Adding jasper_tune.py for tuning of ASR hyper parameters alpha and beta

[harisankarh]

This will help in selecting the appropriate alpha and beta parameters for a given validation set (similar to tune.py in Deepspeech2).

Signed-off-by: Hariankar H [email protected]

[okuchaiev]

I think this should just be merged with "examples/asr/jasper.py" @vsl9 and @blisc what do you think?

[vsl9]

Harisankar, thanks a lot for the PR!
I agree with Oleksii that instead of creating a new script it is worth to modify original jasper_infer.py.
Some nice things to add:

• beam_width parameter to override default 128
• alpha_step, beta_step parameters to define steps for grid search
• rename alpha_from/beta_from to just alpha/beta so that it is possible to do inference with a single set of alpha/beta values (without tuning over a grid)
  There is a similar script in our OpenSeq2Seq project: https://github.com/NVIDIA/OpenSeq2Seq/blob/master/scripts/decode.py
  It would be great to implement a similar command line API with beam_width, alpha, alpha_max, alpha_step, beta, beta_max, beta_step in jasper_infer.py to support a single inference or tuning.

[vsl9]

Please see the comment.

[harisankarh]

Thanks for the feedback. I'll look into it.

[harisankarh]

@vsl9 I've made the changes you suggested. Created fresh pull request #67.

Closing this pull request