WIP: Speaker ID script improvements #1538
Closed
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…e10/v1/local to sre08/v1/sid/nnet2
…cognition in nnet3. Not fully debugged.
…nt online computation.
…d to allow online computation. Add a basic test for multi-segment computation.
…ed/unnecessary 'request' args for optimization).
…ieces to splice).
… add tests, and debug to the extent that the tests succeed.
…internal code for shortcut compilation.
…different-sized egs, and different begin/end l/r context
… the num-frames for examples. This code compiles but is not tested.
…lstm recipe with -1 delay at lowest lstm layer (kaldi-asr#1505) swbd : Added tdnn_lstm recipe with delay -1 at the lowest lstm layer
this improves speed when using hashes.
…nt::Scale() if scale==0.0 (kaldi-asr#1522)
…ectors used in ASR. Results are reported in the default TDNN recipe in AMI. Updating steps/online/nnet2/{train_diag_ubm.sh,train_ivector_extractor.sh} so that they now backup the contents of their destination directory if it already exists. (kaldi-asr#1514)
…(avoid space in version); minor fixes (kaldi-asr#1526)
It appears there may be no good reason to disallow system-wide OpenFst.
…template (kaldi-asr#1530) CuVector::AddColSumMat<float>[no-trans], 16 0.0057 0.0172 3.01x CuVector::AddColSumMat<float>[no-trans], 32 0.0242 0.0668 2.76x CuVector::AddColSumMat<float>[no-trans], 64 0.0992 0.2577 2.60x CuVector::AddColSumMat<float>[no-trans], 128 0.3747 0.9280 2.48x CuVector::AddColSumMat<float>[no-trans], 256 1.4711 3.0541 2.08x CuVector::AddColSumMat<float>[no-trans], 512 5.1709 9.4713 1.83x CuVector::AddColSumMat<float>[no-trans], 1024 12.4352 20.4517 1.64x CuVector::AddColSumMat<double>[no-trans], 16 0.0060 0.0175 2.91x CuVector::AddColSumMat<double>[no-trans], 32 0.0240 0.0672 2.80x CuVector::AddColSumMat<double>[no-trans], 64 0.1006 0.2712 2.70x CuVector::AddColSumMat<double>[no-trans], 128 0.3691 0.9097 2.46x CuVector::AddColSumMat<double>[no-trans], 256 1.4530 3.1044 2.14x CuVector::AddColSumMat<double>[no-trans], 512 4.4524 7.5872 1.70x CuVector::AddColSumMat<double>[no-trans], 1024 11.1212 16.1423 1.45x
- we auto-detect the 'compute capability' problems (these appear as the 'invalid device function'), - we also provide guidelines what to try before posting to forum, and which info to send to us,
…true flags at the same time (kaldi-asr#1541)
…h label delay (kaldi-asr#1540) xconfig : Added delay option for FixedAffineLayer. This will be used for ensuring the model specified in ref.config has at least the context required by the model specified in init.config
david-ryan-snyder
force-pushed
the
sid-fix-2017-02-11
branch
from
1ef0b12
to
6599c9b
Compare
|
Something went wrong during the rebasing. Going to try this again in another PR (#1543). |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This WIP PR is for various improvements to the SID recipes. This is primarily motivated by issues tracked on Git or in the Kaldi forums.
The i-vector scripts that use nnet2 in sid/*_dnn.sh now support using GPU explicitly. This is in response to complaints of slowness in these scripts. This will close Issue SRE10 v2 Improvements #165.
A fix to src/fgmmbin/fgmm-global-init-from-accs.cc so that when a Gaussian has very low occupancy, we don't just crash (See https://groups.google.com/d/msg/kaldi-help/U_L_6IWBN1c/L8oPTcE5AgAJ)
LID and SID scripts now do more cleanup with the --cleanup=true option. This closes Issue cleanup when training iVector extractor #1059.
SID i-vector training scripts now use '--num-threads N' instead of '-pe smp N.' This closes Issue SID and LID recipes should use --num-threads N #1096. (The LID scripts were already doing the right thing).
In sre10/v1/local/plda_scoring.sh added an option for --simple-length-norm (which defaults to 'false' since it gives better performance in SRE10). This closes issue In SID recipes, provide a script-level option for simple-length-normalization #1097.
In egs/sre10/{v1,v2}/run.sh, the PLDA scores are now written to exp instead of local. This is better, since v1 and v2 share the same local directory and would override each other otherwise. Also changing the old-style memory options to the new ones (E.g., --mem 5G))
The scripts to train the DNN for SRE10 have been moved from sre10/v1/local to sre08/v1/sid/nnet2 (this mirrors what we did with lre07/v1/lid/nnet2). This is consist with other setups, and makes them easier to access by new (or user created) SID recipes.
In sre10/v1/local/dnn/run_nnet2_multisplice.sh we now use 8 GPUs to the train the DNN, instead of 18 (which is excessive, and might've been a typo).
Various cosmetic fixes: fixed indentation in several sid and lid scripts. Fixed a typo in src/gmm/full-gmm.cc . Changed wording in egs/sre10/v1/local/dnn/train_dnn.sh so that it, which is an nnet2 pnorm recipe, is no longer referred to as the "current best recipe" but rather as an "older nnet2 recipe," which is now the correct thing to say.