* initial commit Signed-off-by: raver119 <raver119@gmail.com> * - gruCell_bp further Signed-off-by: Yurii <yurii@skymind.io> * - further work on gruCell_bp Signed-off-by: Yurii <yurii@skymind.io> * Inverse matrix cublas implementation. Partial working revision. * Separation of segment ops helpers. Max separation. * Separated segment_min ops. * Separation of segment_mean/sum/prod/sqrtN ops heleprs. * Fixed diagonal processing with LUP decomposition. * Modified inversion approach using current state of LU decomposition. * Implementation of matrix_inverse op with cuda kernels. Working revision. * Implemented sequence_mask cuda helper. Eliminated waste printf with matrix_inverse implementation. Added proper tests. * - further work on gruCell_bp (ff/cuda) Signed-off-by: Yurii <yurii@skymind.io> * comment one test for gruCell_bp Signed-off-by: Yurii <yurii@skymind.io> * - provide cuda static_rnn Signed-off-by: Yurii <yurii@skymind.io> * Refactored random_shuffle op to use new random generator. * Refactored random_shuffle op helper. * Fixed debug tests with random ops tests. * Implement random_shuffle op cuda kernel helper and tests. * - provide cuda scatter_update Signed-off-by: Yurii <yurii@skymind.io> * Implementation of random_shuffle for linear case with cuda kernels and tests. * Implemented random_shuffle with cuda kernels. Final revision. * - finally gruCell_bp is completed Signed-off-by: Yurii <yurii@skymind.io> * Dropout op cuda helper implementation. * Implemented dropout_bp cuda helper. * Implemented alpha_dropout_bp with cuda kernel helpers. * Refactored helper. * Implementation of suppresion helper with cuda kernels. * - provide cpu code fot hsvToRgb, rgbToHsv, adjustHue Signed-off-by: Yurii <yurii@skymind.io> * Using sort by value method. * Implementation of image.non_max_suppression op cuda-based helper. * - correcting and testing adjust_hue, adjust_saturation cpu/cuda code Signed-off-by: Yurii <yurii@skymind.io> * Added cuda device prefixes to declarations. * Implementation of hashcode op with cuda helper. Initital revision. * rnn cu impl removed Signed-off-by: raver119 <raver119@gmail.com>
		
			
				
	
	
		
			91 lines
		
	
	
		
			4.4 KiB
		
	
	
	
		
			C++
		
	
	
	
	
	
			
		
		
	
	
			91 lines
		
	
	
		
			4.4 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 sgazeos@gmail.com
 | |
| //
 | |
| 
 | |
| #include <ops/declarable/helpers/image_suppression.h>
 | |
| //#include <blas/NDArray.h>
 | |
| #include <algorithm>
 | |
| #include <numeric>
 | |
| 
 | |
| namespace nd4j {
 | |
| namespace ops {
 | |
| namespace helpers {
 | |
| 
 | |
|     template <typename T>
 | |
|     static void nonMaxSuppressionV2_(NDArray* boxes, NDArray* scales, int maxSize, double threshold, NDArray* output) {
 | |
|         std::vector<Nd4jLong> indices(scales->lengthOf());
 | |
|         std::iota(indices.begin(), indices.end(), 0);
 | |
| 
 | |
|         std::sort(indices.begin(), indices.end(), [scales](int i, int j) {return scales->e<T>(i) > scales->e<T>(j);});
 | |
| 
 | |
| //        std::vector<int> selected(output->lengthOf());
 | |
|         std::vector<int> selectedIndices(output->lengthOf(), 0);
 | |
|         auto needToSuppressWithThreshold = [] (NDArray& boxes, int previousIndex, int nextIndex, T threshold) -> bool {
 | |
|             T minYPrev = nd4j::math::nd4j_min(boxes.e<T>(previousIndex, 0), boxes.e<T>(previousIndex, 2));
 | |
|             T minXPrev = nd4j::math::nd4j_min(boxes.e<T>(previousIndex, 1), boxes.e<T>(previousIndex, 3));
 | |
|             T maxYPrev = nd4j::math::nd4j_max(boxes.e<T>(previousIndex, 0), boxes.e<T>(previousIndex, 2));
 | |
|             T maxXPrev = nd4j::math::nd4j_max(boxes.e<T>(previousIndex, 1), boxes.e<T>(previousIndex, 3));
 | |
|             T minYNext = nd4j::math::nd4j_min(boxes.e<T>(nextIndex, 0), boxes.e<T>(nextIndex, 2));
 | |
|             T minXNext = nd4j::math::nd4j_min(boxes.e<T>(nextIndex, 1), boxes.e<T>(nextIndex, 3));
 | |
|             T maxYNext = nd4j::math::nd4j_max(boxes.e<T>(nextIndex, 0), boxes.e<T>(nextIndex, 2));
 | |
|             T maxXNext = nd4j::math::nd4j_max(boxes.e<T>(nextIndex, 1), boxes.e<T>(nextIndex, 3));
 | |
|             T areaPrev = (maxYPrev - minYPrev) * (maxXPrev - minXPrev);
 | |
|             T areaNext = (maxYNext - minYNext) * (maxXNext - minXNext);
 | |
| 
 | |
|             if (areaNext <= T(0.f) || areaPrev <= T(0.f)) return false;
 | |
| 
 | |
|             T minIntersectionY = nd4j::math::nd4j_max(minYPrev, minYNext);
 | |
|             T minIntersectionX = nd4j::math::nd4j_max(minXPrev, minXNext);
 | |
|             T maxIntersectionY = nd4j::math::nd4j_min(maxYPrev, maxYNext);
 | |
|             T maxIntersectionX = nd4j::math::nd4j_min(maxXPrev, maxXNext);
 | |
|             T intersectionArea =
 | |
|                     nd4j::math::nd4j_max(T(maxIntersectionY - minIntersectionY), T(0.0f)) *
 | |
|                             nd4j::math::nd4j_max(T(maxIntersectionX - minIntersectionX), T(0.0f));
 | |
|             T intersectionValue = intersectionArea / (areaPrev + areaNext - intersectionArea);
 | |
|             return intersectionValue > threshold;
 | |
| 
 | |
|         };
 | |
| //        int numSelected = 0;
 | |
|         int numBoxes = boxes->sizeAt(0);
 | |
|         int numSelected = 0;
 | |
| 
 | |
|         for (int i = 0; i < numBoxes; ++i) {
 | |
|             bool shouldSelect = numSelected < output->lengthOf();
 | |
|             PRAGMA_OMP_PARALLEL_FOR //_ARGS(firstprivate(numSelected))
 | |
|             for (int j = numSelected - 1; j >= 0; --j) {
 | |
|                 if (shouldSelect)
 | |
|                 if (needToSuppressWithThreshold(*boxes, indices[i], indices[selectedIndices[j]], T(threshold))) {
 | |
|                     shouldSelect = false;
 | |
|                 }
 | |
|             }
 | |
|             if (shouldSelect) {
 | |
|                 output->p(numSelected, indices[i]);
 | |
|                 selectedIndices[numSelected++] = i;
 | |
|             }
 | |
|         }
 | |
|     }
 | |
| 
 | |
|     void nonMaxSuppressionV2(nd4j::LaunchContext * context, NDArray* boxes, NDArray* scales, int maxSize, double threshold, NDArray* output) {
 | |
|         BUILD_SINGLE_SELECTOR(boxes->dataType(), nonMaxSuppressionV2_, (boxes, scales, maxSize, threshold, output), NUMERIC_TYPES);
 | |
|     }
 | |
|     BUILD_SINGLE_TEMPLATE(template void nonMaxSuppressionV2_, (NDArray* boxes, NDArray* scales, int maxSize, double threshold, NDArray* output), NUMERIC_TYPES);
 | |
| 
 | |
| }
 | |
| }
 | |
| } |