Commit Graph

520 Commits (5e152c0d9a9b3029c3539a6d23efebdae605475c)

Author SHA1 Message Date
raver119 5a4d2e8b31
[WIP] SVD (#16)
* - new SVD constructor
- OrthogonalDistribution now uses SVD custom op

Signed-off-by: raver119 <raver119@gmail.com>

* shapes fixed

Signed-off-by: raver119 <raver119@gmail.com>
2019-10-28 12:31:01 +03:00
Yurii Shyrma 029a69a835
Shyrma bn mkl bp (#14)
* - write code for new batchnorm backprop

Signed-off-by: Yurii <iuriish@yahoo.com>

* - testing batchnorm backprop

Signed-off-by: Yurii <iuriish@yahoo.com>

* - write code for batchnorm backprop based on mkl dnn api

Signed-off-by: Yurii <iuriish@yahoo.com>

* - testing and fixing bugs in batchnorm_bp mkl dnn

Signed-off-by: Yurii <iuriish@yahoo.com>

* - made corrections required by reviewer

Signed-off-by: Yurii <iuriish@yahoo.com>

* - change name in java wrapper for batchnorm op

Signed-off-by: Yurii <iuriish@yahoo.com>
2019-10-26 14:14:21 +03:00
Alex Black d333d29099
SameDiff cleanup and fixes (#12)
* #8160 Remove resolvePrepertiesFromSameDiffBeforeExecution

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* SameDiff API cleanup

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* More SameDiff cleanup

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Small fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* #8248 Switch SameDiff variable init from lazy to creation time for more predictable behaviour

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* #8252 TanhDerivative javadoc

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* #8225 Deconvolution2D input validation

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* #8265 Switch SameDiff.outputs() to user settable, instead of unreliable 'best guess'

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* #8224 SameDiff.zero and .one create constants, not variables

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* More cleanup and fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Small test fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Small fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* DL4J SameDiff fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Re-add hack for Deconvolution2DLayer until #8315 is resolved

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* #8270 Move CUDA device/version logging to Java; can be disabled via existing org.nd4j.log.initialization system property

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* All ND4J init logging checks system property

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Small tweak

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Remove redundant device logging

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* One more fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* UX improvements

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Deconv fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Add deconv tests

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Cleanup

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Remove debug code

Signed-off-by: AlexDBlack <blacka101@gmail.com>
2019-10-26 12:38:08 +11:00
Alexander Stoyakin d98784197a
[WIP] Some optimizations in dl4j (#11)
* Minor optimization.

* Reduce number of objects.

* Extend arrays when limit reached

* Test

* Some fixes.

* Small fix

* Wrong condition fixed.

* Fixes of reallocation.

* Small fix.

* Tests

* Clean up

* Test added.

* Tests and some fixes.

* Test

* Test fixed.

* Conflict fixed.

* UX improved
2019-10-24 14:15:50 +03:00
Alex Black 730442ae21
Remove bad import (#13)
Signed-off-by: AlexDBlack <blacka101@gmail.com>
2019-10-24 20:51:15 +11:00
Alex Black 3f0b4a2d4c
SameDiff execution, TF and memory management overhaul (#10)
* SameDiff execution memory management improvements, round 1

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Round 2

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Round 3

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Clear node outputs closed array references; Slight change to OpValidation internals to not rely on cached op outputs

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Next steps

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Next step

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Cleanup

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* More polish

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Add WeakIdentityHashmap

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Session fixes for control ops and next steps

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* More fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* First steps for training session + in-line updating

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Next steps

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fix losses and history during training

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* More fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* BiasAdd and other fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Don't use SDVariable.getArr() in TFGraphTestAllHelper (import tests)

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Small fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Cleanup

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* First steps for new dependency tracking approach

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Start integrating dependency tracking for memory management

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Non-control op dependency tracking works/passes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Switch/merge

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Next steps

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Cleanup and next steps

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Small fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fix issue dependency tracking for initial variables/constants

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Add check for aliases when determining if safe to close array

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* First pass on new TF graph import class

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Import fixes, op fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Cleanup and fixes for new TF import mapper

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* More cleanup

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Partial implementation of new dependency tracker

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Next steps

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* AbstractDependencyTracker for shared code

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Overhaul SameDiff graph execution (dependency tracking)

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* More fixes, cleanup, next steps

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Ad no-op memory manager, cleanup, fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fix switch dependency tracking

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* INDArray.toString: no exception on closed arrays, just note closed

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fix enter and exit dependency tracking

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* TensorArray memory management fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Add unique ID for INDArray instances

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fix memory management for NextIteration outputs in multi-iteration loops

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Remove (now unnecessary) special case handling for nested enters

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Cleanup

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Handle control dependencies during execution; javadoc for memory managers

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Cleanup, polish, code comments, javadoc

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Cleanup and more javadoc

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Add memory validation for all TF import tests - ensure all arrays (except outputs) are released

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Clean up arrays waiting on unexecuted ops at the end of execution

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fixes for enter op memory managent in the context of multiple non-nested loops/frames

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fix order of operation issues for dependency tracker

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Always clear op fields after execution to avoid leaks or unintended array reuse

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Small fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Re-implement dtype conversion

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fix for control dependencies execution (dependency tracking)

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fix TF import overrides and filtering

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fix for constant enter array dependency tracking

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* DL4J Fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* More DL4J fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Cleanup and polish

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* More polish and javadoc

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* More logging level tweaks, small DL4J fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Small fix to DL4J SameDiffLayer

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fix empty array deserialization, add extra deserialization checks

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* FlatBuffers control dep serialization fixes; test serialization as part of all TF import tests

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Variable control dependencies serialization fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fix issue with removing inputs for ops

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* FlatBuffers NDArray deserialization fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* FlatBuffers NDArray deserialization fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Small fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Small fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Final cleanup/polish

Signed-off-by: AlexDBlack <blacka101@gmail.com>
2019-10-23 21:19:50 +11:00
Alexander Stoyakin f31661e13b
Merge pull request #7 from KonduitAI/asto_nd4s_10172019
KDTree optimization
2019-10-23 12:11:25 +03:00
raver119 35e6ffede4
Merge pull request #6 from KonduitAI/shyrma_broadcast
- replace condition isScalar() by condition length ==1 in some NDArra…
2019-10-22 07:56:04 +03:00
Yurii 8f3eaebda5 - replace condition isScalar() by condition length ==1 in some NDArray methodds
Signed-off-by: Yurii <iuriish@yahoo.com>
2019-10-21 16:25:13 +03:00
Alexander Stoyakin 775402c934
Merge pull request #5 from KonduitAI/asto_nd4s_build_10192019
Fix nd4s compilation
2019-10-19 21:57:02 +03:00
Alexander Stoyakin b556519e15 Fix compilation. 2019-10-19 21:48:43 +03:00
Alex Black 7f06df5826
Small URL fix (#8297)
Signed-off-by: AlexDBlack <blacka101@gmail.com>
2019-10-18 11:25:58 +11:00
raver119 a4984d52d7
Merge pull request #4 from KonduitAI/shyrma_lstm
Shyrma lstm
2019-10-17 20:58:06 +03:00
Yurii 99be467f76 - minor change in recurrent.h
Signed-off-by: Yurii <iuriish@yahoo.com>
2019-10-17 20:46:51 +03:00
Yurii 70bd925abd - write 2 versions of new lstmLayer: one is based on own code, second uses mkl dnn api 2019-10-17 20:44:52 +03:00
Alexander Stoyakin 630bb3c9b6
Merge pull request #2 from KonduitAI/asto_ops_wrapper
[WIP] New ops wrapper
2019-10-16 20:21:50 +03:00
Alexander Stoyakin 9e5799847a TF names fixed. 2019-10-16 19:50:18 +03:00
Alexander Stoyakin 502bedf5d5 Register ops for TF import. 2019-10-16 19:39:04 +03:00
Alexander Stoyakin ec722b20ee TF names added 2019-10-16 19:29:19 +03:00
Alexander Stoyakin 99d77e1384 Ops exported for sameDiff 2019-10-16 19:16:47 +03:00
shugeo 3662657d5c
Merge pull request #1 from KonduitAI/shugeo_gamma
Shugeo gamma
2019-10-16 18:49:33 +03:00
shugeo 24a2b2933f Added gamma and lgamma functions. 2019-10-16 18:22:18 +03:00
Alexander Stoyakin 96a9a1a733 Fixed output from operation. 2019-10-16 18:07:52 +03:00
shugeo 7617682a46 Added declarations for igamma and igammac ops. 2019-10-16 14:45:10 +03:00
shugeo 478a0c1f97 Added igamma and igammac broadcastable ops implementations and tests. 2019-10-16 14:02:53 +03:00
shugeo 7103aca8c5 Added broadcastable IGamma and IGammac ops. 2019-10-16 13:58:32 +03:00
shugeo f90e6da97e Added nd4j_gamma, nd4j_igamma and nd4j_igammac functions. 2019-10-16 13:53:31 +03:00
Alexander Stoyakin c4307384f3 Fixed shape for muli 2019-10-16 12:59:25 +03:00
Alexander Stoyakin d5002b14c7 New ops wrappers 2019-10-16 12:59:08 +03:00
Alex Black b9a4b0f25f
Update links (#8292)
Signed-off-by: AlexDBlack <blacka101@gmail.com>
2019-10-16 12:59:52 +11:00
Alexandre Boulanger 171ce51f46 RL4J: Use Nd4j Random instead of java.util.Random (#8282)
* Changed to use Nd4j Random instead of java.util.Random

Signed-off-by: unknown <aboulang2002@yahoo.com>

* Changed to use Nd4j.getRandom() instead of the factory

Signed-off-by: Alexandre Boulanger <aboulang2002@yahoo.com>
2019-10-16 10:56:24 +09:00
shugeo df2448613e Added gamma distribution functions. 2019-10-15 20:00:07 +03:00
AlexDBlack 2d750b69e5 Merge remote-tracking branch 'konduit/master' 2019-10-14 17:21:23 +11:00
raver119 d5568df7b3
Merge pull request #10 from KonduitAI/shugeo_fake_quant2
Shugeo fake quant2
2019-10-10 18:40:02 +03:00
shugeo ace65355c5 Added doc for fake_quant_with_min_max* op helpers cuda implementations. 2019-10-10 18:35:28 +03:00
shugeo c890de5a7b Added doc for fake_quant_with_min_max* op helpers implementations. 2019-10-10 18:31:17 +03:00
shugeo c3f755d975 Refactored helpers both for cuda and cpu platforms. 2019-10-10 18:02:49 +03:00
shugeo a09cb5e2be Added doc for fake_quant_with_min_max_per_channel op declaration. 2019-10-10 17:13:33 +03:00
shugeo 92636b0b86 Eliminated waste operator. 2019-10-10 17:08:59 +03:00
shugeo d5b352273d Implementation of cuda kernel for fake_quant_with_min_max_vars_per_channels op. Final revision. 2019-10-10 16:51:29 +03:00
shugeo 02d8616692 Implementation of cuda kernel for fake_quant_with_min_max_vars_per_channels op. 2019-10-10 16:40:56 +03:00
shugeo 3504b0cda9 Implemented fake_quant_with_min_max_vars_per_channel fop cuda helper. The first working revision. 2019-10-10 15:44:50 +03:00
shugeo 753565145c Refactored fake_quant_with_min_max_vars op cuda implementation. 2019-10-10 14:00:49 +03:00
shugeo c13e945a96 Fixed fake_quant_with_min_max_vars op and tests. 2019-10-10 13:23:11 +03:00
Alexander Stoyakin fa8105bf0f
Merge pull request #9 from KonduitAI/asto_nd4s_indexing
[WIP] Indexing for INDArray in nd4s
2019-10-10 11:03:06 +03:00
Alexander Stoyakin 156fba4f77 Minor fix
Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>
2019-10-10 10:49:07 +03:00
Alexander Stoyakin be70bea359 Operator :: for INDArray
Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>
2019-10-10 10:43:32 +03:00
shugeo 3c0c59ab88 Refactored fake_quant_with_min_max_vars op. 2019-10-09 22:09:33 +03:00
shugeo 352f1eee80 Implemented fake_quant_with_min_max_per_channel helper for cpu platform. The first approach. 2019-10-09 21:39:59 +03:00
Alexander Stoyakin 0de8523ebd Indexing syntax changed
Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>
2019-10-09 16:07:20 +03:00