Ivan Sorokin

icon indicating a website link https://www.linkedin.com/in/lytic/

icon company @5vision @speechpro @sprout-ai @NVIDIA icon location Reinforcement Learning -> Speech Recognition -> Natural Language Processing

Results 30 comments of Ivan Sorokin

Transducer loss leads to memory leak

Hey, @YangHao97 Maybe the issue with the variable `transducer_loss`. Do not store it directly in `ret`. Try to reuse it and return a numpy array, or something like this.

Transducer loss leads to memory leak

I'm still think that issue with the variable `transducer_loss`. You have to free it, instead of store it somewhere else. Anyway, if you have found the workaround, I'm happy for...

RuntimeError: Trying to backward through the graph a second time, but the buffers have already been freed。

Take a look at this example https://github.com/1ytic/open_stt_e2e/blob/master/train_rnnt.py

Exception: CPU version is not implemented

Try to properly specify LD_LIBRARY_PATH.

WARNING: sample 0 [42, 26] has a forward/backward mismatch -52.543503 / 0.000000

Something is wrong with compiled kernels. Basically, this means the wrong computation of the loss.

Different outputs for multiple calls

Thanks for debugging this package. I use this in another project [open_stt_e2e](https://github.com/1ytic/open_stt_e2e). For example, you can see usage of AverageCER, AverageWER in [train_ctc.py](https://github.com/1ytic/open_stt_e2e/blob/master/train_ctc.py). Hope this helps you.

Different outputs for multiple calls

Nope. AverageCER/WER accumulates error statistics (deletions, insertions, substitutions) for the entire dataset. I recommend you look at the test [test.py](https://github.com/1ytic/pytorch-edit-distance/blob/master/torch_edit_distance/test.py). Maybe it helps you to understand the usage of the...

computer_wer returns nan when I use device other than cuda:0

It should handle the selected device. Could you provide the source code for testing?

undefined symbol: _ZNSt19basic_ostringstreamIcSt11char_traitsIcESaIcEEC1Ev

Sorry for late reply. Try to reinstall the new version `pip install warp_rnnt==0.6.0`.

rnnt_loss status 1

Sorry for late reply. I don't know, maybe some rare CUDA hardware? or library mismatch. Try to reinstall packages in the separated environment. And please use a new version `pip...