Getting largest values in a tensor

[chenjus (Justin Chen)]

Say I have a tensor a = [[2,9,0],[4,1,3],[7,6,8]] and I want the largest 4 values and their indices, so [(9, [0, 1]), (8, [2, 2]), (7, [2, 0]), (6, [2, 1])], would it be more efficient to do torch.topk(a.view([1, 9]), 4) and then search through the original tensor for those values to get the positions of each value or would it be more efficient to convert the tensor to a list of lists?

[tom (Thomas V)]

I’d probably go for your topk call on a flattened tensor and unreaveling the indices as described here:

Finding indices of a global maximum value in a variable

Hi, one way to solve it is to flatten the tensor, take the maximum index along dimension 0 and then unravel the index, either using numpy or your own logic (probably you will come up with something less clumsy ): rawmaxidx = mytensor.view(-1).min(0)[1] idx = [] for adim in list(mytensor.size())[::-1]: idx.append(rawmaxidx%adim) rawmaxidx = rawmaxidx / adim idx = torch.cat(idx) (Note that pytorch / on LongTensor is similar to python2 / or python3 // for ints. Best rega…

Best regards

Thomas