following some of the discussions with multinomial in the past, there still seems to be some problem with its implementation in torch. 1.8.1.
The code
sample = torch.multinomial(torch.tensor([[1, 0, 0, 1]], dtype=torch.float), num_samples=3, replacement=False)
print(sample)
does not throw, nor does it for num_samples=4. However,
sample = torch.multinomial(torch.tensor([[1, 0, 0, 1]], dtype=torch.float), num_samples=5, replacement=False)
print(sample)
gives
RuntimeError: cannot sample n_sample > prob_dist.size(-1) samples without replacement
something is wrong here I guess