Alex Black 1170827c18 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 18:37:04 +03:00

276 lines
17 KiB
C++

/*******************************************************************************
* 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
******************************************************************************/
//
// @author Yurii Shyrma (iuriish@yahoo.com)
// @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) {
auto x = INPUT_VARIABLE(0); // input [bS x inSize]
auto hLast = INPUT_VARIABLE(1); // previous cell output [bS x numUnits], that is at previous time step t-1
auto Wru = INPUT_VARIABLE(2); // RU weights - [(nIn+nOut), 2*numUnits] - reset and update gates (input/recurrent weights)
auto Wc = INPUT_VARIABLE(3); // C weights - [(nIn+nOut), numUnits] - cell gate (input/recurrent weights)
auto bru = INPUT_VARIABLE(4); // reset and update biases, [2*numUnits] - reset and update gates
auto bc = INPUT_VARIABLE(5); // cell biases, [numUnits]
auto r = OUTPUT_VARIABLE(0); // Reset gate output [bS, numUnits]
auto u = OUTPUT_VARIABLE(1); // Update gate output [bS, numUnits]
auto c = OUTPUT_VARIABLE(2); // Cell gate output [bS, numUnits]
auto h = OUTPUT_VARIABLE(3); // current cell output [bS, numUnits]
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());
const int rank = x->rankOf();
const auto bS = x->sizeAt(0);
const auto nIn = x->sizeAt(1);
const auto nU = hLast->sizeAt(1);
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());
REQUIRE_TRUE(Wru->sizeAt(0)==(nIn+nU) && Wc->sizeAt(0)==(nIn+nU), 0, "gruCell: Weights size(0) must be equal to inSize + numUnits, 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*numUnits, got %i", Wru->sizeAt(1));
REQUIRE_TRUE(Wc->sizeAt(1)==nU, 0, "gruCell: Weights (cell) size(1) must be equal to numUnits, 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*numUnits");
REQUIRE_TRUE(bc->rankOf()==1 && bc->sizeAt(0)==nU, 0, "gruCell: cell biases must be rank 1, size numUnits");
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,
0, "gruCell: Output arrays must all be rank 2 with size(0) == batchSize and size(1) == numUnits");
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) {
auto x = inputShape->at(0); // input [bS x inSize]
auto hLast = inputShape->at(1); // previous cell output [bS x numUnits], that is at previous time step t-1
auto Wru = inputShape->at(2); // RU weights - [(nIn+nOut), 2*numUnits] - reset and update gates (input/recurrent weights)
auto Wc = inputShape->at(3); // C weights - [(nIn+nOut), numUnits] - cell gate (input/recurrent weights)
auto bru = inputShape->at(4); // reset and update biases, [2*numUnits] - reset and update gates
auto bc = inputShape->at(5); // cell biases, [numUnits]
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];
const auto inSize = x[2];
const auto numUnits = hLast[2];
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));
REQUIRE_TRUE(Wru[1]==(inSize+numUnits) && Wc[1]==(inSize+numUnits), 0, "gruCell: Weights size(0) must be equal to inSize + numUnits, got %i and %i", Wru[1], Wc[1]);
REQUIRE_TRUE(Wru[2]==(2*numUnits), 0, "gruCell: Weights (reset and update) size(1) must be equal to 2*numUnits, got %i", Wru[2]);
REQUIRE_TRUE(Wc[2]==numUnits, 0, "gruCell: Weights (cell) size(1) must be equal to numUnits, got %i", Wc[2]);
REQUIRE_TRUE(shape::rank(bru)==1 && bru[1]==(2*numUnits), 0, "gruCell: reset/update biases must be rank 1, size 2*numUnits");
REQUIRE_TRUE(shape::rank(bc)==1 && bc[1]==numUnits, 0, "gruCell: cell biases must be rank 1, size numUnits");
Nd4jLong *s0(nullptr);
ALLOCATE(s0, block.getWorkspace(), shape::shapeInfoLength(rank), Nd4jLong);// [bS x numUnits]
s0[0] = rank;
s0[1] = bS;
s0[2] = numUnits;
ShapeUtils::updateStridesAndType(s0, x, shape::order(hLast));
auto ts0 = ConstantShapeHelper::getInstance()->createFromExisting(s0, block.workspace());
//4 output shapes, all [bs, numUnits]
return SHAPELIST(ts0, ts0, ts0, ts0);
}
//////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL(gruCell_bp, 6, 5, false, 0, 0) {
auto x = INPUT_VARIABLE(0); // input [bS x iS]
auto hi = INPUT_VARIABLE(1); // previous cell output [bS x nU]
auto Wx = INPUT_VARIABLE(2); // input-to-hidden weights, [iS x 3*nU]
auto Wh = INPUT_VARIABLE(3); // hidden-to-hidden weights, [nU x 3*nU]
auto b = INPUT_VARIABLE(4); // biases, [3*nU]
auto dLdh = INPUT_VARIABLE(5); // gradient wrt output, [bS,nU], that is epsilon_next
auto dLdWxi = block.width() > 6 ? INPUT_VARIABLE(6) : nullptr; // gradient wrt Wx at previous time step, [iS, 3*nU]
auto dLdWhi = block.width() > 7 ? INPUT_VARIABLE(7) : nullptr; // gradient wrt Wh at previous time step, [nU, 3*nU]
auto dLdbi = block.width() > 8 ? INPUT_VARIABLE(8) : nullptr; // gradient wrt b at previous time step, [3*nU]
auto dLdx = OUTPUT_VARIABLE(0); // gradient wrt x, [bS, iS], that is epsilon
auto dLdhi = OUTPUT_VARIABLE(1); // gradient wrt hi, [bS, nU]
auto dLdWx = OUTPUT_VARIABLE(2); // gradient wrt Wx, [iS, 3*nU]
auto dLdWh = OUTPUT_VARIABLE(3); // gradient wrt Wh, [nU, 3*nU]
auto dLdb = OUTPUT_VARIABLE(4); // gradient wrt biases, [3*nU]
const int rank = x->rankOf(); // = 2
const Nd4jLong bS = x->sizeAt(0);
const Nd4jLong iS = x->sizeAt(1);
const Nd4jLong nU = hi->sizeAt(1);
const std::string hiShape = ShapeUtils::shapeAsString(hi);
const std::string hiCorrectShape = ShapeUtils::shapeAsString({bS, nU});
const std::string wxShape = ShapeUtils::shapeAsString(Wx);
const std::string wxCorrectShape = ShapeUtils::shapeAsString({iS, 3*nU});
const std::string whShape = ShapeUtils::shapeAsString(Wh);
const std::string whCorrectShape = ShapeUtils::shapeAsString({nU, 3*nU});
const std::string bShape = ShapeUtils::shapeAsString(b);
const std::string bCorrectShape = ShapeUtils::shapeAsString({3*nU});
const std::string dLdhShape = ShapeUtils::shapeAsString(dLdh);
const std::string dLdhCorrectShape = ShapeUtils::shapeAsString({bS, nU});
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(wxShape == wxCorrectShape, 0, "GRU_CELL_BP op: wrong shape of input-to-hidden weights array, expected is %s, but got %s instead !", wxCorrectShape.c_str(), wxShape.c_str());
REQUIRE_TRUE(whShape == whCorrectShape, 0, "GRU_CELL_BP op: wrong shape of hidden-to-hidden weights array, expected is %s, but got %s instead !", whCorrectShape.c_str(), whShape.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(dLdhShape == dLdhCorrectShape, 0, "GRU_CELL_BP op: wrong shape of dLdh array (epsilon_next), expected is %s, but got %s instead !", dLdhCorrectShape.c_str(), dLdhShape.c_str());
if(dLdWxi != nullptr) {
const std::string dLdWxiShape = ShapeUtils::shapeAsString(dLdWxi);
const std::string dLdWxiCorrectShape = ShapeUtils::shapeAsString({iS, 3*nU});
REQUIRE_TRUE(dLdWxiShape == dLdWxiCorrectShape, 0, "GRU_CELL_BP op: wrong shape of dLdWxi array (gradient wrt Wx at previous time step), expected is %s, but got %s instead !", dLdWxiCorrectShape.c_str(), dLdWxiShape.c_str());
}
if(dLdWhi != nullptr) {
const std::string dLdWhiShape = ShapeUtils::shapeAsString(dLdWhi);
const std::string dLdWhiCorrectShape = ShapeUtils::shapeAsString({nU, 3*nU});
REQUIRE_TRUE(dLdWhiShape == dLdWhiCorrectShape, 0, "GRU_CELL_BP op: wrong shape of dLdWhi array (gradient wrt Wh at previous time step), expected is %s, but got %s instead !", dLdWhiCorrectShape.c_str(), dLdWhiShape.c_str());
}
if(dLdbi != nullptr) {
const std::string dLdbiShape = ShapeUtils::shapeAsString(dLdbi);
const std::string dLdbiCorrectShape = ShapeUtils::shapeAsString({3*nU});
REQUIRE_TRUE(dLdbiShape == dLdbiCorrectShape, 0, "GRU_CELL_BP op: wrong shape of dLdbi array (gradient wrt biases at previous time step), expected is %s, but got %s instead !", dLdbiCorrectShape.c_str(), dLdbiShape.c_str());
}
helpers::gruCellBP(block.launchContext(), x, hi, Wx, Wh, b, dLdh, dLdWxi, dLdWhi, dLdbi, dLdx, dLdhi, dLdWx, dLdWh, dLdb);
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})
->setAllowedOutputTypes({ALL_FLOATS});
}
DECLARE_SHAPE_FN(gruCell_bp) {
auto xShapeInfo = inputShape->at(0); // [bS x iS]
auto hiShapeInfo = inputShape->at(1); // [bS x nU]
auto wxShapeInfo = inputShape->at(2); // [iS x 3*nU]
auto whShapeInfo = inputShape->at(3); // [nU x 3*nU]
auto bShapeInfo = inputShape->at(4); // [3*nU]
auto dLdhShapeInfo = inputShape->at(5); // [bS x nU]
const int rank = xShapeInfo[0]; // = 2
const Nd4jLong bS = xShapeInfo[1];
const Nd4jLong iS = xShapeInfo[2];
const Nd4jLong nU = hiShapeInfo[2];
const std::string hiShape = ShapeUtils::shapeAsString(hiShapeInfo);
const std::string hiCorrectShape = ShapeUtils::shapeAsString({bS, nU});
const std::string wxShape = ShapeUtils::shapeAsString(wxShapeInfo);
const std::string wxCorrectShape = ShapeUtils::shapeAsString({iS, 3*nU});
const std::string whShape = ShapeUtils::shapeAsString(whShapeInfo);
const std::string whCorrectShape = ShapeUtils::shapeAsString({nU, 3*nU});
const std::string bShape = ShapeUtils::shapeAsString(bShapeInfo);
const std::string bCorrectShape = ShapeUtils::shapeAsString({3*nU});
const std::string dLdhShape = ShapeUtils::shapeAsString(dLdhShapeInfo);
const std::string dLdhCorrectShape = ShapeUtils::shapeAsString({bS, nU});
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(wxShape == wxCorrectShape, 0, "GRU_CELL_BP op: wrong shape of input-to-hidden weights array, expected is %s, but got %s instead !", wxCorrectShape.c_str(), wxShape.c_str());
REQUIRE_TRUE(whShape == whCorrectShape, 0, "GRU_CELL_BP op: wrong shape of hidden-to-hidden weights array, expected is %s, but got %s instead !", whCorrectShape.c_str(), whShape.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(dLdhShape == dLdhCorrectShape, 0, "GRU_CELL_BP op: wrong shape of dLdh array (epsilon_next), expected is %s, but got %s instead !", dLdhCorrectShape.c_str(), dLdhShape.c_str());
if(block.width() > 6) {
Nd4jLong* dLdWxiShapeInfo = inputShape->at(6); // [iS x 3*nU]
const std::string dLdWxiShape = ShapeUtils::shapeAsString(dLdWxiShapeInfo);
const std::string dLdWxiCorrectShape = ShapeUtils::shapeAsString({iS, 3*nU});
REQUIRE_TRUE(dLdWxiShape == dLdWxiCorrectShape, 0, "GRU_CELL_BP op: wrong shape of dLdWxi array (gradient wrt Wx at previous time step), expected is %s, but got %s instead !", dLdWxiCorrectShape.c_str(), dLdWxiShape.c_str());
}
if(block.width() > 7) {
Nd4jLong* dLdWhiShapeInfo = inputShape->at(7); // [nU x 3*nU]
const std::string dLdWhiShape = ShapeUtils::shapeAsString(dLdWhiShapeInfo);
const std::string dLdWhiCorrectShape = ShapeUtils::shapeAsString({nU, 3*nU});
REQUIRE_TRUE(dLdWhiShape == dLdWhiCorrectShape, 0, "GRU_CELL_BP op: wrong shape of dLdWhi array (gradient wrt Wh at previous time step), expected is %s, but got %s instead !", dLdWhiCorrectShape.c_str(), dLdWhiShape.c_str());
}
if(block.width() > 8) {
Nd4jLong* dLdbiShapeInfo = inputShape->at(8); // [3*nU]
const std::string dLdbiShape = ShapeUtils::shapeAsString(dLdbiShapeInfo);
const std::string dLdbiCorrectShape = ShapeUtils::shapeAsString({3*nU});
REQUIRE_TRUE(dLdbiShape == dLdbiCorrectShape, 0, "GRU_CELL_BP op: wrong shape of dLdbi array (gradient wrt biases at previous time step), expected is %s, but got %s instead !", dLdbiCorrectShape.c_str(), dLdbiShape.c_str());
}
Nd4jLong *dLdxShapeInfo = nullptr;
COPY_SHAPE(xShapeInfo, dLdxShapeInfo);
Nd4jLong *dLdhiShapeInfo = nullptr;
COPY_SHAPE(hiShapeInfo, dLdhiShapeInfo);
Nd4jLong *dLdWxShapeInfo = nullptr;
COPY_SHAPE(wxShapeInfo, dLdWxShapeInfo);
Nd4jLong *dLdWhShapeInfo = nullptr;
COPY_SHAPE(whShapeInfo, dLdWhShapeInfo);
Nd4jLong *dLdbShapeInfo = nullptr;
COPY_SHAPE(bShapeInfo, dLdbShapeInfo);
return SHAPELIST(dLdxShapeInfo, dLdhiShapeInfo, dLdWxShapeInfo, dLdWhShapeInfo, dLdbShapeInfo);
}
}
}
#endif