73e20eaac2
Allow named arguments to take non literal values in macros.
2015-01-29 10:22:07 -08:00
ac50417b75
fix the build problem in the debug build of CNTKEval project.
2015-01-28 23:21:59 -08:00
0cbab9b3e7
move TimerUtilities.h to common\include. Add TimerUtilities.cpp to CNTKEval project.
2015-01-28 15:59:31 -08:00
97a317e861
force gradient to be sparse if input is sparse in the TimesNode. This is the only dense X sparse combination we suppport right now.
2015-01-28 11:18:16 -08:00
5fd74fe9a3
fill constant node's value in second pass if the size of the constant is known by then.
2015-01-27 15:38:06 -08:00
7baca24d75
changed timing function.
2015-01-27 14:35:20 -08:00
8fbcf8f5b7
added functionality for assymmetric context windows
2015-01-27 09:53:42 -08:00
f9c555830b
added functionality for asymmetric context windows
2015-01-27 09:47:26 -08:00
8c61da007a
Merge branch 'master' of https://git01.codeplex.com/cntk
2015-01-23 15:30:23 -08:00
337033f65d
Change FindSymbol to fix a MEL bug that occurs when using NDL command with dotted names.
2015-01-23 15:30:06 -08:00
8b8f20138b
Fix the bug during loadCheckPointFile. It try to putMarker in the old file.
2015-01-21 23:03:28 -05:00
032a0e8ede
Merge branch 'master' of https://git01.codeplex.com/cntk
2015-01-21 18:25:58 -08:00
6e1876111c
Change Sparse Matrix implementation (esp. GPU Sparse matrix) to make it work for LM training and other sparse features.
...
Note: still many improvements are needed for sparse matrices.
2015-01-21 18:25:41 -08:00
dc6164c06a
Merge branch 'master' of https://git01.codeplex.com/cntk
2015-01-19 16:30:33 -05:00
0879cfef56
Fix the bug for MEL when adding a recurrent loop.
2015-01-19 16:30:25 -05:00
08a9c5e946
Add plumbing for ExpNode in NDL
2015-01-19 09:27:41 -08:00
43ea68b59e
change SparseInputValue node and CPU Sparse matrix to make LM CPU training work.
2015-01-17 13:37:00 -08:00
459ff2ded1
Merge branch 'master' of https://git01.codeplex.com/cntk
2015-01-14 22:51:55 -05:00
97d5fb415a
FIx the bug when Dropout node in recurrent loop. Bug: the funcationValues is unintialized when dropOutNode in a recurrent loop.
2015-01-14 22:23:57 -05:00
eafc86daab
MPI Model Averaging. Added check to short-circuit empty minibatch
2015-01-14 18:51:05 -08:00
3e68d18ad8
Improvements to MPI Model Averaging.
...
DecimateMinibatch now checks that all input matrices have the same number of columns, and throws an exception if they do not.
DecimateMinibatch now checks for empty input matricies (after decimation).
The reduction/averaging code now weights each model by the relative number of samples in it's effective minibatch.
TrainOneEpoch now returns the local value totalEpochSamples, so the proper weighting can be performed during model updating.
2015-01-14 18:09:39 -08:00
9e37c72050
Changed Deviceid=auto to be MPI model-averaging compatable, no more device contention.
2015-01-14 18:09:38 -08:00
6f21c5b388
Added debug output
2015-01-14 18:09:37 -08:00
bd5524720d
Added a check in HTKMLFReader and throw exception if Truncated is false while nbrUttsInEachRecurrentIter is not 1.
2015-01-09 23:09:34 -08:00
4ff3ab53f3
Add SetNZCount to OmmonMatrix to support setting number of Non-zero values externally.
...
Change the Resize function in sparse matrices to make it clear that the numNZ is used to reserve memory only so that we can call Resize repeatedly without affecting the actual number of non-zero values.
Change the cpu sparse matrix's Resize to support keeping existing values when memory is reallocated.
Change SetValue function in cpu sparsematrix to support automatic resizing.
2015-01-09 00:19:17 -08:00
751b1cd744
unify the dense and sparse matrix Resize surrogate at the Matrix class level.
...
fix the bug of the sparse matrix CPU-GPU transfer flag.
change the SetValue(idx, j, v) so that it uses CPU when both CPU and GPU are available.
2015-01-07 23:12:56 -08:00
144db5b70d
bug fix by AdamE in QueryNvmlData(), to skip a check that was found to sometimes fail
2015-01-06 15:27:29 -08:00
6bd6dbeb47
Changes in the CNTK book:
...
1. corrected the parameter list for the Delay command in NDL.
2. added the options for the rmprop gradient update type.
3. added the ConvertDBN command.
2015-01-04 23:54:16 -08:00
e773dc6264
Fix the error during validation stage. Bug: when the anchor node is not the root in a strong component connection, the post visited order may initialize the forward computation incorrect.
2015-01-03 01:55:27 -05:00
12b8ec6e0f
Adding GetFormatString() for char* and wchar_t* in fileutil
2014-12-21 19:48:55 +08:00
2c29a69661
add the option dumpFileName to the NDLNetworkBuilder block so that before validation happens the network structure information can be saved to the file specified by this option. This will help to debug the NDL related issues.
2014-12-20 02:15:09 -08:00
909e8adfb0
minor typo fixes in the CNTK book.
2014-12-11 23:07:15 -08:00
091a785d1f
update the CNTK book to include the learnRateAdjustInterval option
2014-12-10 00:14:23 -08:00
c897f83f70
Added learnRateAdjustInterval option to the automatic learning rate control block for SGD. It allows users to determine the learning rate adjustment frequency when the learning rate is controlled, for example,by a dev set.
2014-12-10 00:06:20 -08:00
3f92c98bfb
changed the implementation of _matrixVectorColumnWiseAdd so that it uses more GPU threads but more memory access as well. Overall the speed is comparable with original implementation for normal usage but significantly faster for CNN bias.
2014-12-09 01:52:37 -08:00
7f8315c8aa
"Correct Fix for bugs in SLU example in CNTKBook"
...
This reverts commit 5e26505d0d
2014-12-04 16:23:14 -08:00
5e26505d0d
Revert "Fixed bugs in SLU example in CNTKBook (section 7.4)"
...
This reverts commit 435d6127d5
2014-12-04 10:29:31 -08:00
435d6127d5
Fixed bugs in SLU example in CNTKBook (section 7.4)
...
1. Problem: RuntimeError if the full path of the output file doesn’t
exist. Fix: make the intermediate dirs for the path of the output file.
2. Problem: Output labels file not written (should be written according
to the documentation). Fix: change the rnnlu.config file.
3. Problem: mismatch between the feature and the output files when
building the evaluation file. Fix: skip empty lines in the feature file
2014-12-04 10:18:52 -08:00
1645818816
Merge branch 'master' of https://git01.codeplex.com/cntk
2014-12-02 20:03:09 -08:00
7a9bb9a64d
fixed the gradient computation bug in CosDistanceNode.
2014-12-02 20:02:49 -08:00
61f37d4ca0
fixed an incorrect argument to a printf format string
2014-12-02 17:24:28 -08:00
89d12a378e
Merge branch 'master' of https://git01.codeplex.com/cntk
2014-12-02 17:01:02 -08:00
1bcb5b7d8f
ACML DLL is now delay-loaded, which is required to be able to control the number of threads of the ACML library (set OMP_NUM_THREADS before first use, which then seems to get picked up upon DLL load)
2014-12-02 17:00:50 -08:00
c157083384
Merge branch 'master' of https://git01.codeplex.com/cntk
2014-12-02 14:42:30 -08:00
737fd2a964
fixed DumpNodeInfo bug
2014-12-02 14:41:39 -08:00
5438ed416d
Fix compilation errors when flag NANCHECK opened
2014-12-02 02:05:07 -08:00
286b109087
change Matrix::SetValue to include matrixFormat.
2014-11-29 21:17:27 -08:00
25107d964f
Merge branch 'master' of https://git01.codeplex.com/cntk into localdev
2014-11-29 19:19:51 -08:00
a56f8ef5aa
Significant improvement to GPUSparseMatrix and CPUSparseMatrix. Also makes device id type consistent across classes.
2014-11-29 19:18:58 -08:00
cdf9cd6d24
remove temp lyx files for CNTK book.
2014-11-25 17:54:52 -08:00
Dong Yu
Mike Seltzer
yzhang87
Jasha Droppo
Frank Seide
fyc0624@gmail.com
guhernan
fcb1899