From 32521430ad67172d372721979e8a32b7d4317620 Mon Sep 17 00:00:00 2001 From: Pablo Tello Date: Thu, 15 Nov 2018 14:43:10 +0000 Subject: COMPMID-1750: Add PriorBox operator to graph API. Change-Id: I5d2ed5dcc342abff8124762f7bdee587cdf20032 --- arm_compute/core/Types.h | 3 +- arm_compute/core/utils/misc/ShapeCalculator.h | 4 +- arm_compute/graph/GraphBuilder.h | 11 ++ arm_compute/graph/INodeVisitor.h | 9 ++ arm_compute/graph/TypePrinter.h | 3 + arm_compute/graph/Types.h | 2 + arm_compute/graph/backends/FunctionHelpers.h | 184 ++++++++++++++++++------- arm_compute/graph/backends/ValidateHelpers.h | 23 ++++ arm_compute/graph/frontend/Layers.h | 54 +++++++- arm_compute/graph/nodes/Nodes.h | 1 + arm_compute/graph/nodes/NodesFwd.h | 1 + arm_compute/graph/nodes/PriorBoxLayerNode.h | 67 +++++++++ scripts/caffe_data_extractor.py | 2 +- src/graph/GraphBuilder.cpp | 15 ++ src/graph/backends/CL/CLFunctionsFactory.cpp | 2 + src/graph/backends/CL/CLNodeValidator.cpp | 2 + src/graph/backends/GLES/GCFunctionsFactory.cpp | 21 ++- src/graph/backends/GLES/GCNodeValidator.cpp | 2 + src/graph/backends/NEON/NEFunctionFactory.cpp | 10 +- src/graph/backends/NEON/NENodeValidator.cpp | 2 + src/graph/detail/ExecutionHelpers.cpp | 3 +- src/graph/nodes/FlattenLayerNode.cpp | 2 +- src/graph/nodes/PriorBoxLayerNode.cpp | 95 +++++++++++++ utils/GraphUtils.cpp | 6 +- utils/GraphUtils.h | 6 +- utils/Utils.h | 8 +- 26 files changed, 461 insertions(+), 77 deletions(-) create mode 100644 arm_compute/graph/nodes/PriorBoxLayerNode.h create mode 100644 src/graph/nodes/PriorBoxLayerNode.cpp diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h index 38094ee56a..ef25dc4150 100644 --- a/arm_compute/core/Types.h +++ b/arm_compute/core/Types.h @@ -857,7 +857,8 @@ public: * @param[in] steps (Optional) Step values. */ PriorBoxLayerInfo(const std::vector &min_sizes, const std::vector &variances, float offset, bool flip = true, bool clip = false, - const std::vector &max_sizes = {}, const std::vector &aspect_ratios = {}, const Coordinates2D &img_size = Coordinates2D{ 0, 0 }, const std::array &steps = { { 0.f, 0.f } }) + const std::vector &max_sizes = {}, const std::vector &aspect_ratios = {}, + const Coordinates2D &img_size = Coordinates2D{ 0, 0 }, const std::array &steps = { { 0.f, 0.f } }) : _min_sizes(min_sizes), _variances(variances), _offset(offset), diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h index 5c9457ed6b..da9ff56fd0 100644 --- a/arm_compute/core/utils/misc/ShapeCalculator.h +++ b/arm_compute/core/utils/misc/ShapeCalculator.h @@ -646,13 +646,13 @@ inline TensorShape compute_space_to_batch_shape(const ITensorInfo *input, const return output_shape; } + inline TensorShape compute_prior_box_shape(const ITensorInfo &input, const PriorBoxLayerInfo &info) { DataLayout data_layout = input.data_layout(); const size_t idx_w = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); const size_t idx_h = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); - - const int num_priors = info.aspect_ratios().size() * info.min_sizes().size() + info.max_sizes().size(); + const int num_priors = info.aspect_ratios().size() * info.min_sizes().size() + info.max_sizes().size(); TensorShape output_shape{}; output_shape.set(0, input.dimension(idx_w) * input.dimension(idx_h) * num_priors * 4); diff --git a/arm_compute/graph/GraphBuilder.h b/arm_compute/graph/GraphBuilder.h index 22fc041684..57ce349984 100644 --- a/arm_compute/graph/GraphBuilder.h +++ b/arm_compute/graph/GraphBuilder.h @@ -308,6 +308,17 @@ public: * @return Node ID of the created node, EmptyNodeID in case of error */ static NodeID add_pooling_node(Graph &g, NodeParams params, NodeIdxPair input, PoolingLayerInfo pool_info); + /** Adds a priorbox layer node to the graph + * + * @param[in] g Graph to add the node to + * @param[in] params Common node parameters + * @param[in] input0 First input to the priorbox layer node as a NodeID-Index pair + * @param[in] input1 Second input to the priorbox layer node as a NodeID-Index pair + * @param[in] prior_info PriorBox parameters + * + * @return Node ID of the created node, EmptyNodeID in case of error + */ + static NodeID add_priorbox_node(Graph &g, NodeParams params, NodeIdxPair input0, NodeIdxPair input1, PriorBoxLayerInfo prior_info); /** Adds a reorg layer node to the graph * * @param[in] g Graph to add the node to diff --git a/arm_compute/graph/INodeVisitor.h b/arm_compute/graph/INodeVisitor.h index a21c9b2d48..2df2574d62 100644 --- a/arm_compute/graph/INodeVisitor.h +++ b/arm_compute/graph/INodeVisitor.h @@ -111,6 +111,11 @@ public: * @param[in] n Node to visit. */ virtual void visit(PoolingLayerNode &n) = 0; + /** Visit PriorBoxLayerNode. + * + * @param[in] n Node to visit. + */ + virtual void visit(PriorBoxLayerNode &n) = 0; /** Visit ReshapeLayerNode. * * @param[in] n Node to visit. @@ -201,6 +206,10 @@ public: { default_visit(); } + virtual void visit(PriorBoxLayerNode &n) override + { + default_visit(); + } virtual void visit(ReshapeLayerNode &n) override { default_visit(); diff --git a/arm_compute/graph/TypePrinter.h b/arm_compute/graph/TypePrinter.h index c66f9cb374..d633091d16 100644 --- a/arm_compute/graph/TypePrinter.h +++ b/arm_compute/graph/TypePrinter.h @@ -113,6 +113,9 @@ inline ::std::ostream &operator<<(::std::ostream &os, const NodeType &node_type) case NodeType::PoolingLayer: os << "PoolingLayer"; break; + case NodeType::PriorBoxLayer: + os << "PriorBoxLayer"; + break; case NodeType::ReorgLayer: os << "ReorgLayer"; break; diff --git a/arm_compute/graph/Types.h b/arm_compute/graph/Types.h index 24c24d328f..b6803c89bc 100644 --- a/arm_compute/graph/Types.h +++ b/arm_compute/graph/Types.h @@ -51,6 +51,7 @@ using arm_compute::FullyConnectedLayerInfo; using arm_compute::PadStrideInfo; using arm_compute::PoolingLayerInfo; using arm_compute::PoolingType; +using arm_compute::PriorBoxLayerInfo; using arm_compute::DimensionRoundingType; using arm_compute::InterpolationPolicy; @@ -141,6 +142,7 @@ enum class NodeType PadLayer, PermuteLayer, PoolingLayer, + PriorBoxLayer, ReorgLayer, ReshapeLayer, ResizeLayer, diff --git a/arm_compute/graph/backends/FunctionHelpers.h b/arm_compute/graph/backends/FunctionHelpers.h index e556e2f284..0d7210f7f8 100644 --- a/arm_compute/graph/backends/FunctionHelpers.h +++ b/arm_compute/graph/backends/FunctionHelpers.h @@ -72,9 +72,9 @@ template void validate_node(const INode &node, size_t num_expected_inputs, size_t num_expected_outputs) { ARM_COMPUTE_LOG_GRAPH_VERBOSE("Creating " << node.type() - << " Target : " << TargetInfo::TargetType - << " ID : " << node.id() - << " Name: " << node.name() + << " Target: " << TargetInfo::TargetType + << " ID: " << node.id() + << node.name() << std::endl); ARM_COMPUTE_ERROR_ON(TargetInfo::TargetType != node.assigned_target()); @@ -105,7 +105,9 @@ std::unique_ptr create_activation_layer(ActivationLayerNode &node) auto func = support::cpp14::make_unique(); func->configure(input, output, act_info); - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() << " Target " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Shape: " << input->info()->tensor_shape() @@ -147,13 +149,15 @@ std::unique_ptr create_batch_normalization_layer(BatchNormalizationLa func->configure(input, output, mean, var, beta, gamma, epsilon, fused_act); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Shape: " << input->info()->tensor_shape() << " Epsilon: " << epsilon << " " << (fused_act.enabled() ? to_string(fused_act.activation()) : "") - << " InPlace : " << is_in_place_operation(input, output) + << " InPlace: " << is_in_place_operation(input, output) << std::endl); return std::move(func); @@ -218,8 +222,10 @@ std::unique_ptr create_channel_shuffle_layer(ChannelShuffleLayerNode auto func = support::cpp14::make_unique(); func->configure(input, output, num_groups); - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Shape: " << input->info()->tensor_shape() << " Num groups: " << num_groups @@ -263,8 +269,10 @@ std::unique_ptr create_concatenate_layer(ConcatenateLaye func->configure(inputs, output, concat_axis); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << output->info()->data_type() << " Shape: " << output->info()->tensor_shape() << " Num Inputs: " << inputs.size() @@ -350,8 +358,10 @@ std::unique_ptr create_convolution_layer(ConvolutionLayerNode &node, << " Weights QuantInfo: " << weights->info()->quantization_info() << " Output QuantInfo: " << output->info()->quantization_info(); } - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << func_name - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << func_name + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Groups: " << num_groups << qss.str() @@ -396,8 +406,10 @@ std::unique_ptr create_deconvolution_layer(DeconvolutionLayerNode &no input, weights, biases, output, deconv_info, inner_border.x(), inner_border.y()); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Input shape: " << input->info()->tensor_shape() << " Weights shape: " << weights->info()->tensor_shape() @@ -462,8 +474,10 @@ std::unique_ptr create_depthwise_convolution_layer(DepthwiseConvoluti << " Weights QuantInfo: " << weights->info()->quantization_info() << " Output QuantInfo: " << output->info()->quantization_info(); } - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << func_name - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << func_name + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << qss.str() << " Input shape: " << input->info()->tensor_shape() @@ -524,11 +538,13 @@ std::unique_ptr create_eltwise_layer(EltwiseLayerNode &node) } // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType - << " Operation " << func_name + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType + << " Operation: " << func_name << " Data Type: " << input1->info()->data_type() - << " Shape : " << input1->info()->tensor_shape() + << " Shape: " << input1->info()->tensor_shape() << std::endl); return func; @@ -560,8 +576,10 @@ std::unique_ptr create_flatten_layer(FlattenLayerNode &node) func->configure(input, output); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Input shape: " << input->info()->tensor_shape() << " Output shape: " << output->info()->tensor_shape() @@ -610,8 +628,10 @@ std::unique_ptr create_fully_connected_layer(FullyConnectedLayerNode << " Weights QuantInfo: " << weights->info()->quantization_info() << " Output QuantInfo: " << output->info()->quantization_info(); } - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << qss.str() << " Input shape: " << input->info()->tensor_shape() @@ -700,8 +720,10 @@ std::unique_ptr create_normalization_layer(NormalizationLayerNode &no func->configure(input, output, norm_info); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Input shape: " << input->info()->tensor_shape() << " Output shape: " << output->info()->tensor_shape() @@ -740,8 +762,10 @@ std::unique_ptr create_normalize_planar_yuv_layer(NormalizePlanarYUVL func->configure(input, output, mean, std); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Shape: " << input->info()->tensor_shape() << std::endl); @@ -775,8 +799,10 @@ std::unique_ptr create_pad_layer(PadLayerNode &node) func->configure(input, output, padding); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Input shape: " << input->info()->tensor_shape() << " Output shape: " << output->info()->tensor_shape() @@ -811,8 +837,10 @@ std::unique_ptr create_permute_layer(PermuteLayerNode &node) func->configure(input, output, perm); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Input shape: " << input->info()->tensor_shape() << " Output shape: " << output->info()->tensor_shape() @@ -848,8 +876,10 @@ std::unique_ptr create_pooling_layer(PoolingLayerNode &node) func->configure(input, output, pool_info); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Input shape: " << input->info()->tensor_shape() << " Output shape: " << output->info()->tensor_shape() @@ -859,6 +889,48 @@ std::unique_ptr create_pooling_layer(PoolingLayerNode &node) return std::move(func); } +/** Create a backend priorbox layer function + * + * @tparam PriorBoxLayerFunction Backend priorbox function + * @tparam TargetInfo Target-specific information + * + * @param[in] node Node to create the backend function for + * + * @return Backend priorbox layer function + */ +template +std::unique_ptr create_priorbox_layer(PriorBoxLayerNode &node) +{ + validate_node(node, 2 /* expected inputs */, 1 /* expected outputs */); + + // Extract IO and info + typename TargetInfo::TensorType *input0 = get_backing_tensor(node.input(0)); + typename TargetInfo::TensorType *input1 = get_backing_tensor(node.input(1)); + typename TargetInfo::TensorType *output = get_backing_tensor(node.output(0)); + const PriorBoxLayerInfo prior_info = node.priorbox_info(); + ARM_COMPUTE_ERROR_ON(input0 == nullptr); + ARM_COMPUTE_ERROR_ON(input1 == nullptr); + ARM_COMPUTE_ERROR_ON(output == nullptr); + + // Create and configure function + auto func = support::cpp14::make_unique(); + func->configure(input0, input1, output, prior_info); + + // Log info + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType + << " Data Type: " << input0->info()->data_type() + << " Input0 shape: " << input0->info()->tensor_shape() + << " Input1 shape: " << input1->info()->tensor_shape() + << " Output shape: " << output->info()->tensor_shape() + << " PriorBoxLayer info: " << prior_info + << std::endl); + + return std::move(func); +} + /** Create a backend reorg layer function * * @tparam ReorgLayerFunction Backend reorg function @@ -884,8 +956,10 @@ std::unique_ptr create_reorg_layer(ReorgLayerNode &node) func->configure(input, output, node.stride()); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Input shape: " << input->info()->tensor_shape() << " Output shape: " << output->info()->tensor_shape() @@ -919,8 +993,10 @@ std::unique_ptr create_reshape_layer(ReshapeLayerNode &node) func->configure(input, output); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Input shape: " << input->info()->tensor_shape() << " Output shape: " << output->info()->tensor_shape() @@ -955,8 +1031,10 @@ std::unique_ptr create_resize_layer(ResizeLayerNode &node) func->configure(input, output, policy, BorderMode::CONSTANT); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Input shape: " << input->info()->tensor_shape() << " Output shape: " << output->info()->tensor_shape() @@ -1034,8 +1112,10 @@ std::unique_ptr create_slice_layer(SliceLayerNode &node) func->configure(input, output, node.starts(), node.ends()); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Input shape: " << input->info()->tensor_shape() << " Output shape: " << output->info()->tensor_shape() @@ -1071,8 +1151,10 @@ std::unique_ptr create_softmax_layer(SoftmaxLayerNode &node, GraphCon func->configure(input, output, beta); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Input shape: " << input->info()->tensor_shape() << " Output shape: " << output->info()->tensor_shape() @@ -1110,8 +1192,10 @@ std::unique_ptr create_upsample_layer(UpsampleLayerNode &node, GraphC func->configure(input, output, info, upsampling_policy); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Input shape: " << input->info()->tensor_shape() << " Output shape: " << output->info()->tensor_shape() @@ -1150,8 +1234,10 @@ std::unique_ptr create_yolo_layer(YOLOLayerNode &node, GraphContext & func->configure(input, output, act_info, num_classes); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Input shape: " << input->info()->tensor_shape() << " Output shape: " << output->info()->tensor_shape() diff --git a/arm_compute/graph/backends/ValidateHelpers.h b/arm_compute/graph/backends/ValidateHelpers.h index 7c31a80967..a6864c2286 100644 --- a/arm_compute/graph/backends/ValidateHelpers.h +++ b/arm_compute/graph/backends/ValidateHelpers.h @@ -300,6 +300,29 @@ Status validate_permute_layer(PermuteLayerNode &node) return PermuteLayer::validate(input, output, perm); } +/** Validates a priorbox layer node + * + * @tparam PriorBoxLayer PriorBox layer type + * + * @param[in] node Node to validate + * + * @return Status + */ +template +Status validate_priorbox_layer(PriorBoxLayerNode &node) +{ + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Validating PriorBoxLayer node with ID : " << node.id() << " and Name: " << node.name() << std::endl); + ARM_COMPUTE_RETURN_ERROR_ON(node.num_inputs() != 2); + ARM_COMPUTE_RETURN_ERROR_ON(node.num_outputs() != 1); + + // Extract IO and info + arm_compute::ITensorInfo *input0 = get_backing_tensor_info(node.input(0)); + arm_compute::ITensorInfo *input1 = get_backing_tensor_info(node.input(1)); + arm_compute::ITensorInfo *output = get_backing_tensor_info(node.output(0)); + const PriorBoxLayerInfo prior_info = node.priorbox_info(); + + return PriorBoxLayer::validate(input0, input1, output, prior_info); +} /** Validates a Reorg layer node * diff --git a/arm_compute/graph/frontend/Layers.h b/arm_compute/graph/frontend/Layers.h index 7ed448e3f2..78a3f20f1f 100644 --- a/arm_compute/graph/frontend/Layers.h +++ b/arm_compute/graph/frontend/Layers.h @@ -225,7 +225,27 @@ public: */ template ConcatLayer(SubStream &&sub_stream1, SubStream &&sub_stream2, Ts &&... rest_sub_streams) - : _sub_streams() + : _sub_streams(), _axis(DataLayoutDimension::CHANNEL) + { + _sub_streams.push_back(arm_compute::support::cpp14::make_unique(std::move(sub_stream1))); + _sub_streams.push_back(arm_compute::support::cpp14::make_unique(std::move(sub_stream2))); + + utility::for_each([&](SubStream && sub_stream) + { + _sub_streams.push_back(arm_compute::support::cpp14::make_unique(std::move(sub_stream))); + }, + std::move(rest_sub_streams)...); + } + /** Construct a concatenation layer + * + * @param[in] axis Axis over the concatenation will be performed + * @param[in] sub_stream1 First graph branch + * @param[in] sub_stream2 Second graph branch + * @param[in] rest_sub_streams Rest sub-graph branches + */ + template + ConcatLayer(DataLayoutDimension axis, SubStream &&sub_stream1, SubStream &&sub_stream2, Ts &&... rest_sub_streams) + : _sub_streams(), _axis(axis) { _sub_streams.push_back(arm_compute::support::cpp14::make_unique(std::move(sub_stream1))); _sub_streams.push_back(arm_compute::support::cpp14::make_unique(std::move(sub_stream2))); @@ -242,7 +262,7 @@ public: */ template ConcatLayer(SubStream &&sub_stream) - : _sub_streams() + : _sub_streams(), _axis(DataLayoutDimension::CHANNEL) { _sub_streams.push_back(arm_compute::support::cpp14::make_unique(std::move(sub_stream))); } @@ -269,13 +289,14 @@ public: } } } - nid = GraphBuilder::add_concatenate_node(s.graph(), common_params, nodes, DataLayoutDimension::CHANNEL); + nid = GraphBuilder::add_concatenate_node(s.graph(), common_params, nodes, _axis); } return nid; } private: std::vector> _sub_streams; + DataLayoutDimension _axis; }; /** Convolution Layer */ @@ -724,6 +745,33 @@ private: PoolingLayerInfo _pool_info; }; +/** PriorBox Layer */ +class PriorBoxLayer final : public ILayer +{ +public: + /** Construct a priorbox layer. + * + * @param[in] sub_stream First graph sub-stream + * @param[in] prior_info PriorBox parameters. + */ + PriorBoxLayer(SubStream &&sub_stream, PriorBoxLayerInfo prior_info) + : _ss(std::move(sub_stream)), _prior_info(prior_info) + { + } + + NodeID create_layer(IStream &s) override + { + NodeParams common_params = { name(), s.hints().target_hint }; + NodeIdxPair input0 = { s.tail_node(), 0 }; + NodeIdxPair input1 = { _ss.tail_node(), 0 }; + return GraphBuilder::add_priorbox_node(s.graph(), common_params, input0, input1, _prior_info); + } + +private: + SubStream _ss; + PriorBoxLayerInfo _prior_info; +}; + /** Reorg Layer */ class ReorgLayer final : public ILayer { diff --git a/arm_compute/graph/nodes/Nodes.h b/arm_compute/graph/nodes/Nodes.h index 342ecbfd3b..5c7599fbfd 100644 --- a/arm_compute/graph/nodes/Nodes.h +++ b/arm_compute/graph/nodes/Nodes.h @@ -45,6 +45,7 @@ #include "arm_compute/graph/nodes/PadLayerNode.h" #include "arm_compute/graph/nodes/PermuteLayerNode.h" #include "arm_compute/graph/nodes/PoolingLayerNode.h" +#include "arm_compute/graph/nodes/PriorBoxLayerNode.h" #include "arm_compute/graph/nodes/ROIAlignLayerNode.h" #include "arm_compute/graph/nodes/ReorgLayerNode.h" #include "arm_compute/graph/nodes/ReshapeLayerNode.h" diff --git a/arm_compute/graph/nodes/NodesFwd.h b/arm_compute/graph/nodes/NodesFwd.h index 8d9bad3771..f956b54c66 100644 --- a/arm_compute/graph/nodes/NodesFwd.h +++ b/arm_compute/graph/nodes/NodesFwd.h @@ -51,6 +51,7 @@ class OutputNode; class PadLayerNode; class PermuteLayerNode; class PoolingLayerNode; +class PriorBoxLayerNode; class ReorgLayerNode; class ReshapeLayerNode; class ResizeLayerNode; diff --git a/arm_compute/graph/nodes/PriorBoxLayerNode.h b/arm_compute/graph/nodes/PriorBoxLayerNode.h new file mode 100644 index 0000000000..901fa0817a --- /dev/null +++ b/arm_compute/graph/nodes/PriorBoxLayerNode.h @@ -0,0 +1,67 @@ +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef __ARM_COMPUTE_GRAPH_PRIORBOX_LAYER_NODE_H__ +#define __ARM_COMPUTE_GRAPH_PRIORBOX_LAYER_NODE_H__ + +#include "arm_compute/graph/INode.h" + +namespace arm_compute +{ +namespace graph +{ +/** PriorBox Layer node */ +class PriorBoxLayerNode final : public INode +{ +public: + /** Constructor + * + * @param[in] prior_info PriorBox Layer information + */ + PriorBoxLayerNode(PriorBoxLayerInfo prior_info); + /** PriorBox metadata accessor + * + * @return PriorBox Layer info + */ + PriorBoxLayerInfo priorbox_info() const; + /** Computes priorbox output descriptor + * + * @param[in] input_descriptor Input descriptor + * @param[in] info PriorBox operation attributes + * + * @return Output descriptor + */ + static TensorDescriptor compute_output_descriptor(const TensorDescriptor &input_descriptor, const PriorBoxLayerInfo &info); + + // Inherited overridden methods: + NodeType type() const override; + bool forward_descriptors() override; + TensorDescriptor configure_output(size_t idx) const override; + void accept(INodeVisitor &v) override; + +private: + PriorBoxLayerInfo _info; +}; +} // namespace graph +} // namespace arm_compute +#endif /* __ARM_COMPUTE_GRAPH_PRIORBOX_LAYER_NODE_H__ */ diff --git a/scripts/caffe_data_extractor.py b/scripts/caffe_data_extractor.py index 65c9938480..47d24b265f 100755 --- a/scripts/caffe_data_extractor.py +++ b/scripts/caffe_data_extractor.py @@ -34,7 +34,7 @@ if __name__ == "__main__": elif i == 1: outname = name + "_b" else: - pass + continue varname = outname if os.path.sep in varname: diff --git a/src/graph/GraphBuilder.cpp b/src/graph/GraphBuilder.cpp index b4c58780bd..b2ca28da57 100644 --- a/src/graph/GraphBuilder.cpp +++ b/src/graph/GraphBuilder.cpp @@ -494,6 +494,21 @@ NodeID GraphBuilder::add_pooling_node(Graph &g, NodeParams params, NodeIdxPair i return create_simple_single_input_output_node(g, params, input, pool_info); } +NodeID GraphBuilder::add_priorbox_node(Graph &g, NodeParams params, NodeIdxPair input0, NodeIdxPair input1, PriorBoxLayerInfo prior_info) +{ + CHECK_NODEIDX_PAIR(input0, g); + CHECK_NODEIDX_PAIR(input1, g); + + // Create priorbox node and connect + NodeID prior_nid = g.add_node(prior_info); + g.add_connection(input0.node_id, input0.index, prior_nid, 0); + g.add_connection(input1.node_id, input1.index, prior_nid, 1); + + set_node_params(g, prior_nid, params); + + return prior_nid; +} + NodeID GraphBuilder::add_reorg_node(Graph &g, NodeParams params, NodeIdxPair input, int stride) { return create_simple_single_input_output_node(g, params, input, stride); diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp index f63aba9ec5..c37a137cf7 100644 --- a/src/graph/backends/CL/CLFunctionsFactory.cpp +++ b/src/graph/backends/CL/CLFunctionsFactory.cpp @@ -113,6 +113,8 @@ std::unique_ptr CLFunctionFactory::create(INode *node, GraphContext & return detail::create_permute_layer(*polymorphic_downcast(node)); case NodeType::PoolingLayer: return detail::create_pooling_layer(*polymorphic_downcast(node)); + case NodeType::PriorBoxLayer: + return detail::create_priorbox_layer(*polymorphic_downcast(node)); case NodeType::ReorgLayer: return detail::create_reorg_layer(*polymorphic_downcast(node)); case NodeType::ReshapeLayer: diff --git a/src/graph/backends/CL/CLNodeValidator.cpp b/src/graph/backends/CL/CLNodeValidator.cpp index 1ea3517467..a070973fd4 100644 --- a/src/graph/backends/CL/CLNodeValidator.cpp +++ b/src/graph/backends/CL/CLNodeValidator.cpp @@ -67,6 +67,8 @@ Status CLNodeValidator::validate(INode *node) return detail::validate_pad_layer(*polymorphic_downcast(node)); case NodeType::PermuteLayer: return detail::validate_permute_layer(*polymorphic_downcast(node)); + case NodeType::PriorBoxLayer: + return detail::validate_priorbox_layer(*polymorphic_downcast(node)); case NodeType::ReorgLayer: return detail::validate_reorg_layer(*polymorphic_downcast(node)); case NodeType::ROIAlignLayer: diff --git a/src/graph/backends/GLES/GCFunctionsFactory.cpp b/src/graph/backends/GLES/GCFunctionsFactory.cpp index 7df659e7b3..2ca453ebde 100644 --- a/src/graph/backends/GLES/GCFunctionsFactory.cpp +++ b/src/graph/backends/GLES/GCFunctionsFactory.cpp @@ -94,7 +94,8 @@ std::unique_ptr create_concatenate_layerconfigure(inputs, output); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() << " Target " << GCTargetInfo::TargetType << " Data Type: " << output->info()->data_type() << " Shape: " << output->info()->tensor_shape() @@ -143,7 +144,9 @@ std::unique_ptr create_convolution_layerinfo()->data_type() << " Input QuantInfo: " << input->info()->quantization_info() << " Weights QuantInfo: " << weights->info()->quantization_info() @@ -191,7 +194,9 @@ std::unique_ptr create_depthwise_convolution_layerinfo()->data_type() << " Input QuantInfo: " << input->info()->quantization_info() @@ -246,11 +251,13 @@ std::unique_ptr create_eltwise_layerinfo()->data_type() - << " Shape : " << input1->info()->tensor_shape() + << " Shape: " << input1->info()->tensor_shape() << std::endl); return func; diff --git a/src/graph/backends/GLES/GCNodeValidator.cpp b/src/graph/backends/GLES/GCNodeValidator.cpp index 9cf39c6675..fe69c7a9ee 100644 --- a/src/graph/backends/GLES/GCNodeValidator.cpp +++ b/src/graph/backends/GLES/GCNodeValidator.cpp @@ -121,6 +121,8 @@ Status GCNodeValidator::validate(INode *node) return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : PadLayer"); case NodeType::PermuteLayer: return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : PermuteLayer"); + case NodeType::PriorBoxLayer: + return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : PriorBoxLayer"); case NodeType::ReorgLayer: return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : ReorgLayer"); case NodeType::ReshapeLayer: diff --git a/src/graph/backends/NEON/NEFunctionFactory.cpp b/src/graph/backends/NEON/NEFunctionFactory.cpp index f03cead2b4..ca8d485f8b 100644 --- a/src/graph/backends/NEON/NEFunctionFactory.cpp +++ b/src/graph/backends/NEON/NEFunctionFactory.cpp @@ -163,8 +163,10 @@ std::unique_ptr create_normalization_layerconfigure(input, output, norm_info); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << NETargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << NETargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Input shape: " << input->info()->tensor_shape() << " Output shape: " << output->info()->tensor_shape() @@ -211,6 +213,8 @@ std::unique_ptr NEFunctionFactory::create(INode *node, GraphContext & return detail::create_permute_layer(*polymorphic_downcast(node)); case NodeType::PoolingLayer: return detail::create_pooling_layer(*polymorphic_downcast(node)); + case NodeType::PriorBoxLayer: + return detail::create_priorbox_layer(*polymorphic_downcast(node)); case NodeType::ReorgLayer: return detail::create_reorg_layer(*polymorphic_downcast(node)); case NodeType::ReshapeLayer: @@ -229,4 +233,4 @@ std::unique_ptr NEFunctionFactory::create(INode *node, GraphContext & } } // namespace backends } // namespace graph -} // namespace arm_compute \ No newline at end of file +} // namespace arm_compute diff --git a/src/graph/backends/NEON/NENodeValidator.cpp b/src/graph/backends/NEON/NENodeValidator.cpp index f2131586b2..a2abc8330c 100644 --- a/src/graph/backends/NEON/NENodeValidator.cpp +++ b/src/graph/backends/NEON/NENodeValidator.cpp @@ -67,6 +67,8 @@ Status NENodeValidator::validate(INode *node) return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : PadLayer"); case NodeType::PermuteLayer: return detail::validate_permute_layer(*polymorphic_downcast(node)); + case NodeType::PriorBoxLayer: + return detail::validate_priorbox_layer(*polymorphic_downcast(node)); case NodeType::ReorgLayer: return detail::validate_reorg_layer(*polymorphic_downcast(node)); case NodeType::ROIAlignLayer: diff --git a/src/graph/detail/ExecutionHelpers.cpp b/src/graph/detail/ExecutionHelpers.cpp index f479963280..f2c381b7df 100644 --- a/src/graph/detail/ExecutionHelpers.cpp +++ b/src/graph/detail/ExecutionHelpers.cpp @@ -254,7 +254,8 @@ bool call_all_output_node_accessors(ExecutionWorkload &workload) bool is_valid = true; std::for_each(std::begin(workload.outputs), std::end(workload.outputs), [&](Tensor * output_tensor) { - is_valid = is_valid && (output_tensor != nullptr) && output_tensor->call_accessor(); + bool valid_output = (output_tensor != nullptr) && output_tensor->call_accessor(); + is_valid = is_valid && valid_output; }); return is_valid; diff --git a/src/graph/nodes/FlattenLayerNode.cpp b/src/graph/nodes/FlattenLayerNode.cpp index 78b45dc305..baae555247 100644 --- a/src/graph/nodes/FlattenLayerNode.cpp +++ b/src/graph/nodes/FlattenLayerNode.cpp @@ -57,7 +57,7 @@ TensorDescriptor FlattenLayerNode::configure_output(size_t idx) const ARM_COMPUTE_ERROR_ON(src == nullptr); TensorDescriptor output_desc = src->desc(); - output_desc.shape.collapse(src->desc().shape.num_dimensions()); + output_desc.shape.collapse(3); return output_desc; } diff --git a/src/graph/nodes/PriorBoxLayerNode.cpp b/src/graph/nodes/PriorBoxLayerNode.cpp new file mode 100644 index 0000000000..edb1fba255 --- /dev/null +++ b/src/graph/nodes/PriorBoxLayerNode.cpp @@ -0,0 +1,95 @@ +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/graph/nodes/PriorBoxLayerNode.h" + +#include "arm_compute/core/Utils.h" +#include "arm_compute/graph/Graph.h" +#include "arm_compute/graph/INodeVisitor.h" +#include "arm_compute/graph/Utils.h" + +namespace arm_compute +{ +namespace graph +{ +PriorBoxLayerNode::PriorBoxLayerNode(PriorBoxLayerInfo prior_info) + : _info(std::move(prior_info)) +{ + _input_edges.resize(2, EmptyEdgeID); + _outputs.resize(1, NullTensorID); +} + +PriorBoxLayerInfo PriorBoxLayerNode::priorbox_info() const +{ + return _info; +} + +TensorDescriptor PriorBoxLayerNode::compute_output_descriptor(const TensorDescriptor &input_descriptor, + const PriorBoxLayerInfo &info) +{ + const unsigned int layer_width = get_dimension_size(input_descriptor, DataLayoutDimension::WIDTH); + const unsigned int layer_height = get_dimension_size(input_descriptor, DataLayoutDimension::HEIGHT); + const unsigned int num_priors = info.aspect_ratios().size() * info.min_sizes().size() + info.max_sizes().size(); + + TensorDescriptor output_descriptor = input_descriptor; + output_descriptor.shape.set(0, layer_width * layer_height * num_priors * 4); + output_descriptor.shape.set(1, 2); + output_descriptor.shape.set(2, 1); + + return output_descriptor; +} + +bool PriorBoxLayerNode::forward_descriptors() +{ + if((input_id(0) != NullTensorID) && (output_id(0) != NullTensorID)) + { + Tensor *dst = output(0); + ARM_COMPUTE_ERROR_ON(dst == nullptr); + dst->desc() = configure_output(0); + return true; + } + return false; +} + +TensorDescriptor PriorBoxLayerNode::configure_output(size_t idx) const +{ + ARM_COMPUTE_UNUSED(idx); + ARM_COMPUTE_ERROR_ON(idx >= _outputs.size()); + + const Tensor *input0 = input(0); + ARM_COMPUTE_ERROR_ON(input0 == nullptr); + + return compute_output_descriptor(input0->desc(), _info); +} + +NodeType PriorBoxLayerNode::type() const +{ + return NodeType::PriorBoxLayer; +} + +void PriorBoxLayerNode::accept(INodeVisitor &v) +{ + v.visit(*this); +} +} // namespace graph +} // namespace arm_compute diff --git a/utils/GraphUtils.cpp b/utils/GraphUtils.cpp index 19fba4c0bf..2f1df7aef2 100644 --- a/utils/GraphUtils.cpp +++ b/utils/GraphUtils.cpp @@ -77,8 +77,8 @@ void TFPreproccessor::preprocess(ITensor &tensor) }); } -CaffePreproccessor::CaffePreproccessor(std::array mean, bool bgr) - : _mean(mean), _bgr(bgr) +CaffePreproccessor::CaffePreproccessor(std::array mean, float scale, bool bgr) + : _mean(mean), _scale(scale), _bgr(bgr) { if(_bgr) { @@ -96,7 +96,7 @@ void CaffePreproccessor::preprocess(ITensor &tensor) execute_window_loop(window, [&](const Coordinates & id) { const float value = *reinterpret_cast(tensor.ptr_to_element(id)) - _mean[id[channel_idx]]; - *reinterpret_cast(tensor.ptr_to_element(id)) = value; + *reinterpret_cast(tensor.ptr_to_element(id)) = value * _scale; }); } diff --git a/utils/GraphUtils.h b/utils/GraphUtils.h index a6d670d761..d7f24afdd8 100644 --- a/utils/GraphUtils.h +++ b/utils/GraphUtils.h @@ -63,14 +63,16 @@ public: /** Default Constructor * * @param mean Mean array in RGB ordering + * @param scale Scale value * @param bgr Boolean specifying if the preprocessing should assume BGR format */ - CaffePreproccessor(std::array mean = std::array { { 0, 0, 0 } }, bool bgr = true); + CaffePreproccessor(std::array mean = std::array { { 0, 0, 0 } }, float scale = 1.f, bool bgr = true); void preprocess(ITensor &tensor) override; private: std::array _mean; - bool _bgr; + float _scale; + bool _bgr; }; /** TF preproccessor */ diff --git a/utils/Utils.h b/utils/Utils.h index 92ab1a30b9..8cac857178 100644 --- a/utils/Utils.h +++ b/utils/Utils.h @@ -607,7 +607,6 @@ void save_to_ppm(T &tensor, const std::string &ppm_filename) /** Template helper function to save a tensor image to a NPY file. * * @note Only F32 data type supported. - * @note Only works with 2D tensors. * @note If the input tensor is a CLTensor, the function maps and unmaps the image * * @param[in] tensor The tensor to save as NPY file @@ -627,9 +626,9 @@ void save_to_npy(T &tensor, const std::string &npy_filename, bool fortran_order) std::vector shape(tensor.info()->num_dimensions()); - for(unsigned int i = 0; i < tensor.info()->num_dimensions(); ++i) + for(unsigned int i = 0, j = tensor.info()->num_dimensions() - 1; i < tensor.info()->num_dimensions(); ++i, --j) { - shape[i] = tensor.info()->tensor_shape()[i]; + shape[i] = tensor.info()->tensor_shape()[!fortran_order ? j : i]; } // Map buffer if creating a CLTensor @@ -802,12 +801,13 @@ int compare_tensor(ITensor &tensor1, ITensor &tensor2) map(tensor1, true); map(tensor2, true); + Iterator itensor1(&tensor1, window); Iterator itensor2(&tensor2, window); execute_window_loop(window, [&](const Coordinates & id) { - if(std::abs(*reinterpret_cast(itensor1.ptr()) - *reinterpret_cast(itensor2.ptr())) > 0.00001) + if(std::abs(*reinterpret_cast(itensor1.ptr()) - *reinterpret_cast(itensor2.ptr())) > 0.0001) { ++num_mismatches; } -- cgit v1.2.1