wehub/deeplearning4j--deeplearning4j
1
0
Fork 0
Code Issues Pull Requests Actions Deployments Models Agent Notes Packages Projects Releases Wiki Activity

chore: import upstream snapshot with attribution

Browse Source
This commit is contained in:
wehub-resource-sync
2026-07-13 12:47:05 +08:00
commit 4f3b7da785
7394 changed files with 2005594 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
libnd4j/include/ops/declarable/generic/nn/pooling/avgpool2d.cpp
+227
View File
@@ -0,0 +1,227 @@
/* ******************************************************************************
*
*
* 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.
*
* See the NOTICE file distributed with this work for additional
* information regarding copyright ownership.
* 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, created on 29/10/17.
// @author Yurii Shyrma (iuriish@yahoo.com), changed on 14.05.2018
//
#include <system/op_boilerplate.h>
#if NOT_EXCLUDED(OP_avgpool2d)
#include <ops/declarable/CustomOperations.h>
#include <ops/declarable/helpers/convolutions.h>
namespace sd {
namespace ops {
CUSTOM_OP_IMPL(avgpool2d, 1, 1, false, 0, 10) {
auto input = INPUT_VARIABLE(0);
auto output = OUTPUT_NULLIFIED(0);
// 0,1 - kernel Height/Width; 2,3 - stride Height/Width; 4,5 - pad Height/Width; 6,7 - dilation Height/Width; 8 - same
// mode;
const LongType kH = INT_ARG(0);
const LongType kW = INT_ARG(1);
const LongType sH = INT_ARG(2);
const LongType sW = INT_ARG(3);
LongType pH = INT_ARG(4);
LongType pW = INT_ARG(5);
const LongType dH = INT_ARG(6);
const LongType dW = INT_ARG(7);
const auto isSameMode = static_cast<bool>(INT_ARG(8));
const auto extraParam0 = INT_ARG(9);
const int isNCHW = block.getIArguments()->size() > 10 ? !INT_ARG(10) : 1; // INT_ARG(10): 0-NCHW, 1-NHWC
REQUIRE_TRUE(input->rankOf() == 4, 0, "AVGPOOL2D op: input should have rank of 4, but got %i instead",
input->rankOf());
REQUIRE_TRUE(dH != 0 && dW != 0, 0, "AVGPOOL2D op: dilation must not be zero, but got instead {%i, %i}", dH, dW);
LongType oH = 0;
LongType oW = 0;
const LongType iH = static_cast<LongType>(isNCHW ? input->sizeAt(2) : input->sizeAt(1));
const LongType iW = static_cast<LongType>(isNCHW ? input->sizeAt(3) : input->sizeAt(2));
if (!isNCHW) {
std::vector<sd::LongType> perm = {0,3,1,2};
input = input->permute(perm, false, false); // [bS, iH, iW, iC] -> [bS, iC, iH, iW] - permute() already returns NDArray*
output = output->permute(perm, false, false); // [bS, oH, oW, iC] -> [bS, iC, oH, oW] - permute() already returns NDArray*
}
ConvolutionUtils::calcOutSizePool2D(oH, oW, kH, kW, sH, sW, pH, pW, dH, dW, iH, iW, isSameMode);
if (isSameMode) ConvolutionUtils::calcPadding2D(pH, pW, oH, oW, iH, iW, kH, kW, sH, sW, dH, dW);
// 0,1 - kernel Height/Width; 2,3 - stride Height/Width; 4,5 - pad Height/Width; 6,7 - dilation Height/Width; 8 -
// poolingMode; 9 - divisor;
ConvolutionUtils::pooling2d(block, *input, *output, kH, kW, sH, sW, pH, pW, dH, dW, AVG_POOL,
extraParam0);
if (!isNCHW) {
delete input;
delete output;
}
return Status::OK;
}
DECLARE_SYN(AvgPool2D, avgpool2d);
DECLARE_SYN(AvgPool, avgpool2d);
DECLARE_SYN(avgpool, avgpool2d);
DECLARE_TYPES(avgpool2d) {
getOpDescriptor()->setAllowedInputTypes(ANY)->setAllowedOutputTypes({ALL_FLOATS});
}
DECLARE_SHAPE_FN(avgpool2d) {
auto inShape = inputShape->at(0);
auto shapeOf = shape::shapeOf(inShape);
// 0,1 - kernel Height/Width; 2,3 - stride Height/Width; 4,5 - pad Height/Width; 6,7 - dilation Height/Width; 8 - same
// mode;
const LongType kH = INT_ARG(0);
const LongType kW = INT_ARG(1);
const LongType sH = INT_ARG(2);
const LongType sW = INT_ARG(3);
const LongType pH = INT_ARG(4);
const LongType pW = INT_ARG(5);
const LongType dH = INT_ARG(6);
const LongType dW = INT_ARG(7);
const int isSameMode = INT_ARG(8);
const int isNCHW = block.getIArguments()->size() > 10 ? !INT_ARG(10) : 1; // INT_ARG(10): 0-NCHW, 1-NHWC
REQUIRE_TRUE(dH != 0 && dW != 0, 0, "AVGPOOL2D op: dilation must not be zero, but got instead {%i, %i}", dH, dW);
const LongType bS = shapeOf[0];
const LongType iD = isNCHW ? shapeOf[1] : shapeOf[3];
const LongType iH = isNCHW ? shapeOf[2] : shapeOf[1];
const LongType iW = isNCHW ? shapeOf[3] : shapeOf[2];
const char order = shape::order(inShape); // output order must be equal to input order
// calculate output Height/Width
LongType oH, oW;
ConvolutionUtils::calcOutSizePool2D(oH, oW, kH, kW, sH, sW, pH, pW, dH, dW, iH, iW, isSameMode);
// allocate memory for new shape
LongType *newShape = new LongType[4];
if (isNCHW) {
newShape[0] = bS;
newShape[1] = iD;
newShape[2] = oH;
newShape[3] = oW;
} else {
newShape[0] = bS;
newShape[1] = oH;
newShape[2] = oW;
newShape[3] = iD;
}
auto ret = SHAPELIST(ConstantShapeHelper::getInstance().bufferForShapeInfo(ArrayOptions::dataType(inShape),
shape::order(inShape),
4,
newShape)->primary());
delete[] newShape;
return ret;
}
DECLARE_TYPES(avgpool2d_bp) {
getOpDescriptor()->setAllowedInputTypes(ANY)->setAllowedOutputTypes({ALL_FLOATS});
}
//////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL(avgpool2d_bp, 2, 1, false, 0, 10) {
auto input = INPUT_VARIABLE(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW)
auto gradO = INPUT_VARIABLE(1); // [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCHW), epsilon_next
auto gradI = OUTPUT_NULLIFIED(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW), epsilon
LongType kH = INT_ARG(0); // filter(kernel) height
LongType kW = INT_ARG(1); // filter(kernel) width
LongType sH = INT_ARG(2); // strides height
LongType sW = INT_ARG(3); // strides width
LongType pH = INT_ARG(4); // paddings height
LongType pW = INT_ARG(5); // paddings width
LongType dH = INT_ARG(6); // dilations height
LongType dW = INT_ARG(7); // dilations width
int isSameMode = INT_ARG(8); // 0-VALID, 1-SAME
int extraParam0 = INT_ARG(9);
int isNCHW = block.getIArguments()->size() > 10 ? !INT_ARG(10) : 1; // INT_ARG(10): 0-NCHW, 1-NHWC
REQUIRE_TRUE(input->rankOf() == 4, 0, "AVGPOOL2D_BP op: input should have rank of 4, but got %i instead",
input->rankOf());
REQUIRE_TRUE(dH != 0 && dW != 0, 0, "AVGPOOL2D_BP op: dilation must not be zero, but got instead {%i, %i}", dH, dW);
LongType bS, iC, iH, iW, oC, oH,
oW; // batch size, input channels, input height/width, output channels, output height/width;
LongType indIOioC, indIiH, indWoC, indWiC, indWkH, indOoH; // corresponding indexes
ConvolutionUtils::getSizesAndIndexesConv2d(isNCHW, 0, *input, *gradO, bS, iC, iH, iW, oC, oH, oW, indIOioC, indIiH,
indWiC, indWoC, indWkH, indOoH);
std::vector<LongType> expectedGradOShape =
ShapeUtils::composeShapeUsingDimsAndIdx({bS, iC, oH, oW, 0, indIOioC, indIiH, indIiH + 1});
std::vector<LongType> expectedGradIShape =
ShapeUtils::composeShapeUsingDimsAndIdx({bS, iC, iH, iW, 0, indIOioC, indIiH, indIiH + 1});
REQUIRE_TRUE(
gradO->isSameShape(expectedGradOShape), 0,
"AVGPOOL2D_BP op: wrong shape of output's gradients array (next epsilon), expected is %s, but got %s instead !",
ShapeUtils::shapeAsString(expectedGradOShape).c_str(), ShapeUtils::shapeAsString(gradO).c_str());
REQUIRE_TRUE(
gradI->isSameShape(expectedGradIShape), 0,
"AVGPOOL2D_BP op: wrong shape of input's gradients array (epsilon), expected is %s, but got %s instead !",
ShapeUtils::shapeAsString(expectedGradIShape).c_str(), ShapeUtils::shapeAsString(gradI).c_str());
if (!isNCHW) {
std::vector<sd::LongType> perm = {0,3,1,2};
input = input->permute(perm, false, false); // [bS, iH, iW, iC] -> [bS, iC, iH, iW] - permute() already returns NDArray*
gradI = gradI->permute(perm, false, false); // [bS, iH, iW, iC] -> [bS, iC, iH, iW] - permute() already returns NDArray*
gradO = gradO->permute(perm, false, false); // [bS, oH, oW, iC] -> [bS, iC, oH, oW] - permute() already returns NDArray*
}
if (isSameMode) // SAME
ConvolutionUtils::calcPadding2D(pH, pW, oH, oW, iH, iW, kH, kW, sH, sW, dH, dW);
// 0,1 - kernel Height/Width; 2,3 - stride Height/Width; 4,5 - pad Height/Width; 6,7 - dilation Height/Width; 8 -
// poolingMode; 9 - divisor;
ConvolutionUtils::pooling2dBP(block, *input, *gradO, *gradI, kH, kW, sH, sW, pH, pW, dH, dW, 1, extraParam0);
if (!isNCHW) {
delete input;
delete gradI;
delete gradO;
}
return Status::OK;
}
DECLARE_SHAPE_FN(avgpool2d_bp) {
REQUIRE_TRUE(inputShape->at(0)[0] == 4, 0, "AVGPOOL2D_BP op: input array must be 4D, but got %i instead!",
inputShape->at(0)[0]);
REQUIRE_TRUE(inputShape->at(1)[0] == 4, 0,
"AVGPOOL2D_BP op: output's gradient array (next epsilon) must be 4D, but got %i instead!",
inputShape->at(1)[0]);
auto desc = new ShapeDescriptor(inputShape->at(0), ArrayOptions::dataType(inputShape->at(1)), false);
return SHAPELIST(ConstantShapeHelper::getInstance().createShapeInfo(desc));
}
} // namespace ops
} // namespace sd
#endif
libnd4j/include/ops/declarable/generic/nn/pooling/avgpool3d.cpp
+241
View File
@@ -0,0 +1,241 @@
/* ******************************************************************************
*
*
* 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.
*
* See the NOTICE file distributed with this work for additional
* information regarding copyright ownership.
* 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), created on 01.03.2018
//
#include <system/op_boilerplate.h>
#if NOT_EXCLUDED(OP_avgpool3dnew)
#include <ops/declarable/CustomOperations.h>
#include <ops/declarable/helpers/convolutions.h>
namespace sd {
namespace ops {
//////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL(avgpool3dnew, 1, 1, false, 0, 14) {
auto input = INPUT_VARIABLE(0); // [bS, iD, iH, iW, iC] (NDHWC) or [bS, iC, iD, iH, iW] (NCDHW)
auto output = OUTPUT_NULLIFIED(0); // [bS, oD, oH, oW, iC] (NDHWC) or [bS, iC, oD, oH, oW] (NCDHW)
LongType kD = INT_ARG(0); // filter(kernel) depth
LongType kH = INT_ARG(1); // filter(kernel) height
LongType kW = INT_ARG(2); // filter(kernel) width
LongType sD = INT_ARG(3); // strides depth
LongType sH = INT_ARG(4); // strides height
LongType sW = INT_ARG(5); // strides width
LongType pD = INT_ARG(6); // paddings depth
LongType pH = INT_ARG(7); // paddings height
LongType pW = INT_ARG(8); // paddings width
LongType dD = INT_ARG(9); // dilations depth
LongType dH = INT_ARG(10); // dilations height
LongType dW = INT_ARG(11); // dilations width
int isSameMode = INT_ARG(12); // 1-SAME, 0-VALID
int extraParam0 = INT_ARG(13);
int isNCDHW = block.getIArguments()->size() > 14 ? !INT_ARG(14) : 1; // 0-NCDHW, 1-NDHWC
REQUIRE_TRUE(input->rankOf() == 5, 0, "AVGPOOL3DNEW OP: rank of input array must be equal to 5, but got %i instead !",
input->rankOf());
REQUIRE_TRUE(dD != 0 && dH != 0 && dW != 0, 0,
"AVGPOOL3DNEW OP: dilation must not be zero, but got instead {%i, %i, %i}", dD, dH, dW);
LongType bS, iC, iD, iH, iW, oC, oD, oH,
oW; // batch size, input channels, input depth/height/width, output channels, output depth/height/width;
LongType indIOioC, indIOioD, indWoC, indWiC, indWkD; // corresponding indexes
ConvolutionUtils::getSizesAndIndexesConv3d(isNCDHW, 0, *input, *output, bS, iC, iD, iH, iW, oC, oD, oH, oW, indIOioC,
indIOioD, indWiC, indWoC, indWkD);
std::vector<LongType> expectedOutputShape =
ShapeUtils::composeShapeUsingDimsAndIdx({bS, iC, oD, oH, oW, 0, indIOioC, indIOioD, indIOioD + 1, indIOioD + 2});
REQUIRE_TRUE(output->isSameShape(expectedOutputShape), 0,
"AVGPOOL3DNEW OP: wrong shape of output array, expected is %s, but got %s instead !",
ShapeUtils::shapeAsString(expectedOutputShape).c_str(), ShapeUtils::shapeAsString(output).c_str());
if (!isNCDHW) {
std::vector<sd::LongType> perm = {0, 4, 1, 2, 3};
input = input->permute(perm, false, false); // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
output =output->permute(perm, false, false); // [bS, oD, oH, oW, iC] -> [bS, iC, oD, oH, oW]
}
if (isSameMode) // SAME
ConvolutionUtils::calcPadding3D(pD, pH, pW, oD, oH, oW, iD, iH, iW, kD, kH, kW, sD, sH, sW, dD, dH, dW);
// T extraParams[] = {};
ConvolutionUtils::pooling3d(block, *input, *output, kD, kH, kW, sD, sH, sW, pD, pH, pW, dD, dH, dW, 1, extraParam0);
if (!isNCDHW) {
delete input;
delete output;
}
return Status::OK;
}
DECLARE_TYPES(avgpool3dnew) {
getOpDescriptor()->setAllowedInputTypes(ANY)->setAllowedOutputTypes({ALL_FLOATS});
}
DECLARE_SHAPE_FN(avgpool3dnew) {
LongType kD = INT_ARG(0); // filter(kernel) depth
LongType kH = INT_ARG(1); // filter(kernel) height
LongType kW = INT_ARG(2); // filter(kernel) width
LongType sD = INT_ARG(3); // strides depth
LongType sH = INT_ARG(4); // strides height
LongType sW = INT_ARG(5); // strides width
LongType pD = INT_ARG(6); // paddings depth
LongType pH = INT_ARG(7); // paddings height
LongType pW = INT_ARG(8); // paddings width
LongType dD = INT_ARG(9); // dilations depth
LongType dH = INT_ARG(10); // dilations height
LongType dW = INT_ARG(11); // dilations width
int isSameMode = INT_ARG(12); // 1-SAME, 0-VALID
int isNCDHW = block.getIArguments()->size() > 14 ? !INT_ARG(14) : 1; // 0-NCDHW, 1-NDHWC
REQUIRE_TRUE(dD != 0 && dH != 0 && dW != 0, 0,
"AVGPOOL3DNEW op: dilation must not be zero, but got instead {%i, %i, %i}", dD, dH, dW);
auto inputShapeInfo = inputShape->at(0);
LongType idxID, idxIC;
if (isNCDHW) {
idxID = 2;
idxIC = 1;
} else {
idxID = 1;
idxIC = 4;
}
LongType bS = inputShapeInfo[1]; // batch size
LongType iC = inputShapeInfo[idxIC + 1]; // input channels
LongType iD = inputShapeInfo[idxID + 1]; // input depth
LongType iH = inputShapeInfo[idxID + 2]; // input height
LongType iW = inputShapeInfo[idxID + 3]; // input width
LongType oD, oH, oW; // output depth, height, width
ConvolutionUtils::calcOutSizePool3D(oD, oH, oW, kD, kH, kW, sD, sH, sW, pD, pH, pW, dD, dH, dW, iD, iH, iW,
isSameMode);
LongType outputShape[5];
outputShape[0] = bS;
if (isNCDHW) {
outputShape[1] = iC;
outputShape[2] = oD;
outputShape[3] = oH;
outputShape[4] = oW;
} else {
outputShape[1] = oD;
outputShape[2] = oH;
outputShape[3] = oW;
outputShape[4] = iC;
}
// TF DOC: A Tensor. Has the same type as input.
// TF DOC: A Tensor. Has the same type as input.
auto ret = SHAPELIST(ConstantShapeHelper::getInstance().bufferForShapeInfo(ArrayOptions::dataType(inputShapeInfo),
shape::order(inputShapeInfo),
5,
outputShape)->primary());
return ret;
}
DECLARE_TYPES(avgpool3dnew_bp) {
getOpDescriptor()->setAllowedInputTypes(ANY)->setAllowedOutputTypes({ALL_FLOATS});
}
//////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL(avgpool3dnew_bp, 2, 1, false, 0, 14) {
auto input = INPUT_VARIABLE(0); // [bS, iD, iH, iW, iC] (NDHWC) or [bS, iC, iD, iH, iW] (NCDHW)
auto gradO = INPUT_VARIABLE(1); // [bS, oD, oH, oW, oC] (NDHWC) or [bS, oC, oD, oH, oW] (NCDHW), epsilon_next
auto gradI = OUTPUT_NULLIFIED(0); // [bS, iD, iH, iW, iC] (NDHWC) or [bS, iC, iD, iH, iW] (NCDHW), epsilon
const LongType kD = INT_ARG(0); // filter(kernel) depth
const LongType kH = INT_ARG(1); // filter(kernel) height
const LongType kW = INT_ARG(2); // filter(kernel) width
const LongType sD = INT_ARG(3); // strides depth
const LongType sH = INT_ARG(4); // strides height
const LongType sW = INT_ARG(5); // strides width
LongType pD = INT_ARG(6); // paddings depth
LongType pH = INT_ARG(7); // paddings height
LongType pW = INT_ARG(8); // paddings width
const LongType dD = INT_ARG(9); // dilations depth
const LongType dH = INT_ARG(10); // dilations height
const LongType dW = INT_ARG(11); // dilations width
const int isSameMode = INT_ARG(12); // 1-SAME, 0-VALID
const int extraParam0 = INT_ARG(13); // define what divisor to use while averaging
const int isNCDHW = block.getIArguments()->size() > 14 ? !INT_ARG(14) : 1; // 0-NCDHW, 1-NDHWC
REQUIRE_TRUE(input->rankOf() == 5, 0, "AVGPOOL3DNEW_BP op: input should have rank of 5, but got %i instead",
input->rankOf());
REQUIRE_TRUE(dD != 0 && dH != 0 && dW != 0, 0,
"AVGPOOL3DNEW_BP op: dilation must not be zero, but got instead {%i, %i, %i}", dD, dH, dW);
LongType bS, iC, iD, iH, iW, oC, oD, oH,
oW; // batch size, input channels, input depth/height/width, output channels, output depth/height/width;
LongType indIOioC, indIOioD, indWoC, indWiC, indWkD; // corresponding indexes
ConvolutionUtils::getSizesAndIndexesConv3d(isNCDHW, 0, *input, *gradO, bS, iC, iD, iH, iW, oC, oD, oH, oW, indIOioC,
indIOioD, indWiC, indWoC, indWkD);
std::vector<LongType> expectedGradOShape =
ShapeUtils::composeShapeUsingDimsAndIdx({bS, iC, oD, oH, oW, 0, indIOioC, indIOioD, indIOioD + 1, indIOioD + 2});
std::vector<LongType> expectedGradIShape =
ShapeUtils::composeShapeUsingDimsAndIdx({bS, iC, iD, iH, iW, 0, indIOioC, indIOioD, indIOioD + 1, indIOioD + 2});
REQUIRE_TRUE(gradO->isSameShape(expectedGradOShape), 0,
"AVGPOOL3DNEW_BP op: wrong shape of output's gradients array (next epsilon), expected is %s, but got %s "
"instead !",
ShapeUtils::shapeAsString(expectedGradOShape).c_str(), ShapeUtils::shapeAsString(gradO).c_str());
REQUIRE_TRUE(
gradI->isSameShape(expectedGradIShape), 0,
"AVGPOOL3DNEW_BP op: wrong shape of input's gradients array (epsilon), expected is %s, but got %s instead !",
ShapeUtils::shapeAsString(expectedGradIShape).c_str(), ShapeUtils::shapeAsString(gradI).c_str());
if (!isNCDHW) {
std::vector<sd::LongType> perm = {0, 4, 1, 2, 3};
input = input->permute(perm, false, false); // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
gradI = gradI->permute(perm, false, false); // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
gradO =gradO->permute(perm, false, false); // [bS, oD, oH, oW, iC] -> [bS, iC, oD, oH, oW]
}
if (isSameMode) // SAME
ConvolutionUtils::calcPadding3D(pD, pH, pW, oD, oH, oW, iD, iH, iW, kD, kH, kW, sD, sH, sW, dD, dH, dW);
// 0,1 - kernel Height/Width; 2,3 - stride Height/Width; 4,5 - pad Height/Width; 6,7 - dilation Height/Width; 8 -
// poolingMode; 9 - divisor;
ConvolutionUtils::pooling3dBP(block, *input, *gradO, *gradI, kD, kH, kW, sD, sH, sW, pD, pH, pW, dD, dH, dW, 1,
extraParam0);
if (!isNCDHW) {
delete input;
delete gradI;
delete gradO;
}
return Status::OK;
}
DECLARE_SHAPE_FN(avgpool3dnew_bp) {
auto ret = SHAPELIST(ConstantShapeHelper::getInstance().castToDataType(inputShape->at(0),
ArrayOptions::dataType(inputShape->at(1))));
return ret;
}
} // namespace ops
} // namespace sd
#endif
libnd4j/include/ops/declarable/generic/nn/pooling/maxpool2d.cpp
+225
View File
@@ -0,0 +1,225 @@
/* ******************************************************************************
*
*
* 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.
*
* See the NOTICE file distributed with this work for additional
* information regarding copyright ownership.
* 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, created on 29/10/17.
// @author Yurii Shyrma (iuriish@yahoo.com), changed on 09.05.2018
//
#include <system/op_boilerplate.h>
#if NOT_EXCLUDED(OP_maxpool2d)
#include <ops/declarable/CustomOperations.h>
#include <ops/declarable/helpers/convolutions.h>
namespace sd {
namespace ops {
//////////////////////////////////////////////////////////////////////////
// maxpool2d corresponds to poolingMode=0
CUSTOM_OP_IMPL(maxpool2d, 1, 1, false, 0, 9) {
auto input = INPUT_VARIABLE(0);
REQUIRE_TRUE(input->rankOf() == 4, 0, "MAXPOOL2D OP: input array should have rank of 4, but got %i instead",
input->rankOf());
// 0,1 - kernel Height/Width; 2,3 - stride Height/Width; 4,5 - pad Height/Width; 6,7 - dilation Height/Width; 8 - same
// mode;
auto output = OUTPUT_NULLIFIED(0);
const LongType kH = INT_ARG(0);
const LongType kW = INT_ARG(1);
const LongType sH = INT_ARG(2);
const LongType sW = INT_ARG(3);
LongType pH = INT_ARG(4);
LongType pW = INT_ARG(5);
const LongType dH = INT_ARG(6);
const LongType dW = INT_ARG(7);
const bool isSameMode = INT_ARG(8);
REQUIRE_TRUE(dH != 0 && dW != 0, 0, "MAXPOOL2D op: dilation must not be zero, but got instead {%i, %i}", dH, dW);
LongType oH = 0;
LongType oW = 0;
int isNCHW = block.getIArguments()->size() > 10 ? !INT_ARG(10) : 1; // INT_ARG(10): 1-NHWC, 0-NCHW
const LongType iH = isNCHW ? input->sizeAt(2) : input->sizeAt(1);
const LongType iW = isNCHW ? input->sizeAt(3) : input->sizeAt(2);
if (!isNCHW) {
std::vector<sd::LongType> perm = {0, 3, 1, 2};
input = input->permute(perm, false, false); // [bS, iH, iW, iC] -> [bS, iC, iH, iW] - permute() already returns NDArray*
output = output->permute(perm, false, false); // [bS, oH, oW, iC] -> [bS, iC, oH, oW] - permute() already returns NDArray*
}
ConvolutionUtils::calcOutSizePool2D(oH, oW, kH, kW, sH, sW, pH, pW, dH, dW, iH, iW, isSameMode);
if (isSameMode) ConvolutionUtils::calcPadding2D(pH, pW, oH, oW, iH, iW, kH, kW, sH, sW, dH, dW);
// 0,1 - kernel Height/Width; 2,3 - stride Height/Width; 4,5 - pad Height/Width; 6,7 - dilation Height/Width;
// poolingMode; 9 - divisor;
ConvolutionUtils::pooling2d(block, *input, *output, kH, kW, sH, sW, pH, pW, dH, dW, MAX_POOL, 1);
if (!isNCHW) {
delete input;
delete output;
}
return Status::OK;
}
DECLARE_SYN(MaxPool2D, maxpool2d);
DECLARE_SYN(MaxPool, maxpool2d);
DECLARE_SYN(maxpool, maxpool2d);
DECLARE_TYPES(maxpool2d) { getOpDescriptor()->setAllowedInputTypes(ANY)->setSameMode(true); }
DECLARE_SHAPE_FN(maxpool2d) {
// NDArray<T> *x = block.getVariables().at(0)->getNDArray();
auto inShape = inputShape->at(0);
auto shapeOf = shape::shapeOf(inShape);
// 0 - number of dimensions; 1,2 - kernel Height/Width; 3,4 - stride Height/Width; 5,6 - pad Height/Width; 7,8 -
// dilation Height/Width; 9,10 - input Height/Width; 11 - batch size; 12 - input depth; 13 - same mode;
LongType kH = INT_ARG(0);
LongType kW = INT_ARG(1);
LongType sH = INT_ARG(2);
LongType sW = INT_ARG(3);
LongType pH = INT_ARG(4);
LongType pW = INT_ARG(5);
LongType dH = INT_ARG(6);
LongType dW = INT_ARG(7);
int isSameMode = INT_ARG(8);
int isNCHW = block.getIArguments()->size() > 10 ? !INT_ARG(10) : 1; // INT_ARG(10): 1-NHWC, 0-NCHW
REQUIRE_TRUE(dH != 0 && dW != 0, 0, "MAXPOOL2D op: dilation must not be zero, but got instead {%i, %i}", dH, dW);
LongType bS = shapeOf[0];
LongType iC = isNCHW ? shapeOf[1] : shapeOf[3];
LongType iH = isNCHW ? shapeOf[2] : shapeOf[1];
LongType iW = isNCHW ? shapeOf[3] : shapeOf[2];
char order = shape::order(inShape); // output order must be equal to input order
// calculate output Height/Width
LongType oH, oW;
ConvolutionUtils::calcOutSizePool2D(oH, oW, kH, kW, sH, sW, pH, pW, dH, dW, iH, iW, isSameMode);
// allocate memory for new shape
LongType newShape[4];
newShape[0] = bS;
if (isNCHW) {
newShape[1] = iC;
newShape[2] = oH;
newShape[3] = oW;
} else {
newShape[1] = oH;
newShape[2] = oW;
newShape[3] = iC;
}
auto ret = SHAPELIST(ConstantShapeHelper::getInstance().bufferForShapeInfo(ArrayOptions::dataType(inShape),
order,
4,
newShape)->primary());
return ret;
}
DECLARE_TYPES(maxpool2d_bp) {
getOpDescriptor()->setAllowedInputTypes(ANY)->setAllowedOutputTypes({ALL_FLOATS});
}
//////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL(maxpool2d_bp, 2, 1, false, 0, 10) {
auto input = INPUT_VARIABLE(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW)
auto gradO = INPUT_VARIABLE(1); // [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCHW), epsilon_next
auto gradI = OUTPUT_NULLIFIED(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW), epsilon
LongType kH = INT_ARG(0); // filter(kernel) height
LongType kW = INT_ARG(1); // filter(kernel) width
LongType sH = INT_ARG(2); // strides height
LongType sW = INT_ARG(3); // strides width
LongType pH = INT_ARG(4); // paddings height
LongType pW = INT_ARG(5); // paddings width
LongType dH = INT_ARG(6); // dilations height
LongType dW = INT_ARG(7); // dilations width
int isSameMode = INT_ARG(8); // 0-VALID, 1-SAME
int isNCHW = block.getIArguments()->size() > 10 ? !INT_ARG(10) : 1; // INT_ARG(10): 1-NHWC, 0-NCHW
REQUIRE_TRUE(input->rankOf() == 4, 0, "MAXPOOL2D_BP op: input should have rank of 4, but got %i instead",
input->rankOf());
REQUIRE_TRUE(dH != 0 && dW != 0, 0, "MAXPOOL2D_BP op: dilation must not be zero, but got instead {%i, %i}", dH, dW);
LongType bS, iC, iH, iW, oC, oH,
oW; // batch size, input channels, input height/width, output channels, output height/width;
LongType indIOioC, indIiH, indWoC, indWiC, indWkH, indOoH; // corresponding indexes
ConvolutionUtils::getSizesAndIndexesConv2d(isNCHW, 0, *input, *gradO, bS, iC, iH, iW, oC, oH, oW, indIOioC, indIiH,
indWiC, indWoC, indWkH, indOoH);
std::vector<LongType> expectedGradOShape =
ShapeUtils::composeShapeUsingDimsAndIdx({bS, iC, oH, oW, 0, indIOioC, indIiH, indIiH + 1});
std::vector<LongType> expectedGradIShape =
ShapeUtils::composeShapeUsingDimsAndIdx({bS, iC, iH, iW, 0, indIOioC, indIiH, indIiH + 1});
REQUIRE_TRUE(
gradO->isSameShape(expectedGradOShape), 0,
"MAXPOOL2D_BP op: wrong shape of output's gradients array (next epsilon), expected is %s, but got %s instead !",
ShapeUtils::shapeAsString(expectedGradOShape).c_str(), ShapeUtils::shapeAsString(gradO).c_str());
REQUIRE_TRUE(
gradI->isSameShape(expectedGradIShape), 0,
"MAXPOOL2D_BP op: wrong shape of input's gradients array (epsilon), expected is %s, but got %s instead !",
ShapeUtils::shapeAsString(expectedGradIShape).c_str(), ShapeUtils::shapeAsString(gradI).c_str());
if (!isNCHW) {
std::vector<sd::LongType> perm = {0, 3, 1, 2};
input = input->permute(perm, false, false); // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
gradI = gradI->permute(perm, false, false); // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
gradO = gradO->permute(perm, false, false); // [bS, oH, oW, iC] -> [bS, iC, oH, oW]
}
if (isSameMode) // SAME
ConvolutionUtils::calcPadding2D(pH, pW, oH, oW, iH, iW, kH, kW, sH, sW, dH, dW);
ConvolutionUtils::pooling2dBP(block, *input, *gradO, *gradI, kH, kW, sH, sW, pH, pW, dH, dW, 0., 1.);
if (!isNCHW) {
delete input;
delete gradI;
delete gradO;
}
return Status::OK;
}
DECLARE_SYN(MaxPool2D_bp, maxpool2d_bp);
DECLARE_SYN(MaxPool_bp, maxpool2d_bp);
DECLARE_SHAPE_FN(maxpool2d_bp) {
REQUIRE_TRUE(inputShape->at(0)[0] == 4, 0, "MAXPOOL2D_BP op: input array must be 4D, but got %i instead!",
inputShape->at(0)[0]);
REQUIRE_TRUE(inputShape->at(1)[0] == 4, 0,
"MAXPOOL2D_BP op: output's gradient array (next epsilon) must be 4D, but got %i instead!",
inputShape->at(1)[0]);
auto desc = new ShapeDescriptor(inputShape->at(0), ArrayOptions::dataType(inputShape->at(1)), false);
return SHAPELIST(ConstantShapeHelper::getInstance().createShapeInfo(desc));
}
} // namespace ops
} // namespace sd
#endif
libnd4j/include/ops/declarable/generic/nn/pooling/maxpool3d.cpp
+241
View File
@@ -0,0 +1,241 @@
/* ******************************************************************************
*
*
* 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.
*
* See the NOTICE file distributed with this work for additional
* information regarding copyright ownership.
* 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), created on 19.02.2018
//
#include <system/op_boilerplate.h>
#if NOT_EXCLUDED(OP_maxpool3dnew)
#include <ops/declarable/CustomOperations.h>
#include <ops/declarable/helpers/convolutions.h>
namespace sd {
namespace ops {
//////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL(maxpool3dnew, 1, 1, false, 0, 14) {
auto input = INPUT_VARIABLE(0); // [bS, iD, iH, iW, iC] (NDHWC) or [bS, iC, iD, iH, iW] (NCDHW)
auto output = OUTPUT_NULLIFIED(0); // [bS, oD, oH, oW, iC] (NDHWC) or [bS, iC, oD, oH, oW] (NCDHW)
LongType kD = INT_ARG(0); // filter(kernel) depth
LongType kH = INT_ARG(1); // filter(kernel) height
LongType kW = INT_ARG(2); // filter(kernel) width
LongType sD = INT_ARG(3); // strides depth
LongType sH = INT_ARG(4); // strides height
LongType sW = INT_ARG(5); // strides width
LongType pD = INT_ARG(6); // paddings depth
LongType pH = INT_ARG(7); // paddings height
LongType pW = INT_ARG(8); // paddings width
LongType dD = INT_ARG(9); // dilations depth
LongType dH = INT_ARG(10); // dilations height
LongType dW = INT_ARG(11); // dilations width
int isSameMode = INT_ARG(12); // 1-SAME, 0-VALID
int extraParam0 = INT_ARG(13); // unnecessary for max case, required only for avg and pnorm cases
int isNCDHW = block.getIArguments()->size() > 14 ? !INT_ARG(14) : 1; // 1-NDHWC, 0-NCDHW
REQUIRE_TRUE(input->rankOf() == 5, 0, "MAXPOOL3DNEW OP: rank of input array must be equal to 5, but got %i instead !",
input->rankOf());
REQUIRE_TRUE(dD != 0 && dH != 0 && dW != 0, 0,
"MAXPOOL3DNEW op: dilation must not be zero, but got instead {%i, %i, %i}", dD, dH, dW);
LongType bS, iC, iD, iH, iW, oC, oD, oH,
oW; // batch size, input channels, input depth/height/width, output channels, output depth/height/width;
LongType indIOioC, indIOioD, indWoC, indWiC, indWkD; // corresponding indexes
ConvolutionUtils::getSizesAndIndexesConv3d(isNCDHW, 0, *input, *output, bS, iC, iD, iH, iW, oC, oD, oH, oW, indIOioC,
indIOioD, indWiC, indWoC, indWkD);
std::vector<LongType> expectedOutputShape =
ShapeUtils::composeShapeUsingDimsAndIdx({bS, iC, oD, oH, oW, 0, indIOioC, indIOioD, indIOioD + 1, indIOioD + 2});
REQUIRE_TRUE(output->isSameShape(expectedOutputShape), 0,
"MAXPOOL3D op: wrong shape of output array, expected is %s, but got %s instead !",
ShapeUtils::shapeAsString(expectedOutputShape).c_str(), ShapeUtils::shapeAsString(output).c_str());
// REQUIRE_TRUE(iD >= kD && iH >= kH && iW >= kW, 0, "MAXPOOL3D OP: the input depth/height/width must be greater
// or equal to kernel(filter) depth/height/width, but got [%i, %i, %i] and [%i, %i, %i] correspondingly !", iD,iH,iW,
// kD,kH,kW); REQUIRE_TRUE(kD/2 >= pD && kH/2 >= pH && kW/2 >= pW, 0, "MAXPOOL3D OP: pad depth/height/width must not
// be greater than half of kernel depth/height/width, but got [%i, %i, %i] and [%i, %i, %i] correspondingly !",
// pD,pH,pW, kD,kH,kW);
if (!isNCDHW) {
std::vector<sd::LongType> perm = {0, 4, 1, 2, 3};
input = input->permute(perm, false, false); // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
output = output->permute(perm, false, false); // [bS, oD, oH, oW, iC] -> [bS, iC, oD, oH, oW]
}
if (isSameMode) // SAME
ConvolutionUtils::calcPadding3D(pD, pH, pW, oD, oH, oW, iD, iH, iW, kD, kH, kW, sD, sH, sW, dD, dH, dW);
ConvolutionUtils::pooling3d(block, *input, *output, kD, kH, kW, sD, sH, sW, pD, pH, pW, dD, dH, dW, 0, 1);
if (!isNCDHW) {
delete input;
delete output;
}
return Status::OK;
}
DECLARE_TYPES(maxpool3dnew) { getOpDescriptor()->setAllowedInputTypes(ANY)->setSameMode(true); }
DECLARE_SHAPE_FN(maxpool3dnew) {
LongType kD = INT_ARG(0); // filter(kernel) depth
LongType kH = INT_ARG(1); // filter(kernel) height
LongType kW = INT_ARG(2); // filter(kernel) width
LongType sD = INT_ARG(3); // strides depth
LongType sH = INT_ARG(4); // strides height
LongType sW = INT_ARG(5); // strides width
LongType pD = INT_ARG(6); // paddings depth
LongType pH = INT_ARG(7); // paddings height
LongType pW = INT_ARG(8); // paddings width
LongType dD = INT_ARG(9); // dilations depth
LongType dH = INT_ARG(10); // dilations height
LongType dW = INT_ARG(11); // dilations width
int isSameMode = INT_ARG(12); // 1-SAME, 0-VALID
// int extraParam0 = INT_ARG(13);
int isNCDHW = block.getIArguments()->size() > 14 ? !INT_ARG(14) : 1; // 1-NDHWC, 0-NCDHW
REQUIRE_TRUE(dD != 0 && dH != 0 && dW != 0, 0,
"MAXPOOL3DNEW op: dilation must not be zero, but got instead {%i, %i, %i}", dD, dH, dW);
auto inputShapeInfo = inputShape->at(0);
LongType idxID, idxIC;
if (isNCDHW) {
idxID = 2;
idxIC = 1;
} else {
idxID = 1;
idxIC = 4;
}
LongType bS = inputShapeInfo[1]; // batch size
LongType iC = inputShapeInfo[idxIC + 1]; // input channels
LongType iD = inputShapeInfo[idxID + 1]; // input depth
LongType iH = inputShapeInfo[idxID + 2]; // input height
LongType iW = inputShapeInfo[idxID + 3]; // input width
LongType oD, oH, oW; // output depth, height, width
ConvolutionUtils::calcOutSizePool3D(oD, oH, oW, kD, kH, kW, sD, sH, sW, pD, pH, pW, dD, dH, dW, iD, iH, iW,
isSameMode);
LongType outputShape[5];
outputShape[0] = bS;
if (isNCDHW) {
outputShape[1] = iC;
outputShape[2] = oD;
outputShape[3] = oH;
outputShape[4] = oW;
} else {
outputShape[1] = oD;
outputShape[2] = oH;
outputShape[3] = oW;
outputShape[4] = iC;
}
auto ret = SHAPELIST(ConstantShapeHelper::getInstance().bufferForShapeInfo(ArrayOptions::dataType(inputShapeInfo),
shape::order(inputShapeInfo),
5,
outputShape)->primary());
return ret;
}
DECLARE_TYPES(maxpool3dnew_bp) {
getOpDescriptor()->setAllowedInputTypes(ANY)->setAllowedOutputTypes({ALL_FLOATS});
}
//////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL(maxpool3dnew_bp, 2, 1, false, 0, 14) {
auto input = INPUT_VARIABLE(0); // [bS, iD, iH, iW, iC] (NDHWC) or [bS, iC, iD, iH, iW] (NCDHW)
auto gradO = INPUT_VARIABLE(1); // [bS, oD, oH, oW, oC] (NDHWC) or [bS, oC, oD, oH, oW] (NCDHW), epsilon_next
auto gradI = OUTPUT_NULLIFIED(0); // [bS, iD, iH, iW, iC] (NDHWC) or [bS, iC, iD, iH, iW] (NCDHW), epsilon
const LongType kD = INT_ARG(0); // filter(kernel) depth
const LongType kH = INT_ARG(1); // filter(kernel) height
const LongType kW = INT_ARG(2); // filter(kernel) width
const LongType sD = INT_ARG(3); // strides depth
const LongType sH = INT_ARG(4); // strides height
const LongType sW = INT_ARG(5); // strides width
LongType pD = INT_ARG(6); // paddings depth
LongType pH = INT_ARG(7); // paddings height
LongType pW = INT_ARG(8); // paddings width
const LongType dD = INT_ARG(9); // dilations depth
const LongType dH = INT_ARG(10); // dilations height
const LongType dW = INT_ARG(11); // dilations width
const int isSameMode = INT_ARG(12); // 1-SAME, 0-VALID
int extraParam0 = INT_ARG(13); // unnecessary for max case, required only for avg and pnorm cases
int isNCDHW = block.getIArguments()->size() > 14 ? !INT_ARG(14) : 1; // 1-NDHWC, 0-NCDHW
REQUIRE_TRUE(input->rankOf() == 5, 0, "MAXPOOL3DNEW_BP op: input should have rank of 5, but got %i instead",
input->rankOf());
REQUIRE_TRUE(dD != 0 && dH != 0 && dW != 0, 0,
"MAXPOOL3DNEW_BP op: dilation must not be zero, but got instead {%i, %i, %i}", dD, dH, dW);
LongType bS, iC, iD, iH, iW, oC, oD, oH,
oW; // batch size, input channels, input depth/height/width, output channels, output depth/height/width;
LongType indIOioC, indIOioD, indWoC, indWiC, indWkD; // corresponding indexes
ConvolutionUtils::getSizesAndIndexesConv3d(isNCDHW, 0, *input, *gradO, bS, iC, iD, iH, iW, oC, oD, oH, oW, indIOioC,
indIOioD, indWiC, indWoC, indWkD);
std::vector<LongType> expectedGradOShape =
ShapeUtils::composeShapeUsingDimsAndIdx({bS, iC, oD, oH, oW, 0, indIOioC, indIOioD, indIOioD + 1, indIOioD + 2});
std::vector<LongType> expectedGradIShape =
ShapeUtils::composeShapeUsingDimsAndIdx({bS, iC, iD, iH, iW, 0, indIOioC, indIOioD, indIOioD + 1, indIOioD + 2});
REQUIRE_TRUE(gradO->isSameShape(expectedGradOShape), 0,
"MAXPOOL3DNEW_BP op: wrong shape of output's gradients array (next epsilon), expected is %s, but got %s "
"instead !",
ShapeUtils::shapeAsString(expectedGradOShape).c_str(), ShapeUtils::shapeAsString(gradO).c_str());
REQUIRE_TRUE(
gradI->isSameShape(expectedGradIShape), 0,
"MAXPOOL3DNEW_BP op: wrong shape of input's gradients array (epsilon), expected is %s, but got %s instead !",
ShapeUtils::shapeAsString(expectedGradIShape).c_str(), ShapeUtils::shapeAsString(gradI).c_str());
if (!isNCDHW) {
std::vector<sd::LongType> perm = {0, 4, 1, 2, 3};
input = input->permute(perm, false, false); // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
gradI = gradI->permute(perm, false, false); // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
gradO = gradO->permute(perm, false, false); // [bS, oD, oH, oW, iC] -> [bS, iC, oD, oH, oW]
}
if (isSameMode) // SAME
ConvolutionUtils::calcPadding3D(pD, pH, pW, oD, oH, oW, iD, iH, iW, kD, kH, kW, sD, sH, sW, dD, dH, dW);
// [bS, iC, kD, kH, kW, oD, oH, oW] is de-convoluted to [bS, iC, iD, iH, iW]
// 0,1 - kernel Height/Width; 2,3 - stride Height/Width; 4,5 - pad Height/Width; 6,7 - dilation Height/Width; 8 -
// poolingMode; 9 - unnecessary;
ConvolutionUtils::pooling3dBP(block, *input, *gradO, *gradI, kD, kH, kW, sD, sH, sW, pD, pH, pW, dD, dH, dW, 0, 1);
if (!isNCDHW) {
delete input;
delete gradI;
delete gradO;
}
return Status::OK;
}
DECLARE_SHAPE_FN(maxpool3dnew_bp) {
auto desc = new ShapeDescriptor(inputShape->at(0), ArrayOptions::dataType(inputShape->at(1)), false);
return SHAPELIST(ConstantShapeHelper::getInstance().createShapeInfo(desc));
}
} // namespace ops
} // namespace sd
#endif
libnd4j/include/ops/declarable/generic/nn/pooling/maxpool_with_argmax.cpp
+67
View File
@@ -0,0 +1,67 @@
/* ******************************************************************************
*
*
* 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.
*
* See the NOTICE file distributed with this work for additional
* information regarding copyright ownership.
* 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
******************************************************************************/
//
// Created by GS <sgazeos@gmail.com> at 2/20/18
//
#include <system/op_boilerplate.h>
#if NOT_EXCLUDED(OP_max_pool_with_argmax)
#include <ops/declarable/CustomOperations.h>
#include <ops/declarable/helpers/convolutions.h>
#include <ops/declarable/helpers/max_pooling.h>
namespace sd {
namespace ops {
CUSTOM_OP_IMPL(max_pool_with_argmax, 1, 2, false, 0, 9) {
auto x = INPUT_VARIABLE(0);
auto z = OUTPUT_NULLIFIED(0);
auto indices = OUTPUT_NULLIFIED(1);
REQUIRE_TRUE(x->rankOf() == 4, 0, "max_pool_with_argmax: Input should have rank of 4, but got %i instead",
x->rankOf());
auto argI = *(block.getIArguments());
helpers::maxPoolingFunctor(block.launchContext(), block, x, z, argI, indices);
return Status::OK;
}
DECLARE_TYPES(max_pool_with_argmax) {
getOpDescriptor()
->setAllowedInputTypes(ANY)
->setAllowedOutputTypes(0, {ALL_FLOATS, ALL_INTS})
->setAllowedOutputTypes(1, {ALL_INDICES});
}
DECLARE_SHAPE_FN(max_pool_with_argmax) {
auto in = inputShape->at(0);
auto dtype = block.numD() ? D_ARG(0) : INT64;
// First shape info uses original data type from 'in'
auto valuesShape = ConstantShapeHelper::getInstance().bufferForShapeInfo(in)->primary();
// Second one needs to be cast to dtype
auto indicesShape = ConstantShapeHelper::getInstance().castToDataType(valuesShape, dtype);
return SHAPELIST(valuesShape, indicesShape);
}
} // namespace ops
} // namespace sd
#endif
libnd4j/include/ops/declarable/generic/nn/pooling/pnormpool2d.cpp
+223
View File
@@ -0,0 +1,223 @@
/* ******************************************************************************
*
*
* 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.
*
* See the NOTICE file distributed with this work for additional
* information regarding copyright ownership.
* 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, created on 29/10/17.
// @author Yurii Shyrma (iuriish@yahoo.com), changed on 14.05.2018
//
#include <system/op_boilerplate.h>
#if NOT_EXCLUDED(OP_pnormpool2d)
#include <ops/declarable/CustomOperations.h>
#include <ops/declarable/helpers/convolutions.h>
namespace sd {
namespace ops {
CUSTOM_OP_IMPL(pnormpool2d, 1, 1, false, 0, 10) {
REQUIRE_OK(this->validateInputLengthMatch(block));
REQUIRE_OK(this->validateInputDimensionsMatch(block));
auto input = INPUT_VARIABLE(0);
auto output = OUTPUT_NULLIFIED(0);
REQUIRE_TRUE(input->rankOf() == 4, 0, "PNORMPOOL2D op: input should have rank of 4, but got %i instead",
input->rankOf());
LongType kY = INT_ARG(0);
LongType kX = INT_ARG(1);
LongType sY = INT_ARG(2);
LongType sX = INT_ARG(3);
LongType pY = INT_ARG(4);
LongType pX = INT_ARG(5);
LongType dY = INT_ARG(6);
LongType dX = INT_ARG(7);
bool isSameMode = static_cast<bool>(INT_ARG(8));
auto extraParam0 = INT_ARG(9);
REQUIRE_TRUE(dY != 0 && dX != 0, 0, "PNORMPOOL2D op: dilation must not be zero, but got instead {%i, %i}", dY, dX);
LongType oY = 0;
LongType oX = 0;
int isNCHW = block.getIArguments()->size() > 10 ? !INT_ARG(10) : 1; // 1-NHWC, 0-NCHW
if (!isNCHW) {
std::vector<sd::LongType> perm = {0, 3, 1, 2};
input = input->permute(perm, false, false); // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
output = output->permute(perm, false, false); // [bS, oH, oW, iC] -> [bS, iC, oH, oW]
}
const LongType inY = static_cast<LongType>(input->sizeAt(2));
const LongType inX = static_cast<LongType>(input->sizeAt(3));
ConvolutionUtils::calcOutSizePool2D(oY, oX, kY, kX, sY, sX, pY, pX, dY, dX, inY, inX, isSameMode);
if (isSameMode) ConvolutionUtils::calcPadding2D(pY, pX, oY, oX, inY, inX, kY, kX, sY, sX, dY, dX);
// 0,1 - kernel Height/Width; 2,3 - stride Height/Width; 4,5 - pad Height/Width; 6,7 - dilation Height/Width; 8 -
// poolingMode; 9 - divisor;
ConvolutionUtils::pooling2d(block, *input, *output, kY, kX, sY, sX, pY, pX, dY, dX, PNORM_POOL,
extraParam0);
if (!isNCHW) {
delete input;
delete output;
}
return Status::OK;
}
DECLARE_SYN(PnormPool2D, pnormpool2d);
DECLARE_SYN(PnormPool, pnormpool2d);
DECLARE_SYN(pnormpool, pnormpool2d);
DECLARE_TYPES(pnormpool2d) {
getOpDescriptor()->setAllowedInputTypes(ANY)->setAllowedOutputTypes({ALL_FLOATS});
}
DECLARE_SHAPE_FN(pnormpool2d) {
auto inShape = inputShape->at(0);
auto shapeOf = shape::shapeOf(inShape);
// 0,1 - kernel Height/Width; 2,3 - stride Height/Width; 4,5 - pad Height/Width; 6,7 - dilation Height/Width; 8 - same
// mode;
std::vector<LongType> argI = *(block.getIArguments());
LongType kH = INT_ARG(0);
LongType kW = INT_ARG(1);
LongType sH = INT_ARG(2);
LongType sW = INT_ARG(3);
LongType pH = INT_ARG(4);
LongType pW = INT_ARG(5);
LongType dH = INT_ARG(6);
LongType dW = INT_ARG(7);
int isSameMode = INT_ARG(8);
int isNCHW = block.getIArguments()->size() > 10 ? !INT_ARG(10) : 1; // 1-NHWC, 0-NCHW
REQUIRE_TRUE(dH != 0 && dW != 0, 0, "PNORMPOOL2D op: dilation must not be zero, but got instead {%i, %i}", dH, dW);
LongType bS = shapeOf[0];
LongType iC = isNCHW ? shapeOf[1] : shapeOf[3];
LongType iH = isNCHW ? shapeOf[2] : shapeOf[1];
LongType iW = isNCHW ? shapeOf[3] : shapeOf[2];
char order = shape::order(inShape); // output order must be equal to input order
// calculate output Height/Width
LongType oH, oW;
ConvolutionUtils::calcOutSizePool2D(oH, oW, kH, kW, sH, sW, pH, pW, dH, dW, iH, iW, isSameMode);
// allocate memory for new shape
LongType newShape[4];
newShape[0] = bS;
if (isNCHW) {
newShape[1] = iC;
newShape[2] = oH;
newShape[3] = oW;
} else {
newShape[1] = oH;
newShape[2] = oW;
newShape[3] = iC;
}
auto ret = SHAPELIST(ConstantShapeHelper::getInstance().bufferForShapeInfo(ArrayOptions::dataType(inShape),
order,
4,
newShape)->primary());
return ret;
}
DECLARE_TYPES(pnormpool2d_bp) {
getOpDescriptor()->setAllowedInputTypes(ANY)->setAllowedOutputTypes({ALL_FLOATS});
}
//////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL(pnormpool2d_bp, 2, 1, false, 1, 10) {
auto input = INPUT_VARIABLE(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW)
auto gradO = INPUT_VARIABLE(1); // [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCHW), epsilon_next
auto gradI = OUTPUT_NULLIFIED(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW), epsilon
LongType kH = INT_ARG(0); // filter(kernel) height
LongType kW = INT_ARG(1); // filter(kernel) width
LongType sH = INT_ARG(2); // strides height
LongType sW = INT_ARG(3); // strides width
LongType pH = INT_ARG(4); // paddings height
LongType pW = INT_ARG(5); // paddings width
LongType dH = INT_ARG(6); // dilations height
LongType dW = INT_ARG(7); // dilations width
int isSameMode = INT_ARG(8); // 0-VALID, 1-SAME
int pnorm = INT_ARG(9);
int isNCHW = block.getIArguments()->size() > 10 ? !INT_ARG(10) : 1; // 1-NHWC, 0-NCHW
// FIXME: double?
double eps = T_ARG(0);
REQUIRE_TRUE(input->rankOf() == 4, 0, "PNORMPOOL2D_BP op: input should have rank of 4, but got %i instead",
input->rankOf());
REQUIRE_TRUE(dH != 0 && dW != 0, 0, "PNORMPOOL2D_BP op: dilation must not be zero, but got instead {%i, %i}", dH, dW);
LongType bS, iC, iH, iW, oC, oH,
oW; // batch size, input channels, input height/width, output channels, output height/width;
LongType indIOioC, indIiH, indWoC, indWiC, indWkH, indOoH; // corresponding indexes
ConvolutionUtils::getSizesAndIndexesConv2d(isNCHW, 0, *input, *gradO, bS, iC, iH, iW, oC, oH, oW, indIOioC, indIiH,
indWiC, indWoC, indWkH, indOoH);
std::vector<LongType> expectedGradOShape =
ShapeUtils::composeShapeUsingDimsAndIdx({bS, iC, oH, oW, 0, indIOioC, indIiH, indIiH + 1});
std::vector<LongType> expectedGradIShape =
ShapeUtils::composeShapeUsingDimsAndIdx({bS, iC, iH, iW, 0, indIOioC, indIiH, indIiH + 1});
REQUIRE_TRUE(
gradO->isSameShape(expectedGradOShape), 0,
"PNORMPOOL2D_BP op: wrong shape of output's gradients array (next epsilon), expected is %s, but got %s instead !",
ShapeUtils::shapeAsString(expectedGradOShape).c_str(), ShapeUtils::shapeAsString(gradO).c_str());
REQUIRE_TRUE(
gradI->isSameShape(expectedGradIShape), 0,
"PNORMPOOL2D_BP op: wrong shape of input's gradients array (epsilon), expected is %s, but got %s instead !",
ShapeUtils::shapeAsString(expectedGradIShape).c_str(), ShapeUtils::shapeAsString(gradI).c_str());
if (!isNCHW) {
std::vector<sd::LongType> perm2 = {0, 3, 1, 2};
input = input->permute(perm2, false, false); // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
gradI = gradI->permute(perm2, false, false); // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
gradO = gradO->permute(perm2, false, false); // [bS, oH, oW, iC] -> [bS, iC, oH, oW]
}
ConvolutionUtils::pooling2dBP(block, *input, *gradO, *gradI, kH, kW, sH, sW, pH, pW, dH, dW, 2, pnorm);
if (!isNCHW) {
delete input;
delete gradI;
delete gradO;
}
return Status::OK;
}
DECLARE_SHAPE_FN(pnormpool2d_bp) {
REQUIRE_TRUE(inputShape->at(0)[0] == 4, 0, "PNORMPOOL2D_BP op: input array must be 4D, but got %i instead!",
inputShape->at(0)[0]);
REQUIRE_TRUE(inputShape->at(1)[0] == 4, 0,
"PNORMPOOL2D_BP op: output's gradient array (next epsilon) must be 4D, but got %i instead!",
inputShape->at(1)[0]);
auto desc = new ShapeDescriptor(inputShape->at(0), ArrayOptions::dataType(inputShape->at(1)), false);
return SHAPELIST(ConstantShapeHelper::getInstance().createShapeInfo(desc));
}
} // namespace ops
} // namespace sd
#endif