Invalid multinomial distribution (encountering probability entry < 0) at /pytorch/aten/src/TH/generic/THTensorRandom.cpp:325

Question

I have a PyTorch tensor called out_probs which is produced like this:

out_probs=F.softmax(out_dec[:,0],dim=0)

Also, the shape of out_probs is [128,20004]

out_probs is the result of a softmax operation and it's not supposed to contain any negative value so naturally the result of out_probs[out_probs<0 is going to be an empty tensor(actually I checked and it was empty) But when I'm running

torch.multinomial(out_probs, 1)

I'm getting :

RuntimeError: invalid argument 2: invalid multinomial distribution (encountering probability entry < 0) at /pytorch/aten/src/TH/generic/THTensorRandom.cpp:325

That implies my tensor has a negative entry and I don't know why this is happening?

Answer 1

I believe you've found a bug in the error reporting for torch.multinomial.

For example

x = torch.ones(128, 1)
x[0] *= 1e100
out_probs = F.softmax(x, dim=0)
print('Negative values:', torch.sum(out_probs < 0).item())
y = torch.multinomial(out_probs, 1)

results in the following output

Negative values: 0
RuntimeError: invalid argument 2: invalid multinomial distribution (encountering probability entry < 0) at /pytorch/aten/src/TH/generic/THTensorRandom.cpp:298

It turns out this is getting triggered because out_probs contains nan entries.

print('nan values:', torch.sum(torch.isnan(out_probs)).item())

gives

nan values: 128

Which are caused by mathematical instabilities in softmax.

Strangely, when the values in out_probs are infinite you get the proper error message

RuntimeError: invalid argument 2: invalid multinomial distribution (encountering probability entry = infinity or NaN) at /pytorch/aten/src/TH/generic/THTensorRandom.cpp:302

This bug should probably be reported at https://github.com/pytorch/pytorch/issues if it hasn't been fixed in the most recent version.

By the way I'm using PyTorch 1.0.1.post2