9ffa9df9c6
Add transfer in getoutput
2015-07-21 06:01:10 -07:00
ae76821812
Set shuffle always true, set m_evalMode under old version
2015-07-21 05:35:16 -07:00
3e46e6e78d
fix dataend issue, and support old-version model for test
...
modified: DataReader/LMSequenceReader/SequenceReader.cpp
modified: MachineLearning/CNTK/TrainingCriterionNodes.h
2015-07-20 20:11:52 -07:00
75acf2b21e
fix some float-int conversion warnings.
2015-07-20 17:28:03 -07:00
c74cdb58f9
Merge branch 'master' of https://git01.codeplex.com/cntk into linux-gcc
2015-07-20 16:57:26 -07:00
7e09cfae6c
Restore the SetSentenceSegBatch call accidentally removed in the prefetch undo
2015-07-20 15:21:35 -07:00
f33224b60e
Undo prefetch
2015-07-20 14:10:19 -07:00
3cbf7fe668
Disable prefetching by default as there are issues in case of RNNs
2015-07-20 14:05:49 -07:00
9bb8d4df3a
minor changes based on code review
2015-07-20 12:50:47 -07:00
45e79ad6d8
fix problems in reader when files have duration mismatch
2015-07-20 12:50:45 -07:00
a054e65894
Fix MPI_SUPPORT build
2015-07-20 10:43:47 -07:00
22dce6513a
Added /bigobj option to CNTKEval.vcxproj to address the 'too many sections' compiltation error in that project
2015-07-19 12:02:30 -07:00
0b78877c77
Merge branch 'master' of https://git01.codeplex.com/cntk into bmitra/Changes
2015-07-19 16:18:55 +10:00
cd34812557
CPU Sparse copy constructor.
2015-07-19 16:18:35 +10:00
f50f0cab50
handle the condition where sentence_start and nolabel happens at the same time.
2015-07-18 18:03:10 -07:00
222e703826
Merge remote-tracking branch 'origin/chenguoguo/kaldi2reader-multi-utterance'
2015-07-19 00:52:33 +00:00
d1e14783b6
changed feature packing flag names from utteranceXYZ and sentence_XYZ to SequenceXYZ to make them consistent and be understood for both speech and text people.
...
Added flag to indicate the condition of no feature (with assumption that there is also no label) to support efficient bi-directional model training.
Updated the windows version of HTKMLFReader, LM and LU Sequence readers.
Caution: this change will not allow for using past and future label information in the recurrent networks. The ultimate solution is to have a flag for each feature read-in. However, this will require huge change and we will do this after reader refactorization to reduce effort.
2015-07-18 17:28:14 -07:00
219ee93501
Merge remote-tracking branch 'origin/linux-gcc' into chenguoguo/kaldi2reader-multi-utterance
2015-07-18 22:34:53 +00:00
8d21003325
Merge remote-tracking branch 'origin/master' into chenguoguo/kaldi2reader-multi-utterance
2015-07-18 20:30:01 +00:00
8830b5a0f9
Merge remote-tracking branch 'origin/master' into linux-gcc
2015-07-18 03:42:51 -04:00
fa3fb05cfc
Fix linux builds.
2015-07-18 00:47:21 -04:00
1726a5944e
Merge branch 'amitaga/prefetcher_packingmask'
2015-07-17 20:40:16 -07:00
9dcc79948d
Fixed the issue of minbatch packing mask not being correctly handled in the minibatch prefetching code
2015-07-17 18:26:38 -07:00
82c7d8cf91
Merge remote-tracking branch 'origin/master' into linux-gcc
2015-07-18 00:32:21 +00:00
967c6babbf
Merge branch 'master' to 'linux-gcc'
2015-07-18 00:30:07 +00:00
bbf49009d1
Merge branch 'master' to 'linux-gcc'
2015-07-17 23:42:22 +00:00
329d3bc6fd
Fix the bug for print correct momentum value.
2015-07-17 19:26:24 -04:00
b5b1c13d68
Fix error when setting perMB momentum.
2015-07-17 17:48:03 -04:00
d54e2d30fe
Adding multi-utterance support in Kalsi2Reader for Truncated == false
2015-07-17 21:25:10 +00:00
4827a728e9
Keep m_SentenceBoundary always in CPU memory; Make CNTKEval compatible with the current SetenceBoundary setting change.
2015-07-16 21:32:00 -07:00
9bed89797a
Fix boundary setting when frameMode=true (For HTKMLFReader_linux); also format the code.
2015-07-16 19:03:42 -04:00
0e65b1f9fd
Fix the boundary setting when framemode = true in HTKMLFReader.
2015-07-16 15:44:05 -07:00
e63927924f
1) Removed iscriterion from checking in forming recurrent loops, because of using a function to check duplicated loops. 2) Check if pairnodes are existing and save them if they are. This aims at backward support of the previous models and network file format.
2015-07-16 12:13:19 -07:00
6a1976895d
Minor reader changes.
2015-07-17 02:48:42 +10:00
5a98803ca4
Bug: MinibatchPrefetcher is not setting matrix type/size.
2015-07-17 02:47:29 +10:00
34b62d10d9
Simplify sigmoid kernel
2015-07-16 09:18:36 -07:00
fdd07c1be5
Make sure Linux and Windows use the same integer type in CUDA kernels
2015-07-16 09:08:22 -07:00
38db72b6ae
Use a function to merge duplicated recurrent loops according to their source node. Removed unused functions in MultiNetworkSGD
2015-07-16 08:59:23 -07:00
65f38063bc
Merge branch 'master' of https://git01.codeplex.com/cntk
2015-07-16 07:36:46 -07:00
fd11f339cd
update g2p bidirectional setup
2015-07-16 07:36:34 -07:00
a356380ada
Merge branch 'master' of https://git01.codeplex.com/cntk into bmitra/Changes
2015-07-16 18:23:21 +10:00
02bb921723
Updated CNTKbook with encoder/decoder framework (first draft)
2015-07-16 00:03:44 -07:00
717aec01e3
Use digit to repsent different networks.
2015-07-16 00:03:06 -07:00
d59b5f37a9
Forming loop using SCC uses isCriterion to true for backward compatibility. Also with this change can create recurrent networks . Update enumeratenodesforeval in pairnodes to ignore its children when search for nodes to be built for a network.
2015-07-16 00:01:11 -07:00
06f3e2de7b
Merge branch 'master' of https://git01.codeplex.com/cntk into bmitra/Changes
2015-07-16 08:42:39 +10:00
f48425b67c
Minor bug fixes.
2015-07-16 08:41:07 +10:00
58855e8fd9
Add returnDense flag.
2015-07-16 08:37:23 +10:00
7a894210bc
MultiNetworkSGD fix linux build
2015-07-15 13:02:10 -07:00
e298bdd00d
fix linux build error and warnings
2015-07-15 12:47:08 -07:00
1fb5918b13
Implement minibatch prefetching
...
This change implements a read-ahead (prefetch) of one minibatch on a separate thread to ensure that main thread always has enough work. This is done through coarse-level parallelism, where GetMinibatch is called from a separate thread and result is cached for main thread to consume. The synchronization is a simple producer/consumer model with a single mutex and a conditional variable.
This change also modifies how we measure time it takes to read and compute a minibatch worth of data.
2015-07-15 11:57:06 -07:00
Yinggong ZHAO
Dong Yu
Amit Agarwal
Marko Radmilac
Mike Seltzer
UnderdogGeek
chenguoguo
Yu
erw
Yu Zhang
kaisheny