From afd38f0c617d6f89b2b4532c6c44f116617e2b6f Mon Sep 17 00:00:00 2001 From: Felix Thomasmathibalan Date: Wed, 27 Sep 2023 17:46:17 +0100 Subject: Apply clang-format on repository Code is formatted as per a revised clang format configuration file(not part of this delivery). Version 14.0.6 is used. Exclusion List: - files with .cl extension - files that are not strictly C/C++ (e.g. Android.bp, Sconscript ...) And the following directories - compute_kernel_writer/validation/ - tests/ - include/ - src/core/NEON/kernels/convolution/ - src/core/NEON/kernels/arm_gemm/ - src/core/NEON/kernels/arm_conv/ - data/ There will be a follow up for formatting of .cl files and the files under tests/ and compute_kernel_writer/validation/. Signed-off-by: Felix Thomasmathibalan Change-Id: Ib7eb1fcf4e7537b9feaefcfc15098a804a3fde0a Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10391 Benchmark: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Gunes Bayir --- .../CPPBoxWithNonMaximaSuppressionLimit.cpp | 157 +++++--- .../CPP/functions/CPPDetectionOutputLayer.cpp | 312 +++++++++------- .../CPP/functions/CPPDetectionPostProcessLayer.cpp | 414 +++++++++++++-------- .../CPP/functions/CPPNonMaximumSuppression.cpp | 21 +- src/runtime/CPP/functions/CPPTopKV.cpp | 5 +- 5 files changed, 557 insertions(+), 352 deletions(-) (limited to 'src/runtime/CPP/functions') diff --git a/src/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.cpp b/src/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.cpp index dccbe4045d..94a1673d59 100644 --- a/src/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.cpp +++ b/src/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.cpp @@ -42,28 +42,37 @@ void dequantize_tensor(const ITensor *input, ITensor *output) Iterator input_it(input, window); Iterator output_it(output, window); - switch(data_type) + switch (data_type) { case DataType::QASYMM8: - execute_window_loop(window, [&](const Coordinates &) - { - *reinterpret_cast(output_it.ptr()) = dequantize(*reinterpret_cast(input_it.ptr()), qinfo.scale, qinfo.offset); - }, - input_it, output_it); + execute_window_loop( + window, + [&](const Coordinates &) + { + *reinterpret_cast(output_it.ptr()) = + dequantize(*reinterpret_cast(input_it.ptr()), qinfo.scale, qinfo.offset); + }, + input_it, output_it); break; case DataType::QASYMM8_SIGNED: - execute_window_loop(window, [&](const Coordinates &) - { - *reinterpret_cast(output_it.ptr()) = dequantize_qasymm8_signed(*reinterpret_cast(input_it.ptr()), qinfo); - }, - input_it, output_it); + execute_window_loop( + window, + [&](const Coordinates &) + { + *reinterpret_cast(output_it.ptr()) = + dequantize_qasymm8_signed(*reinterpret_cast(input_it.ptr()), qinfo); + }, + input_it, output_it); break; case DataType::QASYMM16: - execute_window_loop(window, [&](const Coordinates &) - { - *reinterpret_cast(output_it.ptr()) = dequantize(*reinterpret_cast(input_it.ptr()), qinfo.scale, qinfo.offset); - }, - input_it, output_it); + execute_window_loop( + window, + [&](const Coordinates &) + { + *reinterpret_cast(output_it.ptr()) = + dequantize(*reinterpret_cast(input_it.ptr()), qinfo.scale, qinfo.offset); + }, + input_it, output_it); break; default: ARM_COMPUTE_ERROR("Unsupported data type"); @@ -80,28 +89,37 @@ void quantize_tensor(const ITensor *input, ITensor *output) Iterator input_it(input, window); Iterator output_it(output, window); - switch(data_type) + switch (data_type) { case DataType::QASYMM8: - execute_window_loop(window, [&](const Coordinates &) - { - *reinterpret_cast(output_it.ptr()) = quantize_qasymm8(*reinterpret_cast(input_it.ptr()), qinfo); - }, - input_it, output_it); + execute_window_loop( + window, + [&](const Coordinates &) + { + *reinterpret_cast(output_it.ptr()) = + quantize_qasymm8(*reinterpret_cast(input_it.ptr()), qinfo); + }, + input_it, output_it); break; case DataType::QASYMM8_SIGNED: - execute_window_loop(window, [&](const Coordinates &) - { - *reinterpret_cast(output_it.ptr()) = quantize_qasymm8_signed(*reinterpret_cast(input_it.ptr()), qinfo); - }, - input_it, output_it); + execute_window_loop( + window, + [&](const Coordinates &) + { + *reinterpret_cast(output_it.ptr()) = + quantize_qasymm8_signed(*reinterpret_cast(input_it.ptr()), qinfo); + }, + input_it, output_it); break; case DataType::QASYMM16: - execute_window_loop(window, [&](const Coordinates &) - { - *reinterpret_cast(output_it.ptr()) = quantize_qasymm16(*reinterpret_cast(input_it.ptr()), qinfo); - }, - input_it, output_it); + execute_window_loop( + window, + [&](const Coordinates &) + { + *reinterpret_cast(output_it.ptr()) = + quantize_qasymm16(*reinterpret_cast(input_it.ptr()), qinfo); + }, + input_it, output_it); break; default: ARM_COMPUTE_ERROR("Unsupported data type"); @@ -132,14 +150,23 @@ CPPBoxWithNonMaximaSuppressionLimit::CPPBoxWithNonMaximaSuppressionLimit(std::sh { } -void CPPBoxWithNonMaximaSuppressionLimit::configure(const ITensor *scores_in, const ITensor *boxes_in, const ITensor *batch_splits_in, - ITensor *scores_out, ITensor *boxes_out, ITensor *classes, ITensor *batch_splits_out, - ITensor *keeps, ITensor *keeps_size, const BoxNMSLimitInfo info) +void CPPBoxWithNonMaximaSuppressionLimit::configure(const ITensor *scores_in, + const ITensor *boxes_in, + const ITensor *batch_splits_in, + ITensor *scores_out, + ITensor *boxes_out, + ITensor *classes, + ITensor *batch_splits_out, + ITensor *keeps, + ITensor *keeps_size, + const BoxNMSLimitInfo info) { ARM_COMPUTE_ERROR_ON_NULLPTR(scores_in, boxes_in, scores_out, boxes_out, classes); - ARM_COMPUTE_LOG_PARAMS(scores_in, boxes_in, batch_splits_in, scores_out, boxes_out, classes, batch_splits_out, keeps, keeps_size, info); + ARM_COMPUTE_LOG_PARAMS(scores_in, boxes_in, batch_splits_in, scores_out, boxes_out, classes, batch_splits_out, + keeps, keeps_size, info); - _is_qasymm8 = scores_in->info()->data_type() == DataType::QASYMM8 || scores_in->info()->data_type() == DataType::QASYMM8_SIGNED; + _is_qasymm8 = scores_in->info()->data_type() == DataType::QASYMM8 || + scores_in->info()->data_type() == DataType::QASYMM8_SIGNED; _scores_in = scores_in; _boxes_in = boxes_in; @@ -150,7 +177,7 @@ void CPPBoxWithNonMaximaSuppressionLimit::configure(const ITensor *scores_in, co _batch_splits_out = batch_splits_out; _keeps = keeps; - if(_is_qasymm8) + if (_is_qasymm8) { // Manage intermediate buffers _memory_group.manage(&_scores_in_f32); @@ -160,7 +187,7 @@ void CPPBoxWithNonMaximaSuppressionLimit::configure(const ITensor *scores_in, co _memory_group.manage(&_classes_f32); _scores_in_f32.allocator()->init(scores_in->info()->clone()->set_data_type(DataType::F32)); _boxes_in_f32.allocator()->init(boxes_in->info()->clone()->set_data_type(DataType::F32)); - if(batch_splits_in != nullptr) + if (batch_splits_in != nullptr) { _memory_group.manage(&_batch_splits_in_f32); _batch_splits_in_f32.allocator()->init(batch_splits_in->info()->clone()->set_data_type(DataType::F32)); @@ -168,58 +195,70 @@ void CPPBoxWithNonMaximaSuppressionLimit::configure(const ITensor *scores_in, co _scores_out_f32.allocator()->init(scores_out->info()->clone()->set_data_type(DataType::F32)); _boxes_out_f32.allocator()->init(boxes_out->info()->clone()->set_data_type(DataType::F32)); _classes_f32.allocator()->init(classes->info()->clone()->set_data_type(DataType::F32)); - if(batch_splits_out != nullptr) + if (batch_splits_out != nullptr) { _memory_group.manage(&_batch_splits_out_f32); _batch_splits_out_f32.allocator()->init(batch_splits_out->info()->clone()->set_data_type(DataType::F32)); } - if(keeps != nullptr) + if (keeps != nullptr) { _memory_group.manage(&_keeps_f32); _keeps_f32.allocator()->init(keeps->info()->clone()->set_data_type(DataType::F32)); } - _box_with_nms_limit_kernel.configure(&_scores_in_f32, &_boxes_in_f32, (batch_splits_in != nullptr) ? &_batch_splits_in_f32 : nullptr, + _box_with_nms_limit_kernel.configure(&_scores_in_f32, &_boxes_in_f32, + (batch_splits_in != nullptr) ? &_batch_splits_in_f32 : nullptr, &_scores_out_f32, &_boxes_out_f32, &_classes_f32, - (batch_splits_out != nullptr) ? &_batch_splits_out_f32 : nullptr, (keeps != nullptr) ? &_keeps_f32 : nullptr, - keeps_size, info); + (batch_splits_out != nullptr) ? &_batch_splits_out_f32 : nullptr, + (keeps != nullptr) ? &_keeps_f32 : nullptr, keeps_size, info); } else { - _box_with_nms_limit_kernel.configure(scores_in, boxes_in, batch_splits_in, scores_out, boxes_out, classes, batch_splits_out, keeps, keeps_size, info); + _box_with_nms_limit_kernel.configure(scores_in, boxes_in, batch_splits_in, scores_out, boxes_out, classes, + batch_splits_out, keeps, keeps_size, info); } - if(_is_qasymm8) + if (_is_qasymm8) { _scores_in_f32.allocator()->allocate(); _boxes_in_f32.allocator()->allocate(); - if(_batch_splits_in != nullptr) + if (_batch_splits_in != nullptr) { _batch_splits_in_f32.allocator()->allocate(); } _scores_out_f32.allocator()->allocate(); _boxes_out_f32.allocator()->allocate(); _classes_f32.allocator()->allocate(); - if(batch_splits_out != nullptr) + if (batch_splits_out != nullptr) { _batch_splits_out_f32.allocator()->allocate(); } - if(keeps != nullptr) + if (keeps != nullptr) { _keeps_f32.allocator()->allocate(); } } } -Status validate(const ITensorInfo *scores_in, const ITensorInfo *boxes_in, const ITensorInfo *batch_splits_in, const ITensorInfo *scores_out, const ITensorInfo *boxes_out, const ITensorInfo *classes, - const ITensorInfo *batch_splits_out, const ITensorInfo *keeps, const ITensorInfo *keeps_size, const BoxNMSLimitInfo info) +Status validate(const ITensorInfo *scores_in, + const ITensorInfo *boxes_in, + const ITensorInfo *batch_splits_in, + const ITensorInfo *scores_out, + const ITensorInfo *boxes_out, + const ITensorInfo *classes, + const ITensorInfo *batch_splits_out, + const ITensorInfo *keeps, + const ITensorInfo *keeps_size, + const BoxNMSLimitInfo info) { ARM_COMPUTE_UNUSED(batch_splits_in, batch_splits_out, keeps, keeps_size, info); ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(scores_in, boxes_in, scores_out, boxes_out, classes); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(scores_in, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(scores_in, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, + DataType::F16, DataType::F32); - const bool is_qasymm8 = scores_in->data_type() == DataType::QASYMM8 || scores_in->data_type() == DataType::QASYMM8_SIGNED; - if(is_qasymm8) + const bool is_qasymm8 = + scores_in->data_type() == DataType::QASYMM8 || scores_in->data_type() == DataType::QASYMM8_SIGNED; + if (is_qasymm8) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(boxes_in, 1, DataType::QASYMM16); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(boxes_in, boxes_out); @@ -237,11 +276,11 @@ void CPPBoxWithNonMaximaSuppressionLimit::run() // Acquire all the temporaries MemoryGroupResourceScope scope_mg(_memory_group); - if(_is_qasymm8) + if (_is_qasymm8) { dequantize_tensor(_scores_in, &_scores_in_f32); dequantize_tensor(_boxes_in, &_boxes_in_f32); - if(_batch_splits_in != nullptr) + if (_batch_splits_in != nullptr) { dequantize_tensor(_batch_splits_in, &_batch_splits_in_f32); } @@ -249,16 +288,16 @@ void CPPBoxWithNonMaximaSuppressionLimit::run() Scheduler::get().schedule(&_box_with_nms_limit_kernel, Window::DimY); - if(_is_qasymm8) + if (_is_qasymm8) { quantize_tensor(&_scores_out_f32, _scores_out); quantize_tensor(&_boxes_out_f32, _boxes_out); quantize_tensor(&_classes_f32, _classes); - if(_batch_splits_out != nullptr) + if (_batch_splits_out != nullptr) { quantize_tensor(&_batch_splits_out_f32, _batch_splits_out); } - if(_keeps != nullptr) + if (_keeps != nullptr) { quantize_tensor(&_keeps_f32, _keeps); } diff --git a/src/runtime/CPP/functions/CPPDetectionOutputLayer.cpp b/src/runtime/CPP/functions/CPPDetectionOutputLayer.cpp index 41d875eb97..e6291f973e 100644 --- a/src/runtime/CPP/functions/CPPDetectionOutputLayer.cpp +++ b/src/runtime/CPP/functions/CPPDetectionOutputLayer.cpp @@ -26,9 +26,9 @@ #include "arm_compute/core/Error.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/Validate.h" -#include "src/core/helpers/AutoConfiguration.h" #include "src/common/utils/Log.h" +#include "src/core/helpers/AutoConfiguration.h" #include @@ -36,25 +36,35 @@ namespace arm_compute { namespace { -Status validate_arguments(const ITensorInfo *input_loc, const ITensorInfo *input_conf, const ITensorInfo *input_priorbox, const ITensorInfo *output, DetectionOutputLayerInfo info) +Status validate_arguments(const ITensorInfo *input_loc, + const ITensorInfo *input_conf, + const ITensorInfo *input_priorbox, + const ITensorInfo *output, + DetectionOutputLayerInfo info) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input_loc, input_conf, input_priorbox, output); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input_loc, 1, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_loc, input_conf, input_priorbox); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_loc->num_dimensions() > 2, "The location input tensor should be [C1, N]."); ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_conf->num_dimensions() > 2, "The location input tensor should be [C2, N]."); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_priorbox->num_dimensions() > 3, "The priorbox input tensor should be [C3, 2, N]."); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_priorbox->num_dimensions() > 3, + "The priorbox input tensor should be [C3, 2, N]."); ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.eta() <= 0.f && info.eta() > 1.f, "Eta should be between 0 and 1"); const int num_priors = input_priorbox->tensor_shape()[0] / 4; - ARM_COMPUTE_RETURN_ERROR_ON_MSG(static_cast((num_priors * info.num_loc_classes() * 4)) != input_loc->tensor_shape()[0], "Number of priors must match number of location predictions."); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(static_cast((num_priors * info.num_classes())) != input_conf->tensor_shape()[0], "Number of priors must match number of confidence predictions."); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(static_cast((num_priors * info.num_loc_classes() * 4)) != + input_loc->tensor_shape()[0], + "Number of priors must match number of location predictions."); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(static_cast((num_priors * info.num_classes())) != + input_conf->tensor_shape()[0], + "Number of priors must match number of confidence predictions."); // Validate configured output - if(output->total_size() != 0) + if (output->total_size() != 0) { - const unsigned int max_size = info.keep_top_k() * (input_loc->num_dimensions() > 1 ? input_loc->dimension(1) : 1); + const unsigned int max_size = + info.keep_top_k() * (input_loc->num_dimensions() > 1 ? input_loc->dimension(1) : 1); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), TensorShape(7U, max_size)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_loc, output); } @@ -65,8 +75,7 @@ Status validate_arguments(const ITensorInfo *input_loc, const ITensorInfo *input /** Function used to sort pair in descend order based on the score (first) value. */ template -bool SortScorePairDescend(const std::pair &pair1, - const std::pair &pair2) +bool SortScorePairDescend(const std::pair &pair1, const std::pair &pair2) { return pair1.first > pair2.first; } @@ -82,16 +91,19 @@ bool SortScorePairDescend(const std::pair &pair1, * @param[out] all_location_predictions All the location predictions. * */ -void retrieve_all_loc_predictions(const ITensor *input_loc, const int num, - const int num_priors, const int num_loc_classes, - const bool share_location, std::vector &all_location_predictions) +void retrieve_all_loc_predictions(const ITensor *input_loc, + const int num, + const int num_priors, + const int num_loc_classes, + const bool share_location, + std::vector &all_location_predictions) { - for(int i = 0; i < num; ++i) + for (int i = 0; i < num; ++i) { - for(int c = 0; c < num_loc_classes; ++c) + for (int c = 0; c < num_loc_classes; ++c) { int label = share_location ? -1 : c; - if(all_location_predictions[i].find(label) == all_location_predictions[i].end()) + if (all_location_predictions[i].find(label) == all_location_predictions[i].end()) { all_location_predictions[i][label].resize(num_priors); } @@ -102,19 +114,23 @@ void retrieve_all_loc_predictions(const ITensor *input_loc, const int num, } } } - for(int i = 0; i < num; ++i) + for (int i = 0; i < num; ++i) { - for(int p = 0; p < num_priors; ++p) + for (int p = 0; p < num_priors; ++p) { - for(int c = 0; c < num_loc_classes; ++c) + for (int c = 0; c < num_loc_classes; ++c) { const int label = share_location ? -1 : c; const int base_ptr = i * num_priors * num_loc_classes * 4 + p * num_loc_classes * 4 + c * 4; //xmin, ymin, xmax, ymax - all_location_predictions[i][label][p][0] = *reinterpret_cast(input_loc->ptr_to_element(Coordinates(base_ptr))); - all_location_predictions[i][label][p][1] = *reinterpret_cast(input_loc->ptr_to_element(Coordinates(base_ptr + 1))); - all_location_predictions[i][label][p][2] = *reinterpret_cast(input_loc->ptr_to_element(Coordinates(base_ptr + 2))); - all_location_predictions[i][label][p][3] = *reinterpret_cast(input_loc->ptr_to_element(Coordinates(base_ptr + 3))); + all_location_predictions[i][label][p][0] = + *reinterpret_cast(input_loc->ptr_to_element(Coordinates(base_ptr))); + all_location_predictions[i][label][p][1] = + *reinterpret_cast(input_loc->ptr_to_element(Coordinates(base_ptr + 1))); + all_location_predictions[i][label][p][2] = + *reinterpret_cast(input_loc->ptr_to_element(Coordinates(base_ptr + 2))); + all_location_predictions[i][label][p][3] = + *reinterpret_cast(input_loc->ptr_to_element(Coordinates(base_ptr + 3))); } } } @@ -130,26 +146,28 @@ void retrieve_all_loc_predictions(const ITensor *input_loc, const int num, * @param[out] all_location_predictions All the location predictions. * */ -void retrieve_all_conf_scores(const ITensor *input_conf, const int num, - const int num_priors, const int num_classes, +void retrieve_all_conf_scores(const ITensor *input_conf, + const int num, + const int num_priors, + const int num_classes, std::vector>> &all_confidence_scores) { std::vector tmp_buffer; tmp_buffer.resize(num * num_priors * num_classes); - for(int i = 0; i < num; ++i) + for (int i = 0; i < num; ++i) { - for(int c = 0; c < num_classes; ++c) + for (int c = 0; c < num_classes; ++c) { - for(int p = 0; p < num_priors; ++p) + for (int p = 0; p < num_priors; ++p) { - tmp_buffer[i * num_classes * num_priors + c * num_priors + p] = - *reinterpret_cast(input_conf->ptr_to_element(Coordinates(i * num_classes * num_priors + p * num_classes + c))); + tmp_buffer[i * num_classes * num_priors + c * num_priors + p] = *reinterpret_cast( + input_conf->ptr_to_element(Coordinates(i * num_classes * num_priors + p * num_classes + c))); } } } - for(int i = 0; i < num; ++i) + for (int i = 0; i < num; ++i) { - for(int c = 0; c < num_classes; ++c) + for (int c = 0; c < num_classes; ++c) { all_confidence_scores[i][c].resize(num_priors); all_confidence_scores[i][c].assign(&tmp_buffer[i * num_classes * num_priors + c * num_priors], @@ -168,28 +186,23 @@ void retrieve_all_conf_scores(const ITensor *input_conf, const int num, * @param[out] all_location_predictions All the location predictions. * */ -void retrieve_all_priorbox(const ITensor *input_priorbox, - const int num_priors, - std::vector &all_prior_bboxes, +void retrieve_all_priorbox(const ITensor *input_priorbox, + const int num_priors, + std::vector &all_prior_bboxes, std::vector> &all_prior_variances) { - for(int i = 0; i < num_priors; ++i) + for (int i = 0; i < num_priors; ++i) { - all_prior_bboxes[i] = - { - { - *reinterpret_cast(input_priorbox->ptr_to_element(Coordinates(i * 4))), - *reinterpret_cast(input_priorbox->ptr_to_element(Coordinates(i * 4 + 1))), - *reinterpret_cast(input_priorbox->ptr_to_element(Coordinates(i * 4 + 2))), - *reinterpret_cast(input_priorbox->ptr_to_element(Coordinates(i * 4 + 3))) - } - }; + all_prior_bboxes[i] = {{*reinterpret_cast(input_priorbox->ptr_to_element(Coordinates(i * 4))), + *reinterpret_cast(input_priorbox->ptr_to_element(Coordinates(i * 4 + 1))), + *reinterpret_cast(input_priorbox->ptr_to_element(Coordinates(i * 4 + 2))), + *reinterpret_cast(input_priorbox->ptr_to_element(Coordinates(i * 4 + 3)))}}; } - std::array var({ { 0, 0, 0, 0 } }); - for(int i = 0; i < num_priors; ++i) + std::array var({{0, 0, 0, 0}}); + for (int i = 0; i < num_priors; ++i) { - for(int j = 0; j < 4; ++j) + for (int j = 0; j < 4; ++j) { var[j] = *reinterpret_cast(input_priorbox->ptr_to_element(Coordinates((num_priors + i) * 4 + j))); } @@ -208,13 +221,17 @@ void retrieve_all_priorbox(const ITensor *input_priorbox, * @param[out] decode_bbox The decoded bboxes. * */ -void DecodeBBox(const BBox &prior_bbox, const std::array &prior_variance, - const DetectionOutputLayerCodeType code_type, const bool variance_encoded_in_target, - const bool clip_bbox, const BBox &bbox, BBox &decode_bbox) +void DecodeBBox(const BBox &prior_bbox, + const std::array &prior_variance, + const DetectionOutputLayerCodeType code_type, + const bool variance_encoded_in_target, + const bool clip_bbox, + const BBox &bbox, + BBox &decode_bbox) { // if the variance is encoded in target, we simply need to add the offset predictions // otherwise we need to scale the offset accordingly. - switch(code_type) + switch (code_type) { case DetectionOutputLayerCodeType::CORNER: { @@ -237,10 +254,14 @@ void DecodeBBox(const BBox &prior_bbox, const std::array &prior_varian const float prior_center_x = (prior_bbox[0] + prior_bbox[2]) / 2.; const float prior_center_y = (prior_bbox[1] + prior_bbox[3]) / 2.; - const float decode_bbox_center_x = (variance_encoded_in_target ? bbox[0] : prior_variance[0] * bbox[0]) * prior_width + prior_center_x; - const float decode_bbox_center_y = (variance_encoded_in_target ? bbox[1] : prior_variance[1] * bbox[1]) * prior_height + prior_center_y; - const float decode_bbox_width = (variance_encoded_in_target ? std::exp(bbox[2]) : std::exp(prior_variance[2] * bbox[2])) * prior_width; - const float decode_bbox_height = (variance_encoded_in_target ? std::exp(bbox[3]) : std::exp(prior_variance[3] * bbox[3])) * prior_height; + const float decode_bbox_center_x = + (variance_encoded_in_target ? bbox[0] : prior_variance[0] * bbox[0]) * prior_width + prior_center_x; + const float decode_bbox_center_y = + (variance_encoded_in_target ? bbox[1] : prior_variance[1] * bbox[1]) * prior_height + prior_center_y; + const float decode_bbox_width = + (variance_encoded_in_target ? std::exp(bbox[2]) : std::exp(prior_variance[2] * bbox[2])) * prior_width; + const float decode_bbox_height = + (variance_encoded_in_target ? std::exp(bbox[3]) : std::exp(prior_variance[3] * bbox[3])) * prior_height; decode_bbox[0] = (decode_bbox_center_x - decode_bbox_width / 2.f); decode_bbox[1] = (decode_bbox_center_y - decode_bbox_height / 2.f); @@ -258,10 +279,14 @@ void DecodeBBox(const BBox &prior_bbox, const std::array &prior_varian ARM_COMPUTE_ERROR_ON(prior_width <= 0.f); ARM_COMPUTE_ERROR_ON(prior_height <= 0.f); - decode_bbox[0] = prior_bbox[0] + (variance_encoded_in_target ? bbox[0] : prior_variance[0] * bbox[0]) * prior_width; - decode_bbox[1] = prior_bbox[1] + (variance_encoded_in_target ? bbox[1] : prior_variance[1] * bbox[1]) * prior_height; - decode_bbox[2] = prior_bbox[2] + (variance_encoded_in_target ? bbox[2] : prior_variance[2] * bbox[2]) * prior_width; - decode_bbox[3] = prior_bbox[3] + (variance_encoded_in_target ? bbox[3] : prior_variance[3] * bbox[3]) * prior_height; + decode_bbox[0] = + prior_bbox[0] + (variance_encoded_in_target ? bbox[0] : prior_variance[0] * bbox[0]) * prior_width; + decode_bbox[1] = + prior_bbox[1] + (variance_encoded_in_target ? bbox[1] : prior_variance[1] * bbox[1]) * prior_height; + decode_bbox[2] = + prior_bbox[2] + (variance_encoded_in_target ? bbox[2] : prior_variance[2] * bbox[2]) * prior_width; + decode_bbox[3] = + prior_bbox[3] + (variance_encoded_in_target ? bbox[3] : prior_variance[3] * bbox[3]) * prior_height; break; } @@ -269,9 +294,9 @@ void DecodeBBox(const BBox &prior_bbox, const std::array &prior_varian ARM_COMPUTE_ERROR("Unsupported Detection Output Code Type."); } - if(clip_bbox) + if (clip_bbox) { - for(auto &d_bbox : decode_bbox) + for (auto &d_bbox : decode_bbox) { d_bbox = utility::clamp(d_bbox, 0.f, 1.f); } @@ -289,10 +314,13 @@ void DecodeBBox(const BBox &prior_bbox, const std::array &prior_varian * @param[out] indices The kept indices of bboxes after nms. * */ -void ApplyNMSFast(const std::vector &bboxes, - const std::vector &scores, const float score_threshold, - const float nms_threshold, const float eta, const int top_k, - std::vector &indices) +void ApplyNMSFast(const std::vector &bboxes, + const std::vector &scores, + const float score_threshold, + const float nms_threshold, + const float eta, + const int top_k, + std::vector &indices) { ARM_COMPUTE_ERROR_ON_MSG(bboxes.size() != scores.size(), "bboxes and scores have different size."); @@ -300,9 +328,9 @@ void ApplyNMSFast(const std::vector &bboxes, std::list> score_index_vec; // Generate index score pairs. - for(size_t i = 0; i < scores.size(); ++i) + for (size_t i = 0; i < scores.size(); ++i) { - if(scores[i] > score_threshold) + if (scores[i] > score_threshold) { score_index_vec.emplace_back(std::make_pair(scores[i], i)); } @@ -313,7 +341,7 @@ void ApplyNMSFast(const std::vector &bboxes, // Keep top_k scores if needed. const int score_index_vec_size = score_index_vec.size(); - if(top_k > -1 && top_k < score_index_vec_size) + if (top_k > -1 && top_k < score_index_vec_size) { score_index_vec.resize(top_k); } @@ -322,46 +350,45 @@ void ApplyNMSFast(const std::vector &bboxes, float adaptive_threshold = nms_threshold; indices.clear(); - while(!score_index_vec.empty()) + while (!score_index_vec.empty()) { const int idx = score_index_vec.front().second; bool keep = true; - for(int kept_idx : indices) + for (int kept_idx : indices) { - if(keep) + if (keep) { // Compute the jaccard (intersection over union IoU) overlap between two bboxes. - BBox intersect_bbox = std::array({ 0, 0, 0, 0 }); - if(bboxes[kept_idx][0] > bboxes[idx][2] || bboxes[kept_idx][2] < bboxes[idx][0] || bboxes[kept_idx][1] > bboxes[idx][3] || bboxes[kept_idx][3] < bboxes[idx][1]) + BBox intersect_bbox = std::array({0, 0, 0, 0}); + if (bboxes[kept_idx][0] > bboxes[idx][2] || bboxes[kept_idx][2] < bboxes[idx][0] || + bboxes[kept_idx][1] > bboxes[idx][3] || bboxes[kept_idx][3] < bboxes[idx][1]) { - intersect_bbox = std::array({ { 0, 0, 0, 0 } }); + intersect_bbox = std::array({{0, 0, 0, 0}}); } else { - intersect_bbox = std::array({ { - std::max(bboxes[idx][0], bboxes[kept_idx][0]), - std::max(bboxes[idx][1], bboxes[kept_idx][1]), - std::min(bboxes[idx][2], bboxes[kept_idx][2]), - std::min(bboxes[idx][3], bboxes[kept_idx][3]) - } - }); + intersect_bbox = std::array( + {{std::max(bboxes[idx][0], bboxes[kept_idx][0]), std::max(bboxes[idx][1], bboxes[kept_idx][1]), + std::min(bboxes[idx][2], bboxes[kept_idx][2]), + std::min(bboxes[idx][3], bboxes[kept_idx][3])}}); } float intersect_width = intersect_bbox[2] - intersect_bbox[0]; float intersect_height = intersect_bbox[3] - intersect_bbox[1]; float overlap = 0.f; - if(intersect_width > 0 && intersect_height > 0) + if (intersect_width > 0 && intersect_height > 0) { float intersect_size = intersect_width * intersect_height; - float bbox1_size = (bboxes[idx][2] < bboxes[idx][0] - || bboxes[idx][3] < bboxes[idx][1]) ? - 0.f : - (bboxes[idx][2] - bboxes[idx][0]) * (bboxes[idx][3] - bboxes[idx][1]); //BBoxSize(bboxes[idx]); - float bbox2_size = (bboxes[kept_idx][2] < bboxes[kept_idx][0] - || bboxes[kept_idx][3] < bboxes[kept_idx][1]) ? - 0.f : - (bboxes[kept_idx][2] - bboxes[kept_idx][0]) * (bboxes[kept_idx][3] - bboxes[kept_idx][1]); // BBoxSize(bboxes[kept_idx]); + float bbox1_size = (bboxes[idx][2] < bboxes[idx][0] || bboxes[idx][3] < bboxes[idx][1]) + ? 0.f + : (bboxes[idx][2] - bboxes[idx][0]) * + (bboxes[idx][3] - bboxes[idx][1]); //BBoxSize(bboxes[idx]); + float bbox2_size = + (bboxes[kept_idx][2] < bboxes[kept_idx][0] || bboxes[kept_idx][3] < bboxes[kept_idx][1]) + ? 0.f + : (bboxes[kept_idx][2] - bboxes[kept_idx][0]) * + (bboxes[kept_idx][3] - bboxes[kept_idx][1]); // BBoxSize(bboxes[kept_idx]); overlap = intersect_size / (bbox1_size + bbox2_size - intersect_size); } keep = (overlap <= adaptive_threshold); @@ -371,12 +398,12 @@ void ApplyNMSFast(const std::vector &bboxes, break; } } - if(keep) + if (keep) { indices.push_back(idx); } score_index_vec.erase(score_index_vec.begin()); - if(keep && eta < 1.f && adaptive_threshold > 0.5f) + if (keep && eta < 1.f && adaptive_threshold > 0.5f) { adaptive_threshold *= eta; } @@ -385,13 +412,27 @@ void ApplyNMSFast(const std::vector &bboxes, } // namespace CPPDetectionOutputLayer::CPPDetectionOutputLayer() - : _input_loc(nullptr), _input_conf(nullptr), _input_priorbox(nullptr), _output(nullptr), _info(), _num_priors(), _num(), _all_location_predictions(), _all_confidence_scores(), _all_prior_bboxes(), - _all_prior_variances(), _all_decode_bboxes(), _all_indices() + : _input_loc(nullptr), + _input_conf(nullptr), + _input_priorbox(nullptr), + _output(nullptr), + _info(), + _num_priors(), + _num(), + _all_location_predictions(), + _all_confidence_scores(), + _all_prior_bboxes(), + _all_prior_variances(), + _all_decode_bboxes(), + _all_indices() { } -void CPPDetectionOutputLayer::configure(const ITensor *input_loc, const ITensor *input_conf, const ITensor *input_priorbox, - ITensor *output, DetectionOutputLayerInfo info) +void CPPDetectionOutputLayer::configure(const ITensor *input_loc, + const ITensor *input_conf, + const ITensor *input_priorbox, + ITensor *output, + DetectionOutputLayerInfo info) { ARM_COMPUTE_ERROR_ON_NULLPTR(input_loc, input_conf, input_priorbox, output); ARM_COMPUTE_LOG_PARAMS(input_loc, input_conf, input_priorbox, output, info); @@ -400,11 +441,13 @@ void CPPDetectionOutputLayer::configure(const ITensor *input_loc, const ITensor // Since the number of bboxes to kept is unknown before nms, the shape is set to the maximum // The maximum is keep_top_k * input_loc_size[1] // Each row is a 7 dimension std::vector, which stores [image_id, label, confidence, xmin, ymin, xmax, ymax] - const unsigned int max_size = info.keep_top_k() * (input_loc->info()->num_dimensions() > 1 ? input_loc->info()->dimension(1) : 1); + const unsigned int max_size = + info.keep_top_k() * (input_loc->info()->num_dimensions() > 1 ? input_loc->info()->dimension(1) : 1); auto_init_if_empty(*output->info(), input_loc->info()->clone()->set_tensor_shape(TensorShape(7U, max_size))); // Perform validation step - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input_loc->info(), input_conf->info(), input_priorbox->info(), output->info(), info)); + ARM_COMPUTE_ERROR_THROW_ON( + validate_arguments(input_loc->info(), input_conf->info(), input_priorbox->info(), output->info(), info)); _input_loc = input_loc; _input_conf = input_conf; @@ -420,12 +463,12 @@ void CPPDetectionOutputLayer::configure(const ITensor *input_loc, const ITensor _all_prior_variances.resize(_num_priors); _all_decode_bboxes.resize(_num); - for(int i = 0; i < _num; ++i) + for (int i = 0; i < _num; ++i) { - for(int c = 0; c < _info.num_loc_classes(); ++c) + for (int c = 0; c < _info.num_loc_classes(); ++c) { const int label = _info.share_location() ? -1 : c; - if(label == _info.background_label_id()) + if (label == _info.background_label_id()) { // Ignore background class. continue; @@ -440,7 +483,11 @@ void CPPDetectionOutputLayer::configure(const ITensor *input_loc, const ITensor output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape())); } -Status CPPDetectionOutputLayer::validate(const ITensorInfo *input_loc, const ITensorInfo *input_conf, const ITensorInfo *input_priorbox, const ITensorInfo *output, DetectionOutputLayerInfo info) +Status CPPDetectionOutputLayer::validate(const ITensorInfo *input_loc, + const ITensorInfo *input_conf, + const ITensorInfo *input_priorbox, + const ITensorInfo *output, + DetectionOutputLayerInfo info) { ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input_loc, input_conf, input_priorbox, output, info)); return Status{}; @@ -449,7 +496,8 @@ Status CPPDetectionOutputLayer::validate(const ITensorInfo *input_loc, const ITe void CPPDetectionOutputLayer::run() { // Retrieve all location predictions. - retrieve_all_loc_predictions(_input_loc, _num, _num_priors, _info.num_loc_classes(), _info.share_location(), _all_location_predictions); + retrieve_all_loc_predictions(_input_loc, _num, _num_priors, _info.num_loc_classes(), _info.share_location(), + _all_location_predictions); // Retrieve all confidences. retrieve_all_conf_scores(_input_conf, _num, _num_priors, _info.num_classes(), _all_confidence_scores); @@ -459,75 +507,79 @@ void CPPDetectionOutputLayer::run() // Decode all loc predictions to bboxes const bool clip_bbox = false; - for(int i = 0; i < _num; ++i) + for (int i = 0; i < _num; ++i) { - for(int c = 0; c < _info.num_loc_classes(); ++c) + for (int c = 0; c < _info.num_loc_classes(); ++c) { const int label = _info.share_location() ? -1 : c; - if(label == _info.background_label_id()) + if (label == _info.background_label_id()) { // Ignore background class. continue; } - ARM_COMPUTE_ERROR_ON_MSG_VAR(_all_location_predictions[i].find(label) == _all_location_predictions[i].end(), "Could not find location predictions for label %d.", label); + ARM_COMPUTE_ERROR_ON_MSG_VAR(_all_location_predictions[i].find(label) == _all_location_predictions[i].end(), + "Could not find location predictions for label %d.", label); const std::vector &label_loc_preds = _all_location_predictions[i].find(label)->second; const int num_bboxes = _all_prior_bboxes.size(); ARM_COMPUTE_ERROR_ON(_all_prior_variances[i].size() != 4); - for(int j = 0; j < num_bboxes; ++j) + for (int j = 0; j < num_bboxes; ++j) { - DecodeBBox(_all_prior_bboxes[j], _all_prior_variances[j], _info.code_type(), _info.variance_encoded_in_target(), clip_bbox, label_loc_preds[j], _all_decode_bboxes[i][label][j]); + DecodeBBox(_all_prior_bboxes[j], _all_prior_variances[j], _info.code_type(), + _info.variance_encoded_in_target(), clip_bbox, label_loc_preds[j], + _all_decode_bboxes[i][label][j]); } } } int num_kept = 0; - for(int i = 0; i < _num; ++i) + for (int i = 0; i < _num; ++i) { - const LabelBBox &decode_bboxes = _all_decode_bboxes[i]; - const std::map> &conf_scores = _all_confidence_scores[i]; + const LabelBBox &decode_bboxes = _all_decode_bboxes[i]; + const std::map> &conf_scores = _all_confidence_scores[i]; std::map> indices; - int num_det = 0; - for(int c = 0; c < _info.num_classes(); ++c) + int num_det = 0; + for (int c = 0; c < _info.num_classes(); ++c) { - if(c == _info.background_label_id()) + if (c == _info.background_label_id()) { // Ignore background class continue; } const int label = _info.share_location() ? -1 : c; - if(conf_scores.find(c) == conf_scores.end() || decode_bboxes.find(label) == decode_bboxes.end()) + if (conf_scores.find(c) == conf_scores.end() || decode_bboxes.find(label) == decode_bboxes.end()) { ARM_COMPUTE_ERROR_VAR("Could not find predictions for label %d.", label); } const std::vector &scores = conf_scores.find(c)->second; - const std::vector &bboxes = decode_bboxes.find(label)->second; + const std::vector &bboxes = decode_bboxes.find(label)->second; - ApplyNMSFast(bboxes, scores, _info.confidence_threshold(), _info.nms_threshold(), _info.eta(), _info.top_k(), indices[c]); + ApplyNMSFast(bboxes, scores, _info.confidence_threshold(), _info.nms_threshold(), _info.eta(), + _info.top_k(), indices[c]); num_det += indices[c].size(); } int num_to_add = 0; - if(_info.keep_top_k() > -1 && num_det > _info.keep_top_k()) + if (_info.keep_top_k() > -1 && num_det > _info.keep_top_k()) { std::vector>> score_index_pairs; - for(auto const &it : indices) + for (auto const &it : indices) { const int label = it.first; const std::vector &label_indices = it.second; - if(conf_scores.find(label) == conf_scores.end()) + if (conf_scores.find(label) == conf_scores.end()) { ARM_COMPUTE_ERROR_VAR("Could not find predictions for label %d.", label); } const std::vector &scores = conf_scores.find(label)->second; - for(auto idx : label_indices) + for (auto idx : label_indices) { ARM_COMPUTE_ERROR_ON(idx > static_cast(scores.size())); score_index_pairs.emplace_back(std::make_pair(scores[idx], std::make_pair(label, idx))); @@ -541,7 +593,7 @@ void CPPDetectionOutputLayer::run() // Store the new indices. std::map> new_indices; - for(auto score_index_pair : score_index_pairs) + for (auto score_index_pair : score_index_pairs) { int label = score_index_pair.second.first; int idx = score_index_pair.second.second; @@ -562,25 +614,25 @@ void CPPDetectionOutputLayer::run() _output->info()->set_valid_region(ValidRegion(Coordinates(0, 0), TensorShape(7, num_kept))); int count = 0; - for(int i = 0; i < _num; ++i) + for (int i = 0; i < _num; ++i) { - const std::map> &conf_scores = _all_confidence_scores[i]; - const LabelBBox &decode_bboxes = _all_decode_bboxes[i]; - for(auto &it : _all_indices[i]) + const std::map> &conf_scores = _all_confidence_scores[i]; + const LabelBBox &decode_bboxes = _all_decode_bboxes[i]; + for (auto &it : _all_indices[i]) { const int label = it.first; const std::vector &scores = conf_scores.find(label)->second; const int loc_label = _info.share_location() ? -1 : label; - if(conf_scores.find(label) == conf_scores.end() || decode_bboxes.find(loc_label) == decode_bboxes.end()) + if (conf_scores.find(label) == conf_scores.end() || decode_bboxes.find(loc_label) == decode_bboxes.end()) { // Either if there are no confidence predictions // or there are no location predictions for current label. ARM_COMPUTE_ERROR_VAR("Could not find predictions for the label %d.", label); } const std::vector &bboxes = decode_bboxes.find(loc_label)->second; - const std::vector &indices = it.second; + const std::vector &indices = it.second; - for(auto idx : indices) + for (auto idx : indices) { *(reinterpret_cast(_output->ptr_to_element(Coordinates(count * 7)))) = i; *(reinterpret_cast(_output->ptr_to_element(Coordinates(count * 7 + 1)))) = label; diff --git a/src/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp b/src/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp index ecbc49b3c1..2861d6cacb 100644 --- a/src/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp +++ b/src/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp @@ -26,9 +26,9 @@ #include "arm_compute/core/Error.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/Validate.h" -#include "src/core/helpers/AutoConfiguration.h" #include "src/common/utils/Log.h" +#include "src/core/helpers/AutoConfiguration.h" #include #include @@ -38,53 +38,76 @@ namespace arm_compute { namespace { -Status validate_arguments(const ITensorInfo *input_box_encoding, const ITensorInfo *input_class_score, const ITensorInfo *input_anchors, - ITensorInfo *output_boxes, ITensorInfo *output_classes, ITensorInfo *output_scores, ITensorInfo *num_detection, - DetectionPostProcessLayerInfo info, const unsigned int kBatchSize, const unsigned int kNumCoordBox) +Status validate_arguments(const ITensorInfo *input_box_encoding, + const ITensorInfo *input_class_score, + const ITensorInfo *input_anchors, + ITensorInfo *output_boxes, + ITensorInfo *output_classes, + ITensorInfo *output_scores, + ITensorInfo *num_detection, + DetectionPostProcessLayerInfo info, + const unsigned int kBatchSize, + const unsigned int kNumCoordBox) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input_box_encoding, input_class_score, input_anchors); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input_box_encoding, 1, DataType::F32, DataType::QASYMM8, DataType::QASYMM8_SIGNED); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input_box_encoding, 1, DataType::F32, DataType::QASYMM8, + DataType::QASYMM8_SIGNED); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_box_encoding, input_anchors); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_box_encoding->num_dimensions() > 3, "The location input tensor shape should be [4, N, kBatchSize]."); - if(input_box_encoding->num_dimensions() > 2) + ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_box_encoding->num_dimensions() > 3, + "The location input tensor shape should be [4, N, kBatchSize]."); + if (input_box_encoding->num_dimensions() > 2) { - ARM_COMPUTE_RETURN_ERROR_ON_MSG_VAR(input_box_encoding->dimension(2) != kBatchSize, "The third dimension of the input box_encoding tensor should be equal to %d.", kBatchSize); + ARM_COMPUTE_RETURN_ERROR_ON_MSG_VAR( + input_box_encoding->dimension(2) != kBatchSize, + "The third dimension of the input box_encoding tensor should be equal to %d.", kBatchSize); } - ARM_COMPUTE_RETURN_ERROR_ON_MSG_VAR(input_box_encoding->dimension(0) != kNumCoordBox, "The first dimension of the input box_encoding tensor should be equal to %d.", kNumCoordBox); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_class_score->dimension(0) != (info.num_classes() + 1), - "The first dimension of the input class_prediction should be equal to the number of classes plus one."); - - ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_anchors->num_dimensions() > 3, "The anchors input tensor shape should be [4, N, kBatchSize]."); - if(input_anchors->num_dimensions() > 2) + ARM_COMPUTE_RETURN_ERROR_ON_MSG_VAR(input_box_encoding->dimension(0) != kNumCoordBox, + "The first dimension of the input box_encoding tensor should be equal to %d.", + kNumCoordBox); + ARM_COMPUTE_RETURN_ERROR_ON_MSG( + input_class_score->dimension(0) != (info.num_classes() + 1), + "The first dimension of the input class_prediction should be equal to the number of classes plus one."); + + ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_anchors->num_dimensions() > 3, + "The anchors input tensor shape should be [4, N, kBatchSize]."); + if (input_anchors->num_dimensions() > 2) { - ARM_COMPUTE_RETURN_ERROR_ON_MSG_VAR(input_anchors->dimension(0) != kNumCoordBox, "The first dimension of the input anchors tensor should be equal to %d.", kNumCoordBox); + ARM_COMPUTE_RETURN_ERROR_ON_MSG_VAR(input_anchors->dimension(0) != kNumCoordBox, + "The first dimension of the input anchors tensor should be equal to %d.", + kNumCoordBox); } - ARM_COMPUTE_RETURN_ERROR_ON_MSG((input_box_encoding->dimension(1) != input_class_score->dimension(1)) - || (input_box_encoding->dimension(1) != input_anchors->dimension(1)), + ARM_COMPUTE_RETURN_ERROR_ON_MSG((input_box_encoding->dimension(1) != input_class_score->dimension(1)) || + (input_box_encoding->dimension(1) != input_anchors->dimension(1)), "The second dimension of the inputs should be the same."); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(num_detection->num_dimensions() > 1, "The num_detection output tensor shape should be [M]."); - ARM_COMPUTE_RETURN_ERROR_ON_MSG((info.iou_threshold() <= 0.0f) || (info.iou_threshold() > 1.0f), "The intersection over union should be positive and less than 1."); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.max_classes_per_detection() <= 0, "The number of max classes per detection should be positive."); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(num_detection->num_dimensions() > 1, + "The num_detection output tensor shape should be [M]."); + ARM_COMPUTE_RETURN_ERROR_ON_MSG((info.iou_threshold() <= 0.0f) || (info.iou_threshold() > 1.0f), + "The intersection over union should be positive and less than 1."); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.max_classes_per_detection() <= 0, + "The number of max classes per detection should be positive."); const unsigned int num_detected_boxes = info.max_detections() * info.max_classes_per_detection(); // Validate configured outputs - if(output_boxes->total_size() != 0) + if (output_boxes->total_size() != 0) { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output_boxes->tensor_shape(), TensorShape(4U, num_detected_boxes, 1U)); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output_boxes->tensor_shape(), + TensorShape(4U, num_detected_boxes, 1U)); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_boxes, 1, DataType::F32); } - if(output_classes->total_size() != 0) + if (output_classes->total_size() != 0) { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output_classes->tensor_shape(), TensorShape(num_detected_boxes, 1U)); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output_classes->tensor_shape(), + TensorShape(num_detected_boxes, 1U)); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_classes, 1, DataType::F32); } - if(output_scores->total_size() != 0) + if (output_scores->total_size() != 0) { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output_scores->tensor_shape(), TensorShape(num_detected_boxes, 1U)); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output_scores->tensor_shape(), + TensorShape(num_detected_boxes, 1U)); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_scores, 1, DataType::F32); } - if(num_detection->total_size() != 0) + if (num_detection->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(num_detection->tensor_shape(), TensorShape(1U)); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(num_detection, 1, DataType::F32); @@ -93,15 +116,18 @@ Status validate_arguments(const ITensorInfo *input_box_encoding, const ITensorIn return Status{}; } -inline void DecodeBoxCorner(BBox &box_centersize, BBox &anchor, Iterator &decoded_it, DetectionPostProcessLayerInfo info) +inline void +DecodeBoxCorner(BBox &box_centersize, BBox &anchor, Iterator &decoded_it, DetectionPostProcessLayerInfo info) { const float half_factor = 0.5f; // BBox is equavalent to CenterSizeEncoding [y,x,h,w] const float y_center = box_centersize[0] / info.scale_value_y() * anchor[2] + anchor[0]; const float x_center = box_centersize[1] / info.scale_value_x() * anchor[3] + anchor[1]; - const float half_h = half_factor * static_cast(std::exp(box_centersize[2] / info.scale_value_h())) * anchor[2]; - const float half_w = half_factor * static_cast(std::exp(box_centersize[3] / info.scale_value_w())) * anchor[3]; + const float half_h = + half_factor * static_cast(std::exp(box_centersize[2] / info.scale_value_h())) * anchor[2]; + const float half_w = + half_factor * static_cast(std::exp(box_centersize[3] / info.scale_value_w())) * anchor[3]; // Box Corner encoding boxes are saved as [xmin, ymin, xmax, ymax] auto decoded_ptr = reinterpret_cast(decoded_it.ptr()); @@ -118,12 +144,15 @@ inline void DecodeBoxCorner(BBox &box_centersize, BBox &anchor, Iterator &decode * @param[in] info The detection informations * @param[out] decoded_boxes The decoded bboxes. */ -void DecodeCenterSizeBoxes(const ITensor *input_box_encoding, const ITensor *input_anchors, DetectionPostProcessLayerInfo info, Tensor *decoded_boxes) +void DecodeCenterSizeBoxes(const ITensor *input_box_encoding, + const ITensor *input_anchors, + DetectionPostProcessLayerInfo info, + Tensor *decoded_boxes) { const QuantizationInfo &qi_box = input_box_encoding->info()->quantization_info(); const QuantizationInfo &qi_anchors = input_anchors->info()->quantization_info(); - BBox box_centersize{ {} }; - BBox anchor{ {} }; + BBox box_centersize{{}}; + BBox anchor{{}}; Window win; win.use_tensor_dimensions(input_box_encoding->info()->tensor_shape()); @@ -133,107 +162,155 @@ void DecodeCenterSizeBoxes(const ITensor *input_box_encoding, const ITensor *inp Iterator anchor_it(input_anchors, win); Iterator decoded_it(decoded_boxes, win); - if(input_box_encoding->info()->data_type() == DataType::QASYMM8) + if (input_box_encoding->info()->data_type() == DataType::QASYMM8) { - execute_window_loop(win, [&](const Coordinates &) - { - const auto box_ptr = reinterpret_cast(box_it.ptr()); - const auto anchor_ptr = reinterpret_cast(anchor_it.ptr()); - box_centersize = BBox({ dequantize_qasymm8(*box_ptr, qi_box), dequantize_qasymm8(*(box_ptr + 1), qi_box), - dequantize_qasymm8(*(2 + box_ptr), qi_box), dequantize_qasymm8(*(3 + box_ptr), qi_box) - }); - anchor = BBox({ dequantize_qasymm8(*anchor_ptr, qi_anchors), dequantize_qasymm8(*(anchor_ptr + 1), qi_anchors), - dequantize_qasymm8(*(2 + anchor_ptr), qi_anchors), dequantize_qasymm8(*(3 + anchor_ptr), qi_anchors) - }); - DecodeBoxCorner(box_centersize, anchor, decoded_it, info); - }, - box_it, anchor_it, decoded_it); + execute_window_loop( + win, + [&](const Coordinates &) + { + const auto box_ptr = reinterpret_cast(box_it.ptr()); + const auto anchor_ptr = reinterpret_cast(anchor_it.ptr()); + box_centersize = + BBox({dequantize_qasymm8(*box_ptr, qi_box), dequantize_qasymm8(*(box_ptr + 1), qi_box), + dequantize_qasymm8(*(2 + box_ptr), qi_box), dequantize_qasymm8(*(3 + box_ptr), qi_box)}); + anchor = BBox({dequantize_qasymm8(*anchor_ptr, qi_anchors), + dequantize_qasymm8(*(anchor_ptr + 1), qi_anchors), + dequantize_qasymm8(*(2 + anchor_ptr), qi_anchors), + dequantize_qasymm8(*(3 + anchor_ptr), qi_anchors)}); + DecodeBoxCorner(box_centersize, anchor, decoded_it, info); + }, + box_it, anchor_it, decoded_it); } - else if(input_box_encoding->info()->data_type() == DataType::QASYMM8_SIGNED) + else if (input_box_encoding->info()->data_type() == DataType::QASYMM8_SIGNED) { - execute_window_loop(win, [&](const Coordinates &) - { - const auto box_ptr = reinterpret_cast(box_it.ptr()); - const auto anchor_ptr = reinterpret_cast(anchor_it.ptr()); - box_centersize = BBox({ dequantize_qasymm8_signed(*box_ptr, qi_box), dequantize_qasymm8_signed(*(box_ptr + 1), qi_box), - dequantize_qasymm8_signed(*(2 + box_ptr), qi_box), dequantize_qasymm8_signed(*(3 + box_ptr), qi_box) - }); - anchor = BBox({ dequantize_qasymm8_signed(*anchor_ptr, qi_anchors), dequantize_qasymm8_signed(*(anchor_ptr + 1), qi_anchors), - dequantize_qasymm8_signed(*(2 + anchor_ptr), qi_anchors), dequantize_qasymm8_signed(*(3 + anchor_ptr), qi_anchors) - }); - DecodeBoxCorner(box_centersize, anchor, decoded_it, info); - }, - box_it, anchor_it, decoded_it); + execute_window_loop( + win, + [&](const Coordinates &) + { + const auto box_ptr = reinterpret_cast(box_it.ptr()); + const auto anchor_ptr = reinterpret_cast(anchor_it.ptr()); + box_centersize = BBox({dequantize_qasymm8_signed(*box_ptr, qi_box), + dequantize_qasymm8_signed(*(box_ptr + 1), qi_box), + dequantize_qasymm8_signed(*(2 + box_ptr), qi_box), + dequantize_qasymm8_signed(*(3 + box_ptr), qi_box)}); + anchor = BBox({dequantize_qasymm8_signed(*anchor_ptr, qi_anchors), + dequantize_qasymm8_signed(*(anchor_ptr + 1), qi_anchors), + dequantize_qasymm8_signed(*(2 + anchor_ptr), qi_anchors), + dequantize_qasymm8_signed(*(3 + anchor_ptr), qi_anchors)}); + DecodeBoxCorner(box_centersize, anchor, decoded_it, info); + }, + box_it, anchor_it, decoded_it); } else { - execute_window_loop(win, [&](const Coordinates &) - { - const auto box_ptr = reinterpret_cast(box_it.ptr()); - const auto anchor_ptr = reinterpret_cast(anchor_it.ptr()); - box_centersize = BBox({ *box_ptr, *(box_ptr + 1), *(2 + box_ptr), *(3 + box_ptr) }); - anchor = BBox({ *anchor_ptr, *(anchor_ptr + 1), *(2 + anchor_ptr), *(3 + anchor_ptr) }); - DecodeBoxCorner(box_centersize, anchor, decoded_it, info); - }, - box_it, anchor_it, decoded_it); + execute_window_loop( + win, + [&](const Coordinates &) + { + const auto box_ptr = reinterpret_cast(box_it.ptr()); + const auto anchor_ptr = reinterpret_cast(anchor_it.ptr()); + box_centersize = BBox({*box_ptr, *(box_ptr + 1), *(2 + box_ptr), *(3 + box_ptr)}); + anchor = BBox({*anchor_ptr, *(anchor_ptr + 1), *(2 + anchor_ptr), *(3 + anchor_ptr)}); + DecodeBoxCorner(box_centersize, anchor, decoded_it, info); + }, + box_it, anchor_it, decoded_it); } } -void SaveOutputs(const Tensor *decoded_boxes, const std::vector &result_idx_boxes_after_nms, const std::vector &result_scores_after_nms, const std::vector &result_classes_after_nms, - std::vector &sorted_indices, const unsigned int num_output, const unsigned int max_detections, ITensor *output_boxes, ITensor *output_classes, ITensor *output_scores, - ITensor *num_detection) +void SaveOutputs(const Tensor *decoded_boxes, + const std::vector &result_idx_boxes_after_nms, + const std::vector &result_scores_after_nms, + const std::vector &result_classes_after_nms, + std::vector &sorted_indices, + const unsigned int num_output, + const unsigned int max_detections, + ITensor *output_boxes, + ITensor *output_classes, + ITensor *output_scores, + ITensor *num_detection) { // xmin,ymin,xmax,ymax -> ymin,xmin,ymax,xmax unsigned int i = 0; - for(; i < num_output; ++i) + for (; i < num_output; ++i) { const unsigned int box_in_idx = result_idx_boxes_after_nms[sorted_indices[i]]; - *(reinterpret_cast(output_boxes->ptr_to_element(Coordinates(0, i)))) = *(reinterpret_cast(decoded_boxes->ptr_to_element(Coordinates(1, box_in_idx)))); - *(reinterpret_cast(output_boxes->ptr_to_element(Coordinates(1, i)))) = *(reinterpret_cast(decoded_boxes->ptr_to_element(Coordinates(0, box_in_idx)))); - *(reinterpret_cast(output_boxes->ptr_to_element(Coordinates(2, i)))) = *(reinterpret_cast(decoded_boxes->ptr_to_element(Coordinates(3, box_in_idx)))); - *(reinterpret_cast(output_boxes->ptr_to_element(Coordinates(3, i)))) = *(reinterpret_cast(decoded_boxes->ptr_to_element(Coordinates(2, box_in_idx)))); - *(reinterpret_cast(output_classes->ptr_to_element(Coordinates(i)))) = static_cast(result_classes_after_nms[sorted_indices[i]]); - *(reinterpret_cast(output_scores->ptr_to_element(Coordinates(i)))) = result_scores_after_nms[sorted_indices[i]]; + *(reinterpret_cast(output_boxes->ptr_to_element(Coordinates(0, i)))) = + *(reinterpret_cast(decoded_boxes->ptr_to_element(Coordinates(1, box_in_idx)))); + *(reinterpret_cast(output_boxes->ptr_to_element(Coordinates(1, i)))) = + *(reinterpret_cast(decoded_boxes->ptr_to_element(Coordinates(0, box_in_idx)))); + *(reinterpret_cast(output_boxes->ptr_to_element(Coordinates(2, i)))) = + *(reinterpret_cast(decoded_boxes->ptr_to_element(Coordinates(3, box_in_idx)))); + *(reinterpret_cast(output_boxes->ptr_to_element(Coordinates(3, i)))) = + *(reinterpret_cast(decoded_boxes->ptr_to_element(Coordinates(2, box_in_idx)))); + *(reinterpret_cast(output_classes->ptr_to_element(Coordinates(i)))) = + static_cast(result_classes_after_nms[sorted_indices[i]]); + *(reinterpret_cast(output_scores->ptr_to_element(Coordinates(i)))) = + result_scores_after_nms[sorted_indices[i]]; } - for(; i < max_detections; ++i) + for (; i < max_detections; ++i) { *(reinterpret_cast(output_boxes->ptr_to_element(Coordinates(1, i)))) = 0.0f; *(reinterpret_cast(output_boxes->ptr_to_element(Coordinates(0, i)))) = 0.0f; *(reinterpret_cast(output_boxes->ptr_to_element(Coordinates(3, i)))) = 0.0f; *(reinterpret_cast(output_boxes->ptr_to_element(Coordinates(2, i)))) = 0.0f; - *(reinterpret_cast(output_classes->ptr_to_element(Coordinates(i)))) = 0.0f; - *(reinterpret_cast(output_scores->ptr_to_element(Coordinates(i)))) = 0.0f; + *(reinterpret_cast(output_classes->ptr_to_element(Coordinates(i)))) = 0.0f; + *(reinterpret_cast(output_scores->ptr_to_element(Coordinates(i)))) = 0.0f; } *(reinterpret_cast(num_detection->ptr_to_element(Coordinates(0)))) = num_output; } } // namespace CPPDetectionPostProcessLayer::CPPDetectionPostProcessLayer(std::shared_ptr memory_manager) - : _memory_group(std::move(memory_manager)), _nms(), _input_box_encoding(nullptr), _input_scores(nullptr), _input_anchors(nullptr), _output_boxes(nullptr), _output_classes(nullptr), - _output_scores(nullptr), _num_detection(nullptr), _info(), _num_boxes(), _num_classes_with_background(), _num_max_detected_boxes(), _dequantize_scores(false), _decoded_boxes(), _decoded_scores(), - _selected_indices(), _class_scores(), _input_scores_to_use(nullptr) + : _memory_group(std::move(memory_manager)), + _nms(), + _input_box_encoding(nullptr), + _input_scores(nullptr), + _input_anchors(nullptr), + _output_boxes(nullptr), + _output_classes(nullptr), + _output_scores(nullptr), + _num_detection(nullptr), + _info(), + _num_boxes(), + _num_classes_with_background(), + _num_max_detected_boxes(), + _dequantize_scores(false), + _decoded_boxes(), + _decoded_scores(), + _selected_indices(), + _class_scores(), + _input_scores_to_use(nullptr) { } -void CPPDetectionPostProcessLayer::configure(const ITensor *input_box_encoding, const ITensor *input_scores, - const ITensor *input_anchors, ITensor *output_boxes, ITensor *output_classes, - ITensor *output_scores, ITensor *num_detection, DetectionPostProcessLayerInfo info) +void CPPDetectionPostProcessLayer::configure(const ITensor *input_box_encoding, + const ITensor *input_scores, + const ITensor *input_anchors, + ITensor *output_boxes, + ITensor *output_classes, + ITensor *output_scores, + ITensor *num_detection, + DetectionPostProcessLayerInfo info) { - ARM_COMPUTE_ERROR_ON_NULLPTR(input_box_encoding, input_scores, input_anchors, output_boxes, output_classes, output_scores); + ARM_COMPUTE_ERROR_ON_NULLPTR(input_box_encoding, input_scores, input_anchors, output_boxes, output_classes, + output_scores); ARM_COMPUTE_LOG_PARAMS(input_box_encoding, input_scores, input_anchors, output_boxes, output_classes, output_scores, num_detection, info); _num_max_detected_boxes = info.max_detections() * info.max_classes_per_detection(); - auto_init_if_empty(*output_boxes->info(), TensorInfo(TensorShape(_kNumCoordBox, _num_max_detected_boxes, _kBatchSize), 1, DataType::F32)); - auto_init_if_empty(*output_classes->info(), TensorInfo(TensorShape(_num_max_detected_boxes, _kBatchSize), 1, DataType::F32)); - auto_init_if_empty(*output_scores->info(), TensorInfo(TensorShape(_num_max_detected_boxes, _kBatchSize), 1, DataType::F32)); + auto_init_if_empty(*output_boxes->info(), + TensorInfo(TensorShape(_kNumCoordBox, _num_max_detected_boxes, _kBatchSize), 1, DataType::F32)); + auto_init_if_empty(*output_classes->info(), + TensorInfo(TensorShape(_num_max_detected_boxes, _kBatchSize), 1, DataType::F32)); + auto_init_if_empty(*output_scores->info(), + TensorInfo(TensorShape(_num_max_detected_boxes, _kBatchSize), 1, DataType::F32)); auto_init_if_empty(*num_detection->info(), TensorInfo(TensorShape(1U), 1, DataType::F32)); // Perform validation step - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input_box_encoding->info(), input_scores->info(), input_anchors->info(), output_boxes->info(), output_classes->info(), output_scores->info(), - num_detection->info(), - info, _kBatchSize, _kNumCoordBox)); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments( + input_box_encoding->info(), input_scores->info(), input_anchors->info(), output_boxes->info(), + output_classes->info(), output_scores->info(), num_detection->info(), info, _kBatchSize, _kNumCoordBox)); _input_box_encoding = input_box_encoding; _input_scores = input_scores; @@ -245,13 +322,24 @@ void CPPDetectionPostProcessLayer::configure(const ITensor *input_box_encoding, _info = info; _num_boxes = input_box_encoding->info()->dimension(1); _num_classes_with_background = _input_scores->info()->dimension(0); - _dequantize_scores = (info.dequantize_scores() && is_data_type_quantized(input_box_encoding->info()->data_type())); - - auto_init_if_empty(*_decoded_boxes.info(), TensorInfo(TensorShape(_kNumCoordBox, _input_box_encoding->info()->dimension(1), _kBatchSize), 1, DataType::F32)); - auto_init_if_empty(*_decoded_scores.info(), TensorInfo(TensorShape(_input_scores->info()->dimension(0), _input_scores->info()->dimension(1), _kBatchSize), 1, DataType::F32)); - auto_init_if_empty(*_selected_indices.info(), TensorInfo(TensorShape(info.use_regular_nms() ? info.detection_per_class() : info.max_detections()), 1, DataType::S32)); + _dequantize_scores = (info.dequantize_scores() && is_data_type_quantized(input_box_encoding->info()->data_type())); + + auto_init_if_empty(*_decoded_boxes.info(), + TensorInfo(TensorShape(_kNumCoordBox, _input_box_encoding->info()->dimension(1), _kBatchSize), 1, + DataType::F32)); + auto_init_if_empty( + *_decoded_scores.info(), + TensorInfo(TensorShape(_input_scores->info()->dimension(0), _input_scores->info()->dimension(1), _kBatchSize), + 1, DataType::F32)); + auto_init_if_empty( + *_selected_indices.info(), + TensorInfo(TensorShape(info.use_regular_nms() ? info.detection_per_class() : info.max_detections()), 1, + DataType::S32)); const unsigned int num_classes_per_box = std::min(info.max_classes_per_detection(), info.num_classes()); - auto_init_if_empty(*_class_scores.info(), TensorInfo(info.use_regular_nms() ? TensorShape(_num_boxes) : TensorShape(_num_boxes * num_classes_per_box), 1, DataType::F32)); + auto_init_if_empty( + *_class_scores.info(), + TensorInfo(info.use_regular_nms() ? TensorShape(_num_boxes) : TensorShape(_num_boxes * num_classes_per_box), 1, + DataType::F32)); _input_scores_to_use = _dequantize_scores ? &_decoded_scores : _input_scores; @@ -260,7 +348,9 @@ void CPPDetectionPostProcessLayer::configure(const ITensor *input_box_encoding, _memory_group.manage(&_decoded_scores); _memory_group.manage(&_selected_indices); _memory_group.manage(&_class_scores); - _nms.configure(&_decoded_boxes, &_class_scores, &_selected_indices, info.use_regular_nms() ? info.detection_per_class() : info.max_detections(), info.nms_score_threshold(), info.iou_threshold()); + _nms.configure(&_decoded_boxes, &_class_scores, &_selected_indices, + info.use_regular_nms() ? info.detection_per_class() : info.max_detections(), + info.nms_score_threshold(), info.iou_threshold()); // Allocate and reserve intermediate tensors and vectors _decoded_boxes.allocator()->allocate(); @@ -269,18 +359,28 @@ void CPPDetectionPostProcessLayer::configure(const ITensor *input_box_encoding, _class_scores.allocator()->allocate(); } -Status CPPDetectionPostProcessLayer::validate(const ITensorInfo *input_box_encoding, const ITensorInfo *input_class_score, const ITensorInfo *input_anchors, - ITensorInfo *output_boxes, ITensorInfo *output_classes, ITensorInfo *output_scores, ITensorInfo *num_detection, DetectionPostProcessLayerInfo info) +Status CPPDetectionPostProcessLayer::validate(const ITensorInfo *input_box_encoding, + const ITensorInfo *input_class_score, + const ITensorInfo *input_anchors, + ITensorInfo *output_boxes, + ITensorInfo *output_classes, + ITensorInfo *output_scores, + ITensorInfo *num_detection, + DetectionPostProcessLayerInfo info) { - constexpr unsigned int kBatchSize = 1; - constexpr unsigned int kNumCoordBox = 4; - const TensorInfo _decoded_boxes_info = TensorInfo(TensorShape(kNumCoordBox, input_box_encoding->dimension(1)), 1, DataType::F32); - const TensorInfo _decoded_scores_info = TensorInfo(TensorShape(input_box_encoding->dimension(1)), 1, DataType::F32); - const TensorInfo _selected_indices_info = TensorInfo(TensorShape(info.max_detections()), 1, DataType::S32); - - ARM_COMPUTE_RETURN_ON_ERROR(CPPNonMaximumSuppression::validate(&_decoded_boxes_info, &_decoded_scores_info, &_selected_indices_info, info.max_detections(), info.nms_score_threshold(), - info.iou_threshold())); - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input_box_encoding, input_class_score, input_anchors, output_boxes, output_classes, output_scores, num_detection, info, kBatchSize, kNumCoordBox)); + constexpr unsigned int kBatchSize = 1; + constexpr unsigned int kNumCoordBox = 4; + const TensorInfo _decoded_boxes_info = + TensorInfo(TensorShape(kNumCoordBox, input_box_encoding->dimension(1)), 1, DataType::F32); + const TensorInfo _decoded_scores_info = TensorInfo(TensorShape(input_box_encoding->dimension(1)), 1, DataType::F32); + const TensorInfo _selected_indices_info = TensorInfo(TensorShape(info.max_detections()), 1, DataType::S32); + + ARM_COMPUTE_RETURN_ON_ERROR(CPPNonMaximumSuppression::validate(&_decoded_boxes_info, &_decoded_scores_info, + &_selected_indices_info, info.max_detections(), + info.nms_score_threshold(), info.iou_threshold())); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input_box_encoding, input_class_score, input_anchors, output_boxes, + output_classes, output_scores, num_detection, info, kBatchSize, + kNumCoordBox)); return Status{}; } @@ -293,62 +393,69 @@ void CPPDetectionPostProcessLayer::run() DecodeCenterSizeBoxes(_input_box_encoding, _input_anchors, _info, &_decoded_boxes); // Decode scores if necessary - if(_dequantize_scores) + if (_dequantize_scores) { - if(_input_box_encoding->info()->data_type() == DataType::QASYMM8) + if (_input_box_encoding->info()->data_type() == DataType::QASYMM8) { - for(unsigned int idx_c = 0; idx_c < _num_classes_with_background; ++idx_c) + for (unsigned int idx_c = 0; idx_c < _num_classes_with_background; ++idx_c) { - for(unsigned int idx_b = 0; idx_b < _num_boxes; ++idx_b) + for (unsigned int idx_b = 0; idx_b < _num_boxes; ++idx_b) { *(reinterpret_cast(_decoded_scores.ptr_to_element(Coordinates(idx_c, idx_b)))) = - dequantize_qasymm8(*(reinterpret_cast(_input_scores->ptr_to_element(Coordinates(idx_c, idx_b)))), _input_scores->info()->quantization_info()); + dequantize_qasymm8( + *(reinterpret_cast(_input_scores->ptr_to_element(Coordinates(idx_c, idx_b)))), + _input_scores->info()->quantization_info()); } } } - else if(_input_box_encoding->info()->data_type() == DataType::QASYMM8_SIGNED) + else if (_input_box_encoding->info()->data_type() == DataType::QASYMM8_SIGNED) { - for(unsigned int idx_c = 0; idx_c < _num_classes_with_background; ++idx_c) + for (unsigned int idx_c = 0; idx_c < _num_classes_with_background; ++idx_c) { - for(unsigned int idx_b = 0; idx_b < _num_boxes; ++idx_b) + for (unsigned int idx_b = 0; idx_b < _num_boxes; ++idx_b) { *(reinterpret_cast(_decoded_scores.ptr_to_element(Coordinates(idx_c, idx_b)))) = - dequantize_qasymm8_signed(*(reinterpret_cast(_input_scores->ptr_to_element(Coordinates(idx_c, idx_b)))), _input_scores->info()->quantization_info()); + dequantize_qasymm8_signed(*(reinterpret_cast( + _input_scores->ptr_to_element(Coordinates(idx_c, idx_b)))), + _input_scores->info()->quantization_info()); } } } } // Regular NMS - if(_info.use_regular_nms()) + if (_info.use_regular_nms()) { std::vector result_idx_boxes_after_nms; std::vector result_classes_after_nms; std::vector result_scores_after_nms; std::vector sorted_indices; - for(unsigned int c = 0; c < num_classes; ++c) + for (unsigned int c = 0; c < num_classes; ++c) { // For each boxes get scores of the boxes for the class c - for(unsigned int i = 0; i < _num_boxes; ++i) + for (unsigned int i = 0; i < _num_boxes; ++i) { *(reinterpret_cast(_class_scores.ptr_to_element(Coordinates(i)))) = - *(reinterpret_cast(_input_scores_to_use->ptr_to_element(Coordinates(c + 1, i)))); // i * _num_classes_with_background + c + 1 + *(reinterpret_cast(_input_scores_to_use->ptr_to_element( + Coordinates(c + 1, i)))); // i * _num_classes_with_background + c + 1 } // Run Non-maxima Suppression _nms.run(); - for(unsigned int i = 0; i < _info.detection_per_class(); ++i) + for (unsigned int i = 0; i < _info.detection_per_class(); ++i) { - const auto selected_index = *(reinterpret_cast(_selected_indices.ptr_to_element(Coordinates(i)))); - if(selected_index == -1) + const auto selected_index = + *(reinterpret_cast(_selected_indices.ptr_to_element(Coordinates(i)))); + if (selected_index == -1) { // Nms will return -1 for all the last M-elements not valid break; } result_idx_boxes_after_nms.emplace_back(selected_index); - result_scores_after_nms.emplace_back((reinterpret_cast(_class_scores.buffer()))[selected_index]); + result_scores_after_nms.emplace_back( + (reinterpret_cast(_class_scores.buffer()))[selected_index]); result_classes_after_nms.emplace_back(c); } } @@ -360,49 +467,46 @@ void CPPDetectionPostProcessLayer::run() // Sort selected indices based on result scores sorted_indices.resize(num_selected); std::iota(sorted_indices.begin(), sorted_indices.end(), 0); - std::partial_sort(sorted_indices.data(), - sorted_indices.data() + num_output, + std::partial_sort(sorted_indices.data(), sorted_indices.data() + num_output, sorted_indices.data() + num_selected, [&](unsigned int first, unsigned int second) - { - - return result_scores_after_nms[first] > result_scores_after_nms[second]; - }); + { return result_scores_after_nms[first] > result_scores_after_nms[second]; }); - SaveOutputs(&_decoded_boxes, result_idx_boxes_after_nms, result_scores_after_nms, result_classes_after_nms, sorted_indices, - num_output, max_detections, _output_boxes, _output_classes, _output_scores, _num_detection); + SaveOutputs(&_decoded_boxes, result_idx_boxes_after_nms, result_scores_after_nms, result_classes_after_nms, + sorted_indices, num_output, max_detections, _output_boxes, _output_classes, _output_scores, + _num_detection); } // Fast NMS else { - const unsigned int num_classes_per_box = std::min(_info.max_classes_per_detection(), _info.num_classes()); + const unsigned int num_classes_per_box = + std::min(_info.max_classes_per_detection(), _info.num_classes()); std::vector max_scores; std::vector box_indices; std::vector max_score_classes; - for(unsigned int b = 0; b < _num_boxes; ++b) + for (unsigned int b = 0; b < _num_boxes; ++b) { std::vector box_scores; - for(unsigned int c = 0; c < num_classes; ++c) + for (unsigned int c = 0; c < num_classes; ++c) { - box_scores.emplace_back(*(reinterpret_cast(_input_scores_to_use->ptr_to_element(Coordinates(c + 1, b))))); + box_scores.emplace_back( + *(reinterpret_cast(_input_scores_to_use->ptr_to_element(Coordinates(c + 1, b))))); } std::vector max_score_indices; max_score_indices.resize(_info.num_classes()); std::iota(max_score_indices.data(), max_score_indices.data() + _info.num_classes(), 0); - std::partial_sort(max_score_indices.data(), - max_score_indices.data() + num_classes_per_box, + std::partial_sort(max_score_indices.data(), max_score_indices.data() + num_classes_per_box, max_score_indices.data() + num_classes, [&](unsigned int first, unsigned int second) - { - return box_scores[first] > box_scores[second]; - }); + { return box_scores[first] > box_scores[second]; }); - for(unsigned int i = 0; i < num_classes_per_box; ++i) + for (unsigned int i = 0; i < num_classes_per_box; ++i) { - const float score_to_add = box_scores[max_score_indices[i]]; - *(reinterpret_cast(_class_scores.ptr_to_element(Coordinates(b * num_classes_per_box + i)))) = score_to_add; + const float score_to_add = box_scores[max_score_indices[i]]; + *(reinterpret_cast(_class_scores.ptr_to_element(Coordinates(b * num_classes_per_box + i)))) = + score_to_add; max_scores.emplace_back(score_to_add); box_indices.emplace_back(b); max_score_classes.emplace_back(max_score_indices[i]); @@ -412,10 +516,10 @@ void CPPDetectionPostProcessLayer::run() // Run Non-maxima Suppression _nms.run(); std::vector selected_indices; - for(unsigned int i = 0; i < max_detections; ++i) + for (unsigned int i = 0; i < max_detections; ++i) { // NMS returns M valid indices, the not valid tail is filled with -1 - if(*(reinterpret_cast(_selected_indices.ptr_to_element(Coordinates(i)))) == -1) + if (*(reinterpret_cast(_selected_indices.ptr_to_element(Coordinates(i)))) == -1) { // Nms will return -1 for all the last M-elements not valid break; @@ -425,8 +529,8 @@ void CPPDetectionPostProcessLayer::run() // We select the max detection numbers of the highest score of all classes const auto num_output = std::min(_info.max_detections(), selected_indices.size()); - SaveOutputs(&_decoded_boxes, box_indices, max_scores, max_score_classes, selected_indices, - num_output, max_detections, _output_boxes, _output_classes, _output_scores, _num_detection); + SaveOutputs(&_decoded_boxes, box_indices, max_scores, max_score_classes, selected_indices, num_output, + max_detections, _output_boxes, _output_classes, _output_scores, _num_detection); } } } // namespace arm_compute diff --git a/src/runtime/CPP/functions/CPPNonMaximumSuppression.cpp b/src/runtime/CPP/functions/CPPNonMaximumSuppression.cpp index 6d01b127c0..3217742c6b 100644 --- a/src/runtime/CPP/functions/CPPNonMaximumSuppression.cpp +++ b/src/runtime/CPP/functions/CPPNonMaximumSuppression.cpp @@ -29,9 +29,12 @@ namespace arm_compute { -void CPPNonMaximumSuppression::configure( - const ITensor *bboxes, const ITensor *scores, ITensor *indices, unsigned int max_output_size, - const float score_threshold, const float nms_threshold) +void CPPNonMaximumSuppression::configure(const ITensor *bboxes, + const ITensor *scores, + ITensor *indices, + unsigned int max_output_size, + const float score_threshold, + const float nms_threshold) { ARM_COMPUTE_LOG_PARAMS(bboxes, scores, indices, max_output_size, score_threshold, nms_threshold); @@ -40,10 +43,14 @@ void CPPNonMaximumSuppression::configure( _kernel = std::move(k); } -Status CPPNonMaximumSuppression::validate( - const ITensorInfo *bboxes, const ITensorInfo *scores, const ITensorInfo *indices, unsigned int max_output_size, - const float score_threshold, const float nms_threshold) +Status CPPNonMaximumSuppression::validate(const ITensorInfo *bboxes, + const ITensorInfo *scores, + const ITensorInfo *indices, + unsigned int max_output_size, + const float score_threshold, + const float nms_threshold) { - return CPPNonMaximumSuppressionKernel::validate(bboxes, scores, indices, max_output_size, score_threshold, nms_threshold); + return CPPNonMaximumSuppressionKernel::validate(bboxes, scores, indices, max_output_size, score_threshold, + nms_threshold); } } // namespace arm_compute diff --git a/src/runtime/CPP/functions/CPPTopKV.cpp b/src/runtime/CPP/functions/CPPTopKV.cpp index 62a74735a2..3d64def804 100644 --- a/src/runtime/CPP/functions/CPPTopKV.cpp +++ b/src/runtime/CPP/functions/CPPTopKV.cpp @@ -38,7 +38,10 @@ void CPPTopKV::configure(const ITensor *predictions, const ITensor *targets, ITe _kernel = std::move(kernel); } -Status CPPTopKV::validate(const ITensorInfo *predictions, const ITensorInfo *targets, ITensorInfo *output, const unsigned int k) +Status CPPTopKV::validate(const ITensorInfo *predictions, + const ITensorInfo *targets, + ITensorInfo *output, + const unsigned int k) { return CPPTopKVKernel::validate(predictions, targets, output, k); } -- cgit v1.2.1