Random Choice with Pytorch?

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.

For this size of tensor:

N, D = 386363948, 2k = 190973values = torch.randn(N, D)

The following code works fairly fast. It takes around 0.2s:

indices = torch.tensor(random.sample(range(N), k))indices = torch.tensor(indices)sampled_values = values[indices]

Using torch.randperm, however, would take more than 20s:

sampled_values = values[torch.randperm(N)[:k]]

torch.multinomial provides equivalent behaviour to numpy's random.choice (including sampling with/without replacement):

# Uniform weights for random drawunif = torch.ones(pictures.shape[0])idx = unif.multinomial(10, replacement=True)samples = pictures[idx]

samples.shape>>> torch.Size([10, 28, 28, 3])