search

srvk / eesen

The official repository of the Eesen project
http://arxiv.org/abs/1507.08240
Apache License 2.0
825 stars 343 forks source link

tf_clean swbd v1-tf: illegal argument batch_size passed from run_ctc_phn.sh #189

Closed efosler closed 6 years ago

efosler commented 6 years ago

Trying a run straight out of the box on swbd, data prep went (mostly) fine (have one change I'll suggest at some point). Ran into this with run_ctc_phn.sh

...

            Training AM with the Full Set

===================================================================== generating train labels... generating cv labels... steps/train_ctc_tf.sh --nlayer 4 --nhidden 320 --batch_size 16 --learn_rate 0.005 --half_after 6 --model deepbilstm --window 3 --ninitproj 80 --nproj 60 --nfinalproj 100 --norm false data/train_nodup data/train_dev exp/train_phn_l4_c320_mdeepbilstm_w3_nfalse_p60_ip80_fp80 steps/train_ctc_tf.sh: invalid option --batch_size

Sure enough, the train_ctc_tf.sh script does not define a default batch_size as well as learn_rate (may be lr_rate?), window, or norm. Is there some sort of version skew between the scripts?

efosler commented 6 years ago

FWIW, it does look like run_ctc_char.sh has lr_rate rather than learn_rate, but also declares batch_size.

ramonsanabria commented 6 years ago

Hi,

Thanks for pointing this out. I think now should be solved ( https://github.com/srvk/eesen/commit/52d8974d1e5e43b51a309d05f44f55a39ec16f0e). We had a version disagreement with train_ctc_tf.sh. There are few parameters defined outside the .sh script and some in the python script (that we pas as arguments straight away). We should come up with some type of agreement. For now is clean but it would be good to have them either in the python side or in the bash side.

Any thoughts/suggestions @fmetze @efosler ?

Also, please remember to change path.sh accordingly (i.e. add your hostname in the path.sh with $PYTHONPATH and source your virtualenv).

Please keep us letting us know you experience with the tf_clean branch. I am very happy to keep on solving possible bugs.

Thanks!

2018-06-21 13:26 GMT-04:00 Eric Fosler-Lussier notifications@github.com:

FWIW, it does look like run_ctc_char.sh has lr_rate rather than learn_rate, but also declares batch_size.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/srvk/eesen/issues/189#issuecomment-399181855, or mute the thread https://github.com/notifications/unsubscribe-auth/AMlwPUbPi0Lx4-BIQNdELKNo5yn3erY5ks5t-9dKgaJpZM4Uyd7V .