[WIP] sMBR on LFR with shifted feats

[hhadian]

This is the result (the experiment with shifted feats has _fs_smbr as the end):

# System tdnn_lfr1c_sp_smbr:1 tdnn_lfr1c_sp_smbr:2 tdnn_lfr1c_sp_smbr:3 tdnn_lfr1c_sp_fs_smbr:1 tdnn_lfr1c_sp_fs_smbr:2 tdnn_lfr1c_sp_fs_smbr:3 # WER on train_dev(tg) 16.26 16.11 16.02 16.02 15.77 15.78 # WER on train_dev(fg) 15.01 14.91 14.80 14.79 14.58 14.50 # WER on eval2000(tg) 18.9 18.7 18.6 18.6 18.5 18.5 # WER on eval2000(fg) 17.4 17.2 17.1 17.1 17.0 16.9 

Overall there is 0.2-0.3 improvement due to shifted features.

[danpovey]

cool.
can you add a soft-link from one level up?
I'd be interested to know this compares with the chain TDNN system.

[hhadian]

OK.
Do you want to merge it? If yes, first I need to change the code to expand ivector scp using a new script copy_ivector_dir.sh as we discussed (currently I am doing it with a for loop in the main script).

So, I will now do these experiments on a chain-TDNN baseline.

[danpovey]

yes I want to merge it- I'll wait for those changes.

…

On Mon, Mar 6, 2017 at 1:41 PM, Hossein Hadian ***@***.***> wrote: OK. Do you want to merge it? If yes, first I need to change the code to expand ivector scp using a new script copy_ivector_dir.sh as we discussed (currently I am doing it with a for loop in the main script). So, I will now do these experiments on a chain-TDNN baseline. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#1477 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADJVu9X7UsMSOaX9o4ki4278kNz-VK-3ks5rjFNLgaJpZM4MUgHV> .

[hhadian]

I write utt2orig file in ${train_data_dir}/utt2orig and then delete it. I guess this is not very good practice because the source directory might be read-only. Do you want me to write it to the current directory and delete it?

[danpovey]

yes please. [you can use some dir that you know you are outputting to, as a temp dir.]

…

On Mon, Mar 6, 2017 at 11:03 PM, Hossein Hadian ***@***.***> wrote: ***@***.**** commented on this pull request. ------------------------------ In egs/swbd/s5c/local/nnet3/tuning/run_tdnn_lfr1c_disc.sh <#1477 (comment)>: > +if [ ! -f ${srcdir}/final.mdl ]; then + echo "$0: expected ${srcdir}/final.mdl to exist" + exit 1; +fi + + +frame_subsampling_factor=1 +if [ -f $srcdir/frame_subsampling_factor ]; then + frame_subsampling_factor=$(cat $srcdir/frame_subsampling_factor) +fi + +affix= # Will be set if doing input frame shift +if [[ "$shift_feats" = true && $frame_subsampling_factor -ne 1 ]]; then + if [ $stage -le 0 ]; then + utils/data/shift_and_combine_feats.sh --write-utt2orig ${train_data_dir}/utt2orig \ + $frame_subsampling_factor $train_data_dir ${train_data_dir}_fs I write utt2orig file in ${train_data_dir}/utt2orig and then delete it. I guess this is not very good practice because the source directory might be read-only. Do you want me to write it to the current directory and delete it? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#1477 (review)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADJVu3S-SOt1SlCuB31OvDNbOwsCB4vjks5rjOUYgaJpZM4MUgHV> .

[hhadian]

This is ready [unless there is anything I need to fix about copy_ivector_dir.sh]

[danpovey]

thanks- merging.