c7c2547f55
CNTK v2 library: Fixed handling of references of existing netowrk matrix storage handed out from Forward/Backward
2017-04-14 01:33:24 -07:00
1f23f6e161
CNTK v2 library: Add ability to register and instantiate native C++ user-defined functions from python
2017-04-14 01:33:24 -07:00
43bd6cc25a
CNTK v2 library: Report an error on passing an empty parameters list to a learner
2017-04-14 01:32:55 -07:00
20d4bf2747
factor out asMultipleOf
2017-04-13 17:42:59 -07:00
1565fca563
Merge remote-tracking branch 'origin/master' into nikosk/workaround-curand
2017-04-13 17:42:25 -07:00
5a25d3ecce
Integrate jqian/wip/gdr into master
2017-04-13 16:23:44 -07:00
bc04458bda
Do gradients allreduce aggregation with GPUDirect RDMA
...
1. Iallreduce with cuda-aware not supported
ref: https://www.open-mpi.org/faq/?category=runcuda#mpi-apis-no-cuda
2017-04-13 09:16:37 -07:00
f96df7c6b5
Simplify bilinear upsampling layer interface
...
Bilinear upsampling only makes sense if the number of input
channels equals the number of output channels.
2017-04-13 10:11:52 +02:00
621052af8f
Add option to feed mean gradient into learners
2017-04-12 15:08:44 -07:00
2bec9f2c31
work around some curand limitations
2017-04-12 14:04:20 -07:00
3c4b0e2911
CNTK v2 library: Improve handling of certain error conditions
...
1) Better error handling in Axis normalization
2) Beter handling of bad convolution operand shapes
3) Switch NDShape::operator[] to checked indexing
2017-04-11 13:33:46 -07:00
49d4398062
Integrate alrezni/log_hyperparameter_changes into master
2017-04-11 02:44:19 -07:00
2ef2b7e8e3
Adding ability to specify sequece key in the output format
2017-04-11 09:18:40 +02:00
41dca1b206
Output learner hyperparameters with ProgressWriters
...
Use ProgressWriter logging facilities to emit a log message
every time the hyperparameters change.
2017-04-10 16:01:41 +02:00
dc2b0d6276
Integrate sagalic/acc_share_fix_commit into master
2017-04-10 06:53:29 -07:00
e8a8de17de
Fix _ftelli64 return type on Linux
2017-04-10 13:23:33 +02:00
41614a434e
Fix accumulator matrix sharing
...
Currently, accumulator matrix is allocated form the pool of shared
matrices. However, this matrix is meant to be persistent across
forward-backward loops. Since it is shared it may be overwritten by
some other node.
Fix for this issue is not to use memory pool but to allocate matrix
internally.
2017-04-10 10:04:22 +02:00
e1f48f4c14
Integrate kedeng/addLr into master
2017-04-07 18:29:17 -07:00
e7c16095c0
Add learning rate to adadelta
2017-04-07 16:24:50 -07:00
9330def020
NcclComm V2: make NcclComm part of the class
2017-04-07 15:40:11 -07:00
1b2a7a041d
Enable NCCL support for Python in V2 gradients aggregation
2017-04-07 15:40:11 -07:00
27039409f0
Change convolution to be deterministic.
...
Change convolution to be deterministic.
Fix warning.
Fix based on CR.
Code fix for CR.
Test deterministic behaviors. Commented out, waiting some bug fixes on set_fixed_random_seed().
2017-04-07 11:23:12 -07:00
59a856c7ff
Integrate eldak/optimizingHTK3 into master
2017-04-07 09:36:02 -07:00
0ec0a2af87
Some memory optimizations and refactoring of HTK deserializer
2017-04-07 15:56:47 +02:00
8ffa010a0d
Fix aggregation of EpochAccumulator nodes
...
When propagating aggregated values of accumulator nodes to
evaluation nodes, forward pass should be performed only on the
nodes between accumulator nodes and evaluation nodes, not on all
nodes in the network.
2017-04-07 09:32:30 +02:00
0c51a8de88
CNTK v2 library: Enable feeding SparseBlock data into dense input variables.
2017-04-06 23:11:50 -07:00
dd97e7534c
Integrate amitaga/rcBugFixes into master
2017-04-06 14:45:21 -07:00
39138b9c2d
CNTK v2 library: Fix Value::Create to handle >1D sparse NDArrayViews.
2017-04-06 12:19:06 -07:00
6738818091
Reducing size of indexer
2017-04-06 16:52:51 +02:00
9af17fa354
Introducing chunked mlf deserializer
2017-04-06 16:52:51 +02:00
3b4efc3f6d
First implementation of distributed evaluation
2017-04-06 12:01:50 +02:00
d6383ce062
Fix the bug in OneHotNode that m_value might be used in memory pool when it's sparse
2017-04-05 13:06:32 -07:00
99a51808c2
Integrate kedeng/fixAdadeltaCPU into master
2017-04-04 13:20:22 -07:00
fe9b401146
Fix Adadelta in CPU Sparse
2017-04-04 11:29:15 -07:00
82030c48d7
Add V2 api to get/set trace level
...
* Add yet another setter for the V1 component (math lib).
* Add global getter and setter to the V2 API. If the V1 trace level is set
to Info (the maximum possible value), tracing is also automatically enabled
in all V1 components (network, GPU memory allocator and math lib).
* Print an error message if cudaGetDeviceCount fails in BestGpu::Init().
2017-04-04 18:35:46 +02:00
94993f3c81
Integrate alrezni/v2_dropout into master
2017-04-04 09:16:31 -07:00
859296557d
Add local worker's state checkpointing
...
Add a mechanism to preserve the local state of a distributed worker
(together with the external state aggregated by the main node and saved
inside the checkpoint).
2017-04-04 16:07:48 +02:00
d192653c11
Integrate vadimma/fix_fb into master
2017-04-04 00:58:14 -07:00
ce7d96d3c7
Normalize pre-release version number
...
Conflicts:
Source/CNTK/CNTK.cpp
bindings/python/cntk/__init__.py
bindings/python/doc/gettingstarted.rst
2017-04-03 18:18:09 +02:00
861bfbd714
Bump versions
2017-04-03 18:16:31 +02:00
fba356fde0
CNTK v2 library: Several sparse data handling fixes
...
1) Handle conversion of a single scipy.csr_matrix to a Value object
2) Better error handling for conversion to csr matrices and fixed a perf
issue in the conversion.
3) Fixed the AssignOneHot sparse matrix implementations
4) Add sparse/dense value type inference in the computation network and
centralize matrix shape determination for sparse vs. dense matrices.
2017-04-02 18:40:26 -07:00
8148cab1c7
CNTK v2 library: Fix per_dim_mean_variance_normalize op.
2017-04-02 11:34:12 -07:00
e61af8c426
CNTK v2 library: Add str() for Axis
2017-04-02 11:34:12 -07:00
f863ff6dfc
Add brainscript parsing to ForwardBackward node
2017-04-02 11:23:26 -07:00
5eca042341
Integrate chenta/one_hot_bug_fix into master
2017-03-31 17:48:43 -07:00
cac9fb181c
Integrate kedeng/fasterTimes2 into master
2017-03-31 15:16:55 -07:00
72f7a41f01
remove useless code
2017-03-31 11:47:20 -07:00
59a2c26d7d
Make sparse label CE backprop faster
...
When minibatch size is big (like 10000), the diagonal matrix in times node for gradient could be big. This change implements a ColumnwiseScaleAndWeightedAdd to reduce the cost in that case.
2017-03-31 11:07:11 -07:00
4ddb7f18dd
Bug fix and check back in NVidia's change for convolution workspace.
2017-03-30 22:21:20 -07:00
f32fde505b
CNTK v2 library: Add support for reporting distributed training synchronization updates in the ProgressWriter.
2017-03-30 12:11:20 -07:00
Amit Agarwal
Nikos Karampatziakis
Project Philly
Junjie Qian
Nikola Milosavljevic
KeDengMS
Eldar Akchurin
Alexey Reznichenko
Sasa Galic
Cha Zhang
Mark Hillebrand
Vadim Mazalov
Cheng Tang