[FIXED] Random Choice with Pytorch?

Issue

I have a tensor of pictures, and would like to randomly select from it. I'm looking for the equivalent of np.random.choice().

import torch

pictures = torch.randint(0, 256, (1000, 28, 28, 3))

Let's say I want 10 of these pictures.

Solution

torch has no equivalent implementation of np.random.choice(), see the discussion here. The alternative is indexing with a shuffled index or random integers.

To do it with replacement:

1. Generate n random indices
2. Index your original tensor with these indices

pictures[torch.randint(len(pictures), (10,))]  

To do it without replacement:

1. Shuffle the index
2. Take the n first elements

indices = torch.randperm(len(pictures))[:10]

pictures[indices]

Read more about torch.randint and torch.randperm. Second code snippet is inspired by this post in PyTorch Forums.

Answered By - Nicolas Gervais