f7865837e3
Merge branch 'master' of https://github.com/Microsoft/CNTK into fseide/s2s
2016-03-01 13:32:44 -08:00
673506e8ca
some clean-up of sparse matrices;
...
disabled meaningless elementwise sparse ops such as log();
sparse matrice slices now keep track of their original matrix (not used yet);
bug fix: ~SyncGuard() should not swallow errors since CUDA errors may get discovered here;
bug fix: HTKMLFReader test 16 must use consistent precision
2016-03-01 10:43:50 -08:00
2d0c7eff3c
Merge branch 'master' of https://github.com/Microsoft/CNTK into amitaga/meAllocFix
2016-03-01 00:15:52 -08:00
5304784b6c
fixed NoGPU build re do_sync;
...
updated ExperimentalHTKMLFReader w.r.t. file naming
2016-02-29 21:27:49 -08:00
2b48da7de9
changed do_sync mechanism to an RAII class
2016-02-29 20:51:43 -08:00
eb98d333c2
removed <ElemType> from ExperimentalHTKMLFReader
2016-02-29 19:27:52 -08:00
7d3e84e773
merged from master
2016-02-29 19:17:35 -08:00
7f7fce4c6e
RNN.cntk "write" example now writes out the log sentence probabilities, which is more useful, e.g. for N-best rescoring;
...
bug fix: SimpleOutputWriter must handle no-MBLayout case;
bug fix to LaunchUnaryTensorOp(), must free the CUDA event or else leak memory--thanks to Yu Zhang for tracking it down!
2016-02-29 18:16:55 -08:00
eaa1dd0791
Added support for cuDNN v4 prod.
2016-02-29 14:57:30 -08:00
eeaf79103c
Addressed review comments.
2016-02-29 14:09:28 -08:00
9db0cd375d
Fixed gcc build issues.
2016-02-29 14:09:18 -08:00
834fab956c
Convolution engine refactoring. Preparing for ConvolveGeometry.
2016-02-29 14:09:07 -08:00
d0aa673ea7
Updated image samples.
2016-02-29 14:08:57 -08:00
daddc24c2e
Updated samples that use batch norm. Fixed convo auto-tuner issue.
2016-02-29 14:08:46 -08:00
dbb8b3f918
Added time const to batch norm.
2016-02-29 14:08:25 -08:00
e886a4dfed
Added CPU BN inference implementation.
2016-02-29 14:08:15 -08:00
9bb0fd9d7b
Integrate 367db5e into master
2016-02-29 13:53:40 -08:00
aa141188c2
more detailed logging to GetInputMatrix()
2016-02-29 12:07:45 -08:00
67f12739bc
Fixed Issue #176 : random number used for random_shuffle() had an int overflow on gcc, which caused uncaught out-of-bounds accesses.
2016-02-29 11:58:13 -08:00
a699567ce5
Integrate mahilleb/HtkmlfProtoSquashedAgain into master
2016-02-29 11:43:12 -08:00
c2167a3eeb
Small fix of mblayout copy
2016-02-29 17:23:30 +01:00
4d5566d5ac
Experimental HTKMLF Reader (currently only supports frame mode)
2016-02-29 15:54:39 +01:00
e54b352822
removed template parameter ElemType from (I)DataReader and (I)DataWriter
2016-02-28 19:01:07 -08:00
81fa019c22
Fixed a memory allocation bug in GPUSparseMatrix
2016-02-28 17:58:03 -08:00
5bdc86e5de
removed <ElemType> parameter from StreamMinibatchInputs
2016-02-28 17:04:45 -08:00
6249935000
StreamMinibatchInputs now stores shared_ptrs, simplifying ownership;
...
UCIFastReader::m_prefetchMatrices changed to be StreamMinibatchInputs as well
2016-02-28 15:05:14 -08:00
0a0b388b3d
replaced all calls to GetInputMatrixPtr() by GetInputMatrix()
2016-02-28 13:52:49 -08:00
f3a720e5ce
replaced StreamMinibatchInputs::operator[] by named searchable accessor functions
2016-02-28 13:34:02 -08:00
24c88c04fd
abstracted the map<wstring, Matrix*> into a class, for future enrichment
2016-02-28 12:25:36 -08:00
ea8babf571
updated rnn.cntk config
2016-02-28 11:40:38 -08:00
20d5f02e89
implemented make mode in "writeWordAndClass" action
2016-02-28 11:13:02 -08:00
ddeb324bb7
Implement an option to explicitly exclude specific devices from deviceId=auto selection algorithm
2016-02-28 01:27:31 -08:00
12d79a420e
added "write" command to rnn.cnkt;
...
bug fix in ClassBasedCrossEntropyWithSoftmaxNode, a wrong index if #parseq > 1;
bug fix in TransposeTime dimension inference;
bug fix in simple writer, must call DataEnd()
2016-02-27 23:52:27 -08:00
570b6a8f7b
mini checkin
2016-02-27 22:23:22 -08:00
d55374d199
some clean-up of RNN sample and DoWriteWordAndClassInfo()
2016-02-27 18:15:35 -08:00
e514ae0a93
merged from origin/master
2016-02-27 00:04:51 -08:00
20d1f3f803
bug fix: InitRandom() must use GetNumCols() for inputSize
2016-02-26 23:40:08 -08:00
123f94c49b
(refined an error message)
2016-02-26 23:23:18 -08:00
15a50e12b9
updated AlexNet to work with tensor Times
2016-02-26 22:39:43 -08:00
1cc857de0f
MBLayout support for "write" command;
...
added sample to simple2d for "write"
2016-02-26 17:27:27 -08:00
1a618bc708
bug fix: ImageParameter must understand >3-dim tensors, for Times operation
2016-02-26 16:25:21 -08:00
33c7c10d09
gcc 4.8.4 does not implement regex correctly. Replace the HTKMLFReader uses.
2016-02-26 16:06:10 -08:00
81884e31db
fixed 02_Convolutional as well for new tensor TimesNode
2016-02-26 14:14:43 -08:00
10c4b7d3de
fixed Image/QuickE2E for new TimesNode
2016-02-26 13:45:18 -08:00
3011bcf76c
made gcc happy
2016-02-25 16:19:19 -08:00
94932b2668
TimesNode now supports tensors
2016-02-25 15:44:49 -08:00
abdc9d3068
Remove FUSE workaround now that FUSE is updated
2016-02-25 13:30:13 -08:00
246907dbb6
Addressed code review comments.
2016-02-25 10:56:42 -08:00
e982a6748e
Enabled ImageReader unit tests in Jenkins.
2016-02-25 09:47:14 -08:00
467520a851
Fixed no libzip build.
2016-02-25 09:47:13 -08:00
U-FAREAST\fseide
Frank Seide
Amit Agarwal
Alexey Kamenev
Alexey Kamenev
Project Philly
Eldar Akchurin
Mark Hillebrand
Chris Basoglu