2019-06-06 14:21:15 +02:00
/*******************************************************************************
* Copyright ( c ) 2015 - 2018 Skymind , Inc .
*
* This program and the accompanying materials are made available under the
* terms of the Apache License , Version 2.0 which is available at
* https : //www.apache.org/licenses/LICENSE-2.0.
*
* Unless required by applicable law or agreed to in writing , software
* distributed under the License is distributed on an " AS IS " BASIS , WITHOUT
* WARRANTIES OR CONDITIONS OF ANY KIND , either express or implied . See the
* License for the specific language governing permissions and limitations
* under the License .
*
* SPDX - License - Identifier : Apache - 2.0
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
Merge master to upstream (#7945)
* Shugeo strided slice zeros (#14)
* Modified strided_slice op to properly work with empty-like shapes.
* Fixed test for reduce_mean with empty-like input.
* [WIP] Last merge (#15)
* correct logsoftmax looss (#2)
* Small SameDiff listener fix (#4)
* Various fixes (#6)
* #7839 Fix for asXMatrix and tests
* #7866 EmbeddingSequenceLayer dtype fix + test
* #7856 SameDiff save/load stream methods
* #7859 RegressionEvaluation rank 4 fix + tests + axis configuration
* EvaluationBinary 3d/4d
* More evaluation 3d/4d tests
* #7847 Evaluation empty checks
* Small test ifx
* #7848 Fix median edge case
* Improve DL4J samediff layer tests
* [WIP] FastText wrapper implemented (#8)
* FastText implemented
* Some fixes
* Fix shapes for wordsNearest
* Validation of input vectors
* Fixes
* Fixed test
* Thread tagged
* Some tweaks
* setContextClassLoader for DeallocatorServiceThread
* Numpy format tests (#1)
* Various fixes (#11)
* #7852 SameDiff gather fix
* #7892 SameDiff placeholder to constant conversion
* #7890 validate input rank for MLN/CG init methods
* Fix broken permute shape calculation
* Permute and gather fixes
* Tests
* #7850 LogSumExp fix + test
* Handful of test fixes
* Empty arrays with non-scalar shapes (#10)
* minor rearrangements for lambdas
* empty tensors with non-scalar shapes
* numpy empty tensors with non-scalar shapes
* few more empty tweaks
* Small fixes
* conv3d signature update
* micro fix in batchnorm mkldnn
* Import fixes
* Fix
* MKL-DNN update
* Small fill fix
* fill with empty input + test
* Fixes
* Small error improvement
* Fix
* one special test
* couple of fixes for lstm
* Rewrite TFGraphMapper.getNDArrayFromTensor to be maintainable and less error prone
* Fixes
* FP16
* Unsigned
* BFloat16
* Fill op - empty tweaks
* - couple of fixes for empty arrays construction
- stack updated
* strided slice fix
* one transform test
* provide method for reducing shapeInfo in case of input array is empty
* Fixed reduceAlongDimensions to use empty input properly.
* couple of broadcast tests
* couple of tests broadcast tests + tweak to make them pass
* add check of non-empty to methods producing sub-arrays
* Fixed reshapeC with zeros in shape.
* complete empty check in reduce_... legacy ops
* Concat and cumsum/prod
* Tweak to empty shape inference on import
* add empty check to the rest of reduce legacy ops
* one more test
* correct typo in evalReduceShapeInfoEmpty
* Added tests for reduce_* ops to tests with zero shapes.
* few more tests for empty reductions
* Fixed strided_slice op with empty case and tests.
* one more empty reduction test
* Fixed strided_slice test.
* add empty check to NDArray::reshapei
* infOrMax
* empty min/max with infinity tests
* made unstack working correctly with empty arrays
* few IndexReduce tests + tweaks for empty shapes
* add test for empty concat
* few tests fixed
* Validation fix for reductions on empty shapes
* Reverse fix
* Reduction shape calc fixes
* SameDiff.generateOutputVariable: don't use shape function to determine number of outputs
* Range fix
* - NDArray constructor updated for scalars/empty arrays
- few tests fixed
* More fixes
* Empty creator fixes
* concat fix
* concat fix
* TF import tests: allow 'both all NaN' and 'both all inf' to pass
* Slice, zero fraction, and reshape fixes
* transpose, gather
* Zero fraction
* scalar cast fix
* Empty reduction axis support
* few more tests fixed
* Fixed input checks conforming with TF for concat op and tests.
* few tests fixed
* matmul scalar shape fix
* Fixed checkout for data type and scalarity with concat to allow non-empty scalars with vector concats.
* broadcast bool fix
* few more tests
* few more tests
* correct evalReduceShapeInfoEmpty
* argmax/argmin + tests
* one more empty edge case + one more test
* argmax/argmin/realdiv_bp tweaks
* empty reshape test + fix
* Helper fixes
* Small fixes
* Gather test fix
* Gather test fix
* Small fixes
* reduce scalar zero values
* scalar mean workaround
* Remove debug code
* along dim mean workaround
* one more test
* - equalsTo() tweak for empty arrays
- one more test
* broadcast tweaks
* [WIP] Fixing outstanding issues for NLP (#9)
* Avoid using not-inited objects
* Test fixed.
* Redundant method avoided for models like FastText
* KMeans++ implementation
* KMeans++ implementation
* Disable parallel execution
* KMeans++
* Tests
* Dev branch merge (#16)
* SameDiff: convertDataType and gradient check util improvements (#12)
* GradCheck util improvements
* StopGradient constructor + test
* SameDiff: Add datatype conversion
* Javadoc and add DataType.isNumerical()
* Small fix
* Fix SameDiff TF import test cases intermediate naming (workaround for bad default)
* TFGraphTestAllHelper: check intermediates in execution order
* Add missing debug listener
* [WIP] lstmBlock fix + other changes (#13)
- fixes lstmBlock issue
- changes NDArray method reshape(), permute(), transpose() by making them return instance instead of pointer
- CheckNumerics op
- fixes for ReduceBool IsInfOrNan & IsFinite
* Small test fix
* CheckNumerics op wrapper
* Fix some issues on master (#17)
* Fix DataVec test issue
* Fix issue with dl4j SameDiff output layer
* Dtype fix for lambda layers
* #7912 BertIterator dtype fix (use float32 not global default)
* [WIP] Next set of CUDA stuff (#7)
New CUDA implementations and improvements
* bad file
* Dev branch master merge (#23)
* SameDiff: convertDataType and gradient check util improvements (#12)
* GradCheck util improvements
* StopGradient constructor + test
* SameDiff: Add datatype conversion
* Javadoc and add DataType.isNumerical()
* Small fix
* Fix SameDiff TF import test cases intermediate naming (workaround for bad default)
* TFGraphTestAllHelper: check intermediates in execution order
* Add missing debug listener
* [WIP] lstmBlock fix + other changes (#13)
- fixes lstmBlock issue
- changes NDArray method reshape(), permute(), transpose() by making them return instance instead of pointer
- CheckNumerics op
- fixes for ReduceBool IsInfOrNan & IsFinite
* Small test fix
* CheckNumerics op wrapper
* Compatibility of deserialization (#18)
Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>
* SameDiff: add activation gradient checking support for debugging (#19)
* SameDiff gradient checker: first pass on activation gradient checks
* Fixes + tests for activation gradient checking
* Javadoc
* [WIP] Some nd4j data type corrections (#20)
* Adjust data type
* Set correct Data type.
* Size of proper data type.
* fix averaged cpu load (#22)
* SameDiff ops, TF import and fixes (#24)
* CheckNumerics tests + fixes + misc fixes
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* Fake quant
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* Fixes
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* FakeQuantWithMinMaxArgs
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* CheckNumerics fix
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* Fix libnd4j ALL_INTS and ALL_FLOATS declaration (uint and bfloat types)
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* Small fix
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* Javadoc
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* Exception tweak
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* fix
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* Fix for out of scope stack allocated var use
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* Ignores
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* Ignore for known failing test (already logged issue)
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* Merge upstream to fork (#25)
* Add thousand-separator commas to TotalParams (#7915)
* Add thousand-separator commas to TotalParams
The number of parameters can be quite large, and it would help the reading of the summary printout to have the TotalParams column & values at the bottom have thousand-separator-commas in them.
* Add thousand-separator commas to MultiLayerNetwork
Corresponding change to MultiLayerNetwork
Signed-off-by: Jxtps Jxtps <jxtps435@gmail.com>
* Update contributing and issue/PR templates (#7934)
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* Fix link to AdaDelta paper (#7942)
Fix link to AdaDelta paper hosted on matthewzeiler.com
Signed-off-by: Jxtps
* Fixes, and ignores for known/logged failing issues (#7943)
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* SameDiff + DL4J/SameDiff: Multiple fixes (#28)
* #7919 HDF5 attribute buffer length fix
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* #7909 Arbiter constructor exception ux improvements
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* #7925 RNN output layer length checks
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* #7939 Add listener for validating inputs are not incorrectly modified
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* #7939 Integrate NonInplaceValidationListener into tests
* #7844 DL4J SameDiff fixes for variable minibatch size
* DL4J SameDiff fixes - ensure gradient for input placeholder is available
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* Tweaks to ExternalErrorsFunction - use placeholders, make more robust
* Another fix
* More fixes
* More SameDiff/DL4J fixes
* Scope out scalar array creation in BaseScalarOp
* Remove debug code
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* [WIP] Final dev branch merge (#29)
* SameDiff: convertDataType and gradient check util improvements (#12)
* GradCheck util improvements
* StopGradient constructor + test
* SameDiff: Add datatype conversion
* Javadoc and add DataType.isNumerical()
* Small fix
* Fix SameDiff TF import test cases intermediate naming (workaround for bad default)
* TFGraphTestAllHelper: check intermediates in execution order
* Add missing debug listener
* [WIP] lstmBlock fix + other changes (#13)
- fixes lstmBlock issue
- changes NDArray method reshape(), permute(), transpose() by making them return instance instead of pointer
- CheckNumerics op
- fixes for ReduceBool IsInfOrNan & IsFinite
* Small test fix
* CheckNumerics op wrapper
* Compatibility of deserialization (#18)
Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>
* SameDiff: add activation gradient checking support for debugging (#19)
* SameDiff gradient checker: first pass on activation gradient checks
* Fixes + tests for activation gradient checking
* Javadoc
* [WIP] Some nd4j data type corrections (#20)
* Adjust data type
* Set correct Data type.
* Size of proper data type.
* fix averaged cpu load (#22)
* [WIP] Multiple dataset iterators (#27)
* Splitting dataset into arbitrary number
* Fixes
* Multiple split of iterator
* Test
* Test
* Some fixes
* signature change
* one more tweak
Signed-off-by: raver119 <raver119@gmail.com>
* one more test for sequential use of DataSetIteratorSplitter
Signed-off-by: raver119 <raver119@gmail.com>
* Fixes
* Fixes
* one more test for Alexander
Signed-off-by: raver119 <raver119@gmail.com>
* Some fixes
* Some fixes
* one more test for Alexander
Signed-off-by: raver119 <raver119@gmail.com>
* minor test fix
Signed-off-by: raver119 <raver119@gmail.com>
* Some fixes
* Some fixes
* couple of assertions tweaked
Signed-off-by: raver119 <raver119@gmail.com>
* MDS splitter test :/
Signed-off-by: raver119 <raver119@gmail.com>
* Minor refactoring
* Multi dataset
* Some fixes
* More tests
* Small number of test fixes/improvements (failures on CI) (#31)
Signed-off-by: AlexDBlack <blacka101@gmail.com>
* [WIP] More CUDA stuff (#26)
* initial commit
Signed-off-by: raver119 <raver119@gmail.com>
* LRN BP CUDA
Signed-off-by: raver119 <raver119@gmail.com>
* less memory
Signed-off-by: raver119 <raver119@gmail.com>
* Fixed bug with crop_and_resize op helper.
* get rid of unnecessary index-calculation dunction
Signed-off-by: Yurii <yurii@skymind.io>
* Fixed sort with nth_element cuda-based helper.
* Refactored nth_element.
* Refactored nth_element op and tests.
* Modified usage of dim array with sortTad routine.
* Refactored main routine of helper for non_max_image_suppression op.
* non_max_image_suppression op helper with cuda kernel implementation. Initial revision.
* fix vol2col cuda kernel
* meh
Signed-off-by: raver119 <raver119@gmail.com>
* topK concept
Signed-off-by: raver119 <raver119@gmail.com>
* unsorted topK with scanWitdh of 1
Signed-off-by: raver119 <raver119@gmail.com>
* correct vol2col tests
* sorted/unsorted topK
Signed-off-by: raver119 <raver119@gmail.com>
* implementation and fixing col2im/col2vol
* Corrected usage flags with input/output with reverse op.
* dup is const now
Signed-off-by: raver119 <raver119@gmail.com>
* percentile op
Signed-off-by: raver119 <raver119@gmail.com>
* group tests for mapool2d
Signed-off-by: Yurii <yurii@skymind.io>
* special test for george
Signed-off-by: raver119 <raver119@gmail.com>
* less threads for sortTad
Signed-off-by: raver119 <raver119@gmail.com>
* provide conv2d for cuda
Signed-off-by: Yurii <yurii@skymind.io>
* remove auther in sort tad kernel code
Signed-off-by: Yurii <yurii@skymind.io>
* provide depthwise_conv2d for cuda
Signed-off-by: Yurii <yurii@skymind.io>
* - max_pooling_with_argmax
- null check for special use
Signed-off-by: raver119 <raver119@gmail.com>
* dts cuda
Signed-off-by: raver119 <raver119@gmail.com>
* provide sconv2d for cuda
Signed-off-by: Yurii <yurii@skymind.io>
* std cuda
Signed-off-by: raver119 <raver119@gmail.com>
* Refactored non_max_suppression op to conform TF implementation.
* Improved suppression helper.
* provide pooling3d for cuda
Signed-off-by: Yurii <yurii@skymind.io>
* minor lstm rearrangements
Signed-off-by: raver119 <raver119@gmail.com>
* more of minor lstm rearrangements
Signed-off-by: raver119 <raver119@gmail.com>
* (bi)dynamic_rnn
Signed-off-by: raver119 <raver119@gmail.com>
* templates init order
Signed-off-by: raver119 <raver119@gmail.com>
* Refactored non_max_suppression op.
* Added cuda kernel for non_max_suppression.
* CPU sort by key/value
Signed-off-by: raver119 <raver119@gmail.com>
* CPU sort TAD by key/value
Signed-off-by: raver119 <raver119@gmail.com>
* CPU sort TAD by key/value tests
Signed-off-by: raver119 <raver119@gmail.com>
* Eliminate compiler error with cuda implementation.
* - repaired gradCheck in cuda
- provide conv2d_bp for cuda
Signed-off-by: Yurii <yurii@skymind.io>
* missed signature
Signed-off-by: raver119 <raver119@gmail.com>
* provide depthwise_conv2d_bp for cuda
Signed-off-by: Yurii <yurii@skymind.io>
* Implementation of lup helper with cuda kernel. Initial commit.
* further work on backprops for convolutions
Signed-off-by: Yurii <yurii@skymind.io>
* CUDA linear sort by key/val
Signed-off-by: raver119 <raver119@gmail.com>
* CUDA tad sort by key/val
Signed-off-by: raver119 <raver119@gmail.com>
* start providing of backprop for pooling2d/3d
Signed-off-by: Yurii <yurii@skymind.io>
* Added atomicAdd for bool datatype.
* dynamic partition concept
Signed-off-by: raver119 <raver119@gmail.com>
* dynamic partition concept
Signed-off-by: raver119 <raver119@gmail.com>
* dynamic partition scalar CUDA
Signed-off-by: raver119 <raver119@gmail.com>
* important comment
Signed-off-by: raver119 <raver119@gmail.com>
* fix pooling2d/3d backprop helpers
Signed-off-by: Yurii <yurii@skymind.io>
* Added non-linear test with dynamic_partition.
* Improved test for dynamic_partition.
* dynamic_partition TAD concept
Signed-off-by: raver119 <raver119@gmail.com>
* - dynamic_partition TAD CUDA impl
- dynamic_partition TAD CPU fix
Signed-off-by: raver119 <raver119@gmail.com>
* - rewrite cpu code for usampling2d/3d
- write cuda code for usampling2d/3d
Signed-off-by: Yurii <yurii@skymind.io>
* dynamic_stitch CUDA vector case
Signed-off-by: raver119 <raver119@gmail.com>
* dynamic_stitch CUDA TAD case concept
Signed-off-by: raver119 <raver119@gmail.com>
* dynamic_stitch CUDA TAD case impl
Signed-off-by: raver119 <raver119@gmail.com>
* Added tests for dynamic_stitch 3D-4D cases.
* minor tests tweaks
Signed-off-by: raver119 <raver119@gmail.com>
* Fixed type check for dynamic stitch.
* min/max bp
Signed-off-by: raver119 <raver119@gmail.com>
* rewrite code for upsampling2d/3d cpu
Signed-off-by: Yurii <yurii@skymind.io>
* reduce min/max/norm_max bp
Signed-off-by: raver119 <raver119@gmail.com>
* lup implementation. Additional enhancements.
* provide code for upsamling2d/3d backprop
Signed-off-by: Yurii <yurii@skymind.io>
* weightedCrossEntropyWithLogits
Signed-off-by: raver119 <raver119@gmail.com>
* Fixed template math atomicMul for 64bit ints.
* Refactored dynamic_partition_bp op.
* inverseBroadcast fix
Signed-off-by: raver119 <raver119@gmail.com>
* DynamicPartitionBP test datatype fixed.
* - nd4j_atomicMul Windows fix
- cpu/NDArrayLambda.hpp excluded from CUDA
Signed-off-by: raver119 <raver119@gmail.com>
2019-06-27 17:37:04 +02:00
//
// @author Yurii Shyrma (iuriish@yahoo.com)
2019-06-06 14:21:15 +02:00
// @author Alex Black
//
# include <op_boilerplate.h>
# if NOT_EXCLUDED(OP_gruCell)
# include <ops/declarable/CustomOperations.h>
# include <ops/declarable/helpers/gru.h>
namespace nd4j {
namespace ops {
//////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL ( gruCell , 6 , 4 , false , 0 , 0 ) {
2019-07-20 07:58:44 +02:00
auto x = INPUT_VARIABLE ( 0 ) ; // input [bS, nIn], nIn - input size
auto hLast = INPUT_VARIABLE ( 1 ) ; // previous cell output [bS, nU], that is at previous time step t-1, nU - number of units
auto Wru = INPUT_VARIABLE ( 2 ) ; // RU weights - [nIn+nU, 2*nU] - reset and update gates (input/recurrent weights)
auto Wc = INPUT_VARIABLE ( 3 ) ; // C weights - [nIn+nU, nU] - cell gate (input/recurrent weights)
auto bru = INPUT_VARIABLE ( 4 ) ; // reset and update biases, [2*nU] - reset and update gates
auto bc = INPUT_VARIABLE ( 5 ) ; // cell biases, [nU]
auto r = OUTPUT_VARIABLE ( 0 ) ; // Reset gate output [bS, nU]
auto u = OUTPUT_VARIABLE ( 1 ) ; // Update gate output [bS, nU]
auto c = OUTPUT_VARIABLE ( 2 ) ; // Cell gate output [bS, nU]
auto h = OUTPUT_VARIABLE ( 3 ) ; // current cell output [bS, nU]
2019-06-06 14:21:15 +02:00
REQUIRE_TRUE ( x - > rankOf ( ) = = 2 & & hLast - > rankOf ( ) = = 2 , 0 , " gruCell: Input ranks must be 2 for inputs 0 and 1 (x, hLast) - got %i, %i " , x - > rankOf ( ) , hLast - > rankOf ( ) ) ;
2019-07-12 10:51:51 +02:00
const int rank = x - > rankOf ( ) ;
const auto bS = x - > sizeAt ( 0 ) ;
const auto nIn = x - > sizeAt ( 1 ) ;
const auto nU = hLast - > sizeAt ( 1 ) ;
2019-06-06 14:21:15 +02:00
REQUIRE_TRUE ( x - > sizeAt ( 0 ) = = hLast - > sizeAt ( 0 ) , 0 , " gruCell: Input minibatch sizes (dimension 0) must be same for x and hLast " ) ;
REQUIRE_TRUE ( Wru - > rankOf ( ) = = 2 & & Wc - > rankOf ( ) = = 2 , 0 , " gruCell: weight arrays (Wru, Wc) arrays must be 2, got %i and %i " , Wru - > rankOf ( ) , Wc - > rankOf ( ) ) ;
2019-07-12 10:51:51 +02:00
REQUIRE_TRUE ( Wru - > sizeAt ( 0 ) = = ( nIn + nU ) & & Wc - > sizeAt ( 0 ) = = ( nIn + nU ) , 0 , " gruCell: Weights size(0) must be equal to nIn + nU, got %i " , Wru - > sizeAt ( 0 ) ) ;
REQUIRE_TRUE ( Wru - > sizeAt ( 1 ) = = ( 2 * nU ) , 0 , " gruCell: Weights (reset and update) size(1) must be equal to 2*nU, got %i " , Wru - > sizeAt ( 1 ) ) ;
REQUIRE_TRUE ( Wc - > sizeAt ( 1 ) = = nU , 0 , " gruCell: Weights (cell) size(1) must be equal to nU, got %i " , Wc - > sizeAt ( 1 ) ) ;
REQUIRE_TRUE ( bru - > rankOf ( ) = = 1 & & bru - > sizeAt ( 0 ) = = ( 2 * nU ) , 0 , " gruCell: reset/update biases must be rank 1, size 2*nU " ) ;
REQUIRE_TRUE ( bc - > rankOf ( ) = = 1 & & bc - > sizeAt ( 0 ) = = nU , 0 , " gruCell: cell biases must be rank 1, size nU " ) ;
2019-06-06 14:21:15 +02:00
REQUIRE_TRUE ( r - > rankOf ( ) = = 2 & & u - > rankOf ( ) = = 2 & & c - > rankOf ( ) = = 2 & & h - > rankOf ( ) = = 2 & &
r - > sizeAt ( 0 ) = = bS & & u - > sizeAt ( 0 ) = = bS & & c - > sizeAt ( 0 ) = = bS & & h - > sizeAt ( 0 ) = = bS & &
r - > sizeAt ( 1 ) = = nU & & u - > sizeAt ( 1 ) = = nU & & c - > sizeAt ( 1 ) = = nU & & h - > sizeAt ( 1 ) = = nU ,
2019-07-12 10:51:51 +02:00
0 , " gruCell: Output arrays must all be rank 2 with size(0) == batchSize and size(1) == nU " ) ;
2019-06-06 14:21:15 +02:00
helpers : : gruCell ( block . launchContext ( ) , x , hLast , Wru , Wc , bru , bc , r , u , c , h ) ;
return Status : : OK ( ) ;
}
DECLARE_TYPES ( gruCell ) {
getOpDescriptor ( )
- > setAllowedInputTypes ( 0 , nd4j : : DataType : : ANY )
- > setAllowedInputTypes ( 1 , { ALL_FLOATS } )
- > setAllowedInputTypes ( 2 , { ALL_FLOATS } )
- > setAllowedInputTypes ( 3 , { ALL_FLOATS } )
- > setAllowedInputTypes ( 4 , { ALL_FLOATS } )
- > setAllowedOutputTypes ( { ALL_FLOATS } ) ;
}
DECLARE_SHAPE_FN ( gruCell ) {
2019-07-12 10:51:51 +02:00
auto x = inputShape - > at ( 0 ) ; // input [bS x nIn]
auto hLast = inputShape - > at ( 1 ) ; // previous cell output [bS x nU], that is at previous time step t-1
auto Wru = inputShape - > at ( 2 ) ; // RU weights - [(nIn+nU), 2*nU] - reset and update gates (input/recurrent weights)
auto Wc = inputShape - > at ( 3 ) ; // C weights - [(nIn+nU), nU] - cell gate (input/recurrent weights)
auto bru = inputShape - > at ( 4 ) ; // reset and update biases, [2*nU] - reset and update gates
auto bc = inputShape - > at ( 5 ) ; // cell biases, [nU]
2019-06-06 14:21:15 +02:00
REQUIRE_TRUE ( shape : : rank ( x ) = = 2 & & shape : : rank ( hLast ) = = 2 , 0 , " gruCell: Input ranks must be 2 for inputs 0 and 1 (x, hLast) - got %i, %i " , shape : : rank ( x ) , shape : : rank ( hLast ) ) ;
const int rank = x [ 0 ] ;
const auto bS = x [ 1 ] ;
2019-07-12 10:51:51 +02:00
const auto nIn = x [ 2 ] ;
const auto nU = hLast [ 2 ] ;
2019-06-06 14:21:15 +02:00
REQUIRE_TRUE ( x [ 1 ] = = hLast [ 1 ] , 0 , " gruCell: Input minibatch sizes (dimension 0) must be same for x and hLast " ) ;
REQUIRE_TRUE ( shape : : rank ( Wru ) = = 2 & & shape : : rank ( Wc ) = = 2 , 0 , " gruCell: weight arrays (Wru, Wc) arrays must be 2, got %i and %i " , shape : : rank ( Wru ) , shape : : rank ( Wc ) ) ;
2019-07-12 10:51:51 +02:00
REQUIRE_TRUE ( Wru [ 1 ] = = ( nIn + nU ) & & Wc [ 1 ] = = ( nIn + nU ) , 0 , " gruCell: Weights size(0) must be equal to nIn + nU, got %i and %i " , Wru [ 1 ] , Wc [ 1 ] ) ;
REQUIRE_TRUE ( Wru [ 2 ] = = ( 2 * nU ) , 0 , " gruCell: Weights (reset and update) size(1) must be equal to 2*nU, got %i " , Wru [ 2 ] ) ;
REQUIRE_TRUE ( Wc [ 2 ] = = nU , 0 , " gruCell: Weights (cell) size(1) must be equal to nU, got %i " , Wc [ 2 ] ) ;
REQUIRE_TRUE ( shape : : rank ( bru ) = = 1 & & bru [ 1 ] = = ( 2 * nU ) , 0 , " gruCell: reset/update biases must be rank 1, size 2*nU " ) ;
REQUIRE_TRUE ( shape : : rank ( bc ) = = 1 & & bc [ 1 ] = = nU , 0 , " gruCell: cell biases must be rank 1, size nU " ) ;
2019-06-06 14:21:15 +02:00
Nd4jLong * s0 ( nullptr ) ;
2019-07-12 10:51:51 +02:00
ALLOCATE ( s0 , block . getWorkspace ( ) , shape : : shapeInfoLength ( rank ) , Nd4jLong ) ; // [bS x nU]
2019-06-06 14:21:15 +02:00
s0 [ 0 ] = rank ;
s0 [ 1 ] = bS ;
2019-07-12 10:51:51 +02:00
s0 [ 2 ] = nU ;
2019-06-06 14:21:15 +02:00
ShapeUtils : : updateStridesAndType ( s0 , x , shape : : order ( hLast ) ) ;
auto ts0 = ConstantShapeHelper : : getInstance ( ) - > createFromExisting ( s0 , block . workspace ( ) ) ;
2019-07-12 10:51:51 +02:00
//4 output shapes, all [bs, nU]
2019-06-06 14:21:15 +02:00
return SHAPELIST ( ts0 , ts0 , ts0 , ts0 ) ;
}
//////////////////////////////////////////////////////////////////////////
2019-07-20 07:58:44 +02:00
CUSTOM_OP_IMPL ( gruCell_bp , 10 , 6 , false , 0 , 0 ) {
auto x = INPUT_VARIABLE ( 0 ) ; // input [bS x iS]
auto hi = INPUT_VARIABLE ( 1 ) ; // previous cell output [bS x nU]
auto W = INPUT_VARIABLE ( 2 ) ; // weights, [iS+nU x 2*nU]
auto Wc = INPUT_VARIABLE ( 3 ) ; // c weights, [iS+nU x nU]
auto b = INPUT_VARIABLE ( 4 ) ; // biases, [2*nU]
auto bc = INPUT_VARIABLE ( 5 ) ; // biases, [nU]
auto dLdr = INPUT_VARIABLE ( 6 ) ; // gradient wrt reset gate, [bS, nU]
auto dLdu = INPUT_VARIABLE ( 7 ) ; // gradient wrt update gate, [bS, nU]
auto dLdc = INPUT_VARIABLE ( 8 ) ; // gradient wrt cell state, [bS, nU]
auto dLdh = INPUT_VARIABLE ( 9 ) ; // gradient wrt current cell output, [bS, nU]
auto dLdx = OUTPUT_VARIABLE ( 0 ) ; // gradient wrt x, [bS, iS]
auto dLdhi = OUTPUT_VARIABLE ( 1 ) ; // gradient wrt hi, [bS, nU]
auto dLdW = OUTPUT_VARIABLE ( 2 ) ; // gradient wrt W, [iS+nU x 2*nU]
auto dLdWc = OUTPUT_VARIABLE ( 3 ) ; // gradient wrt Wc, [iS+nU x nU]
auto dLdb = OUTPUT_VARIABLE ( 4 ) ; // gradient wrt biases, [2*nU]
auto dLdbc = OUTPUT_VARIABLE ( 5 ) ; // gradient wrt c biases, [nU]
const Nd4jLong bS = x - > sizeAt ( 0 ) ;
const Nd4jLong iS = x - > sizeAt ( 1 ) ;
const Nd4jLong nU = hi - > sizeAt ( 1 ) ;
REQUIRE_TRUE ( x - > rankOf ( ) = = 2 , 0 , " GRU_CELL_BP: rank of input array x must be 2, but got %i instead " , x - > rankOf ( ) ) ;
const std : : string hiShape = ShapeUtils : : shapeAsString ( hi ) ;
const std : : string hiCorrectShape = ShapeUtils : : shapeAsString ( { bS , nU } ) ;
const std : : string wShape = ShapeUtils : : shapeAsString ( W ) ;
const std : : string wCorrectShape = ShapeUtils : : shapeAsString ( { iS + nU , 2 * nU } ) ;
const std : : string wcShape = ShapeUtils : : shapeAsString ( Wc ) ;
const std : : string wcCorrectShape = ShapeUtils : : shapeAsString ( { iS + nU , nU } ) ;
const std : : string bShape = ShapeUtils : : shapeAsString ( b ) ;
const std : : string bCorrectShape = ShapeUtils : : shapeAsString ( { 2 * nU } ) ;
const std : : string bcShape = ShapeUtils : : shapeAsString ( bc ) ;
const std : : string bcCorrectShape = ShapeUtils : : shapeAsString ( { nU } ) ;
const std : : string dLdrShape = ShapeUtils : : shapeAsString ( dLdr ) ;
const std : : string dLduShape = ShapeUtils : : shapeAsString ( dLdu ) ;
const std : : string dLdcShape = ShapeUtils : : shapeAsString ( dLdc ) ;
const std : : string dLdhShape = ShapeUtils : : shapeAsString ( dLdh ) ;
REQUIRE_TRUE ( hiShape = = hiCorrectShape , 0 , " GRU_CELL_BP op: wrong shape of previous cell output array, expected is %s, but got %s instead ! " , hiCorrectShape . c_str ( ) , hiShape . c_str ( ) ) ;
REQUIRE_TRUE ( wShape = = wCorrectShape , 0 , " GRU_CELL_BP op: wrong shape of weights array, expected is %s, but got %s instead ! " , wCorrectShape . c_str ( ) , wShape . c_str ( ) ) ;
REQUIRE_TRUE ( wcShape = = wcCorrectShape , 0 , " GRU_CELL_BP op: wrong shape of c weights array, expected is %s, but got %s instead ! " , wcCorrectShape . c_str ( ) , wcShape . c_str ( ) ) ;
REQUIRE_TRUE ( bShape = = bCorrectShape , 0 , " GRU_CELL_BP op: wrong shape of biases array, expected is %s, but got %s instead ! " , bCorrectShape . c_str ( ) , bShape . c_str ( ) ) ;
REQUIRE_TRUE ( bcShape = = bcCorrectShape , 0 , " GRU_CELL_BP op: wrong shape of c biases array, expected is %s, but got %s instead ! " , bcCorrectShape . c_str ( ) , bcShape . c_str ( ) ) ;
REQUIRE_TRUE ( dLdrShape = = hiCorrectShape , 0 , " GRU_CELL_BP op: wrong shape of dLdr array (gradient wrt reset gate), expected is %s, but got %s instead ! " , hiCorrectShape . c_str ( ) , dLdrShape . c_str ( ) ) ;
REQUIRE_TRUE ( dLduShape = = hiCorrectShape , 0 , " GRU_CELL_BP op: wrong shape of dLdu array (gradient wrt update gate), expected is %s, but got %s instead ! " , hiCorrectShape . c_str ( ) , dLduShape . c_str ( ) ) ;
REQUIRE_TRUE ( dLdcShape = = hiCorrectShape , 0 , " GRU_CELL_BP op: wrong shape of dLdc array (gradient wrt cell state), expected is %s, but got %s instead ! " , hiCorrectShape . c_str ( ) , dLdcShape . c_str ( ) ) ;
REQUIRE_TRUE ( dLdhShape = = hiCorrectShape , 0 , " GRU_CELL_BP op: wrong shape of dLdh array (gradient wrt current cell output), expected is %s, but got %s instead ! " , hiCorrectShape . c_str ( ) , dLdhShape . c_str ( ) ) ;
helpers : : gruCellBP ( block . launchContext ( ) , x , hi , W , Wc , b , bc , dLdr , dLdu , dLdc , dLdh , dLdx , dLdhi , dLdW , dLdWc , dLdb , dLdbc ) ;
2019-06-06 14:21:15 +02:00
return Status : : OK ( ) ;
}
DECLARE_TYPES ( gruCell_bp ) {
getOpDescriptor ( )
- > setAllowedInputTypes ( 0 , nd4j : : DataType : : ANY )
- > setAllowedInputTypes ( 1 , { ALL_FLOATS } )
- > setAllowedInputTypes ( 2 , { ALL_FLOATS } )
- > setAllowedInputTypes ( 3 , { ALL_FLOATS } )
- > setAllowedInputTypes ( 4 , { ALL_FLOATS } )
- > setAllowedInputTypes ( 5 , { ALL_FLOATS } )
- > setAllowedInputTypes ( 6 , { ALL_FLOATS } )
- > setAllowedInputTypes ( 7 , { ALL_FLOATS } )
- > setAllowedInputTypes ( 8 , { ALL_FLOATS } )
2019-07-20 07:58:44 +02:00
- > setAllowedInputTypes ( 9 , { ALL_FLOATS } )
2019-06-06 14:21:15 +02:00
- > setAllowedOutputTypes ( { ALL_FLOATS } ) ;
}
DECLARE_SHAPE_FN ( gruCell_bp ) {
2019-07-20 07:58:44 +02:00
auto xShapeInfo = inputShape - > at ( 0 ) ; // [bS x iS]
auto hiShapeInfo = inputShape - > at ( 1 ) ; // [bS x nU]
auto wShapeInfo = inputShape - > at ( 2 ) ; // [iS+nU x 2*nU]
auto wcShapeInfo = inputShape - > at ( 3 ) ; // [iS+nU x nU]
auto bShapeInfo = inputShape - > at ( 4 ) ; // [2*nU]
auto bcShapeInfo = inputShape - > at ( 5 ) ; // [nU]
auto dLdrShapeInfo = inputShape - > at ( 6 ) ; // [bS, nU]
auto dLduShapeInfo = inputShape - > at ( 7 ) ; // [bS, nU]
auto dLdcShapeInfo = inputShape - > at ( 8 ) ; // [bS, nU]
auto dLdhShapeInfo = inputShape - > at ( 9 ) ; // [bS, nU]
2019-06-06 14:21:15 +02:00
const int rank = xShapeInfo [ 0 ] ; // = 2
const Nd4jLong bS = xShapeInfo [ 1 ] ;
const Nd4jLong iS = xShapeInfo [ 2 ] ;
const Nd4jLong nU = hiShapeInfo [ 2 ] ;
2019-07-20 07:58:44 +02:00
REQUIRE_TRUE ( xShapeInfo [ 0 ] = = 2 , 0 , " GRU_CELL_BP: rank of input array x must be 2, but got %i instead " , xShapeInfo [ 0 ] ) ;
const std : : string hiShape = ShapeUtils : : shapeAsString ( hiShapeInfo ) ;
const std : : string hiCorrectShape = ShapeUtils : : shapeAsString ( { bS , nU } ) ;
const std : : string wShape = ShapeUtils : : shapeAsString ( wShapeInfo ) ;
const std : : string wCorrectShape = ShapeUtils : : shapeAsString ( { iS + nU , 2 * nU } ) ;
const std : : string wcShape = ShapeUtils : : shapeAsString ( wcShapeInfo ) ;
const std : : string wcCorrectShape = ShapeUtils : : shapeAsString ( { iS + nU , nU } ) ;
const std : : string bShape = ShapeUtils : : shapeAsString ( bShapeInfo ) ;
const std : : string bCorrectShape = ShapeUtils : : shapeAsString ( { 2 * nU } ) ;
const std : : string bcShape = ShapeUtils : : shapeAsString ( bcShapeInfo ) ;
const std : : string bcCorrectShape = ShapeUtils : : shapeAsString ( { nU } ) ;
const std : : string dLdrShape = ShapeUtils : : shapeAsString ( dLdrShapeInfo ) ;
const std : : string dLduShape = ShapeUtils : : shapeAsString ( dLduShapeInfo ) ;
const std : : string dLdcShape = ShapeUtils : : shapeAsString ( dLdcShapeInfo ) ;
const std : : string dLdhShape = ShapeUtils : : shapeAsString ( dLdhShapeInfo ) ;
REQUIRE_TRUE ( hiShape = = hiCorrectShape , 0 , " GRU_CELL_BP op: wrong shape of previous cell output array, expected is %s, but got %s instead ! " , hiCorrectShape . c_str ( ) , hiShape . c_str ( ) ) ;
REQUIRE_TRUE ( wShape = = wCorrectShape , 0 , " GRU_CELL_BP op: wrong shape of weights array, expected is %s, but got %s instead ! " , wCorrectShape . c_str ( ) , wShape . c_str ( ) ) ;
REQUIRE_TRUE ( wcShape = = wcCorrectShape , 0 , " GRU_CELL_BP op: wrong shape of c weights array, expected is %s, but got %s instead ! " , wcCorrectShape . c_str ( ) , wcShape . c_str ( ) ) ;
REQUIRE_TRUE ( bShape = = bCorrectShape , 0 , " GRU_CELL_BP op: wrong shape of biases array, expected is %s, but got %s instead ! " , bCorrectShape . c_str ( ) , bShape . c_str ( ) ) ;
REQUIRE_TRUE ( bcShape = = bcCorrectShape , 0 , " GRU_CELL_BP op: wrong shape of c biases array, expected is %s, but got %s instead ! " , bcCorrectShape . c_str ( ) , bcShape . c_str ( ) ) ;
REQUIRE_TRUE ( dLdrShape = = hiCorrectShape , 0 , " GRU_CELL_BP op: wrong shape of dLdr array (gradient wrt reset gate), expected is %s, but got %s instead ! " , hiCorrectShape . c_str ( ) , dLdrShape . c_str ( ) ) ;
REQUIRE_TRUE ( dLduShape = = hiCorrectShape , 0 , " GRU_CELL_BP op: wrong shape of dLdu array (gradient wrt update gate), expected is %s, but got %s instead ! " , hiCorrectShape . c_str ( ) , dLduShape . c_str ( ) ) ;
REQUIRE_TRUE ( dLdcShape = = hiCorrectShape , 0 , " GRU_CELL_BP op: wrong shape of dLdc array (gradient wrt cell state), expected is %s, but got %s instead ! " , hiCorrectShape . c_str ( ) , dLdcShape . c_str ( ) ) ;
REQUIRE_TRUE ( dLdhShape = = hiCorrectShape , 0 , " GRU_CELL_BP op: wrong shape of dLdh array (gradient wrt current cell output), expected is %s, but got %s instead ! " , hiCorrectShape . c_str ( ) , dLdhShape . c_str ( ) ) ;
2019-06-06 14:21:15 +02:00
Nd4jLong * dLdxShapeInfo = nullptr ;
COPY_SHAPE ( xShapeInfo , dLdxShapeInfo ) ;
Nd4jLong * dLdhiShapeInfo = nullptr ;
COPY_SHAPE ( hiShapeInfo , dLdhiShapeInfo ) ;
2019-07-20 07:58:44 +02:00
Nd4jLong * dLdWShapeInfo = nullptr ;
COPY_SHAPE ( wShapeInfo , dLdWShapeInfo ) ;
2019-06-06 14:21:15 +02:00
2019-07-20 07:58:44 +02:00
Nd4jLong * dLdWcShapeInfo = nullptr ;
COPY_SHAPE ( wcShapeInfo , dLdWcShapeInfo ) ;
2019-06-06 14:21:15 +02:00
Nd4jLong * dLdbShapeInfo = nullptr ;
COPY_SHAPE ( bShapeInfo , dLdbShapeInfo ) ;
2019-07-20 07:58:44 +02:00
Nd4jLong * dLdbcShapeInfo = nullptr ;
COPY_SHAPE ( bcShapeInfo , dLdbcShapeInfo ) ;
2019-06-06 14:21:15 +02:00
2019-08-14 17:28:25 +02:00
return SHAPELIST ( CONSTANT ( dLdxShapeInfo ) , CONSTANT ( dLdhiShapeInfo ) , CONSTANT ( dLdWShapeInfo ) , CONSTANT ( dLdWcShapeInfo ) , CONSTANT ( dLdbShapeInfo ) , CONSTANT ( dLdbcShapeInfo ) ) ;
2019-06-06 14:21:15 +02:00
}
}
}
# endif