Yurii Shyrma 5d9b2a16e5 Shyrma temp (#131)
* - specifying template instantiation for certain types in float16 and bloat16

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

* - polishing bfloat16 and float16 member functions template specialization

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

* - rewrite and overload array +-*/ scalar and scalar +-*/ arr in NDAray class

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

* - make corrections which have to do with and rvalue lvalue conversions

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

* - provide move semantic in NDArray operators array +-/* array

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

* float16/bfloat16 tweaks

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

* one more tweak

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

* - make float16 and bfloat16 to compile successfully on cuda

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

* - do not use resources of view-like arrays when move semantics is applied

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

* - get rid of pointers in signatures NDArray methods 1

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

* - correction of signature of NDArray::dup method

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

* - correction of signature of NDArray::reduceAlongDimension method

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

* - signature correction of NDArray::applyIndexReduce and applyTrueBroadcast methods

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

* - signature correction of NDArray::applyReduce3 and varianceAlongDimension methods

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

* - signature correction of NDArray::tensorsAlongDimension and diagonal methods

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

* - signature correction of NDArray::allTensorsAlongDimension

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

* - signature correction of NDArray::reduceAlongDimension 2

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

* - signature correction of NDArray::applyTransform 2

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

* - signature correction of NDArray::applyPairwiseTransform 2

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

* - signature correction of NDArray::applyBroadcast 2

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

* - signature correction of NDArray::applyTrueBroadcast 2

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

* - signature correction of NDArray::applyScalar and applyScalarArr

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

* - signature correction of NDArray::lambda methods

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

* - signature correction of NDArray::reduce3 methods 2

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

* - signature correction of following NDArray methods: add/sub/mul/div row/column and fillAsTriangular

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

* - signature correction of NDArray::tileToShape methods

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

* - signature correction of NDArray::isShapeSameStrict method

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

* minor corrections in tests

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

* - replace reduce op in batchnorm mkldnn

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

* - add explicit templates instantiations for operator+(NDArray&&. const scalar)

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

* - corrections of casts in float16/bfloat16

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

* - provide move semantics in following NDArray methods: transform, applyTrueBroadcast, transpose, reshape, permute

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

* - get rid of input array A duplicate in svd cuda op

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

* - avoid available bug in svd cuda API

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

* - add temporary global memory buffer in svd cuda when calcUV = false and  m != n

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

* - remove test with blfoat16 type for betainC

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

* - resolve conflicts after master has been merged in

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

* - changed type of affected input array in fused_batch_norm

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

* - add several explicit type castings

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

* - add ND4J_EXPORT to operators

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

* - add explicit template types in instantiations of template arithm operators of NDArray class

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

* - one more test fix

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

Co-authored-by: raver119 <raver119@gmail.com>
2019-12-20 22:35:39 +03:00

116 lines
4.1 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 raver119@gmail.com
//
#include <op_boilerplate.h>
#if NOT_EXCLUDED(OP_crelu)
#include <ops/declarable/CustomOperations.h>
#include<ops/declarable/helpers/transforms.h>
#include <ops/declarable/helpers/legacy_helpers.h>
namespace nd4j {
namespace ops {
CUSTOM_OP_IMPL(crelu, 1, 1, false, 0, 0) {
auto x = INPUT_VARIABLE(0);
REQUIRE_TRUE(x->isR(), 0, "CRELU: input must be real type");
auto tmp = x->dup();
tmp.applyTransform(nd4j::transform::Neg, tmp);
auto z = OUTPUT_VARIABLE(0);
helpers::concat(block.launchContext(), {x, &tmp}, *z, x->rankOf()-1);
// NDArrayFactory<T>::concat({x, tmp}, -1, z);
// TODO: make this configurable?
double threshold = 0.0;
z->applyScalar(nd4j::scalar::RELU, threshold, *z);
STORE_RESULT(z);
return Status::OK();
}
DECLARE_TYPES(crelu) {
getOpDescriptor()
->setAllowedInputTypes(0, DataType::ANY)
->setSameMode(true);
}
DECLARE_SHAPE_FN(crelu) {
auto inShape = inputShape->at(0);
std::vector<Nd4jLong> shape;
for (int e = 0; e < shape::rank(inShape); e++)
shape.emplace_back(shape::shapeOf(inShape)[e]);
shape[shape.size()-1] *= 2;
auto newShape = ConstantShapeHelper::getInstance()->createShapeInfo(ArrayOptions::dataType(inShape), shape::order(inShape), shape);
return SHAPELIST(newShape);
}
CUSTOM_OP_IMPL(crelu_bp, 2, 1, false, 0, 0) {
auto input = INPUT_VARIABLE(0);
auto epsilonNext = INPUT_VARIABLE(1);
auto epsilon = OUTPUT_VARIABLE(0);
// at first step we build fwd activation
nd4j::ops::crelu op;
auto tmpResult = op.execute({input}, {}, {}, {});
if (tmpResult->status() != ND4J_STATUS_OK)
return tmpResult->status();
auto actv = tmpResult->at(0);
// now we do RELU backward pass
//actv->applyPairwiseTransform(pairwise::RELUDerivativeE, *epsilon, nullptr);
helpers::reluDerivative(block.launchContext(), actv, epsilonNext);
// now we split updated array into 2 chunks along last dimension
nd4j::ops::concat_bp opc;
auto dec = opc.execute({input, input, actv}, {}, {-1}, {});
if (dec->status() != ND4J_STATUS_OK)
return dec->status();
// and now we subtract two parts of epsilons and pass result out
auto pos = dec->at(0);
auto neg = dec->at(1);
pos->applyPairwiseTransform(nd4j::pairwise::Subtract, *neg, *epsilon);
delete tmpResult;
delete dec;
return ND4J_STATUS_OK;
}
DECLARE_TYPES(crelu_bp) {
getOpDescriptor()
->setAllowedInputTypes(0, DataType::ANY)
->setAllowedInputTypes(1, {DataType::FLOAT32, DataType ::DOUBLE, DataType::HALF})
->setAllowedOutputTypes(0, {DataType::FLOAT32, DataType ::DOUBLE, DataType::HALF});
}
DECLARE_SHAPE_FN(crelu_bp) {
auto inShape = inputShape->at(0);
return SHAPELIST(ConstantShapeHelper::getInstance()->createShapeInfo(ShapeDescriptor(inShape)));
}
}
}
#endif