change --frame argument

[LvHang]

@danpovey
Related to #1370 . Attention, this is in "shortcut" branch, not "master" branch.

Hi Dan,

1. Following you interpretation, I change the code about how to compute the --frame option.
2. I rewrite the tedlium/s5_r2/local/nnet3/run_tdnn.sh so that it will call steps/nnet3/train_dnn.py to train the model. The annotation (line 78-88) is the original shell version. The code (line 90-130) is what I converted.
3. The WER result is followed:
   | after-change | ori
   decode-dev | 11.7 | 12.6
   decode_dev_rescore | 10.9 | 11.5
   decode_test | 11.8 | 11.9
   decode_test_rescore | 11.2 | 10.9
   (The result in "ori" column is extracted from "RESULT" file in tedlium/s5_r2)

Hang

[danpovey]

Interesting. I'm not sure what run_tdnn_py.sh was modified from; anyway it is not following the correct pattern of these filenames; get Yiming to explain the system to you.
I think it's better if you rerun the baseline; numbers in the RESULTS file may be out of date.
Also the system that you are using is not up-to-date at all (it uses the deprecated config-generating scripts) so I would not want to check this in. But it doesn't matter for purposes of testing the change. Just run it again without the change. If the egs were not deleted, then re-use them via the common-egs-dir option.

[LvHang]

1. The "run_tdnn_py.sh" was modified from "tedlium/s5_r2/local/nnet3/run_tdnn.sh". In the 12th stage of "run_tdnn.sh", it call the shell script--"steps/nnet3/tdnn/train.sh". But in Change how the --frame argument is chosen in TDNN training. #1370 , I think you let me modify the "train_dnn.py". So I rewrite it. I asked Yiming and move them to tuning directory.
2. I will rerun the baseline right now.
3. You said the config-generating scripts is out-of-date. However, I check the "master" and "shortcut" branch in your repository, all of them use the "steps/nnet3/tdnn/make_configs.py". I guess you mean we need to write xconfig and use "steps/nnet3/xconfig_to_configs.py", right?
   Hang

[danpovey]

yes it should use the xconfigs, that tdnn sript in tedlium was out of date

…

On Tue, Jan 31, 2017 at 10:27 PM, LvHang ***@***.***> wrote: 1. The "run_tdnn_py.sh" was modified from "tedlium/s5_r2/local/nnet3/run_tdnn.sh". In the 12th stage of "run_tdnn.sh", it call the shell script--"steps/nnet3/tdnn/train.sh". But in #1370 <#1370> , I think you let me modify the "train_dnn.py". So I rewrite it. I asked Yiming and move them to tuning directory. 2. I will rerun the baseline right now. 3. You said the config-generating scripts is out-of-date. However, I check the "master" and "shortcut" branch in your repository, all of them use the "steps/nnet3/tdnn/make_configs.py". I guess you mean we need to write xconfig and use "steps/nnet3/xconfig_to_configs.py", right? Hang — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1389 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADJVu88NYe-us5Vgk1alah2LS_MgB-9aks5rX_uOgaJpZM4Lzb1-> .

[LvHang]

Hi Dan,
I have already run the baseline.The WER result is below:
　　　　　　　　　　| after-change | baseline
decode-dev 　　　　　|　 11.7 　　| 11.7
decode_dev_rescore　 | 10.9 　　　| 11.0
decode_test 　　　　　| 11.8 　　　| 11.6
decode_test_rescore 　| 11.2 　　　| 11.0
Do you think this is what you expect?
If you think it's ok, I'll use the xconfigs.

[danpovey]

I think those differences are just the normal differences we see when we rerun.
I'll merge the change but I don't need any example script, just leave the one-line change
of common.py.

[LvHang]

Ok. I need to modify it on "master" branch or "shortcut" branch? Thanks

[danpovey]

shortcut

…

On Thu, Feb 2, 2017 at 8:11 PM, LvHang ***@***.***> wrote: Ok. I need to modify it on "master" branch or "shortcut" branch? Thanks — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1389 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADJVuwhJp9dKIQzp0c9aezv4i_9SJJfaks5rYn6sgaJpZM4Lzb1-> .

[LvHang]

Update.
Thank you for giving me the chance to read the process of nnet3. Can I help something others? Thanks!
Hang