GitLab

- supports arbitrary number of languages,
- uses 'multi-task' training with per-language weights (one softmax per language),
- (the decoding script not yet merged)

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@5224 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

- 'blocksoftmax' is better name as it has same name as the used output component,
- we use multitask objective function, so we can see performance on individual tasks,
- also adding some older LSTM results on tedlium, not working on this recipe now,



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@5222 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

- it works, but it needs further analysis, so far it did not outperform the LSTM.



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@5183 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@5170 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

- faster cnn2d by pre-computing scaling factors used on gradients of cnn2d weights
- unit tests for all methods of cnn2d, max-pool-2d, avg-pool-2d components
- added cnn2d recipe in Resource Management, local/nnet/run_cnn2d.sh
bugfix steps/nnet/make_priors.sh : in importing cmvn-option in preparing feature-extraction pipeline 
 


git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@5165 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: standardizing all perl and python scripts to start with a shebang /usr/bin/env perl|python, and to be executable.  Minor script change to train_mono.sh to print better info in case of error.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@5110 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk,nnet1: updating paste-post.cc, improving info-print for MultiSoftmax, fixing multi-database example script to use less RAM



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@5106 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4966 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4961 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4906 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: various script changes to better support Mac OS X.  Avoiding the '-T' option to 'cp', and the '>/dev/stderr' syntax in awk, which is only supported by GNU awk.  Also providing an alternative to 'du -b' for getting file sizes, to failover to 'stat' on Mac

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4877 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4875 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4872 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4862 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4816 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

a) options --apply-cmvn --norm-vars are replaced by --cmvn-opts
b) option --delta-oder replaced by --delta-opts

- the files with setup of cmvn/deltas in exp/dnn.../ changed too, while keeping backward compatibility for nnet/decode.sh nnet/align.sh etc (please let me know if you have problems).



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4815 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4814 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

(trunk/rm) small fix and removing the scoring script as a preparation for replacement with the kaldi default scoring script.


git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4813 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: add example for multi-language discriminative nnet training to egs/rm/s5/.  Misc minor fixes.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4766 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4761 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: Adding example script for 'multi-lingual' nnet2-online training setup, using RM and WSJ as examples.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4760 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4755 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: a lot of minor fixes, cosmetic changes and revisions (most significantly: making conjugate gradient more robust.)

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4681 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4673 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4624 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4619 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: nnet2 setup: Code and script changes to support dumping multi-frame egs to disk.  See get_egs2.sh, train_pnorm_simple2.sh, etc.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4614 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

several nnet2-online changes: make it easier to get the feature extraction options right in cross-system training; add train_pnorm_simple.sh script (simplified learning-rate schedule and improved combination at the end, supersedes train_pnorm_fast.sh); modifying big-data online-nnet2 recipes to use 40-dimensional MFCC rather than 13 as input (will add results soon, but they are improved).  Modified filter_scp.pl to have one-based, not zero-based, field index.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4493 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

- updating the CNN recipe (results not better than DNN)
- removing old and unused initialization scripts



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4470 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: Replacing extract_ivectors_online.sh with (a modified version of) extract_ivectors_online2.sh and deleting the latter from svn. The new version provides the option of creating fake-speakers (see script). Some example scripts using the online setup were changed to remove calls to extract_ivectors_online2.sh.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4441 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

sandbox/online: committing a bunch of changes to the online setup, including code and scripts for retraining models on top of models for other setups.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/sandbox/online@4342 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: various mostly cosmetic changes to nnet scripts; add scripts in RM setup to demonstrate conversion nnet1->nnet2 followed by training.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4317 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

sandbox/online: add scripts for VTLN/frame-shift-perturbed online-nnet2 training on RM (works really well) and reorganize the naems of run_* scripts for online-decoding.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/sandbox/online@4300 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/sandbox/online@4285 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: automatically removing un-needed newlines in logging statements; adding more conversion scripts for neural nets (which will be needed for online-decoding with nnet1 models); improvement to linear CGD implementation.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4283 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

sandbox/online: demonstrating cross-system model transfer for online nnet2 decoding, see run_online_decoding_nnet2_wsj.sh

git-svn-id: https://svn.code.sf.net/p/kaldi/code/sandbox/online@4278 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/sandbox/online@4267 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: finish example script for nnet1->nnet2 model conversion, see demo in egs/rm/s5/local/run_dnn.sh

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4257 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: checking in some partial work towards nnet1->nnet2 conversion (still compiles but the program doesn't run.).

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4246 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8