Per row/column clamping

1

Hello,

I am wondering what is the most efficient way to implement per-row/column clamping? For instance,

input = torch.tensor([[0,3,2,-1],[2,1,3,4]])
input_clamp = torch.Tensor([2,3])
output = per_row_clamp(input, input_clamp)

The output then should be [[0,2,2,-1],[2,1,3,3]]

One approach I can think of is to do a per-row division, then clamp everything to 1, then multiply the per-row clamp value back.

Thanks!

2

You can try below code. just do some boolean indexing in the data. After comparator, do the replacement in the data. You will get the idea.

import torch

input = torch.tensor([[0,3,2,-1],[2,1,3,4]],dtype = torch.float)
input_clamp = torch.Tensor([2,3])

tiled = input_clamp.repeat(4,1).t()

bigger = input > tiled
input[bigger] = tiled[bigger]

print(input)```