Skip to content
Snippets Groups Projects
Select Git revision
  • master default protected
1 result
You can move around the graph by using the arrow keys.
Created with Raphaël 2.2.011Sep1098765432131Aug302928272624232221201917161514131110986542131Jul3029272625242322212019181716141211109765432130Jun2928272625242120191817161514137654230May2928252423222120191817161514111098630Apr292827252423222119181716141312111098765432131Mar30292827262524232120191817161514131211109875432sandbox/dan2: wrote comments for Splice and Copy in cu-math.h, still need to finish unit tests for those functionssandbox/dan2: Fixed a bug in AddMatMat function in cu-matrix.cc, which doesn't check 0-dimensionality for CPU caseAdding description for CompObjfAndDerivsandbox/dan2: CUDA-lize calculation in NnetUpdater::ComputeObjfAndDeriv() function. Ugly pointers are replaced by cleaner codestrunk/nnet : fixing memory leak in destructor of CuRand (big thanks goes to Dan)(trunk/egs/wsj) Fix the train_sgmm2_group.sh to correctly handle the ci_phones optionstrunk:egs/hkust Update with GPU-based DNN experimentstrunk:egs/hkust Update with GPU-based DNN experiments and resultssandbox/dan2: CUDA-lize calculation in NnetUpdater::ComputeObjfAndDeriv() function. Documentation and Tests added. There are a few lines of REALLY ugly codes (lots of pointers) to be changedsandbox/dan2: Rework CuBlockMatrix to make initialization more efficient; add memory checking code; fix memory leaks in packed-matrix copy; compilation fix in nnet2/nnet-update.ccsandbox/dan2: CUDA-lize calculation in NnetUpdater::ComputeObjfAndDeriv() function. Documentation and Tests coming soon...sandbox/dan2: Committing initial version of CuBlockMatrix code (for use in MixtureProbComponent)(BABEL): follow the release of the IndusDB for the new langs (102 and 103), fixes, include example configs with FFV featurestrunk/nnetbin : Introducing length constrain of 1 min on segments that are used for mini-batch frame cross-entropy DNN training. This will avoid possible out-of-memory issues by discarding the very long segments which are infrequent.sandbon/dan2: Added documentation to Randomize in cu-math.hfix a bug - the last argument was checked by ClassifyRSpecifier, but it should be the first argument, since the last is a W specifierfix to babel/s5/local/make_ffv.shsandbox/dan2: rework of CuStlVector.sandbox/dan2: cudamatrix bug-fixes and more tests; modifying scripts to use single-threaded program for GPU (multi-threaded code is not easily compatible with CUDA).sandbox/dan2: add test for cuda-matrix copy across types; fix bugs in cross-type cuda-matrix copying, and streamline that code.sandbox/dan2: minor improvements to the code; bug fixes in matrix copying RE zero dimsandbox/dan2: add speed test for Sum() of CuVectortrunk: small fix to nnet training script (remove call to deprecated script)sandbox/dan2: various cudamatrix optimizations and fixes. neural net training job now running with cudamatrix (training script not yet fully tested).sandbox/dan2: fixes to tests for random-number generators, cosmetic changes; mov cu-rand-inl.h to cu-rand.cc.sandbox/dan2: Adding SetRandUniform() function for matrix; adding tests for SetRandn() and SetRandUniform(), cudamatrix SetRandUniform() test failing currently.sandbox/dan2: Added efficient CUDA implementation for one case of CopyRowsFromVec.trunk: code change that should fix memory leak in bin/latgen-faster-mapped-parallel [fix other branch of the code, omitted first time]trunk: code change that should fix memory leak in bin/latgen-faster-mapped-parallelsandbox/dan2: Fixing more errors; nnet-component-test now passes.sandbox/dan2: various fixes.sandbox/dan2: Fixes and extension to CopyRows()/CopyCols() for matrix/cudamatrixsandbox/dan2: some fairly minor changes.sandbox/dan2: fixing the bug where src.dim() is passed into cuda kernel instead of this->dim()sandbox/dan2: cudamatrix bug fix; improve test programs to test with and without GPU (would have caught the bug I just fixed).sandbox/dan2: Various cudamatrix extensions. nnet-component test code currently failing.sandbox/dan2:add cu-vector-speed-test.ccopt softmax in cu-vectortrunk/nnet : alignment script can optionally generate output in lattice format (handy for getting word alignments)sandbox/dan2: fix the previous equal-stride constraints. Not completed yet, might be bugs