raver119 3c4e959e21 [WIP] More of CUDA (#95)
* initial commit

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

* Implementation of hashcode cuda helper. Working edition.

* Fixed parallel test input arangements.

* Fixed tests for hashcode op.

* Fixed shape calculation for image:crop_and_resize op and test.

* NativeOps tests. Initial test suite.

* Added tests for indexReduce methods.

* Added test on execBroadcast with NDArray as dimensions.

* Added test on execBroadcastBool with NDArray as dimensions.

* Added tests on execPairwiseTransform and execPairwiseTransofrmBool.

* Added tests for execReduce with scalar results.

* Added reduce tests for non-empty dims array.

* Added tests for reduce3.

* Added tests for execScalar.

* Added tests for execSummaryStats.

* - provide cpu/cuda code for batch_to_space
- testing it

Signed-off-by: Yurii <yurii@skymind.io>

* - remove old test for batch_to_space (had wrong format and numbers were not checked)

Signed-off-by: Yurii <yurii@skymind.io>

* Fixed complilation errors with test.

* Added test for execTransformFloat.

* Added test for execTransformSame.

* Added test for execTransformBool.

* Added test for execTransformStrict.

* Added tests for execScalar/execScalarBool with TADs.

* Added test for flatten.

* - provide cpu/cuda code for space_to_Batch operaion

Signed-off-by: Yurii <yurii@skymind.io>

* Added test for concat.

* comment unnecessary stuff in s_t_b

Signed-off-by: Yurii <yurii@skymind.io>

* Added test for specialConcat.

* Added tests for memcpy/set routines.

* Fixed pullRow cuda test.

* Added pullRow test.

* Added average test.

* - correct typo in NDArray::applyPairwiseTransform(nd4j::pairwise::BoolOps op...)

Signed-off-by: Yurii <yurii@skymind.io>

* - debugging and fixing cuda tests in JavaInteropTests file

Signed-off-by: Yurii <yurii@skymind.io>

* - correct some tests

Signed-off-by: Yurii <yurii@skymind.io>

* Added test for shuffle.

* Fixed ops declarations.

* Restored omp and added shuffle test.

* Added convertTypes test.

* Added tests for execRandom. Eliminated usage of RandomBuffer with NativeOps.

* Added sort tests.

* Added tests for execCustomOp.

* - further debuging and fixing tests terminated with crash

Signed-off-by: Yurii <yurii@skymind.io>

* Added tests for calculateOutputShapes.

* Addded Benchmarks test.

* Commented benchmark tests.

* change assertion

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

* Added tests for apply_sgd op. Added cpu helper for that op.

* Implement cuda helper for aplly_sgd op. Fixed tests for NativeOps.

* Added test for assign broadcastable.

* Added tests for assign_bp op.

* Added tests for axpy op.

* - assign/execScalar/execTransformAny signature change
- minor test fix

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

* Fixed axpy op.

* meh

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

* - fix tests for nativeOps::concat

Signed-off-by: Yurii <yurii@skymind.io>

* sequential transform/scalar

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

* allow nested parallelism

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

* assign_bp leak fix

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

* block setRNG fix

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

* enable parallelism by default

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

* enable nested parallelism by default

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

* Added cuda implementation for row_count helper.

* Added implementation for tnse gains op helper.

* - take into account possible situations when input arrays are empty in reduce_ cuda stuff

Signed-off-by: Yurii <yurii@skymind.io>

* Implemented tsne/edge_forces op cuda-based helper. Parallelized cpu-based helper for edge_forces.

* Added kernel for tsne/symmetrized op heleper.

* Implementation of tsne/symmetrized op cuda helper. Working edition.

* Eliminated waste printfs.

* Added test for broadcastgradientargs op.

* host-only fallback for empty reduce float

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

* - some tests fixes

Signed-off-by: Yurii <yurii@skymind.io>

* - correct the rest of reduce_ stuff

Signed-off-by: Yurii <yurii@skymind.io>

* - further correction of reduce_ stuff

Signed-off-by: Yurii <yurii@skymind.io>

* Added test for Cbow op. Also added cuda implementation for cbow helpers.

* - improve code of stack operation for scalar case

Signed-off-by: Yurii <yurii@skymind.io>

* - provide cuda kernel for gatherND operation

Signed-off-by: Yurii <yurii@skymind.io>

* Implementation of cbow helpers with cuda kernels.

* minor tests tweaks

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

* minor tests tweaks

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

* - further correction of cuda stuff

Signed-off-by: Yurii <yurii@skymind.io>

* Implementatation of cbow op helper with cuda kernels. Working edition.

* Skip random testing for cudablas case.

* lstmBlockCell context fix

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

* Added tests for ELU and ELU_BP ops.

* Added tests for eq_scalar, gt_scalar, gte_scalar and lte_scalar ops.

* Added tests for neq_scalar.

* Added test for noop.

* - further work on clipbynorm_bp

Signed-off-by: Yurii <yurii@skymind.io>

* - get rid of concat op call, use instead direct concat helper call

Signed-off-by: Yurii <yurii@skymind.io>

* lstmBlockCell context fix

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

* Added tests for lrelu and lrelu_bp.

* Added tests for selu and selu_bp.

* Fixed lrelu derivative helpers.

* - some corrections in lstm

Signed-off-by: Yurii <yurii@skymind.io>

* operator * result shape fix

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

* - correct typo in lstmCell

Signed-off-by: Yurii <yurii@skymind.io>

* few tests fixed

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

* CUDA inverse broadcast bool fix

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

* disable MMAP test for CUDA

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

* BooleanOp syncToDevice

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

* meh

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

* additional data types for im2col/col2im

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

* Added test for firas_sparse op.

* one more RandomBuffer test excluded

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

* Added tests for flatten op.

* Added test for Floor op.

* bunch of tests fixed

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

* mmulDot tests fixed

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

* more tests fixed

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

* Implemented floordiv_bp op and tests.

* Fixed scalar case with cuda implementation for bds.

* - work on cuda kernel for clip_by_norm backprop op is completed

Signed-off-by: Yurii <yurii@skymind.io>

* Eliminate cbow crach.

* more tests fixed

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

* more tests fixed

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

* Eliminated abortion with batched nlp test.

* more tests fixed

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

* Fixed shared flag initializing.

* disabled bunch of cpu workspaces tests

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

* scalar operators fix: missing registerSpecialUse call

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

* Fixed logdet for cuda and tests.

* - correct clipBynorm_bp

Signed-off-by: Yurii <yurii@skymind.io>

* Fixed crop_and_resize shape datatype.

* - correct some mmul tests

Signed-off-by: Yurii <yurii@skymind.io>
2019-08-05 11:27:05 +10:00

141 lines
6.0 KiB
C++

/*******************************************************************************
* Copyright (c) 2015-2019 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, created on 14.02.2018
//
// implementation of operation for LSTM cell with peep hole connections:
// http://www.bioinf.jku.at/publications/older/2604.pdf
// S. Hochreiter and J. Schmidhuber. "Long Short-Term Memory". Neural Computation, 9(8):1735-1780, 1997.
// and
// https://research.google.com/pubs/archive/43905.pdf
// Hasim Sak, Andrew Senior, and Francoise Beaufays. "Long short-term memory recurrent neural network architectures for large scale acoustic modeling." INTERSPEECH, 2014.
#include <ops/declarable/helpers/lstm.h>
#include <VariableSpace.h>
#include <ops/declarable/CustomOperations.h>
#include <ops/declarable/helpers/transforms.h>
#include <ops/declarable/helpers/legacy_helpers.h>
#include <array/NDArrayList.h>
#include <iterator>
#include <MmulHelper.h>
namespace nd4j {
namespace ops {
namespace helpers {
/////////////////////////////////////////////////////////////////////////////
void lstmBlockTimeLoop(const NDArray* maxSeqLength, const NDArray* xSeq, const NDArray* c0, const NDArray* y0,
const NDArray* W, const NDArray* Wci, const NDArray* Wcf, const NDArray* Wco, const NDArray* b,
const NDArray* iSeq, const NDArray* cSeq, const NDArray* fSeq, const NDArray* oSeq, const NDArray* zSeq,
const NDArray* hSeq, const NDArray* ySeq, const std::vector<double>& params, const int dataFormat){
int seqLen, bS, nIn, nOut;
if(dataFormat == 0) {
seqLen = xSeq->sizeAt(0);
bS = xSeq->sizeAt(1);
nIn = xSeq->sizeAt(2);
nOut = iSeq->sizeAt(2);
}
else if(dataFormat == 1) {
seqLen = xSeq->sizeAt(2);
bS = xSeq->sizeAt(0);
nIn = xSeq->sizeAt(1);
nOut = iSeq->sizeAt(1);
}
else if(dataFormat == 2) {
seqLen = xSeq->sizeAt(1);
bS = xSeq->sizeAt(0);
nIn = xSeq->sizeAt(2);
nOut = iSeq->sizeAt(2);
}
const std::vector<Nd4jLong> inSliceShape({bS,nIn});
const std::vector<Nd4jLong> outSliceShape({bS,nOut});
auto c_t1 = const_cast<NDArray*>(c0);
auto y_t1 = const_cast<NDArray*>(y0);
// loop through time steps
for (int t = 0; t < seqLen; ++t) {
auto xt = timeSubset(xSeq, t, dataFormat);
auto it = timeSubset(iSeq, t, dataFormat);
auto ct = timeSubset(cSeq, t, dataFormat);
auto ft = timeSubset(fSeq, t, dataFormat);
auto ot = timeSubset(oSeq, t, dataFormat);
auto zt = timeSubset(zSeq, t, dataFormat);
auto ht = timeSubset(hSeq, t, dataFormat);
auto yt = timeSubset(ySeq, t, dataFormat);
helpers::lstmBlockCell(&xt, c_t1, y_t1, W, Wci, Wcf, Wco, b, &it, &ct, &ft, &ot, &zt, &ht, &yt, params);
if(t != 0) {
delete c_t1;
delete y_t1;
}
if(t < seqLen - 1) {
c_t1 = new NDArray(std::move(ct));
y_t1 = new NDArray(std::move(yt));
}
}
}
//////////////////////////////////////////////////////////////////////////
void lstmTimeLoop(nd4j::LaunchContext * context, const NDArray* x, const NDArray* h0, const NDArray* c0, const NDArray* Wx, const NDArray* Wh, const NDArray* Wc, const NDArray* Wp, const NDArray* b,
NDArray* h, NDArray* c, const std::vector<double>& params) {
// x input [time x bS x nIn]
// h0 initial cell output (at time step = 0) [bS x numProj], in case of projection=false -> numProj == numUnits !!!
// c0 initial cell state (at time step = 0) [bS x numUnits],
// Wx input-to-hidden weights, [nIn x 4*numUnits]
// Wh hidden-to-hidden weights, [numProj x 4*numUnits]
// Wc diagonal weights for peephole connections [3*numUnits]
// Wp projection weights [numUnits x numProj]
// b biases, [4*numUnits]
// h cell outputs [time x bS x numProj], that is per each time step
// c cell states [time x bS x numUnits] that is per each time step
const int time = x->sizeAt(0);
NDArray currentH(*h0);
NDArray currentC(*c0);
// loop through time steps
for (int t = 0; t < time; ++t) {
auto xt = (*x)({t,t+1, 0,0, 0,0});
auto ht = (*h)({t,t+1, 0,0, 0,0});
auto ct = (*c)({t,t+1, 0,0, 0,0});
helpers::lstmCell(context, &xt,&currentH,&currentC, Wx,Wh,Wc,Wp, b, &ht, &ct, params);
currentH.assign(ht);
currentC.assign(ct);
}
}
}
}
}