-1).
For the channel dimension, I think you can just use a single integer 0, but you may need to create a vector of zeros of the same length as bselect.