/******************************************************************************* * 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 // @author Yurii Shyrma // #include #if NOT_EXCLUDED(OP_deconv2d) #include #include #include #include #include #include namespace sd { namespace ops { CUSTOM_OP_IMPL(deconv2d, 2, 1, false, 0, 9) { auto input = INPUT_VARIABLE(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW) auto weights = INPUT_VARIABLE(1); // [kH, kW, oC, iC], [iC, oC, kH, kW], [iC, kH, kW, oC] auto bias = block.width() > 2 ? INPUT_VARIABLE(2) : nullptr; // [oC] auto output = OUTPUT_NULLIFIED(0); // [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCHW) REQUIRE_TRUE(input->rankOf() == 4, 0, "CUSTOM DECONV2D OP: rank of input array must be equal to 4, but got %i instead !", input->rankOf()); REQUIRE_TRUE(weights->rankOf() == 4, 0, "CUSTOM DECONV2D OP: rank of weights array must be equal to 4, but got %i instead !", weights->rankOf()); int kH = INT_ARG(0) > 0 ? INT_ARG(0) : static_cast(weights->sizeAt(0));// filter(kernel) height int kW = INT_ARG(1) > 0 ? INT_ARG(1) : static_cast(weights->sizeAt(1));// filter(kernel) width int sH = INT_ARG(2); // strides height int sW = INT_ARG(3); // strides width int pH = INT_ARG(4); // paddings height int pW = INT_ARG(5); // paddings width int dH = INT_ARG(6); // dilations height int dW = INT_ARG(7); // dilations width int isSameMode = INT_ARG(8); // 0-VALID, 1-SAME int isNCHW = block.getIArguments()->size() > 9 ? !INT_ARG(9) : 1; // INT_ARG(9): 0-NCHW, 1-NHWC int wFormat = block.getIArguments()->size() > 10 ? INT_ARG(10) : 0; // 0 - [kH, kW, oC, iC], 1 - [iC, oC, kH, kW], 2 - [iC, kH, kW, oC] int bS, iC, iH, iW, oC, oH, oW; // batch size, input channels, input height/width, output channels, output height/width; int indIOioC, indIiH, indWoC, indWiC, indWkH, indOoH; // corresponding indexes ConvolutionUtils::getSizesAndIndexesConv2d(isNCHW, wFormat, *input, *output, bS, iC, iH, iW, oC, oH, oW, indIOioC, indIiH, indWoC, indWiC, indWkH, indOoH); std::vector expectedWeightsShape = ConvolutionUtils::expectWeightsShape(wFormat, kH, kW, oC, iC); REQUIRE_TRUE(weights->isSameShape(expectedWeightsShape), 0, "CUSTOM DECONV2D OP: wrong shape of weights array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedWeightsShape).c_str(), ShapeUtils::shapeAsString(weights).c_str()); if (bias) REQUIRE_TRUE(bias->rankOf() <= 2 && oC == bias->lengthOf(), 0, "CUSTOM DECONV2D OP: wrong shape of array with biases, expected rank, length: <=2, %i, but got %i, %i instead !", oC, bias->rankOf(), bias->lengthOf()); if(!isNCHW) output = new NDArray(output->permute({0, 3, 1, 2})); // [bS, oH, oW, oC] -> [bS, oC, oH, oW] std::vector colPermut; if(1 == wFormat) colPermut = {1, 2, 3, 0, 4, 5}; else colPermut = {2, 3, 1, 0, 4, 5}; if(isSameMode) // Note: we're intentionally swapping iH and oH, to calculated the padding for a"normal" conv (not deconv) forward pass ConvolutionUtils::calcPadding2D(pH, pW, iH, iW, oH, oW, kH, kW, sH, sW, dH, dW); NDArray columns(input->ordering(), {bS, oC, kH, kW, iH, iW}, input->dataType(), block.launchContext()); //----- calculation of output -----// // NHWC: [kH, kW, oC, iC] x [bS, iH, iW, iC] = [kH, kW, oC, bS, iH, iW] // NHWC: [iC, oC, kH, kW] x [bS, iH, iW, iC] = [oC, kH, kW, bS, iH, iW] // NHWC: [iC, kH, kW, oC] x [bS, iH, iW, iC] = [kH, kW, oC, bS, iH, iW] sd::MmulHelper::tensorDot(weights, input, &columns, {indWiC}, {indIOioC}, colPermut); LaunchContext* ctx = block.launchContext(); helpers::col2im(*ctx, columns, *output, sH, sW, pH, pW, oH, oW, dH, dW); // [bS, oC, kH, kW, iH, iW] is de-convoluted to [bS, oC, oH, oW] //----- add biases if required -----// if(bias) // output->applyBroadcast(broadcast::Add, {1}, bias); helpers::addBias(block, *output, *bias, *output, true); if(!isNCHW) delete output; return Status::OK(); } DECLARE_TYPES(deconv2d) { getOpDescriptor() ->setAllowedInputTypes(sd::DataType::ANY) ->setAllowedOutputTypes({ALL_FLOATS}); } DECLARE_SHAPE_FN(deconv2d) { auto inputShapeInfo = inputShape->at(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCHW) auto weightsShapeInfo = inputShape->at(1); // [kH, kW, oC, iC], [iC, oC, kH, kW], [iC, kH, kW, oC] auto biasShapeInfo = block.width() > 2 ? inputShape->at(2) : nullptr; // [oC] const int rank = 4; REQUIRE_TRUE(shape::rank(inputShapeInfo) == rank, 0, "CUSTOM DECONV2D OP: rank of input array must be equal to %i, but got %i instead !", rank, shape::rank(inputShapeInfo)); REQUIRE_TRUE(shape::rank(weightsShapeInfo) == rank, 0, "CUSTOM DECONV2D OP: rank of weights array must be equal to %i, but got %i instead !", rank, shape::rank(weightsShapeInfo)); int kH = INT_ARG(0) > 0 ? INT_ARG(0) : static_cast(shape::sizeAt(weightsShapeInfo, 0));// filter(kernel) height int kW = INT_ARG(1) > 0 ? INT_ARG(1) : static_cast(shape::sizeAt(weightsShapeInfo, 1));// filter(kernel) width int sH = INT_ARG(2); // strides height int sW = INT_ARG(3); // strides width int pH = INT_ARG(4); // paddings height int pW = INT_ARG(5); // paddings width int dH = INT_ARG(6); // dilations height int dW = INT_ARG(7); // dilations width int isSameMode = INT_ARG(8); // 0-VALID, 1-SAME int isNCHW = block.getIArguments()->size() > 9 ? !INT_ARG(9) : 1; // INT_ARG(9): 1-NHWC, 0-NCHW int wFormat = block.getIArguments()->size() > 10 ? INT_ARG(10) : 0; // 0 - [kH, kW, oC, iC], 1 - [iC, oC, kH, kW], 2 - [iC, kH, kW, oC] int indIOioC, indIiH, indWoC(0 == wFormat ? 2 : (1 == wFormat ? 1 : 3)); if(!isNCHW) { indIOioC = 3; indIiH = 1; } else { indIOioC = 1; indIiH = 2; } const int bS = inputShapeInfo[1]; // batch size const int iH = inputShapeInfo[indIiH+1]; // input height const int iW = inputShapeInfo[indIiH+2]; // input width const int iC = inputShapeInfo[indIOioC+1]; // input channels const int oC = weightsShapeInfo[indWoC+1]; // output channels std::vector expectedWeightsShape = ConvolutionUtils::expectWeightsShape(wFormat, kH, kW, oC, iC); REQUIRE_TRUE(shape::shapeEquals(4, expectedWeightsShape.data(), shape::rank(weightsShapeInfo), shape::shapeOf(weightsShapeInfo)), 0, "CUSTOM DECONV2D OP: wrong shape of weights array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedWeightsShape).c_str(), ShapeUtils::shapeAsString(weightsShapeInfo).c_str()); if (biasShapeInfo) REQUIRE_TRUE(shape::rank(biasShapeInfo) <= 2 && oC == shape::length(biasShapeInfo), 0, "CUSTOM DECONV2D OP: wrong shape of array with biases, expected rank, length: <=2, %i, but got %i, %i instead !", oC, biasShapeInfo[0], shape::length(biasShapeInfo)); int oH, oW; // output height, width ConvolutionUtils::calcOutSizeDeconv2D(oH, oW, kH, kW, sH, sW, pH, pW, dH, dW, iH, iW, isSameMode); Nd4jLong outputShape[4]; outputShape[0] = bS; if (isNCHW) { outputShape[1] = oC; outputShape[2] = oH; outputShape[3] = oW; } else { outputShape[1] = oH; outputShape[2] = oW; outputShape[3] = oC; } return SHAPELIST(ConstantShapeHelper::getInstance()->createShapeInfo(ShapeDescriptor(ArrayOptions::dataType(weightsShapeInfo), shape::order(inputShapeInfo), outputShape, 4))); } DECLARE_TYPES(deconv2d_bp) { getOpDescriptor() ->setAllowedInputTypes(sd::DataType::ANY) ->setAllowedOutputTypes({ALL_FLOATS}); } ////////////////////////////////////////////////////////////////////////// CUSTOM_OP_IMPL(deconv2d_bp, 3, 2, false, 0, 9) { auto input = INPUT_VARIABLE(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCDHW) auto weights = INPUT_VARIABLE(1); // [kH, kW, oC, iC], [iC, oC, kH, kW], [iC, kH, kW, oC] auto bias = block.width() > 3 ? INPUT_VARIABLE(2) : nullptr; // [oC] auto gradO = block.width() > 3 ? INPUT_VARIABLE(3) : INPUT_VARIABLE(2); // [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCDHW), epsilon_next auto gradI = OUTPUT_VARIABLE(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCDHW), gradI auto gradW = OUTPUT_VARIABLE(1); // [kH, kW, oC, iC], [iC, oC, kH, kW], [iC, kH, kW, oC] auto gradB = block.width() > 3 ? OUTPUT_VARIABLE(2) : nullptr; // [oC] REQUIRE_TRUE(input->rankOf() == 4, 0, "CUSTOM DECONV2D_BP OP: rank of input array must be equal to 4, but got %i instead !", input->rankOf()); REQUIRE_TRUE(weights->rankOf() == 4, 0, "CUSTOM DECONV2D_BP OP: rank of weights array must be equal to 4 , but got %i instead !", weights->rankOf()); REQUIRE_TRUE(gradO->rankOf() == 4, 0, "CUSTOM DECONV2D_BP OP: rank of output gradients (next epsilon) array must be equal to 4, but got %i instead !", gradO->rankOf()); int kH = INT_ARG(0) > 0 ? INT_ARG(0) : static_cast(weights->sizeAt(0));// filter(kernel) height int kW = INT_ARG(1) > 0 ? INT_ARG(1) : static_cast(weights->sizeAt(1));// filter(kernel) width int sH = INT_ARG(2); // strides height int sW = INT_ARG(3); // strides width int pH = INT_ARG(4); // paddings height int pW = INT_ARG(5); // paddings width int dH = INT_ARG(6); // dilations height int dW = INT_ARG(7); // dilations width int isSameMode = INT_ARG(8); // 0-VALID, 1-SAME int isNCHW = block.getIArguments()->size() > 9 ? !INT_ARG(9) : 1; // INT_ARG(9): 1-NHWC, 0-NCHW int wFormat = block.getIArguments()->size() > 10 ? INT_ARG(10) : 0; // 0 - [kH, kW, oC, iC], 1 - [iC, oC, kH, kW], 2 - [iC, kH, kW, oC] int bS, iC, iH, iW, oC, oH, oW; // batch size, input channels, input height/width, output channels, output height/width; int indIOioC, indIiH, indWoC, indWiC, indWkH, indOoH; // corresponding indexes ConvolutionUtils::getSizesAndIndexesConv2d(isNCHW, wFormat, *input, *gradO, bS, iC, iH, iW, oC, oH, oW, indIOioC, indIiH, indWoC, indWiC, indWkH, indOoH); int trueoH, trueoW; // true output height, width ConvolutionUtils::calcOutSizeDeconv2D(trueoH, trueoW, kH, kW, sH, sW, pH, pW, dH, dW, iH, iW, isSameMode); std::vector expectedGradOShape = ShapeUtils::composeShapeUsingDimsAndIdx({bS,oC,trueoH,trueoW, 0,indIOioC,indOoH,indOoH+1}); std::vector expectedWeightsShape = ConvolutionUtils::expectWeightsShape(wFormat, kH, kW, oC, iC); REQUIRE_TRUE(gradO->isSameShape(expectedGradOShape), 0, "CUSTOM DECONV2D_BP OP: wrong shape of output gradients (next epsilon) array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedGradOShape).c_str(), ShapeUtils::shapeAsString(gradO).c_str()); REQUIRE_TRUE(weights->isSameShape(expectedWeightsShape), 0, "CUSTOM DECONV2D_BP OP: wrong shape of weights array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedWeightsShape).c_str(), ShapeUtils::shapeAsString(weights).c_str()); if(bias) REQUIRE_TRUE(bias->rankOf() <= 2 && oC == bias->lengthOf(), 0, "CUSTOM DECONV2D_BP OP: wrong shape of array with biases, expected rank, length: <=2, %i, but got %i, %i instead !", oC, bias->rankOf(), bias->lengthOf()); if(isSameMode){ // SAME //Note: we're intentionally swapping iH and oH, to calculated the padding for a"normal" conv (not deconv) forward pass ConvolutionUtils::calcPadding2D(pH, pW, iH, iW, oH, oW, kH, kW, sH, sW, dH, dW); } // ----- calculation of gradI -> pass it through conv2d_ff ----- // sd::ops::conv2d conv2d; const Nd4jStatus status = conv2d.execute({gradO, weights}, {gradI}, {}, {kH,kW, sH,sW, pH,pW, dH,dW, isSameMode, !isNCHW, wFormat}, {}); if (status != ND4J_STATUS_OK) return status; // -----prepare permutation arrays and axes for dot product ----- // std::vector inputAxes; if(!isNCHW) { gradO = new NDArray(gradO->permute({0, 3, 1, 2})); // [bS, oH, oW, oC] -> [bS, oC, oH, oW] inputAxes = {0, 1, 2}; // bS, iH, iW } else inputAxes = {0, 2, 3}; // bS, iH, iW std::vector gradWAxes; // empty for wFormat = 1 if(0 == wFormat) gradWAxes = {3, 2, 0, 1}; else if(2 == wFormat) gradWAxes = {0, 3, 1, 2}; // ----- calculation of gradW ----- // NDArray columns(input->ordering(), {bS, oC, kH, kW, iH, iW}, input->dataType(), block.launchContext()); LaunchContext* ctx = block.launchContext(); helpers::im2col(*ctx, *gradO, columns, kH, kW, sH, sW, pH, pW, dH, dW, NDArrayFactory::create(0.f, input->getContext())); // [bS, oC, oH, oW] is convoluted to [bS, oC, kH, kW, iH, iW] MmulHelper::tensorDot(input, &columns, gradW, inputAxes, {0, 4, 5}, gradWAxes); // [bS, iC, iH, iW]/[bS, iH, iW, iC] x [bS, oC, kH, kW, iH, iW] = [iC, oC, kH, kW] // ----- calculation of gradB ----- // if(gradB) { if(gradB->rankOf() == 2) gradB = new NDArray(gradB->reshape(gradB->ordering(), {gradB->lengthOf()}, false)); gradO->reduceAlongDimension(reduce::Sum, *gradB, {0, 2, 3}); // sum over bS, oH, oW if(gradB != OUTPUT_VARIABLE(2)) delete gradB; } if(!isNCHW) delete gradO; return Status::OK(); } DECLARE_SHAPE_FN(deconv2d_bp) { auto inputShapeInfo = inputShape->at(0); // [bS, iH, iW, iC] (NHWC) or [bS, iC, iH, iW] (NCDHW) auto weightsShapeInfo = inputShape->at(1); // [kH, kW, oC, iC], [iC, oC, kH, kW], [iC, kH, kW, oC] Nd4jLong const* biasShapeInfo = block.width() > 3 ? inputShape->at(2) : nullptr; // [oC] auto gradOShapeInfo = block.width() > 3 ? inputShape->at(3) : inputShape->at(2); // [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCDHW), epsilon_next const int rank = 4; REQUIRE_TRUE(shape::rank(inputShapeInfo) == rank, 0, "CUSTOM DECONV2D_BP OP: rank of input array must be equal to %i, but got %i instead !", rank, shape::rank(inputShapeInfo)); REQUIRE_TRUE(shape::rank(weightsShapeInfo) == rank, 0, "CUSTOM DECONV2D_BP OP: rank of weights array must be equal to %i , but got %i instead !", rank, shape::rank(weightsShapeInfo)); REQUIRE_TRUE(shape::rank(gradOShapeInfo) == rank, 0, "CUSTOM DECONV2D_BP OP: rank of output gradients (next epsilon) array must be equal to %i, but got %i instead !", rank, shape::rank(gradOShapeInfo)); int kH = INT_ARG(0) > 0 ? INT_ARG(0) : static_cast(shape::sizeAt(weightsShapeInfo, 0));// filter(kernel) height int kW = INT_ARG(1) > 0 ? INT_ARG(1) : static_cast(shape::sizeAt(weightsShapeInfo, 1));// filter(kernel) width int sH = INT_ARG(2); // strides height int sW = INT_ARG(3); // strides width int pH = INT_ARG(4); // paddings height int pW = INT_ARG(5); // paddings width int dH = INT_ARG(6); // dilations height int dW = INT_ARG(7); // dilations width int isSameMode = INT_ARG(8); // 0-VALID, 1-SAME int isNCHW = block.getIArguments()->size() > 9 ? !INT_ARG(9) : 1; // INT_ARG(9): 1-NHWC, 0-NCHW int wFormat = block.getIArguments()->size() > 10 ? INT_ARG(10) : 0; // 0 - [kH, kW, oC, iC], 1 - [iC, oC, kH, kW], 2 - [iC, kH, kW, oC] int indIOioC, indIiH, indOoH, indWoC(0 == wFormat ? 2 : (1 == wFormat ? 1 : 3)); if(!isNCHW) { indIOioC = 3; indIiH = 1; indOoH = 1; } else { indIOioC = 1; indIiH = 2; indOoH = 2; } const int bS = inputShapeInfo[1]; // batch size const int iH = inputShapeInfo[indIiH+1]; // input height const int iW = inputShapeInfo[indIiH+2]; // input width const int iC = inputShapeInfo[indIOioC+1]; // input channels const int oC = weightsShapeInfo[indWoC+1]; // output channels int trueoH, trueoW; // true output height, width ConvolutionUtils::calcOutSizeDeconv2D(trueoH, trueoW, kH, kW, sH, sW, pH, pW, dH, dW, iH, iW, isSameMode); std::vector expectedGradOShape = ShapeUtils::composeShapeUsingDimsAndIdx({bS,oC,trueoH,trueoW, 0,indIOioC,indOoH,indOoH+1}); std::vector expectedWeightsShape = ConvolutionUtils::expectWeightsShape(wFormat, kH, kW, oC, iC); REQUIRE_TRUE(shape::shapeEquals(4, expectedGradOShape.data(), shape::rank(gradOShapeInfo), shape::shapeOf(gradOShapeInfo)), 0, "CUSTOM DECONV2D_BP OP: wrong shape of output gradients next epsilon) array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedGradOShape).c_str(), ShapeUtils::shapeAsString(gradOShapeInfo).c_str()); REQUIRE_TRUE(shape::shapeEquals(4, expectedWeightsShape.data(), shape::rank(weightsShapeInfo), shape::shapeOf(weightsShapeInfo)), 0, "CUSTOM DECONV2D_BP OP: wrong shape of weights array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedWeightsShape).c_str(), ShapeUtils::shapeAsString(weightsShapeInfo).c_str()); if(biasShapeInfo) REQUIRE_TRUE(biasShapeInfo[0] <= 2 && oC == shape::length(biasShapeInfo), 0, "CUSTOM DECONV2D_BP OP: wrong shape of array with biases, expected rank, length: <=2, %i, but got %i, %i instead !", oC, biasShapeInfo[0], shape::length(biasShapeInfo)); auto gradIShapeInfo = ShapeBuilders::copyShapeInfoAndType(inputShapeInfo, gradOShapeInfo, false, block.getWorkspace()); auto gradWShapeInfo = ShapeBuilders::copyShapeInfoAndType(weightsShapeInfo, gradOShapeInfo, false, block.getWorkspace()); auto shapes = SHAPELIST(CONSTANT(gradIShapeInfo), CONSTANT(gradWShapeInfo)); if (biasShapeInfo != nullptr) { auto gradBShapeInfo = ShapeBuilders::copyShapeInfoAndType(biasShapeInfo, gradOShapeInfo, false, block.getWorkspace()); shapes->push_back(CONSTANT(gradBShapeInfo)); } return shapes; } } } #endif