Embed defaultlm

[qmpzzpmq]

add to parameter to control embed unit in default rnnlm, method same with transformer LM.

[qmpzzpmq]

I don't know what's going on with

espnet/test/test_e2e_compatibility.py

Line 40 in 8db0c29

subprocess.run(cmd, check=True)

espnet/test/test_e2e_compatibility.py

Line 50 in 8db0c29

r'<a id="uc-download-link" [^>]* href="([^"]*)">', out)[0].replace('&amp;', '&'))

espnet/test/test_e2e_compatibility.py

Line 74 in 8db0c29

model_path = download_zip_from_google_drive(tmpdir, download_info[1])

it seems a downloading protein model problem.
can you guys give me a clue?

[sw005320]

This happens on the other PR as well.
I could manually download the zip file, so it might be a tentative issue in the Google drive.
Please ignore it for now.

[codecov]

Codecov Report

Merging #1623 into master will increase coverage by 0.36%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master    #1623      +/-   ##
==========================================
+ Coverage   78.08%   78.45%   +0.36%     
==========================================
  Files         127      127              
  Lines       11774    11991     +217     
==========================================
+ Hits         9194     9407     +213     
- Misses       2580     2584       +4

Impacted Files	Coverage Δ
espnet/nets/pytorch_backend/lm/default.py	90.83% <100%> (+0.36%)	⬆️
espnet/nets/pytorch_backend/e2e_asr_transformer.py	94.79% <0%> (-0.17%)	⬇️
espnet/nets/pytorch_backend/e2e_st_transformer.py	85.5% <0%> (+0.76%)	⬆️
espnet/nets/pytorch_backend/rnn/decoders.py	95.04% <0%> (+1.42%)	⬆️
espnet/nets/pytorch_backend/e2e_mt_transformer.py	78.34% <0%> (+1.69%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 8db0c29...097b51d. Read the comment docs.

[sw005320]

Now it's working (this kind of thing happens several times...)

[sw005320]

Currently, the default uses args.unit, so could you also set it like that instead of default=128? (maybe we can just set with default=None?

[qmpzzpmq]

Actually, I think about it last night. I agree with change default with 650 keeping same with unit.

[sw005320]

I think it's better to make default=None.
Then, it just uses the same dimension as the number of hidden units according to

espnet/espnet/nets/pytorch_backend/lm/default.py

Line 242 in 860cbd2

if n_embed is None:

Thus, this becomes the exact same behavior unless your new option of --embed-unit is specified.
If you agree, please change it, and also add such comments in the argument help.

[qmpzzpmq]

ok, I can do it, but which place you suggest my add some comment? in parser.add_argument 's help parameter?

[sw005320]

yes!

[sw005320]

Thanks a lot!