Tdnn pool #3

…ing gmm-latgen-faster The binary complains that there is no such option and exits abnormally, which in turn causes local/test_decoders.sh to fail with a scary error message.

rm/s5: Don't use the no-longer-existing '--max-arcs' option when call…

12 hours is the estimate I got when I ran the recipe: ``` root@27af4331c701:/opt/kaldi/egs/swbd/s5c/data/train_30kshort $ awk '{s+=$4-$3}END{print s}' segments 43022.7 ``` (That's 11.95 hours.)

Fix swbd/s5c recipe comment

… with online decoding (and to enable fix to --snip-edges=false bug).

…o write online feature-extraction code that would respect the snip-edges=false option.

…rturb

- the stays same functionality, but now it is more 'correct', as we don't do the triple cast with 'bool'.

moving the src/path.sh into tools/config/common_path.sh

Fixing stuff and renaming into am_nnet.

Updated arpa2fst usage message

WIP: fix for utt2dur when applied on speed-pertubed data

…oef'

…he correct corpra.

…es option; includes rewrite of window extraction code.

… bn-music-speech.

…the HUB4 corpus, update to local/make_bn.py so that it is more flexible with regard to differences in source directory format

…ch-fix BN music speech fixes

smbr: Avoid extra epochs if frame shift is not used during training

Adding error message if $(tree-info | grep) fails. Exit correctly when the command fails.

…t context from 5 to 0

modified run_tdnn_6v.sh by appending online decoding.

vadfix: fixed bug in voice activity detection context, changed default context from 5 to 0

utils/mkgraph.sh finds the information about the context in the tree.

pipe added

…some utility scripts

Update arpa2fst invocations in individual egs/*/local scripts

Update arpa2fst invocations in common (wsj/{utils,steps}) files

smbr_fix: Set modify-learning-rate false in chain smbr

Reorganization of feature code in order to correctly handle --snip-edges=false in online feature generation

adding clarification about recipes not working under windows

changed beam to 11, commented online decoding block and added online decoding results to 6v_sp script.

Fixing a few compile warnings, each occuring only once in the whole codebase.

Update run_lstm.sh

- Now in Component::ReadData() we can read variant number of <Tokens>, without the danger of 'eating' the next component. - We keep the backward compatibility with the existing models,

Online nnet3 code

Non-test code only.

Singular warning fixes

Fixed one option in nnet/run_lstm.sh

force to generate utt2dur for every srcdir while calling combine_data.sh

…ents

Subset on utt2spk when feats.scp is not necessary.

Modify analyze-counts.cc: add check for negative inputs on the alignm…

SetLogHandler() to redirect Kaldi loggigng to custom function

nnet1: creating tool 'nnet-set-learnrate', some structural changes in I/O,

…i-asr#708

…A 5.5

Nnet3 steps fixes

Resolving some coding errors in cnfigure script. Proposing adding CUDA 5.5

fix hkust locale

Add table of examples (kaldi-asr#469)

Fix typo, remove extra quote

…chain add fisher_swbd nnet3 and chain recipe

- new implementation of Nnet::GetParams, Nnet::SetParams, Nnet::GetGradient - slight change of API for Nnet - chaning the code to be closer to Google style

…eatures to /tmp (due to filling up the disk), the features are removed.

…frame shifts.

Update cmd.sh

Chime4 egs

- Now 'analyze-couts' obtains the dimensionality of the counts by option. - The problem was reported in: https://groups.google.com/forum/#!topic/kaldi-help/seUrycQ7nQc

- adding I/O functions for CuArray<int32>, - new method for adding constant to CuArray<int32>, - simplified I/O for nnet1 components with CuArray<int32>, - in nnet1 added function for building 'CuArray<int32>' from vector<vector<int32> > holding 'matlab-like' indices

Iban recipe

…learning rate. This allows it to be updated via nnet3-copy during training.

….t. removing $tmpdir)

…g (should make no difference to anything)

Fix an apparent typo in nnet3::ComputeNnetComputationEpochs()

…nc-comp-fix nnet3: fix to ConstantFunctionComponent

…o clarify it's likely a Makefile issue.

…n-smbr Update fisher swbd tdnn smbr

…ons.cc into two pieces; add flag in makefile. Plus a cosmetic change in nnet-computation-graph.cc.

- reordering cudamatrix/cu-array-inl.h, so the methods appear in the same order as in cudamatrix/cu-array.h - adding asserts for 0-dim of CuArray

… scoring if needed.

Modify steps/nnet{2,3}/decode.sh to only pass in the --iter option to…

iban recipe -- adding path.sh that was accidentally removed today

…he olabel (which shouldn't matter in our normal use-cases), and make final-probs application more efficient.

…cleanup) to word-align-lattice.cc, that affects the --test=true option of lattice-align-words for some topologies including chain models.

…roperly Also add an additional unit test for the topological sorting function

- e.what() contains stackttrace or is empty string - we should also consider changing: 'std::cerr << e.what();' -> 'fprintf(stderr, e.what().c_str());' - fprintf is thread-safe and it is better not to mix 'std::cerr' and 'stderr', and 'stderr' is already used for logging...

…t shown to the user.

fix small bug in train_mono.sh regarding how errors in feat-to-dim ge…

nnet3: Undo the damage done by 2310a19 and fix the original problem properly

Update to ComposeCompactLatticeDeterministic to fix a bug regarding t…

Small fixes

…peex_usage Added --speex-* options to configure.

…city.

Fix inversion of input and output of open2.

…eline in nnet2 and nnet3 decoders to allow more flexible feature pipelines.

…local/wsj_format_data.sh

Sw005320 chime4 egs

update to validate_dict_dir.pl to check for ^M (Issue 729)

…efer to each other; improve messages printed out by 'configure'

Use OnlineFeatureInterface in nnet2 and nnet3 decoders

some changes to fix compilation issues on cygwin: broke up kws-functi…

chime4: minor bug-fixes

…efer to each other; improve messages printed out by 'configure'

…w script supports both nnet2 and nnet3. Probably nnet1.

Adding nnet3/chain support to steps/cleanup/find_bad_utts_nnet.sh, no…

- some TODO's are to be decided: - Can we remove the: 'IsKaldiError()'? (It's very 'dirty' function. And it's used only in the table-I/O to suppress printing 'what' messages from KALDI_ERR. IMHO, it may not be a good idea to suppress this.) - With Kirill's log-handler, the log is sent and then there's no abort() for errors/asserts (seems like a bad idea, but it is the way it worked previously).

…t3 setup) regarding how silence-weighting is applied in iVector estimation. (thanks: xiang li)

Cpplint style

- fixing bug when i-vectors are used with 'steps/nnet/train_mpe.sh', - cosmetic changes,

nnet1: updating the MMI / sMBR scripts,

…BM when there are too few Gaussians and some zero counts.

…t addressed (RE zero weights)

…w when speakers have many utterances.

Remove old memory alignment flags

Add left biphone recipe for swbd

* tedlium: Make scoring scripts correctly remove <unk> from ctm files. Previous scripts incorrectly added double quotes within the single quotes, so that grep would try to match only quoted words. They also mistakenly used \<, and \>. < and >, when escaped by backslashes, indicate word boundaries, not the characters "<" and ">". * tedlium: scripts to reproduce chain paper results. This includes nnet3 CE and sMBR model baseline scripts.

Minor bugfix in utils/data/perturb_data_dir_volume.sh

…ailures for chain models) and to another bug found by ilya platonov regarding extra disambig symbols.

…anks to remi francis for noticing the issue)

Now GetParams requires vector allocated already. Changes in nnet\nnet-various.h at Apr 21, 2016

Add seed for random number generator in utils/data/perturb_data_dir.sh

Update perturb_data_dir_volume.sh

Update nnet1-to-raw-nnet.cc

Old: LOG (TestCuFindRowMaxId():cu-matrix-speed-test.cc:264) For CuMatrix::FindRowMaxId<float>, for dim = 1024, speed was 3.99218 gigaflops. LOG (TestCuFindRowMaxId():cu-matrix-speed-test.cc:264) For CuMatrix::FindRowMaxId<double>, for dim = 1024, speed was 3.46283 gigaflops. New: LOG (TestCuFindRowMaxId():cu-matrix-speed-test.cc:264) For CuMatrix::FindRowMaxId<float>, for dim = 1024, speed was 66.2965 gigaflops. LOG (TestCuFindRowMaxId():cu-matrix-speed-test.cc:264) For CuMatrix::FindRowMaxId<double>, for dim = 1024, speed was 58.442 gigaflops.

…s not.

align-equal-compiled.cc: correct the usage description

…ming more generic, - the binary can be replaced (so we could eventually append posteriors, features, etc.)

A new CUDA kernel for CuMatrixBase<Real>::FindRowMaxId;

base/kaldi_error : the error messages are no longer printed 2x

…[email protected])

Add barrier for correct timing. Original performance: LOG (TestCuMatrixTransposeCross():cu-matrix-speed-test.cc:91) For CuMatrix::TransposeCross<float>, for dim = 1024, speed was 4.26727 gigaflops. LOG (TestCuMatrixTransposeS():cu-matrix-speed-test.cc:72) For CuMatrix::TransposeS<float>, for dim = 1024, speed was 5.97203 gigaflops. LOG (TestCuMatrixTransposeNS():cu-matrix-speed-test.cc:56) For CuMatrix::TransposeNS<float>, for dim = 1024, speed was 3.0816 gigaflops. LOG (TestCuMatrixTransposeCross():cu-matrix-speed-test.cc:91) For CuMatrix::TransposeCross<double>, for dim = 1024, speed was 3.95059 gigaflops. LOG (TestCuMatrixTransposeS():cu-matrix-speed-test.cc:72) For CuMatrix::TransposeS<double>, for dim = 1024, speed was 4.36189 gigaflops. LOG (TestCuMatrixTransposeNS():cu-matrix-speed-test.cc:56) For CuMatrix::TransposeNS<double>, for dim = 1024, speed was 2.39275 gigaflops.

LOG (TestCuMatrixTransposeCross():cu-matrix-speed-test.cc:91) For CuMatrix::TransposeCross<float>, for dim = 1024, speed was 14.0498 gigaflops. LOG (TestCuMatrixTransposeS():cu-matrix-speed-test.cc:72) For CuMatrix::TransposeS<float>, for dim = 1024, speed was 16.845 gigaflops. LOG (TestCuMatrixTransposeNS():cu-matrix-speed-test.cc:56) For CuMatrix::TransposeNS<float>, for dim = 1024, speed was 14.2464 gigaflops. LOG (TestCuMatrixTransposeCross():cu-matrix-speed-test.cc:91) For CuMatrix::TransposeCross<double>, for dim = 1024, speed was 10.4523 gigaflops. LOG (TestCuMatrixTransposeS():cu-matrix-speed-test.cc:72) For CuMatrix::TransposeS<double>, for dim = 1024, speed was 9.65529 gigaflops. LOG (TestCuMatrixTransposeNS():cu-matrix-speed-test.cc:56) For CuMatrix::TransposeNS<double>, for dim = 1024, speed was 8.52148 gigaflops.

add new results for Multi-splice version of online recipe of Librispeech, including those on test set.

…size. New: LOG (TestCuMatrixTraceMatMat():cu-matrix-speed-test.cc:458) For CuMatrix::TraceMatMat<float>, for dim = 1024, speed was 10.1076 gigaflops. LOG (TestCuMatrixTraceMatMat():cu-matrix-speed-test.cc:458) For CuMatrix::TraceMatMat<float> [transposed], for dim = 1024, speed was 11.8711 gigaflops. LOG (TestCuMatrixTraceMatMat():cu-matrix-speed-test.cc:458) For CuMatrix::TraceMatMat<double>, for dim = 1024, speed was 7.10019 gigaflops. LOG (TestCuMatrixTraceMatMat():cu-matrix-speed-test.cc:458) For CuMatrix::TraceMatMat<double> [transposed], for dim = 1024, speed was 7.81977 gigaflops. Old: LOG (TestCuMatrixTraceMatMat():cu-matrix-speed-test.cc:458) For CuMatrix::TraceMatMat<float>, for dim = 1024, speed was 4.57783 gigaflops. LOG (TestCuMatrixTraceMatMat():cu-matrix-speed-test.cc:458) For CuMatrix::TraceMatMat<float> [transposed], for dim = 1024, speed was 7.96795 gigaflops. LOG (TestCuMatrixTraceMatMat():cu-matrix-speed-test.cc:458) For CuMatrix::TraceMatMat<double>, for dim = 1024, speed was 3.61182 gigaflops. LOG (TestCuMatrixTraceMatMat():cu-matrix-speed-test.cc:458) For CuMatrix::TraceMatMat<double> [transposed], for dim = 1024, speed was 6.39571 gigaflops.

2 CUDA kernels for TraceMatMat with/without transpose for all matrix size.

smbr: Fixed minor bug in generating diagnostics egs

Speed up CuMatrix<Real>::Transpose() and transposed copy from matrix

some cosmetic changes: add comments to RNNLM rescoring utilities to r…

added utils/combine_ali_dirs.sh (fixes kaldi-asr#553).

…ningful error messages.

$ ./configure --mkl-root=/opt/intel/mkl --static-math=yes ... Configuring MKL library directory: Found: /opt/intel/mkl/lib/intel64 MKL configured with threading: sequential, libs: -Wl,--start-group /opt/intel/mkl/lib/intel64/libmkl_intel_lp64.a /opt/intel/mkl/lib/intel64/libmkl_core.a /opt/intel/mkl/lib/intel64/libmkl_sequential.a -Wl,--end-group MKL include directory configured as: /opt/intel/mkl/include Configuring MKL threading as sequential MKL threading libraries configured as -lpthread -lm Using Intel MKL as the linear algebra library. /opt/intel/mkl/lib/intel64/libmkl_core.a(mkl_memory_patched.o): In function `mkl_serv_set_memory_limit': mkl_memory.c:(.text+0x49c): undefined reference to `dlsym' mkl_memory.c:(.text+0x4b2): undefined reference to `dlsym' mkl_memory.c:(.text+0x4c8): undefined reference to `dlsym' /opt/intel/mkl/lib/intel64/libmkl_core.a(mkl_memory_patched.o): In function `mkl_serv_allocate': mkl_memory.c:(.text+0x1251): undefined reference to `dlsym' mkl_memory.c:(.text+0x1267): undefined reference to `dlsym' ...

Add missing dependencies to Makefiles

Add dimension check in online-nnet3 decoding code, so we get more mea…

Fix bug: static link to MKL 11.3.2 failed.

Tdnn pool #3

Are you sure you want to change the base?

Tdnn pool #3

Commits on Apr 8, 2016

Commits on Apr 9, 2016

Commits on Apr 10, 2016

Commits on Apr 11, 2016

Commits on Apr 12, 2016

Commits on Apr 13, 2016

Commits on Apr 14, 2016

Commits on Apr 15, 2016

Commits on Apr 16, 2016

Commits on Apr 17, 2016

Commits on Apr 18, 2016

Commits on Apr 19, 2016

Commits on Apr 20, 2016

Commits on Apr 21, 2016

Commits on Apr 22, 2016

Commits on Apr 23, 2016

Commits on Apr 24, 2016

Commits on Apr 25, 2016

Commits on Apr 26, 2016

Commits on Apr 27, 2016

Commits on Apr 28, 2016

Commits on Apr 29, 2016

Commits on Apr 30, 2016

Commits on May 1, 2016

Commits on May 2, 2016

Commits on May 3, 2016

Commits on May 4, 2016

Commits on May 5, 2016

Commits on May 6, 2016

Commits on May 7, 2016

Commits on May 9, 2016

Commits on May 10, 2016

Commits on May 11, 2016

Commits on May 12, 2016

Commits on May 13, 2016

Commits on May 14, 2016

Commits on May 15, 2016

Commits on May 16, 2016

Commits on May 17, 2016

Commits on May 18, 2016

Commits on May 19, 2016

Commits on May 20, 2016

Commits on May 21, 2016