Potential bug when sampling from categorical distribution

[christinaheinze]

Hi,

I would like to sample from a categorical distribution where the probabilities passed are the columns of a tensor p_dG:

G = 3
D = 2
p_dG = torch.Tensor(G, D)
p_dG[:, 0] = torch.Tensor([0.1, 0.8, 0.1])
p_dG[:, 1] = torch.Tensor([0.1, 0.8, 0.1])

z_list = []
p_dg = p_dG[:, 0]
print(p_dg)
dis_Z = dis.Categorical(p_dg)
for _ in range(250):
    z = dis_Z.sample()
    z_list.append(z)

z = torch.cat(z_list, dim=0)

true_z_np = z.numpy()
v, c = np.unique(true_z_np, return_counts=True)
print(v)
print(c)

The output from print(c) is [ 29 25 196]. So the distribution is completely off since the bulk of the observations should be equal to 1 (instead of being equal to 2).

In contrast, the following code works as expected:

z_list = []
dis_Z = dis.Categorical(torch.Tensor([0.1, 0.8, 0.1]))
for _ in range(250):
    z = dis_Z.sample()
    z_list.append(z)

z = torch.cat(z_list, dim=0)
true_z_np = z.numpy()
v, c = np.unique(true_z_np, return_counts=True)
print(v)
print(c)

Is this a bug or am I missing something?

I am using:

• OS: macOS High Sierra
• PyTorch version: 0.3.0
• How you installed PyTorch (conda, pip, source): conda
• Python version: Python 3.6.3 |Anaconda custom (64-bit)| (default, Oct 6 2017, 12:04:38)
• CUDA/cuDNN version: --
• GPU models and configuration: --
• GCC version (if compiling from source): --

Thanks!

[ssnl]

Yes this is a bug. Fortunately it seems to have been fixed on master. Thanks for reporting!

[soumith]

@christinaheinze you can follow instructions here to easily compile for master: https://github.com/pytorch/pytorch#from-source

[christinaheinze]

Ok, thanks!

[christinaheinze]

On master the issue has been solved for the categorial distribution but not for the multinomial distribution as it seems:

import torch.distributions as dis
import torch
import numpy as np

G = 3
D = 2
p_dG = torch.Tensor(G, D)
p_dG[:, 0] = torch.Tensor([0.1, 0.8, 0.1])
p_dG[:, 1] = torch.Tensor([0.1, 0.8, 0.1])

p_dg = p_dG[:, 0]
z = p_dg.multinomial(250, replacement=True)
true_z_np = z.numpy()
v, c = np.unique(true_z_np, return_counts=True)
print(v)
print(c)

and also

z = torch.multinomial(p_dg, 250, replacement=True)
true_z_np = z.numpy()
v, c = np.unique(true_z_np, return_counts=True)
print(v)
print(c)

[ssnl]

I'll take a look at this tomorrow.

[alicanb]

Same problem is in distributions.Multinomial as well, even though it doesn't use torch.multinomial for sampling. I'll check that out

[ssnl]

@alicanb It's maintaining a _categorical object internally. So it's probably that.

[alicanb]

On second thought, what is the bug here?

[ssnl]

@alicanb It's CPU torch.multinomial not working on noncontiguous tensors. I'm writing a fix now.

[ssnl]

@alicanb Btw, do you know why distributions.Multinomial use a distributions.Categorical and a scatter_add, instead of just using torch.multinomial? What's it that I am missing here?

[alicanb]

@ssnl torch.multinomial is misnomer actually, it's equivalent to torch.distributions.Categorical(...).sample(). We thought about correcting it, but it's too much of a breaking change. Here is the discussion probtorch#46

[alicanb]

@christinaheinze if you're trying to implement multinomial distribution btw, we have distributions.Multinomial(total_count, probs) on master that behaves correctly