Pytorch Operation to detect NaNs


Is there a Pytorch-internal procedure to detect NaNs in Tensors? Tensorflow has the tf.is_nan and the tf.check_numerics operations … Does Pytorch have something similar, somewhere? I could not find something like this in the docs…

I am looking specifically for a Pytorch internal routine, since I would like this to happen on the GPU as well as on the CPU. This excludes numpy – based solutions (like np.isnan(sometensor.numpy()).any()) …

Asked By: cleros



You can always leverage the fact that nan != nan:

>>> x = torch.tensor([1, 2, np.nan])
tensor([  1.,   2., nan.])
>>> x != x
tensor([ 0,  0,  1], dtype=torch.uint8)

With pytorch 0.4 there is also torch.isnan:

>>> torch.isnan(x)
tensor([ 0,  0,  1], dtype=torch.uint8)
Answered By: nemo

Starting with PyTorch 0.4.1 there is the detect_anomaly context manager, which automatically inserts assertions equivalent to assert not torch.isnan(grad).any() between all steps of backward propagation. It’s very useful when issues arise during backward pass.

Answered By: Jatentaki

As suggested by @cleros in the comment on @nemo’s answer, you can get this as a boolean using the any() operator:

Answered By: crypdick

True if any value is nan:


True if all is nan:

Answered By: joydeba

If you want to call it on a tensor directly:

import torch

x = torch.randn(5, 4)


import torch
x = torch.randn(5, 4)
Answered By: Charlie Parker
Categories: questions Tags: ,
Answers are sorted by their score. The answer accepted by the question owner as the best is marked with
at the top-right corner.