Eldar Akchurin
f3289b0f84
First version of distributed aggregator for v2
2016-10-19 13:49:10 +02:00
yuxiao.guo
8c360f27d9
Merge branch 'master' into v-yuxgu/resnet-rc
2016-10-18 15:02:13 +08:00
yuxiao.guo
9cb329a0da
Optimize the flow of post batch normalization statistics, and allow disable regularization terms in batch normalization
2016-10-18 13:53:36 +08:00
Alexey Reznichenko
e0fea5b687
Auto-enable distributed mb reading and parallel train
...
* the default value for parallelTrain must automatically be true
when running under MPI with >1 node, else false
* same for distributedReading when running with a new (v2) reader
2016-10-02 20:18:50 +01:00
Alexey Reznichenko
1dd9216db0
In Eval and Write, randomize should default to false
2016-09-21 18:25:42 -07:00
Mark Hillebrand
74611be601
Fix path differing only in case. Fix hard tabs.
2016-09-19 09:13:31 +02:00
Frank Seide
53fb3a7736
merged from master;
...
added a few TODO comments
2016-09-08 15:28:22 -07:00
yuxiao.guo
64b407e406
using space instead of tab
2016-09-05 15:19:38 +08:00
Frank Seide
35c44227cb
bug fixes: traceLevel parameters must be treated as optional;
...
fixed a left-over %zd
2016-09-04 15:56:55 -07:00
yuxiao.guo
d0016cd035
add support of multi-gpu on PBN
2016-09-01 11:11:59 +08:00
yuxiaoguo
a3a5d2f770
add export path & add BN weight decay disable
2016-08-26 17:02:33 +08:00
yuxiaoguo
7e09a99ce8
add post batch normalization process
2016-08-23 15:08:40 +08:00
Frank Seide
3bf23b386a
temporarily hiding BrainScript in MNIST 01_ example
2016-06-08 19:10:12 -07:00
Frank Seide
f50f6a78e8
changed 01_OneHidden to BrainScript & dealt with fallout
2016-06-06 18:22:10 -07:00
Alexey Reznichenko
b42e1967b1
Fix the second SimpleEvaluator ctor call
...
Close #485 .
2016-06-06 10:14:01 +02:00
Alexey Reznichenko
02216f2621
Fix a bug in DoEvalBase
...
Add a missing argument to SimpleEvaluator ctor call, so
that all parameters are aligned with proper values.
2016-06-02 13:46:10 +02:00
Amit Agarwal
08747ebdab
Fix distributed reading handling in the Eval code
2016-04-25 19:13:47 -07:00
Frank Seide
4c9f91868e
merged from master and resolved conflicts
2016-04-16 14:29:11 -07:00
Frank Seide
1224dfc9bc
refactored criterion reporting in SGD and SimpleEvaluator to carry a per-criterion-node sample count, in preparation for supporting criterion nodes with different sequence lengths for seq-2-seq
2016-04-12 23:09:08 -07:00
Alexey Reznichenko
9fe696f556
Merge branch 'master' into deepbliscore/vnext
...
Conflicts:
.gitignore
Source/ComputationNetworkLib/ComputationNetwork.h
Source/ComputationNetworkLib/ComputationNetworkEvaluation.cpp
Source/ComputationNetworkLib/ReshapingNodes.cpp
Source/SGDLib/DataReaderHelpers.h
Tests/UnitTests/ReaderTests/Common/ReaderTestHelper.h
2016-04-06 17:28:14 +02:00
Frank Seide
531f5538b6
updated seq-2-seq decoder to match updated training;
...
new operation Trace() (TraceNode) as a debugging aid;
WriteMinibatchWithFormatting() can now log by FrameRange;
WriteFormattingOptions now in ComputationNode.h
2016-04-02 23:29:40 -07:00
Frank Seide
539f6e63fd
added node tracing to 'eval' and 'write', and added a sparse option
2016-03-29 15:32:54 -07:00
Frank Seide
d5271b352e
merged with master;
...
restored right 1-bit submodule version
2016-03-24 10:14:03 -07:00
Clemens Marschner
70f1954268
Enable testing of backward pass in write action
...
to enable, set nodeUnitTest to true inside write action block
- Relax requirement of 1x1 matrix on output node for that. It will just
match the output matrix and do a backprop of 1.
- introduce Identity() node (right now only in C++ API and NDL) to work
around the fact that inputs don't get gradients updated.
- cherry-picked from clemensm/GradientTest. Need some more changes to
integrate with Frank's latest refactorings
2016-03-22 18:16:16 +01:00
Frank Seide
76543e8708
bug fix: InferMBLayoutFromInputsForStandardCase() must test isFinalValidationPass
2016-03-21 21:34:41 -07:00
Clemens Marschner
bcbdd2ca63
Enable testing of backward pass in write action
...
to enable, set nodeUnitTest to true inside write action block
- Relax requirement of 1x1 matrix on output node for that. It will just
match the output matrix and do a backprop of 1.
- introduce Identity() node (right now only in C++ API and NDL) to work
around the fact that inputs don't get gradients updated.
- cherry-picked from clemensm/GradientTest. Need some more changes to
integrate with Frank's latest refactorings
2016-03-21 15:44:51 +01:00
Alexey Reznichenko
ed2831ebb5
Merge complete (merged with master)
2016-03-18 16:07:35 +01:00
Alexey Reznichenko
5073754aae
Fix write mode (add layouts to StreamMinibatchInputs, disable traceLevel overriding)
2016-03-18 14:08:52 +01:00
Amit Agarwal
0339ce7a24
Merge branch 'master' of https://github.com/Microsoft/CNTK into amitaga/parallelizationFix
2016-03-14 08:54:54 -07:00
Gaizka Navarro
6d57746934
Applies CR feedback
...
Cleanup
2016-03-14 09:55:08 +01:00
Gaizka Navarro
6603f55e32
Added GetModelFromConfig to encapsulate the logic for deciding to either load the model or create the network
2016-03-14 09:55:04 +01:00
Gaizka Navarro
d238daaccd
Relocates NetworkFactory header content to inside Actions.h and removes the NetworkFactory.h file
...
Assigns the correct build configuration to the solution platform for the NetworkTests project
2016-03-14 09:54:59 +01:00
Gaizka Navarro
e2f31ff658
Addresses CR : Rename NetworkCreator to NetworkFactory
2016-03-14 09:54:57 +01:00
Gaizka Navarro
f2c6ad01fa
Implements building the network from the network definition and evaluating the network through the "write" command without the need of a model.
2016-03-14 09:54:29 +01:00
Amit Agarwal
877725d15e
Bugfix: Disable parallelization for commands other than ones for train, test and cv
2016-03-12 00:04:52 -08:00
Wayne Xiong
b6e5992676
Fix the build error after merge.
2016-03-03 11:26:00 -08:00
Wayne Xiong
b2b669111c
Apply code review feedback
2016-03-03 11:25:58 -08:00
Wayne Xiong
b0c3c5f7cf
Make the cross validation part supports multiple GPUs
2016-03-03 11:25:57 -08:00
Frank Seide
e54b352822
removed template parameter ElemType from (I)DataReader and (I)DataWriter
2016-02-28 19:01:07 -08:00
Frank Seide
26c850322b
bug fix in Transpose();
...
prepared file format for RowStack and RowSlice to store their dimension along which they operate (currently limited to 1);
bug fix in "write" command "format" parsing;
ComputationNetwork::Read() now checks for too new file version
2016-02-23 11:14:36 -08:00
Frank Seide
d4a251c4fb
made gcc happy
2016-02-15 12:47:08 -08:00
Frank Seide
343893e882
"write" command can now apply a labelMappingFile to print classification results as text labels
2016-02-15 11:45:54 -08:00
Frank Seide
6ea2892967
"write" command now accepts to say type="category"
2016-02-15 11:03:53 -08:00
Frank Seide
27820eb9f0
added formatting options to the "write" command, and allow pipes and "-" as the output name;
...
bug fix in FlattenInPlace()
2016-02-15 10:43:03 -08:00
thhoens
a930d1cfb4
Updated EvalActions to fix support for OutputNodeNames in the config.
2016-02-04 13:11:13 -08:00
Frank Seide
a4cf44544a
normalized inconsistent commenting style from "//text" to "// text"
2016-01-22 13:58:47 -08:00
Frank Seide
a4a20183aa
made everything build again after deleting the unused MultiNetworks functions, incl. deleting related entries in SimpleNetworkBuilder;
...
deleted PairNetworkNode;
renamed EsotericActions.cpp to SpecialPurposeActions.cpp
2016-01-22 09:13:52 -08:00
Mark Hillebrand
dcfc0e9679
License change
2016-01-18 09:36:17 +01:00
Mark Hillebrand
2f39fe3f01
Re-format code using clang-format (plus some post-processing)
2016-01-18 09:36:14 +01:00
Frank Seide
08e7d59059
moved all actions (DoXXX()) from CNTK.cpp to ActionsLib (no code change);
...
reenabled tensor lib (undid accidental commit)
2016-01-09 18:13:31 -08:00