GitLab

- 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@4861 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk,nnet1: some cosmetic changes, adding stubs of minor experimental features (block-softmax,frame-pooling) 



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4472 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk,nnet1 : changing CuMatrix -> CuMatrixBase and CuVector -> CuVectorBase in interfaces, where applicable. Note the 'Base' classes mean that the dimensions are fixed and object has pre-allocated data.



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4272 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4082 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

nnet,trunk: adding linear transform without bias (for bottlenecks), each component can rescale learning-rate


git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@4004 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/sandbox/language_id@3586 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

- removing 'Cache' class, replaced by 'Randomizer'
- removing obsolete training tools using 'Cache' class
- adding 2d convolutional components from Harish
- removing scaled inv-prior-weighted cross-entropy Class, not helpful




git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3551 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/sandbox/tacc_kaldi@3477 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk,nnet: updating training scripts (now using prototypes, initializing with 'nnet-initialize'; NN training target represented as 'Posterior'), preparing NNs to allow parallel nested networks.



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3415 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk,nnet: committing latest nnet code, preparation for CNNs, a binary tool for initialization (will be used in examples later)



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3402 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3241 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/sandbox/dan2@3172 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

- new per-frame randomizers (future plan to remove Cache, and use lightweight randomizers instead, already in rbm-train-cd1-frmshuff and nnet-train-frmshuff)
- nnet training tools added verbose level3 for analyzing graidents, or feedforward data
- cross entropy loss now has progressive reporting 



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3171 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

sandbox/dan2: merge various changes from trunk, plus some changes to enable efficient GPU-based training after 'mix-up'.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/sandbox/dan2@3108 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk,nnet1: warning! changes to Component/Nnet interaface to be closer to Dan's, you may need to fix your extensions to nnet1 code, main changes are:
- nnet files now have optional begin/end tokens <Nnet></Nnet> (backward compatibility)
- in the Nnet API, "*Layer" was replaced by "*Component", which better corresponds to code
- removed *Nnet pointer from Component and constructors, it was unused
- addded Component::Copy() method, simplified nnet consistency checks Nnet::Check()
- fixed "Copyright" notices



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3102 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

Update COPYING notice to further clarify our conventions regarding copyright and multiple authors; update Apache headers to point to COPYING notice.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@2988 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@2703 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@2558 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@2557 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@2556 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

Committing a lot of minor changes to Ondra's sandbox (will soon merge to trunk, I hope): fix issues with creating dynamic libraries by reorganizing some code and moving some files; fix a lot of minor issues such as wrong usage messages and include guards; adding Ondra's name to copyright on files he modified.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/sandbox/oplatek@2533 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

- NN-training hyper-parameters now in classes NnetTrainOptions RbmTrainOptions
- removing nnet-trim-n-last-transforms (functionality now in nnet-copy --remove-last-layers)



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@2504 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/sandbox/dan@1708 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@1686 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

fixing the GPU build process which now uses multiple CUDA compute capabilities



git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@1600 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@985 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@983 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@982 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@965 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@962 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@943 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/sandbox/karel@765 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@761 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@736 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

the CUDA toolkit is automatically located



git-svn-id: https://svn.code.sf.net/p/kaldi/code/sandbox/karel@624 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/sandbox/karel@277 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/sandbox/karel@275 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@52 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8