Wav2letter: Pre-trained models, CPU-only

[JanX2]

After a few smaller patches (forthcoming) I got through the whole process describe in the README for testing the pre-trained models. Up to the point where I can finally run test.lua. I’m pretty confident that this would run, but I can’t get test.lua to run CPU-only. I don’t have an nVidia card handy and so didn’t install any CUDA-related libraries.

So I’m running test.lua with -gpu 0 (which requires another forthcoming patch) and this is where I’m stuck:

| number of classes (network) = 30
| reloading model </Volumes/Rest/Temp/prem 201X-XX-XX/LibriSpeech-11/librispeech-glu-highdropout.bin>
luajit: ...ment/Projects/torch/install/share/lua/5.1/torch/File.lua:343: unknown Torch class <torch.CudaTensor>
stack traceback:
    [C]: in function 'error'
    ...ment/Projects/torch/install/share/lua/5.1/torch/File.lua:343: in function 'readObject'
    ...ment/Projects/torch/install/share/lua/5.1/torch/File.lua:369: in function 'readObject'
    ...pment/Projects/torch/install/share/lua/5.1/nn/Module.lua:192: in function 'read'
    ...ment/Projects/torch/install/share/lua/5.1/torch/File.lua:351: in function 'readObject'
    ...ment/Projects/torch/install/share/lua/5.1/torch/File.lua:369: in function 'readObject'
    ...ment/Projects/torch/install/share/lua/5.1/torch/File.lua:369: in function 'readObject'
    ...pment/Projects/torch/install/share/lua/5.1/nn/Module.lua:192: in function 'read'
    ...ment/Projects/torch/install/share/lua/5.1/torch/File.lua:351: in function 'readObject'
    ...ment/Projects/torch/install/share/lua/5.1/torch/File.lua:369: in function 'readObject'
    ...orch/install/share/lua/5.1/wav2letter/runtime/serial.lua:57: in function 'loadmodel'
    ...do/Development/Projects/wav2letter-Facebook-STT/test.lua:108: in main chunk
    [C]: at 0x010319be60

The issue appears to be similar to https://github.com/soumith/dcgan.torch/issues/28, but I don’t know enough of the architecture to make progress here quick.

It looks like the network was saved in GPU-only mode and requires some kind of conversion to run CPU-only. Is this correct?

In that case, an you supply a CPU-only or hybrid model?

Thanks for all the work you put into this project!

[JanX2]

PR #5 has my small patches.

[VitaliyLi]

Hi, yes, we are working on converting our GPU model and supplying CPU-only model.

[JanX2]

That should have been PR #7.

[soswow]

Duplicate of #4 but I like that this one is still open

[VitaliyLi]

CPU model has been uploaded:
https://s3.amazonaws.com/wav2letter/models/librispeech-glu-highdropout-cpu.bin