Recipe update to be compatible with pytorch

[kan-bayashi]

Under construction.

[ShigekiKarita]

I think it is better to merge #9 before this PR to avoid new conflicts.

[sw005320]

@kan-bayashi is this ready for merging?

[kan-bayashi]

@sw005320 Now it can be merged.

[sw005320]

• @ShigekiKarita I didn't notice that the stage structure in the WSJ recipe was changed. I want to stick to the original structure (data preparation for test data and scoring are merged in one recognition stage). If it is OK, I will change it to the original structure. Then, with @kan-bayashi 's changes, all recipes have the consistent stage structure.
• pytorch backend does not currently support vggblstnp architecture, and better to mention it and possibly make blstmp as default architecture for the pytorch case.

[sw005320]

Can you also change chainer/pytorch switch as follows?

# switch backend
if [[ ${backend} == chainer ]]; then
    train_script=asr_train.py
    decode_script=asr_recog.py
else
    train_script=asr_train_th.py
    decode_script=asr_recog_th.py
fi

see https://github.com/espnet/espnet/blob/master/egs/wsj/asr1/run.sh

[kan-bayashi]

Finished.
And I changed tab with space to clean up the code.

[sw005320]

Thanks for fixing the indent. Can you also fix that of WSJ run.sh (or already done?)

[kan-bayashi]

Done.

[sw005320]

@ShigekiKarita can you also review src/nets/e2e_asr_attctc_th.py?

[ShigekiKarita]

OK

[ShigekiKarita]

LGTM