search

k2-fsa / k2

FSA/FST algorithms, differentiable, with PyTorch compatibility.
https://k2-fsa.github.io/k2
Apache License 2.0
1.08k stars 211 forks source link

Build from source, rnnt_loss_test_py Failed #1191

Open tomato18463 opened 1 year ago

tomato18463 commented 1 year ago

My system is Ubuntu 18.04, CUDA 11.8, pytorch 2.0.0. I built k2 from source, it worked well. Then I run the test case as instructed here. I have all test cases passed but rnnt_loss_test_py failed. I have attached the test log below.

Please can you have a look at what's wrong? Is it a problem with my system? Would it influence normal use of k2, and if yes, how to fix it?

Thanks!

LastTest.log

yfyeung commented 1 year ago

You can change test_rnnt_loss_empty_reference to _test_rnnt_loss_empty_reference to disable this test.

tomato18463 commented 1 year ago

Will it cause some errorous result in some specific function if I ignore this test? Is there any function I should avoid with this test failed?

You can change test_rnnt_loss_empty_reference to _test_rnnt_loss_empty_reference to disable this test.

yfyeung commented 1 year ago

Currently, we didn't find any errorous results caused by rnn_loss.py. When S is 0, torchaudio.functional.rnnt_loss gets non-zeros value, which is different with k2.rnnt_loss.

tomato18463 commented 1 year ago

Currently, we didn't find any errorous results caused by rnn_loss.py.

I see. Thanks!

tomato18463 commented 1 year ago

Is it this line that I need to change?

yfyeung commented 1 year ago

Is it this line that I need to change?

You can refer this PR: https://github.com/k2-fsa/k2/pull/1193