Skip to content

P
Paddle

Crayon鑫 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 049c9c7d 编写于 作者: P phlrain
浏览文件
操作

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into add_dropout_att_new

上级 ffb24a73 6b4056bb
展开全部 隐藏空白更改
内联 并排
Showing with 6743 addition and 882 deletion
README.md
浏览文件 @ 049c9c7d
......@@ -2,8 +2,8 @@
[![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/index_cn.html)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.0/getstarted/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html)
[![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
[![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
......@@ -19,7 +19,7 @@ Our vision is to enable deep learning for everyone via PaddlePaddle.
Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle.
### Latest PaddlePaddle Release: [Fluid 1.0.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.0.0)
### Latest PaddlePaddle Release: [Fluid 1.0.1](https://github.com/PaddlePaddle/Paddle/tree/release/1.0.0)
### Install Latest Stable Release:
```
# Linux CPU
......@@ -27,9 +27,9 @@ pip install paddlepaddle
# Linux GPU cuda9cudnn7
pip install paddlepaddle-gpu
# Linux GPU cuda8cudnn7
pip install paddlepaddle-gpu==0.15.0.post87
pip install paddlepaddle-gpu==1.0.1.post87
# Linux GPU cuda8cudnn5
pip install paddlepaddle-gpu==0.15.0.post85
pip install paddlepaddle-gpu==1.0.1.post85
# For installation on other platform, refer to http://paddlepaddle.org/
```
......
cmake/generic.cmake
浏览文件 @ 049c9c7d
......@@ -261,6 +261,13 @@ function(cc_library TARGET_NAME)
add_dependencies(${TARGET_NAME} mklml)
target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed")
endif()
# remove link to python, see notes at:
# https://github.com/pybind/pybind11/blob/master/docs/compiling.rst#building-manually
if("${cc_library_DEPS};" MATCHES "python;")
list(REMOVE_ITEM cc_library_DEPS python)
add_dependencies(${TARGET_NAME} python)
target_link_libraries(${TARGET_NAME} "-Wl,-undefined,dynamic_lookup")
endif()
target_link_libraries(${TARGET_NAME} ${cc_library_DEPS})
add_dependencies(${TARGET_NAME} ${cc_library_DEPS})
endif()
......
paddle/fluid/API.spec
浏览文件 @ 049c9c7d
......@@ -116,6 +116,7 @@ paddle.fluid.layers.pad ArgSpec(args=['x', 'paddings', 'pad_value', 'name'], var
paddle.fluid.layers.pad_constant_like ArgSpec(args=['x', 'y', 'pad_value', 'name'], varargs=None, keywords=None, defaults=(0.0, None))
paddle.fluid.layers.label_smooth ArgSpec(args=['label', 'prior_dist', 'epsilon', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, 0.1, 'float32', None))
paddle.fluid.layers.roi_pool ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1, 1, 1.0))
paddle.fluid.layers.roi_align ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale', 'sampling_ratio', 'name'], varargs=None, keywords=None, defaults=(1, 1, 1.0, -1, None))
paddle.fluid.layers.dice_loss ArgSpec(args=['input', 'label', 'epsilon'], varargs=None, keywords=None, defaults=(1e-05,))
paddle.fluid.layers.image_resize ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR'))
paddle.fluid.layers.image_resize_short ArgSpec(args=['input', 'out_short_len', 'resample'], varargs=None, keywords=None, defaults=('BILINEAR',))
......
paddle/fluid/framework/details/var_handle.h
浏览文件 @ 049c9c7d
......@@ -49,6 +49,8 @@ struct VarHandleBase {
void AddOutput(OpHandleBase* out, ir::Node* node) {
if (pending_ops_.find(out) == pending_ops_.end()) {
PADDLE_ENFORCE(out != nullptr, "The output of %s should not be nullptr",
this->Node()->Name());
pending_ops_.insert(out);
node_->outputs.push_back(node);
}
......
paddle/fluid/framework/ir/CMakeLists.txt
浏览文件 @ 049c9c7d
......@@ -10,7 +10,7 @@ function(pass_library TARGET DEST)
set(oneValueArgs "")
set(multiValueArgs SRCS DEPS)
cmake_parse_arguments(op_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
cc_library(${TARGET} SRCS ${TARGET}.cc DEPS graph_pattern_detector pass ${op_library_DEPS})
cc_library(${TARGET} SRCS ${TARGET}.cc DEPS graph_pattern_detector pass fuse_pass_base ${op_library_DEPS})
# add more DEST here, such as train, dist and collect USE_PASS into a file automatically.
if (${DEST} STREQUAL "base" OR ${DEST} STREQUAL "inference")
message(STATUS "add pass ${TARGET} ${DEST}")
......@@ -25,13 +25,11 @@ cc_library(graph_helper SRCS graph_helper.cc DEPS graph)
cc_library(pass SRCS pass.cc DEPS graph node graph_helper)
cc_library(graph_traits SRCS graph_traits.cc DEPS graph)
cc_library(graph_pattern_detector SRCS graph_pattern_detector.cc DEPS graph graph_helper graph_traits)
cc_library(fuse_pass_base SRCS fuse_pass_base.cc DEPS pass)
pass_library(graph_to_program_pass base)
pass_library(graph_viz_pass base)
pass_library(fc_fuse_pass inference)
if (WITH_MKLDNN)
pass_library(conv_relu_mkldnn_fuse_pass inference)
endif ()
pass_library(attention_lstm_fuse_pass inference)
pass_library(infer_clean_graph_pass inference)
pass_library(fc_lstm_fuse_pass inference)
......@@ -39,6 +37,13 @@ pass_library(embedding_fc_lstm_fuse_pass inference)
pass_library(fc_gru_fuse_pass inference)
pass_library(seq_concat_fc_fuse_pass inference)
pass_library(conv_bn_fuse_pass inference)
pass_library(seqconv_eltadd_relu_fuse_pass inference)
if(WITH_MKLDNN)
pass_library(mkldnn_placement_pass base)
pass_library(conv_bias_mkldnn_fuse_pass inference)
pass_library(conv_relu_mkldnn_fuse_pass inference)
pass_library(conv_elementwise_add_mkldnn_fuse_pass inference)
endif()
cc_library(fuse_elewise_add_act_pass SRCS fuse_elewise_add_act_pass.cc DEPS pass graph_pattern_detector )
......@@ -54,4 +59,5 @@ cc_test(test_graph_pattern_detector SRCS graph_pattern_detector_tester.cc DEPS g
cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framework_proto)
if (WITH_MKLDNN)
cc_test(test_conv_relu_mkldnn_fuse_pass SRCS conv_relu_mkldnn_fuse_pass_tester.cc DEPS conv_relu_mkldnn_fuse_pass)
cc_test(test_conv_elementwise_add_mkldnn_fuse_pass SRCS conv_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS conv_elementwise_add_mkldnn_fuse_pass)
endif ()
paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc 0 → 100644
浏览文件 @ 049c9c7d
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h"
#include <functional>
#include <string>
#include <vector>
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/platform/enforce.h"
namespace paddle {
namespace framework {
namespace ir {
template <typename BinaryOperation>
LoDTensor tensor_apply_eltwise(const LoDTensor& vec_a, const LoDTensor& vec_b,
BinaryOperation f) {
PADDLE_ENFORCE_EQ(vec_a.dims(), vec_b.dims());
LoDTensor vec_y;
vec_y.Resize(vec_a.dims());
const float* a = vec_a.data<float>();
const float* b = vec_b.data<float>();
float* y = vec_y.mutable_data<float>(platform::CPUPlace());
for (int i = 0; i < vec_a.numel(); i++) {
y[i] = f(a[i], b[i]);
}
return vec_y;
}
std::unique_ptr<ir::Graph> ConvBiasFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
PADDLE_ENFORCE(graph.get());
FusePassBase::Init(name_scope_, graph.get());
auto* scope = param_scope();
PADDLE_ENFORCE(scope);
GraphPatternDetector gpd;
auto* conv_input =
gpd.mutable_pattern()
->NewNode(patterns::PDNodeName(name_scope_, "conv_input"))
->AsInput()
->assert_is_op_input("conv2d", "Input");
patterns::ConvBias conv_bias_pattern(gpd.mutable_pattern(), name_scope_);
conv_bias_pattern(conv_input);
int found_conv_bias_count = 0;
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
VLOG(4) << "handle ConvBias fuse";
GET_IR_NODE_FROM_SUBGRAPH(conv_weight, conv_weight,
conv_bias_pattern); // Filter
GET_IR_NODE_FROM_SUBGRAPH(conv_out, conv_out, conv_bias_pattern); // tmp
GET_IR_NODE_FROM_SUBGRAPH(conv, conv, conv_bias_pattern); // CONV op
// bias
GET_IR_NODE_FROM_SUBGRAPH(eltwise_bias, eltwise_bias, conv_bias_pattern);
// output
GET_IR_NODE_FROM_SUBGRAPH(eltwise_out, eltwise_out, conv_bias_pattern);
// elementwise_add op
GET_IR_NODE_FROM_SUBGRAPH(eltwise, eltwise, conv_bias_pattern);
PADDLE_ENFORCE(subgraph.count(conv_input));
// check if fuse can be done and if MKL-DNN should be used
FuseOptions fuse_option = FindFuseOption(*conv, *eltwise);
if (fuse_option == DO_NOT_FUSE || fuse_option == FUSE_NATIVE) {
VLOG(3) << "do not perform conv+bias fuse";
return;
}
auto* eltwise_bias_tensor =
scope->FindVar(eltwise_bias->Name())->GetMutable<LoDTensor>();
auto input_names = conv->Op()->InputNames();
bool has_bias = std::find(input_names.begin(), input_names.end(), "Bias") !=
input_names.end();
if (has_bias && conv->Op()->Input("Bias").size() > 0) {
auto conv_bias_names = conv->Op()->Input("Bias");
// add eltwise bias to existing conv bias
PADDLE_ENFORCE_EQ(conv_bias_names.size(), 1);
auto* conv_bias_var = scope->FindVar(conv_bias_names[0]);
auto* conv_bias_tensor = conv_bias_var->GetMutable<LoDTensor>();
PADDLE_ENFORCE_EQ(conv_bias_tensor->dims(), eltwise_bias_tensor->dims());
*conv_bias_tensor = tensor_apply_eltwise(
*conv_bias_tensor, *eltwise_bias_tensor, std::plus<float>());
conv->Op()->SetOutput("Output",
std::vector<std::string>({eltwise_out->Name()}));
GraphSafeRemoveNodes(graph.get(), {eltwise, conv_out});
IR_NODE_LINK_TO(conv, eltwise_out);
} else {
// take eltwise bias as conv bias
OpDesc desc;
desc.SetInput(
"Input", std::vector<std::string>({subgraph.at(conv_input)->Name()}));
desc.SetInput("Filter", std::vector<std::string>({conv_weight->Name()}));
desc.SetInput("Bias", std::vector<std::string>({eltwise_bias->Name()}));
desc.SetOutput("Output", std::vector<std::string>({eltwise_out->Name()}));
desc.SetType("conv2d");
for (auto& attr : conv->Op()->GetAttrMap()) {
desc.SetAttr(attr.first, attr.second);
}
auto conv_bias_node = g->CreateOpNode(&desc);
IR_NODE_LINK_TO(subgraph.at(conv_input), conv_bias_node);
IR_NODE_LINK_TO(conv_weight, conv_bias_node);
IR_NODE_LINK_TO(eltwise_bias, conv_bias_node);
IR_NODE_LINK_TO(conv_bias_node, eltwise_out);
GraphSafeRemoveNodes(graph.get(), {conv, eltwise, conv_out});
}
found_conv_bias_count++;
};
gpd(graph.get(), handler);
AddStatis(found_conv_bias_count);
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(conv_bias_mkldnn_fuse_pass,
paddle::framework::ir::ConvBiasFusePass);
paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h 0 → 100644
浏览文件 @ 049c9c7d
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <string>
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace ir {
/*
* Fuse the Conv and Elementwise_add to a ConvBiasOp.
*/
class ConvBiasFusePass : public FusePassBase {
public:
virtual ~ConvBiasFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
const std::string name_scope_{"conv_bias_mkldnn_fuse"};
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/conv_bn_fuse_pass.cc
浏览文件 @ 049c9c7d
......@@ -126,12 +126,21 @@ std::unique_ptr<ir::Graph> ConvBNFusePass::ApplyImpl(
// conv, batch_norm,
// conv_weight, conv_out,
// bn_scale, bn_bias, bn_mean, bn_variance,
// bn_out, bn_mean_out, bn_variance_out, bn_saved_mean, bn_saved_variance
// bn_out, bn_mean_out, bn_variance_out, bn_saved_mean,
// bn_saved_variance
GET_CONV_BN_NODES(conv_bn_pattern);
// check if fuse can be done and if MKL-DNN should be used
FuseOptions fuse_option = FindFuseOption(*conv, *batch_norm);
if (fuse_option == DO_NOT_FUSE) {
VLOG(3) << "do not perform conv+bn fuse";
return;
}
// Create eltwise_y (conv bias) variable
VarDesc eltwise_y_in_desc(
patterns::PDNodeName(name_scope_, "eltwise_y_in"));
eltwise_y_in_desc.SetPersistable(true);
auto* eltwise_y_in_node = g->CreateVarNode(&eltwise_y_in_desc);
auto* eltwise_y_in_tensor =
scope->Var(eltwise_y_in_node->Name())->GetMutable<LoDTensor>();
......@@ -151,27 +160,59 @@ std::unique_ptr<ir::Graph> ConvBNFusePass::ApplyImpl(
*bn_mean, *bn_variance, eltwise_y_in_tensor,
epsilon);
// Create an elementwise add node
OpDesc desc;
desc.SetInput("X", std::vector<std::string>({conv_out->Name()}));
desc.SetInput("Y", std::vector<std::string>({eltwise_y_in_node->Name()}));
desc.SetOutput("Out", std::vector<std::string>({bn_out->Name()}));
desc.SetType("elementwise_add");
desc.SetAttr("axis", 1);
bool a = boost::get<bool>(conv->Op()->GetAttr("use_mkldnn"));
desc.SetAttr("use_mkldnn", a);
auto eltwise_op = g->CreateOpNode(&desc); // OpDesc will be copied.
GraphSafeRemoveNodes(graph.get(), {bn_scale, bn_bias, bn_mean, bn_variance,
batch_norm, bn_mean_out, bn_variance_out,
bn_saved_mean, bn_saved_variance});
PADDLE_ENFORCE(subgraph.count(conv_input));
IR_NODE_LINK_TO(conv_out, eltwise_op);
IR_NODE_LINK_TO(eltwise_y_in_node, eltwise_op);
IR_NODE_LINK_TO(eltwise_op, bn_out);
found_conv_bn_count++;
// with MKL-DNN fuse conv+bn into conv with bias
// without MKL-DNN fuse conv+bn into conv+elementwise_add
if (fuse_option == FUSE_MKLDNN) {
auto input_names = conv->Op()->InputNames();
bool has_bias = std::find(input_names.begin(), input_names.end(),
"Bias") != input_names.end();
if (has_bias && conv->Op()->Input("Bias").size() > 0) {
// reuse existing conv bias node
auto conv_bias_names = conv->Op()->Input("Bias");
PADDLE_ENFORCE_EQ(conv_bias_names.size(), 1);
auto* conv_bias_var = scope->FindVar(conv_bias_names[0]);
auto* conv_bias_tensor = conv_bias_var->GetMutable<LoDTensor>();
PADDLE_ENFORCE_EQ(conv_bias_tensor->dims(),
eltwise_y_in_tensor->dims());
auto eigen_conv_bias = EigenVector<float>::From(*conv_bias_tensor);
eigen_conv_bias += EigenVector<float>::From(*eltwise_y_in_tensor);
} else {
// add new conv_bias node
conv->Op()->SetInput(
"Bias", std::vector<std::string>({eltwise_y_in_node->Name()}));
IR_NODE_LINK_TO(eltwise_y_in_node, conv);
}
conv->Op()->SetOutput("Output",
std::vector<std::string>({bn_out->Name()}));
GraphSafeRemoveNodes(
graph.get(),
{conv_out, bn_scale, bn_bias, bn_mean, bn_variance, batch_norm,
bn_mean_out, bn_variance_out, bn_saved_mean, bn_saved_variance});
IR_NODE_LINK_TO(conv, bn_out);
found_conv_bn_count++;
} else { // fuse_option == FUSE_NATIVE
// create an elementwise add node.
OpDesc desc;
desc.SetInput("X", std::vector<std::string>({conv_out->Name()}));
desc.SetInput("Y", std::vector<std::string>({eltwise_y_in_node->Name()}));
desc.SetOutput("Out", std::vector<std::string>({bn_out->Name()}));
desc.SetType("elementwise_add");
desc.SetAttr("axis", 1);
auto eltwise_op = g->CreateOpNode(&desc); // OpDesc will be copied.
GraphSafeRemoveNodes(
graph.get(),
{bn_scale, bn_bias, bn_mean, bn_variance, batch_norm, bn_mean_out,
bn_variance_out, bn_saved_mean, bn_saved_variance});
IR_NODE_LINK_TO(conv_out, eltwise_op);
IR_NODE_LINK_TO(eltwise_y_in_node, eltwise_op);
IR_NODE_LINK_TO(eltwise_op, bn_out);
found_conv_bn_count++;
}
};
gpd(graph.get(), handler);
......@@ -237,7 +278,6 @@ std::unique_ptr<ir::Graph> ConvEltwiseAddBNFusePass::ApplyImpl(
{bn_scale, bn_bias, bn_mean, bn_variance, batch_norm, bn_mean_out,
bn_variance_out, bn_saved_mean, bn_saved_variance, eltwise_out});
PADDLE_ENFORCE(subgraph.count(conv_input));
IR_NODE_LINK_TO(eltwise, bn_out);
found_conv_bn_count++;
......
paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc 0 → 100644
浏览文件 @ 049c9c7d
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h"
#include <functional>
#include <utility>
#include "paddle/fluid/framework/ir/graph_traits.h"
namespace paddle {
namespace framework {
namespace ir {
namespace {
// The function keeps the graph consistent by replacing
// a node 'from' in the set of inputs nodes
// of the visited node by a node 'to'.
void CorrectGraphEdges(Graph* graph, Node* from, Node* to) {
for (auto& node : GraphTraits::DFS(*graph)) {
auto from_in_inputs =
std::find(std::begin(node.inputs), std::end(node.inputs), from);
if (from_in_inputs != std::end(node.inputs)) {
IR_NODE_LINK_TO(to, (&node));
auto inputs = node.Op()->Inputs();
using input_type = VariableNameMap::value_type;
std::for_each(std::begin(inputs), std::end(inputs),
[from, to, &node](const input_type& i) -> void {
auto param_names = i.second;
auto pi = std::find(std::begin(param_names),
std::end(param_names), from->Name());
if (pi != std::end(param_names)) {
node.Op()->SetInput(i.first, {to->Name()});
}
});
}
}
}
} // namespace
using graph_ptr = std::unique_ptr<ir::Graph>;
graph_ptr ConvElementwiseAddMKLDNNFusePass::ApplyImpl(graph_ptr graph) const {
FusePassBase::Init(name_scope_, graph.get());
GraphPatternDetector gpd;
auto pattern = gpd.mutable_pattern();
patterns::Conv conv_pattern{pattern, name_scope_};
auto conv_output = conv_pattern();
patterns::ElementwiseAdd elementwise_add_pattern{pattern, name_scope_};
elementwise_add_pattern(conv_output);
conv_output->AsIntermediate();
auto conv_op_has_bias = [](const Node& conv_op) -> std::pair<bool, Node*> {
auto bias_input_names = conv_op.Op()->Inputs();
auto bias_it = bias_input_names.find("Bias");
if (bias_it != std::end(bias_input_names)) {
bool has_bias = !bias_it->second.empty();
if (has_bias) {
auto conv_bias_names = bias_it->second;
auto conv_bias_names_it =
std::find_if(std::begin(conv_op.inputs), std::end(conv_op.inputs),
[&conv_bias_names](Node* n) -> bool {
return n->Name() == conv_bias_names[0];
});
return std::make_pair(has_bias, *conv_bias_names_it);
}
}
return std::make_pair(false, nullptr);
};
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
GET_IR_NODE_FROM_SUBGRAPH(conv_op, conv_op, conv_pattern);
GET_IR_NODE_FROM_SUBGRAPH(conv_input, conv_input, conv_pattern);
GET_IR_NODE_FROM_SUBGRAPH(conv_filter, conv_filter, conv_pattern);
GET_IR_NODE_FROM_SUBGRAPH(conv_output, conv_output, conv_pattern);
GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_op, elementwise_add_op,
elementwise_add_pattern);
GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_x, elementwise_add_x,
elementwise_add_pattern);
GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_out, elementwise_add_out,
elementwise_add_pattern);
if (FindFuseOption(*conv_op, *elementwise_add_op) != FUSE_MKLDNN) return;
OpDesc op_desc;
op_desc.SetType("conv2d");
op_desc.SetInput("Input", {conv_input->Name()});
op_desc.SetInput("Filter", {conv_filter->Name()});
op_desc.SetInput("ResidualData", {elementwise_add_x->Name()});
op_desc.SetOutput("Output", {conv_output->Name()});
bool has_bias;
Node* conv_bias;
std::tie(has_bias, conv_bias) = conv_op_has_bias(*conv_op);
if (has_bias) {
op_desc.SetInput("Bias", {conv_bias->Name()});
}
for (const auto& attr : conv_op->Op()->GetAttrMap()) {
op_desc.SetAttr(attr.first, attr.second);
}
op_desc.SetAttr("fuse_residual_connection", true);
auto fused_conv_op = g->CreateOpNode(&op_desc);
IR_NODE_LINK_TO(conv_input, fused_conv_op);
IR_NODE_LINK_TO(conv_filter, fused_conv_op);
IR_NODE_LINK_TO(elementwise_add_x, fused_conv_op);
IR_NODE_LINK_TO(fused_conv_op, conv_output);
if (has_bias) {
IR_NODE_LINK_TO(conv_bias, fused_conv_op);
}
CorrectGraphEdges(g, elementwise_add_out, conv_output);
GraphSafeRemoveNodes(g, {elementwise_add_out, conv_op, elementwise_add_op});
};
gpd(graph.get(), handler);
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(conv_elementwise_add_mkldnn_fuse_pass,
paddle::framework::ir::ConvElementwiseAddMKLDNNFusePass);
paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h 0 → 100644
浏览文件 @ 049c9c7d
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <string>
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
namespace paddle {
namespace framework {
namespace ir {
class ConvElementwiseAddMKLDNNFusePass : public FusePassBase {
public:
virtual ~ConvElementwiseAddMKLDNNFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
const std::string name_scope_{"residual_connections_fuse_pass"};
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass_tester.cc 0 → 100644
浏览文件 @ 049c9c7d
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include <gtest/gtest.h>
#include <string>
#include "paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h"
#include "paddle/fluid/framework/ir/graph_traits.h"
namespace paddle {
namespace framework {
namespace ir {
namespace {
constexpr int nodes_removed = 3;
constexpr int nodes_added = 1;
void SetOp(ProgramDesc* prog, const std::string& type,
const std::vector<std::pair<std::string, std::string>>& inputs,
const std::pair<std::string, std::string>& output) {
auto op = prog->MutableBlock(0)->AppendOp();
op->SetType(type);
op->SetAttr("use_mkldnn", true);
for (const auto& input : inputs) {
op->SetInput(input.first, {input.second});
}
op->SetOutput(output.first, {output.second});
}
struct IsReachable {
using func = std::function<bool(const std::string&, const std::string&)>;
auto operator()(const std::unique_ptr<ir::Graph>& graph) -> func {
auto find_node = [](const std::unique_ptr<ir::Graph>& graph,
const std::string& name) -> Node* {
for (auto& node : GraphTraits::DFS(*graph)) {
if (name == node.Name()) {
return &node;
}
}
return nullptr;
};
return [&](std::string from, const std::string to) -> bool {
if (from == to) return true;
std::map<std::string, bool> visited;
for (auto& node : GraphTraits::DFS(*graph)) {
visited[node.Name()] = false;
}
visited[from] = true;
std::list<std::string> queue;
queue.push_back(from);
while (!queue.empty()) {
auto cur = find_node(graph, queue.front());
queue.pop_front();
if (cur == nullptr) return false;
for (auto n : cur->outputs) {
if (n->Name() == to) return true;
if (!visited[n->Name()]) {
visited[n->Name()] = true;
queue.push_back(n->Name());
}
}
}
return false;
};
}
};
void AssertOpsCount(const std::unique_ptr<ir::Graph>& graph) {
int conv_count = 0;
int elementwise_add_count = 0;
for (auto* node : graph->Nodes()) {
if (node->IsOp() && node->Op()->Type() == "conv2d") {
++conv_count;
}
if (node->IsOp() && node->Op()->Type() == "elementwise_add") {
++elementwise_add_count;
}
}
EXPECT_EQ(conv_count, 1);
EXPECT_EQ(elementwise_add_count, 0);
}
ProgramDesc BuildProgramDesc(const std::vector<std::string>& transient_vars,
const std::vector<std::string>& persistent_vars) {
ProgramDesc prog;
auto add_var_to_prog = [&prog](const std::string& var_name) -> VarDesc* {
auto var = prog.MutableBlock(0)->Var(var_name);
var->SetType(proto::VarType::LOD_TENSOR);
return var;
};
for (const auto& v : transient_vars) {
add_var_to_prog(v);
}
for (const auto& v : persistent_vars) {
auto var = add_var_to_prog(v);
var->SetPersistable(true);
}
return prog;
}
} // namespace
TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionWithElementwiseAddRelu) {
auto prog =
BuildProgramDesc({"a", "b", "c", "d", "e", "f"}, {"bias", "weights"});
SetOp(&prog, "conv2d",
{{"Input", "a"}, {"Bias", "bias"}, {"Filter", "weights"}},
{"Output", "b"});
SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"});
SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
IsReachable is_reachable;
EXPECT_TRUE(is_reachable(graph)("a", "relu"));
auto pass =
PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
int original_nodes_num = graph->Nodes().size();
graph = pass->Apply(std::move(graph));
int current_nodes_num = graph->Nodes().size();
EXPECT_TRUE(is_reachable(graph)("a", "relu"));
EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
current_nodes_num);
AssertOpsCount(graph);
}
TEST(ConvElementwiseAddMKLDNNFusePass,
ConvolutionWithElementwiseAddReluNoBias) {
auto prog = BuildProgramDesc({"a", "b", "c", "d", "e"}, {"weights"});
SetOp(&prog, "conv2d", {{"Input", "a"}, {"Filter", "weights"}},
{"Output", "b"});
SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"});
SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"});
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
IsReachable is_reachable;
EXPECT_TRUE(is_reachable(graph)("a", "relu"));
auto pass =
PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
int original_nodes_num = graph->Nodes().size();
graph = pass->Apply(std::move(graph));
int current_nodes_num = graph->Nodes().size();
EXPECT_TRUE(is_reachable(graph)("a", "relu"));
EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
current_nodes_num);
AssertOpsCount(graph);
}
TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionElementwiseAdd) {
auto prog = BuildProgramDesc({"a", "b", "c", "d"}, {"bias", "weights"});
SetOp(&prog, "conv2d",
{{"Input", "a"}, {"Bias", "bias"}, {"Filter", "weights"}},
{"Output", "b"});
SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"});
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
IsReachable is_reachable;
EXPECT_TRUE(is_reachable(graph)("a", "d"));
auto pass =
PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
int original_nodes_num = graph->Nodes().size();
graph = pass->Apply(std::move(graph));
int current_nodes_num = graph->Nodes().size();
EXPECT_FALSE(is_reachable(graph)("a", "d"));
EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
current_nodes_num);
AssertOpsCount(graph);
}
TEST(ConvElementwiseAddMKLDNNFusePass, SigmoidConvolutionAddElementwiseRelu) {
auto prog =
BuildProgramDesc({"a", "b", "c", "d", "e", "f"}, {"bias", "weights"});
SetOp(&prog, "sigmoid", {{"X", "a"}}, {"Out", "b"});
SetOp(&prog, "conv2d",
{{"Input", "b"}, {"Bias", "bias"}, {"Filter", "weights"}},
{"Output", "c"});
SetOp(&prog, "elementwise_add", {{"X", "c"}, {"Y", "d"}}, {"Out", "e"});
SetOp(&prog, "relu", {{"X", "e"}}, {"Out", "f"});
std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
IsReachable is_reachable;
EXPECT_TRUE(is_reachable(graph)("a", "f"));
auto pass =
PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass");
int original_nodes_num = graph->Nodes().size();
graph = pass->Apply(std::move(graph));
int current_nodes_num = graph->Nodes().size();
EXPECT_TRUE(is_reachable(graph)("a", "f"));
EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added,
current_nodes_num);
AssertOpsCount(graph);
}
} // namespace ir
} // namespace framework
} // namespace paddle
USE_PASS(conv_elementwise_add_mkldnn_fuse_pass);
paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc
浏览文件 @ 049c9c7d
......@@ -46,6 +46,12 @@ std::unique_ptr<ir::Graph> ConvReLUFusePass::ApplyImpl(
GET_IR_NODE_FROM_SUBGRAPH(relu_out, relu_out, conv_relu_pattern); // Out
GET_IR_NODE_FROM_SUBGRAPH(relu, relu, conv_relu_pattern); // ReLU op
FuseOptions fuse_option = FindFuseOption(*conv, *relu);
if (fuse_option == DO_NOT_FUSE) {
VLOG(3) << "do not perform conv+relu fuse";
return;
}
// Transform Conv node into ConvReLU node.
OpDesc* desc = conv->Op();
desc->SetOutput("Output", std::vector<std::string>({relu_out->Name()}));
......
paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc
浏览文件 @ 049c9c7d
......@@ -20,17 +20,19 @@ namespace paddle {
namespace framework {
namespace ir {
void SetOp(ProgramDesc* prog, const std::string& type,
void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
const std::vector<std::string>& inputs,
const std::vector<std::string>& outputs) {
const std::vector<std::string>& outputs, bool use_mkldnn = false) {
auto* op = prog->MutableBlock(0)->AppendOp();
op->SetType(type);
if (type == "conv2d") {
op->SetAttr("use_mkldnn", true);
op->SetAttr("use_mkldnn", use_mkldnn);
op->SetAttr("name", name);
op->SetInput("Input", {inputs[0]});
op->SetInput("Filter", {inputs[1]});
op->SetInput("Bias", {inputs[2]});
} else if (type == "relu") {
op->SetAttr("use_mkldnn", use_mkldnn);
op->SetInput("X", inputs);
}
op->SetOutput("Out", outputs);
......@@ -43,7 +45,8 @@ void SetOp(ProgramDesc* prog, const std::string& type,
ProgramDesc BuildProgramDesc() {
ProgramDesc prog;
for (auto& v :
std::vector<std::string>({"a", "b", "c", "weights", "bias", "f", "g"})) {
std::vector<std::string>({"a", "b", "c", "weights", "bias", "f", "g",
"h", "weights2", "bias2", "k", "l"})) {
auto* var = prog.MutableBlock(0)->Var(v);
var->SetType(proto::VarType::SELECTED_ROWS);
if (v == "weights" || v == "bias") {
......@@ -51,14 +54,24 @@ ProgramDesc BuildProgramDesc() {
}
}
SetOp(&prog, "OP0", std::vector<std::string>({"a"}),
SetOp(&prog, "OP0", "op0", std::vector<std::string>({"a"}),
std::vector<std::string>({"b"}));
SetOp(&prog, "OP1", std::vector<std::string>({"b"}),
SetOp(&prog, "OP1", "op1", std::vector<std::string>({"b"}),
std::vector<std::string>({"c"}));
SetOp(&prog, "conv2d", std::vector<std::string>({"c", "weights", "bias"}),
std::vector<std::string>({"f"}));
SetOp(&prog, "relu", std::vector<std::string>({"f"}),
std::vector<std::string>({"g"}));
// conv+relu, both with MKL-DNN
SetOp(&prog, "conv2d", "conv1",
std::vector<std::string>({"c", "weights", "bias"}),
std::vector<std::string>({"f"}), true);
SetOp(&prog, "relu", "relu1", std::vector<std::string>({"f"}),
std::vector<std::string>({"g"}), true);
SetOp(&prog, "OP3", "op3", std::vector<std::string>({"g"}),
std::vector<std::string>({"h"}));
// conv+relu, only one with MKL-DNN
SetOp(&prog, "conv2d", "conv2",
std::vector<std::string>({"h", "weights2", "bias2"}),
std::vector<std::string>({"k"}), true);
SetOp(&prog, "relu", "relu2", std::vector<std::string>({"k"}),
std::vector<std::string>({"l"}));
return prog;
}
......@@ -88,10 +101,16 @@ TEST(ConvReLUFusePass, basic) {
auto* op = node->Op();
ASSERT_TRUE(op->HasAttr("use_mkldnn"));
EXPECT_TRUE(boost::get<bool>(op->GetAttr("use_mkldnn")));
ASSERT_TRUE(op->HasAttr("fuse_relu"));
bool fuse_relu = boost::get<bool>(op->GetAttr("fuse_relu"));
if (fuse_relu) {
++conv_relu_count;
// check if only "conv1" convolution is fused
auto op_name = boost::get<std::string>(op->GetAttr("name"));
if (op_name == "conv1") {
ASSERT_TRUE(op->HasAttr("fuse_relu"));
bool fuse_relu = boost::get<bool>(op->GetAttr("fuse_relu"));
if (fuse_relu) {
++conv_relu_count;
}
} else if (op_name == "conv2") {
ASSERT_FALSE(op->HasAttr("fuse_relu"));
}
}
}
......
paddle/fluid/framework/ir/fuse_pass_base.cc 0 → 100644
浏览文件 @ 049c9c7d
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
namespace paddle {
namespace framework {
namespace ir {
void FusePassBase::Init(const std::string& repr, Graph* graph) const {
repr_ = repr;
graph_ = graph;
}
Scope* FusePassBase::param_scope() const {
PADDLE_ENFORCE(graph_->Has(kParamScopeAttr));
return graph_->Get<framework::Scope*>(kParamScopeAttr);
}
void FusePassBase::AddStatis(int count_of_fused) const {
PADDLE_ENFORCE(graph_);
PADDLE_ENFORCE(!repr_.empty());
if (!graph_->Has(kFuseStatisAttr)) {
graph_->Set(kFuseStatisAttr, new std::unordered_map<std::string, int>);
}
auto& info =
graph_->Get<std::unordered_map<std::string, int>>(kFuseStatisAttr);
info[repr_] = count_of_fused;
}
FuseOptions FusePassBase::FindFuseOption(const Node& node1,
const Node& node2) const {
#ifdef PADDLE_WITH_MKLDNN
bool node1_mkldnn = node1.Op()->HasAttr("use_mkldnn") &&
boost::get<bool>(node1.Op()->GetAttr("use_mkldnn"));
bool node2_mkldnn = node2.Op()->HasAttr("use_mkldnn") &&
boost::get<bool>(node2.Op()->GetAttr("use_mkldnn"));
if (node1_mkldnn && node2_mkldnn)
return FUSE_MKLDNN;
else if (!node1_mkldnn && !node2_mkldnn)
return FUSE_NATIVE;
else
return DO_NOT_FUSE;
#else
return FUSE_NATIVE;
#endif
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/fuse_pass_base.h
浏览文件 @ 049c9c7d
......@@ -25,32 +25,24 @@ namespace ir {
static const char kParamScopeAttr[] = "__param_scope__";
static const char kFuseStatisAttr[] = "__fuse_statis__";
enum FuseOptions {
DO_NOT_FUSE, // fusing will not be done
FUSE_NATIVE, // fusing will be done without MKL-DNN
FUSE_MKLDNN // fusing will be done with MKL-DNN
};
class FusePassBase : public Pass {
public:
void Init(const std::string& repr, Graph* graph) const {
repr_ = repr;
graph_ = graph;
}
Scope* param_scope() const {
PADDLE_ENFORCE(graph_->Has(kParamScopeAttr));
return graph_->Get<framework::Scope*>(kParamScopeAttr);
}
void AddStatis(int count_of_fused) const {
PADDLE_ENFORCE(graph_);
PADDLE_ENFORCE(!repr_.empty());
if (!graph_->Has(kFuseStatisAttr)) {
graph_->Set(kFuseStatisAttr, new std::unordered_map<std::string, int>);
}
auto& info =
graph_->Get<std::unordered_map<std::string, int>>(kFuseStatisAttr);
info[repr_] = count_of_fused;
}
void Init(const std::string& repr, Graph* graph) const;
Scope* param_scope() const;
void AddStatis(int count_of_fused) const;
virtual ~FusePassBase() {}
protected:
virtual FuseOptions FindFuseOption(const Node& node1,
const Node& node2) const;
mutable Graph* graph_;
mutable std::string repr_;
};
......
paddle/fluid/framework/ir/graph_helper_test.cc
浏览文件 @ 049c9c7d
......@@ -200,15 +200,15 @@ TEST(GraphHelperTest, GraphNum) {
Graph g(prog);
BuildZeroGraph(&g);
ASSERT_EQ(GraphNum(g), 0);
ASSERT_EQ(GraphNum(g), 0UL);
Graph g2(prog);
BuildOneGraph(&g2);
ASSERT_EQ(GraphNum(g2), 1);
ASSERT_EQ(GraphNum(g2), 1UL);
Graph g3(prog);
BuildTwoGraphs(&g3);
ASSERT_EQ(GraphNum(g3), 2);
ASSERT_EQ(GraphNum(g3), 2UL);
}
} // namespace ir
......
paddle/fluid/framework/ir/graph_pattern_detector.cc
浏览文件 @ 049c9c7d
......@@ -761,6 +761,51 @@ PDNode *patterns::ConvReLU::operator()(
return relu_out_var;
}
PDNode *patterns::SeqConvEltAddRelu::operator()(
paddle::framework::ir::PDNode *seqconv_input) {
// Create Operators
seqconv_input->assert_is_op_input("sequence_conv", "X");
auto *seqconv_op = pattern->NewNode(seqconv_repr())
->assert_is_op("sequence_conv")
->assert_op_attr<bool>("paddingTrainable", false)
->assert_op_attr<int>("contextStride", 1);
auto *eltadd_op =
pattern->NewNode(eltadd_repr())->assert_is_op("elementwise_add");
auto *relu_op = pattern->NewNode(relu_repr())->assert_is_op("relu");
// Create variables
// Filter
auto *seqconv_weight_var =
pattern->NewNode(seqconv_weight_repr())
->AsInput()
->assert_is_persistable_var()
->assert_is_op_input("sequence_conv", "Filter");
// Bias
auto *eltadd_bias_var = pattern->NewNode(eltadd_bias_repr())
->AsInput()
->assert_is_op_input("elementwise_add");
// intermediate variable, will be removed in the IR after fuse.
auto *seqconv_out_var = pattern->NewNode(seqconv_out_repr())
->AsIntermediate()
->assert_is_only_output_of_op("sequence_conv")
->assert_is_op_input("elementwise_add");
auto *eltadd_out_var = pattern->NewNode(eltadd_out_repr())
->AsIntermediate()
->assert_is_only_output_of_op("elementwise_add")
->assert_is_only_input_of_op("relu");
// output
auto *relu_out_var = pattern->NewNode(relu_out_repr())
->AsOutput()
->assert_is_op_output("relu");
seqconv_op->LinksFrom({seqconv_input, seqconv_weight_var})
.LinksTo({seqconv_out_var});
eltadd_op->LinksFrom({seqconv_out_var, eltadd_bias_var})
.LinksTo({eltadd_out_var});
relu_op->LinksFrom({eltadd_out_var}).LinksTo({relu_out_var});
return relu_out_var;
}
PDNode *patterns::FC::operator()(paddle::framework::ir::PDNode *x,
bool with_bias) {
// Create shared nodes.
......@@ -966,6 +1011,79 @@ PDNode *patterns::ElewiseAddActInplaceGrad::operator()(
return ele_add_grad;
}
PDNode *patterns::ConvBias::operator()(
paddle::framework::ir::PDNode *conv_input) {
// Create Operators
conv_input->assert_is_op_input("conv2d", "Input");
auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op("conv2d");
auto *eltiwse_op =
pattern->NewNode(eltwise_repr())->assert_is_op("elementwise_add");
// Create variables
// Filter
auto *conv_weight_var = pattern->NewNode(conv_weight_repr())
->AsInput()
->assert_is_persistable_var()
->assert_is_op_input("conv2d", "Filter");
// intermediate variable, will be removed in the IR after fuse.
auto *conv_out_var = pattern->NewNode(conv_out_repr())
->AsIntermediate()
->assert_is_only_output_of_op("conv2d")
->assert_is_op_input("elementwise_add");
// Bias stored in elementwise_add
auto *eltwise_bias_var = pattern->NewNode(eltwise_bias_repr())
->AsInput()
->assert_is_persistable_var()
->assert_is_op_input("elementwise_add", "Y");
// output
auto *eltwise_out_var = pattern->NewNode(eltwise_out_repr())
->AsOutput()
->assert_is_op_output("elementwise_add");
conv_op->LinksFrom({conv_input, conv_weight_var}).LinksTo({conv_out_var});
eltiwse_op->LinksFrom({conv_out_var, eltwise_bias_var})
.LinksTo({eltwise_out_var});
return eltwise_out_var;
}
PDNode *patterns::Conv::operator()() {
auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d");
auto input_var = pattern->NewNode(conv_input_repr())
->AsInput()
->assert_is_op_input("conv2d", "Input");
auto filter_var = pattern->NewNode(conv_filter_repr())
->AsInput()
->assert_is_op_input("conv2d", "Filter");
auto output_var = pattern->NewNode(conv_output_repr())
->AsOutput()
->assert_is_op_output("conv2d", "Output");
conv_op->LinksFrom({input_var, filter_var});
conv_op->LinksTo({output_var});
return output_var;
}
PDNode *patterns::ElementwiseAdd::operator()(PDNode *x_var) {
auto elementwise_add_op = pattern->NewNode(elementwise_add_op_repr())
->assert_is_op("elementwise_add");
x_var->assert_is_op_input("elementwise_add", "X");
auto y_var = pattern->NewNode(elementwise_add_x_repr())
->AsInput()
->assert_is_op_input("elementwise_add", "Y");
auto out_var = pattern->NewNode(elementwise_add_out_repr())
->AsOutput()
->assert_is_op_output("elementwise_add", "Out");
elementwise_add_op->LinksFrom({x_var, y_var});
elementwise_add_op->LinksTo({out_var});
return out_var;
}
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/graph_pattern_detector.h
浏览文件 @ 049c9c7d
......@@ -128,6 +128,15 @@ struct PDNode {
const std::unordered_set<std::string>& op_types,
const std::string& argument, int nth);
template <typename T>
PDNode* assert_op_attr(const std::string& attr_name, const T& attr) {
asserts_.emplace_back([=](Node* x) {
return x && x->IsOp() && x->Op()->HasAttr(attr_name) &&
boost::get<T>(x->Op()->GetAttr(attr_name)) == attr;
});
return this;
}
private:
PDNode(PDPattern* pattern, const std::string& name = "",
Type type = Type::kVar)
......@@ -434,6 +443,31 @@ struct ConvReLU : public PatternBase {
PATTERN_DECL_NODE(relu_out);
};
// SEQCONV with Elementwise_Add ReLU
// op: seqconv + elementwise_add + relu
// named nodes:
// seqconv_input, seqconv_weight,
// seqconv_out, seqconv,
// elementwise_add_bias, elementwise_add_out, elementwise_add
// relu_out, relu
struct SeqConvEltAddRelu : public PatternBase {
SeqConvEltAddRelu(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "seqconv_eltadd_relu") {}
PDNode* operator()(PDNode* seqconv_input);
// declare operator node's name
PATTERN_DECL_NODE(seqconv);
PATTERN_DECL_NODE(eltadd);
PATTERN_DECL_NODE(relu);
// declare variable node's name
PATTERN_DECL_NODE(seqconv_weight);
PATTERN_DECL_NODE(seqconv_out);
PATTERN_DECL_NODE(eltadd_bias);
PATTERN_DECL_NODE(eltadd_out);
PATTERN_DECL_NODE(relu_out);
};
// FC with bias
// op: mul + elementwise_add
// named nodes:
......@@ -578,6 +612,65 @@ struct ElewiseAddActInplaceGrad : public PatternBase {
PATTERN_DECL_NODE(d_ele_y);
PATTERN_DECL_NODE(ele_y);
};
// Conv with Elementwise_add as bias
// op: conv + elementwise_add
// named nodes:
// conv_input, conv_weight,
// conv_out, conv,
// eltwise_bias, eltwise_out,
// elementwise_add
struct ConvBias : public PatternBase {
ConvBias(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "conv_bias") {}
PDNode* operator()(PDNode* conv_input);
// declare operator node's name
PATTERN_DECL_NODE(conv);
PATTERN_DECL_NODE(eltwise);
// declare variable node's name
PATTERN_DECL_NODE(conv_weight);
PATTERN_DECL_NODE(conv_out);
PATTERN_DECL_NODE(eltwise_bias);
PATTERN_DECL_NODE(eltwise_out);
};
// Convolution op
// Forward pass for convolution.
// conv_input, conv_bias and conv_filter are inputs.
// conv_output is a result of the operator.
// residual_data is data used by skip connection.
// If residual connection fusion is on, the formula is:
// conv_output = conv_op(conv_filter, conv_input, conv_bias)
// + conv_residual_data
// If the fusion is off, conv_residual_data is not added.
struct Conv : public PatternBase {
Conv(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "convolution") {}
PDNode* operator()();
PATTERN_DECL_NODE(conv_op);
PATTERN_DECL_NODE(conv_input);
PATTERN_DECL_NODE(conv_filter);
PATTERN_DECL_NODE(conv_residual_data);
PATTERN_DECL_NODE(conv_output);
};
// ElementwiseAdd used in residual connections.
// y_var is used and convolution output.
// The operator is removed, when residual
// connection fusion is on.
struct ElementwiseAdd : public PatternBase {
ElementwiseAdd(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "elementwise_add") {}
PDNode* operator()(PDNode* x_var);
PATTERN_DECL_NODE(elementwise_add_op);
PATTERN_DECL_NODE(elementwise_add_x);
PATTERN_DECL_NODE(elementwise_add_y);
PATTERN_DECL_NODE(elementwise_add_out);
};
} // namespace patterns
// Link two ir::Nodes from each other.
......
paddle/fluid/framework/ir/graph_test.cc
浏览文件 @ 049c9c7d
......@@ -124,7 +124,7 @@ TEST(GraphTest, Basic) {
ASSERT_EQ(n->outputs.size(), 0UL);
}
}
ASSERT_EQ(nodes.size(), 5);
ASSERT_EQ(nodes.size(), 5UL);
}
TEST(GraphTest, WriteAfterRead) {
......
paddle/fluid/framework/ir/mkldnn_placement_pass.cc 0 → 100644
浏览文件 @ 049c9c7d
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/ir/mkldnn_placement_pass.h"
namespace paddle {
namespace framework {
namespace ir {
std::unique_ptr<ir::Graph> MKLDNNPlacementPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
VLOG(3) << "Aplies MKL-DNN placement strategy.";
for (const Node* n : graph->Nodes()) {
if (n->IsOp() && n->Op()->HasAttr("use_mkldnn")) {
n->Op()->SetAttr("use_mkldnn", true);
}
}
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(mkldnn_placement_pass,
paddle::framework::ir::MKLDNNPlacementPass);
paddle/fluid/framework/ir/mkldnn_placement_pass.h 0 → 100644
浏览文件 @ 049c9c7d
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace ir {
class MKLDNNPlacementPass : public Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc 0 → 100644
浏览文件 @ 049c9c7d
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h"
#include <string>
#include "paddle/fluid/framework/lod_tensor.h"
namespace paddle {
namespace framework {
namespace ir {
int BuildFusion(Graph* graph, const std::string& name_scope, Scope* scope) {
GraphPatternDetector gpd;
auto* pattern = gpd.mutable_pattern();
PDNode* x = pattern->NewNode(patterns::PDNodeName(name_scope, "X"))
->assert_is_op_input("sequence_conv")
->assert_var_not_persistable();
patterns::SeqConvEltAddRelu fuse_pattern(pattern, name_scope);
fuse_pattern(x);
// Create New OpDesc
auto fuse_creator = [&](Node* seqconv, Node* input, Node* seqconv_weight,
Node* eltadd_bias, Node* relu_out) {
OpDesc op_desc;
op_desc.SetType("fusion_seqconv_eltadd_relu");
op_desc.SetInput("X", {input->Name()});
op_desc.SetInput("Filter", {seqconv_weight->Name()});
op_desc.SetInput("Bias", {eltadd_bias->Name()});
op_desc.SetAttr("contextLength", seqconv->Op()->GetAttr("contextLength"));
op_desc.SetAttr("contextStart", seqconv->Op()->GetAttr("contextStart"));
op_desc.SetAttr("contextStride", seqconv->Op()->GetAttr("contextStride"));
PADDLE_ENFORCE(graph->Has(kParamScopeAttr));
auto* scope = graph->Get<Scope*>(kParamScopeAttr);
const std::string ColMat = patterns::UniqueKey("SeqConvColMat");
op_desc.SetOutput("ColMat", {ColMat});
op_desc.SetOutput("Out", {relu_out->Name()});
scope->Var(ColMat)->GetMutable<LoDTensor>();
auto* op = graph->CreateOpNode(&op_desc);
IR_NODE_LINK_TO(input, op);
IR_NODE_LINK_TO(seqconv_weight, op);
IR_NODE_LINK_TO(eltadd_bias, op);
IR_NODE_LINK_TO(op, relu_out);
return op;
};
int fusion_count{0};
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
VLOG(4) << "handle SeqConv EltAdd Relu fuse";
GET_IR_NODE_FROM_SUBGRAPH(seqconv, seqconv, fuse_pattern);
GET_IR_NODE_FROM_SUBGRAPH(seqconv_weight, seqconv_weight, fuse_pattern);
GET_IR_NODE_FROM_SUBGRAPH(seqconv_out, seqconv_out, fuse_pattern);
GET_IR_NODE_FROM_SUBGRAPH(eltadd, eltadd, fuse_pattern);
GET_IR_NODE_FROM_SUBGRAPH(eltadd_bias, eltadd_bias, fuse_pattern);
GET_IR_NODE_FROM_SUBGRAPH(eltadd_out, eltadd_out, fuse_pattern);
GET_IR_NODE_FROM_SUBGRAPH(relu, relu, fuse_pattern);
GET_IR_NODE_FROM_SUBGRAPH(relu_out, relu_out, fuse_pattern);
fuse_creator(seqconv, subgraph.at(x), seqconv_weight, eltadd_bias,
relu_out);
std::unordered_set<const Node*> marked_nodes(
{seqconv, seqconv_out, eltadd, eltadd_out, relu});
GraphSafeRemoveNodes(graph, marked_nodes);
++fusion_count;
};
gpd(graph, handler);
return fusion_count;
}
std::unique_ptr<ir::Graph> SeqConvEltAddReluFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
FusePassBase::Init(name_scope_, graph.get());
int fusion_count = BuildFusion(graph.get(), name_scope_, param_scope());
AddStatis(fusion_count);
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(seqconv_eltadd_relu_fuse_pass,
paddle::framework::ir::SeqConvEltAddReluFusePass);
paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h 0 → 100644
浏览文件 @ 049c9c7d
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <string>
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
namespace paddle {
namespace framework {
namespace ir {
class SeqConvEltAddReluFusePass : public FusePassBase {
public:
virtual ~SeqConvEltAddReluFusePass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
const std::string name_scope_{"seqconv_eltadd_relu_fuse"};
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/op_desc.cc
浏览文件 @ 049c9c7d
......@@ -515,20 +515,14 @@ void OpDesc::InferShape(const BlockDesc &block) const {
}
void OpDesc::InferVarType(BlockDesc *block) const {
// There are a few places that var type can be set.
// When VarDesc is created, default set to LOD_TENSOR.
// When output variable is created, default is defaut set to LOD_TENSOR.
// We limit here to be the only place that operator defines its customized
// var type inference. Hence, we don't do any "default" setting here.
auto &info = OpInfoMap::Instance().Get(this->Type());
if (info.infer_var_type_) {
info.infer_var_type_(*this, block);
} else {
// all output type is LoDTensor by default
VLOG(10) << this->Type()
<< " has not registered InferVarType. Set output variables to "
"LOD_TENSOR";
for (auto &out_pair : this->outputs_) {
for (auto &out_var_name : out_pair.second) {
block->FindRecursiveOrCreateVar(out_var_name)
.SetType(proto::VarType::LOD_TENSOR);
}
}
}
}
......
paddle/fluid/framework/parallel_executor.cc
浏览文件 @ 049c9c7d
......@@ -299,6 +299,12 @@ void ParallelExecutor::FeedAndSplitTensorIntoLocalScopes(
}
ParallelExecutor::~ParallelExecutor() {
const auto dev_ctxs =
platform::DeviceContextPool::Instance().GetAllDeviceContexts();
for (auto &dev_ctx : dev_ctxs) {
dev_ctx->Wait();
}
if (member_->own_local_scope_) {
for (size_t i = 1; i < member_->local_scopes_.size(); ++i) {
Scope *local_scope = member_->local_scopes_[i];
......
paddle/fluid/framework/program_desc_test.cc
浏览文件 @ 049c9c7d
......@@ -103,7 +103,7 @@ TEST(ProgramDesc, copy_ctor) {
ASSERT_EQ(1, op->GetBlockAttrId("sub_block"));
found_sub_block = true;
ASSERT_EQ(2, op->GetBlocksAttrIds("sub_blocks").size());
ASSERT_EQ(2UL, op->GetBlocksAttrIds("sub_blocks").size());
found_sub_blocks = true;
}
}
......
paddle/fluid/framework/reader_test.cc
浏览文件 @ 049c9c7d
......@@ -40,7 +40,7 @@ TEST(READER, decorate_chain) {
auto endpoints = root->GetEndPoints();
ASSERT_EQ(endpoints.size(), 2U);
ASSERT_NE(endpoints.count(end_point1.get()), 0UL);
ASSERT_NE(endpoints.count(end_point2.get()), 0);
ASSERT_NE(endpoints.count(end_point2.get()), 0UL);
}
{
......
paddle/fluid/inference/analysis/analyzer.cc
浏览文件 @ 049c9c7d
......@@ -101,7 +101,13 @@ Analyzer::Analyzer() { Register("manager1", new DfgPassManagerImpl); }
void Analyzer::Run(Argument* argument) {
std::vector<std::string> passes;
for (auto& pass : all_ir_passes_) {
#ifdef PADDLE_WITH_MKLDNN
if (use_mkldnn_) {
VLOG(3) << "Adding MKL-DNN placement pass";
passes.push_back("mkldnn_placement_pass");
}
#endif
for (auto& pass : ir_passes_) {
if (!disabled_ir_passes_.count(pass)) {
passes.push_back(pass);
passes.push_back("graph_viz_pass"); // add graphviz for debug.
......@@ -117,11 +123,26 @@ void Analyzer::Run(Argument* argument) {
}
}
Analyzer& Analyzer::IncludeAllIrPasses() {
ir_passes_ = all_ir_passes_;
return *this;
}
Analyzer& Analyzer::DisableIrPasses(const std::vector<std::string>& passes) {
disabled_ir_passes_.insert(passes.begin(), passes.end());
return *this;
}
Analyzer& Analyzer::IncludeIrPasses(const std::vector<std::string>& passes) {
ir_passes_ = passes;
return *this;
}
Analyzer& Analyzer::SetUseMkldnn(bool use_mkldnn) {
use_mkldnn_ = use_mkldnn;
return *this;
}
} // namespace analysis
} // namespace inference
} // namespace paddle
paddle/fluid/inference/analysis/analyzer.h
浏览文件 @ 049c9c7d
......@@ -54,6 +54,9 @@ class Analyzer : public OrderedRegistry<PassManager> {
void Run(Argument* argument);
Analyzer& DisableIrPasses(const std::vector<std::string>& passes);
Analyzer& IncludeIrPasses(const std::vector<std::string>& passes);
Analyzer& IncludeAllIrPasses();
Analyzer& SetUseMkldnn(bool use_mkldnn);
DISABLE_COPY_AND_ASSIGN(Analyzer);
......@@ -64,23 +67,29 @@ class Analyzer : public OrderedRegistry<PassManager> {
// larger fusion.
const std::vector<std::string> all_ir_passes_{{
// Manual update the passes here.
"infer_clean_graph_pass", //
"attention_lstm_fuse_pass", //
"embedding_fc_lstm_fuse_pass", //
"fc_lstm_fuse_pass", //
"mul_lstm_fuse_pass", //
"fc_gru_fuse_pass", //
"mul_gru_fuse_pass", //
"seq_concat_fc_fuse_pass", //
"fc_fuse_pass", //
"conv_bn_fuse_pass", //
"conv_eltwiseadd_bn_fuse_pass", //
"infer_clean_graph_pass", //
"attention_lstm_fuse_pass", //
"seqconv_eltadd_relu_fuse_pass", //
"embedding_fc_lstm_fuse_pass", //
"fc_lstm_fuse_pass", //
"mul_lstm_fuse_pass", //
"fc_gru_fuse_pass", //
"mul_gru_fuse_pass", //
"seq_concat_fc_fuse_pass", //
"fc_fuse_pass", //
"conv_bn_fuse_pass", //
"conv_eltwiseadd_bn_fuse_pass", //
#ifdef PADDLE_WITH_MKLDNN
"conv_relu_mkldnn_fuse_pass", //
"conv_bias_mkldnn_fuse_pass", //
"conv_relu_mkldnn_fuse_pass", //
"conv_elementwise_add_mkldnn_fuse_pass", //
#endif
}};
std::unordered_set<std::string> disabled_ir_passes_;
// Ir passes to run
std::vector<std::string> ir_passes_;
bool use_mkldnn_;
};
} // namespace analysis
......
paddle/fluid/inference/api/analysis_predictor.cc
浏览文件 @ 049c9c7d
......@@ -77,10 +77,6 @@ bool AnalysisPredictor::Init(
inference_program_ = program;
}
if (config_._use_mkldnn) {
executor_->EnableMKLDNN(*inference_program_);
}
executor_->Prepare(scope_.get(), *inference_program_, 0,
config_.use_feed_fetch_ops);
......@@ -225,10 +221,24 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
argument_.origin_program_desc.reset(
new ProgramDesc(*inference_program_->Proto()));
PADDLE_ENFORCE(
config_.ir_mode == contrib::AnalysisConfig::IrPassMode::kExclude,
"Only kExclude is supported yet.");
Analyzer().DisableIrPasses(config_.ir_passes).Run(&argument_);
switch (config_.ir_mode) {
case contrib::AnalysisConfig::IrPassMode::kExclude:
Analyzer()
.IncludeAllIrPasses()
.SetUseMkldnn(config_._use_mkldnn)
.DisableIrPasses(config_.ir_passes)
.Run(&argument_);
break;
case contrib::AnalysisConfig::IrPassMode::kInclude:
Analyzer()
.SetUseMkldnn(config_._use_mkldnn)
.IncludeIrPasses(config_.ir_passes)
.Run(&argument_);
break;
default:
LOG(ERROR) << "Only kExclude and kInclude modes are supoorted yet.";
}
CHECK(argument_.transformed_program_desc);
VLOG(5) << "to prepare executor";
......
paddle/fluid/inference/api/paddle_inference_api.h
浏览文件 @ 049c9c7d
......@@ -259,10 +259,17 @@ struct AnalysisConfig : public NativeConfig {
kExclude // Specify the disabled passes in `ir_passes`.
};
void SetIncludeMode() {
ir_mode = IrPassMode::kInclude;
// this pass has to be run at the beginning of all fuse passes
ir_passes = {"infer_clean_graph_pass"};
}
// Determine whether to perform graph optimization.
bool enable_ir_optim = true;
// Manually determine the IR passes to run.
IrPassMode ir_mode{IrPassMode::kExclude};
// passes to be excluded/included
std::vector<std::string> ir_passes{"embedding_fc_lstm_fuse_pass"};
// NOT stable yet.
......
paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc
浏览文件 @ 049c9c7d
......@@ -71,7 +71,7 @@ void profile(bool use_mkldnn = false) {
}
TEST(Analyzer_resnet50, profile) { profile(); }
#ifndef PADDLE_WITH_MKLDNN
#ifdef PADDLE_WITH_MKLDNN
TEST(Analyzer_resnet50, profile_mkldnn) { profile(true /* use_mkldnn */); }
#endif
......
paddle/fluid/inference/tests/api/analyzer_rnn2_tester.cc
浏览文件 @ 049c9c7d
......@@ -18,12 +18,12 @@ namespace paddle {
namespace inference {
using namespace framework; // NOLINT
static std::vector<float> result_data;
struct DataRecord {
std::vector<std::vector<std::vector<float>>> link_step_data_all;
std::vector<size_t> lod;
std::vector<std::vector<float>> rnn_link_data;
std::vector<float> result_data;
size_t num_samples; // total number of samples
size_t batch_iter{0};
size_t batch_size{1};
......@@ -57,6 +57,7 @@ struct DataRecord {
std::ifstream file(path);
std::string line;
int num_lines = 0;
result_data.clear();
while (std::getline(file, line)) {
num_lines++;
std::vector<std::string> data;
......@@ -135,13 +136,12 @@ TEST(Analyzer_rnn2, profile) {
if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) {
// the first inference result
DataRecord data(FLAGS_infer_data, FLAGS_batch_size);
PADDLE_ENFORCE_GT(outputs.size(), 0);
size_t size = GetSize(outputs[0]);
PADDLE_ENFORCE_GT(size, 0);
float *result = static_cast<float *>(outputs[0].data.data());
for (size_t i = 0; i < size; i++) {
EXPECT_NEAR(result[i], data.result_data[i], 1e-3);
EXPECT_NEAR(result[i], result_data[i], 1e-3);
}
}
}
......
paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc
浏览文件 @ 049c9c7d
......@@ -183,7 +183,13 @@ TEST(Analyzer_seq_conv1, fuse_statis) {
SetConfig(&cfg);
int num_ops;
auto predictor = CreatePaddlePredictor<AnalysisConfig>(cfg);
GetFuseStatis(predictor.get(), &num_ops);
auto fuse_statis = GetFuseStatis(predictor.get(), &num_ops);
ASSERT_TRUE(fuse_statis.count("fc_fuse"));
ASSERT_TRUE(fuse_statis.count("seqconv_eltadd_relu_fuse"));
EXPECT_EQ(fuse_statis.at("fc_fuse"), 2);
EXPECT_EQ(fuse_statis.at("seqconv_eltadd_relu_fuse"), 6);
EXPECT_EQ(num_ops, 32);
}
// Compare result of NativeConfig and AnalysisConfig
......
paddle/fluid/inference/tests/api/tester_helper.h
浏览文件 @ 049c9c7d
......@@ -50,7 +50,7 @@ void CompareResult(const std::vector<PaddleTensor> &outputs,
auto &ref_out = ref_outputs[i];
size_t size = VecReduceToInt(out.shape);
size_t ref_size = VecReduceToInt(ref_out.shape);
EXPECT_GT(size, 0);
EXPECT_GT(size, 0UL);
EXPECT_EQ(size, ref_size);
EXPECT_EQ(out.dtype, ref_out.dtype);
switch (out.dtype) {
......
paddle/fluid/operators/CMakeLists.txt
浏览文件 @ 049c9c7d
......@@ -86,7 +86,7 @@ function(op_library TARGET)
# remove windows unsupported op, because windows has no nccl, no warpctc such ops.
foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op" "warpctc_op" "hierarchical_sigmoid_op"
"crf_decoding_op" "select_op" "lstmp_op" "gru_op" "fusion_gru_op" "lstm_op" "fusion_lstm_op" "cumsum_op"
"channel_send_op" "channel_create_op" "channel_close_op" "channel_recv_op")
"fusion_seqconv_eltadd_relu_op" "channel_send_op" "channel_create_op" "channel_close_op" "channel_recv_op")
if ("${TARGET}" STREQUAL "${windows_unsupport_op}")
return()
endif()
......@@ -284,10 +284,10 @@ op_library(max_sequence_len_op DEPS lod_rank_table)
op_library(sequence_conv_op DEPS context_project)
op_library(sequence_pool_op DEPS sequence_pooling)
if (NOT WIN32)
op_library(lstm_op DEPS sequence2batch lstm_compute)
op_library(hierarchical_sigmoid_op DEPS matrix_bit_code)
op_library(lstmp_op DEPS sequence2batch lstm_compute)
op_library(gru_op DEPS sequence2batch gru_compute)
op_library(lstm_op DEPS sequence2batch lstm_compute)
op_library(hierarchical_sigmoid_op DEPS matrix_bit_code)
op_library(lstmp_op DEPS sequence2batch lstm_compute)
op_library(gru_op DEPS sequence2batch gru_compute)
endif(NOT WIN32)
op_library(recurrent_op DEPS executor)
op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale)
......@@ -316,7 +316,7 @@ op_library(save_op DEPS lod_tensor)
op_library(load_op DEPS lod_tensor)
op_library(save_combine_op DEPS lod_tensor)
op_library(load_combine_op DEPS lod_tensor)
op_library(concat_op DEPS concat)
op_library(concat_op DEPS concat_and_split)
list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
......@@ -348,6 +348,6 @@ cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor memory)
cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op)
cc_test(save_load_combine_op_test SRCS save_load_combine_op_test.cc DEPS save_combine_op load_combine_op)
if(NOT WIN32)
nv_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
nv_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
endif()
nv_test(dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor)
paddle/fluid/operators/array_to_lod_tensor_op.cc
浏览文件 @ 049c9c7d
......@@ -11,7 +11,7 @@ distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include <paddle/fluid/operators/math/concat.h>
#include <paddle/fluid/operators/math/concat_and_split.h>
#include <numeric>
#include "paddle/fluid/framework/lod_rank_table.h"
......
paddle/fluid/operators/concat_op.h
浏览文件 @ 049c9c7d
......@@ -17,7 +17,7 @@ limitations under the License. */
#include <utility>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/concat.h"
#include "paddle/fluid/operators/math/concat_and_split.h"
#include "paddle/fluid/operators/strided_memcpy.h"
namespace paddle {
......@@ -89,29 +89,17 @@ class ConcatGradKernel : public framework::OpKernel<T> {
outputs.push_back(nullptr);
}
}
auto& dev_ctx = ctx.template device_context<DeviceContext>();
// Sometimes direct copies will be faster, this maybe need deeply analysis.
if (axis == 0 && outs.size() < 10) {
size_t input_offset = 0;
const auto in_stride = framework::stride_numel(out_grad->dims());
for (size_t i = 0; i < outs.size(); ++i) {
auto out_stride = framework::stride_numel(ins[i]->dims());
auto* out = outputs[i];
if (out != nullptr) {
StridedNumelCopyWithAxis<T>(
ctx.device_context(), axis, out->data<T>(), out_stride,
out_grad->data<T>() + input_offset, in_stride, out_stride[axis]);
}
input_offset += out_stride[axis];
}
std::vector<const framework::Tensor*> ref_shape;
ref_shape.insert(ref_shape.begin(), ins.begin(), ins.end());
StridedMemcpyWithAxis0<T>(dev_ctx, *out_grad, ref_shape, &outputs);
} else {
auto& dev_ctx = ctx.template device_context<DeviceContext>();
paddle::operators::math::ConcatGradFunctor<DeviceContext, T>
concat_grad_functor;
concat_grad_functor(dev_ctx, *out_grad,
ctx.MultiInput<framework::Tensor>("X"),
static_cast<int>(axis), &outputs);
math::SplitFunctor<DeviceContext, T> split_functor;
split_functor(dev_ctx, *out_grad, ctx.MultiInput<framework::Tensor>("X"),
static_cast<int>(axis), &outputs);
}
}
};
......
paddle/fluid/operators/conv_mkldnn_op.cc
浏览文件 @ 049c9c7d
......@@ -300,10 +300,10 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
bool fuse_relu = ctx.Attr<bool>("fuse_relu");
bool fuse_eltwise = ctx.Attr<bool>("fuse_eltwise");
bool fuse_residual_conn = ctx.Attr<bool>("fuse_residual_connection");
int groups = ctx.Attr<int>("groups");
// TODO: add support for dilation
// TODO(tpatejko): add support for dilation
PADDLE_ENFORCE(
dilations.size() == 2 && dilations[0] == 1 && dilations[1] == 1,
"dilation in convolution is not implemented yet");
......@@ -369,11 +369,11 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
bias_tz, platform::MKLDNNGetDataType<T>(), memory::format::x);
conv_pd = ConvFwdPrimitiveDesc(src_md, weights_md, bias_md, dst_md,
strides, paddings, mkldnn_engine,
fuse_relu, fuse_eltwise);
fuse_relu, fuse_residual_conn);
} else {
conv_pd =
ConvFwdPrimitiveDesc(src_md, weights_md, dst_md, strides, paddings,
mkldnn_engine, fuse_relu, fuse_eltwise);
mkldnn_engine, fuse_relu, fuse_residual_conn);
}
// Save conv_pd/src_memory/weights_memory for backward pass
dev_ctx.SetBlob(key_conv_pd, conv_pd);
......@@ -386,8 +386,26 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
auto user_weights_memory_p = handler.AcquireWeightsMemory(
user_weights_md, to_void_cast<T>(filter_data));
T* output_data =
output->mutable_data<T>(ctx.GetPlace(), handler.GetDstMemorySize());
T* output_data = nullptr;
if (fuse_residual_conn) {
auto residual_param = ctx.Input<Tensor>("ResidualData");
auto residual_param_data = residual_param->data<T>();
PADDLE_ENFORCE(
residual_param_data != nullptr,
"Provide data if you want MKLDNN conv+elementwise_add fusion");
PADDLE_ENFORCE_EQ(output->dims(), residual_param->dims(),
"Output and elementwise parameter need to have the "
"same dimension sizes");
output->ShareDataWith(*residual_param);
output_data = output->mutable_data<T>(ctx.GetPlace());
} else {
output_data =
output->mutable_data<T>(ctx.GetPlace(), handler.GetDstMemorySize());
}
// create reorder primitive if the input format is not the preferred one
auto src_memory_p =
handler.AcquireSrcMemoryFromPrimitive(user_src_memory_p, pipeline);
......@@ -424,14 +442,15 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
private:
mkldnn::primitive_attr CreatePostOps(bool fuse_relu,
bool fuse_eltwise) const {
bool fuse_residual_conn) const {
mkldnn::primitive_attr conv_attr;
mkldnn::post_ops post_operations;
// Fusion with Elementwise layer relies on adding a sum post-operation with
// the scale parameter. It is assumed that when fuse_eltwise is true, the
// Output tensor contains the data coming from residual connection. The
// result of this post_op is: Output = scale * Output + Conv_Out.
if (fuse_eltwise) {
// the scale parameter. It is assumed that when fuse_residual_connection is
// true, the output tensor contains the data coming from residual
// connection. The result of this post_op is:
// Output = scale * Output + Conv_Out.
if (fuse_residual_conn) {
post_operations.append_sum(1.0f);
}
// Fusion with ReLU layer is executed through the PostOps feature. Create a
......@@ -452,7 +471,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
const memory::desc& dst, const std::vector<int>& strides,
const std::vector<int>& paddings,
const mkldnn::engine& engine, const bool fuse_relu,
const bool fuse_eltwise) const {
const bool fuse_residual_conn) const {
memory::dims stride_dims = {strides[0], strides[1]};
memory::dims padding_dims = {paddings[0], paddings[1]};
......@@ -461,7 +480,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
dst, stride_dims, padding_dims, padding_dims,
mkldnn::padding_kind::zero);
mkldnn::primitive_attr conv_attr = CreatePostOps(fuse_relu, fuse_eltwise);
mkldnn::primitive_attr conv_attr =
CreatePostOps(fuse_relu, fuse_residual_conn);
auto p_conv_pd = new mkldnn::convolution_forward::primitive_desc(
conv_desc, conv_attr, engine);
......@@ -476,7 +496,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
const std::vector<int>& strides,
const std::vector<int>& paddings,
const mkldnn::engine& engine, const bool fuse_relu,
const bool fuse_eltwise) const {
const bool fuse_residual_conn) const {
memory::dims stride_dims = {strides[0], strides[1]};
memory::dims padding_dims = {paddings[0], paddings[1]};
......@@ -485,7 +505,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
bias, dst, stride_dims, padding_dims, padding_dims,
mkldnn::padding_kind::zero);
mkldnn::primitive_attr conv_attr = CreatePostOps(fuse_relu, fuse_eltwise);
mkldnn::primitive_attr conv_attr =
CreatePostOps(fuse_relu, fuse_residual_conn);
auto p_conv_pd = new mkldnn::convolution_forward::primitive_desc(
conv_desc, conv_attr, engine);
......
paddle/fluid/operators/conv_op.cc
浏览文件 @ 049c9c7d
......@@ -132,6 +132,11 @@ void Conv2DOpMaker::Make() {
"(Tensor) The output tensor of convolution operator. "
"The format of output tensor is also NCHW.")
.Reuse("Input");
AddInput("ResidualData",
"(Tensor) Tensor with residual data "
"to which convolution output will be added."
"Used with fuse_residual_connection fusion.")
.AsDispensable();
AddAttr<std::vector<int>>("strides",
"(vector<int> default:{1, 1}), the "
"strides(h_stride, w_stride) of "
......@@ -164,10 +169,10 @@ void Conv2DOpMaker::Make() {
.SetDefault(false);
AddAttr<bool>("fuse_relu", "(bool, default false) Only used in mkldnn kernel")
.SetDefault(false);
AddAttr<bool>("fuse_eltwise",
AddAttr<bool>("fuse_residual_connection",
"(bool, default false) Only used in mkldnn kernel. Used "
"whenever convolution output is connected via skip connection "
"to a previous layer.")
"whenever convolution output is as an input to residual "
"connection.")
.SetDefault(false);
AddAttr<std::string>(
"data_format",
......
paddle/fluid/operators/detection/CMakeLists.txt
浏览文件 @ 049c9c7d
......@@ -20,7 +20,7 @@ detection_library(box_coder_op SRCS box_coder_op.cc box_coder_op.cu)
detection_library(iou_similarity_op SRCS iou_similarity_op.cc
iou_similarity_op.cu)
detection_library(mine_hard_examples_op SRCS mine_hard_examples_op.cc)
detection_library(multiclass_nms_op SRCS multiclass_nms_op.cc)
detection_library(multiclass_nms_op SRCS multiclass_nms_op.cc poly_util.cc gpc.cc)
detection_library(prior_box_op SRCS prior_box_op.cc prior_box_op.cu)
detection_library(anchor_generator_op SRCS anchor_generator_op.cc
anchor_generator_op.cu)
......
paddle/fluid/operators/detection/generate_proposal_labels_op.cc
浏览文件 @ 049c9c7d
......@@ -16,7 +16,7 @@ limitations under the License. */
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/detection/bbox_util.h"
#include "paddle/fluid/operators/gather.h"
#include "paddle/fluid/operators/math/concat.h"
#include "paddle/fluid/operators/math/concat_and_split.h"
#include "paddle/fluid/operators/math/math_function.h"
namespace paddle {
......
paddle/fluid/operators/detection/generate_proposals_op.cc
浏览文件 @ 049c9c7d
......@@ -12,10 +12,12 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include <cmath>
#include <cstring>
#include <string>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/var_type.h"
#include "paddle/fluid/operators/detail/safe_ref.h"
#include "paddle/fluid/operators/gather.h"
#include "paddle/fluid/operators/math/math_function.h"
......@@ -25,21 +27,17 @@ namespace operators {
using Tensor = framework::Tensor;
using LoDTensor = framework::LoDTensor;
struct AppendProposalsFunctor {
LoDTensor *out_;
int64_t offset_;
Tensor *to_add_;
static const double kBBoxClipDefault = std::log(1000.0 / 16.0);
AppendProposalsFunctor(LoDTensor *out, int64_t offset, Tensor *to_add)
: out_(out), offset_(offset), to_add_(to_add) {}
template <typename T>
void apply() const {
auto *out_data = out_->data<T>();
auto *to_add_data = to_add_->data<T>();
memcpy(out_data + offset_, to_add_data, to_add_->numel() * sizeof(T));
}
};
static void AppendProposals(Tensor *dst, int64_t offset, const Tensor &src) {
auto *out_data = dst->data<void>();
auto *to_add_data = src.data<void>();
size_t size_of_t = framework::SizeOfType(src.type());
offset *= size_of_t;
std::memcpy(
reinterpret_cast<void *>(reinterpret_cast<uintptr_t>(out_data) + offset),
to_add_data, src.numel() * size_of_t);
}
class GenerateProposalsOp : public framework::OperatorWithKernel {
public:
......@@ -75,8 +73,9 @@ class GenerateProposalsOp : public framework::OperatorWithKernel {
};
template <class T>
void BoxCoder(const platform::DeviceContext &ctx, Tensor *all_anchors,
Tensor *bbox_deltas, Tensor *variances, Tensor *proposals) {
static inline void BoxCoder(const platform::DeviceContext &ctx,
Tensor *all_anchors, Tensor *bbox_deltas,
Tensor *variances, Tensor *proposals) {
T *proposals_data = proposals->mutable_data<T>(ctx.GetPlace());
int64_t row = all_anchors->dims()[0];
......@@ -108,11 +107,11 @@ void BoxCoder(const platform::DeviceContext &ctx, Tensor *all_anchors,
anchor_center_y;
bbox_width = std::exp(std::min<T>(variances_data[i * len + 2] *
bbox_deltas_data[i * len + 2],
std::log(1000.0 / 16.0))) *
kBBoxClipDefault)) *
anchor_width;
bbox_height = std::exp(std::min<T>(variances_data[i * len + 3] *
bbox_deltas_data[i * len + 3],
std::log(1000.0 / 16.0))) *
kBBoxClipDefault)) *
anchor_height;
} else {
bbox_center_x =
......@@ -120,10 +119,10 @@ void BoxCoder(const platform::DeviceContext &ctx, Tensor *all_anchors,
bbox_center_y =
bbox_deltas_data[i * len + 1] * anchor_height + anchor_center_y;
bbox_width = std::exp(std::min<T>(bbox_deltas_data[i * len + 2],
std::log(1000.0 / 16.0))) *
kBBoxClipDefault)) *
anchor_width;
bbox_height = std::exp(std::min<T>(bbox_deltas_data[i * len + 3],
std::log(1000.0 / 16.0))) *
kBBoxClipDefault)) *
anchor_height;
}
......@@ -136,30 +135,32 @@ void BoxCoder(const platform::DeviceContext &ctx, Tensor *all_anchors,
}
template <class T>
void ClipTiledBoxes(const platform::DeviceContext &ctx, const Tensor &im_info,
Tensor *boxes) {
static inline void ClipTiledBoxes(const platform::DeviceContext &ctx,
const Tensor &im_info, Tensor *boxes) {
T *boxes_data = boxes->mutable_data<T>(ctx.GetPlace());
const T *im_info_data = im_info.data<T>();
T zero(0);
for (int64_t i = 0; i < boxes->numel(); ++i) {
if (i % 4 == 0) {
boxes_data[i] =
std::max(std::min(boxes_data[i], im_info_data[1] - 1), 0.0f);
std::max(std::min(boxes_data[i], im_info_data[1] - 1), zero);
} else if (i % 4 == 1) {
boxes_data[i] =
std::max(std::min(boxes_data[i], im_info_data[0] - 1), 0.0f);
std::max(std::min(boxes_data[i], im_info_data[0] - 1), zero);
} else if (i % 4 == 2) {
boxes_data[i] =
std::max(std::min(boxes_data[i], im_info_data[1] - 1), 0.0f);
std::max(std::min(boxes_data[i], im_info_data[1] - 1), zero);
} else {
boxes_data[i] =
std::max(std::min(boxes_data[i], im_info_data[0] - 1), 0.0f);
std::max(std::min(boxes_data[i], im_info_data[0] - 1), zero);
}
}
}
template <class T>
void FilterBoxes(const platform::DeviceContext &ctx, Tensor *boxes,
float min_size, const Tensor &im_info, Tensor *keep) {
static inline void FilterBoxes(const platform::DeviceContext &ctx,
Tensor *boxes, float min_size,
const Tensor &im_info, Tensor *keep) {
const T *im_info_data = im_info.data<T>();
T *boxes_data = boxes->mutable_data<T>(ctx.GetPlace());
T im_scale = im_info_data[2];
......@@ -185,24 +186,24 @@ void FilterBoxes(const platform::DeviceContext &ctx, Tensor *boxes,
keep->Resize({keep_len});
}
bool SortScorePairDescend(const std::pair<float, int> &pair1,
const std::pair<float, int> &pair2) {
return pair1.first > pair2.first;
}
template <class T>
void GetMaxScoreIndex(const std::vector<T> &scores,
std::vector<std::pair<T, int>> *sorted_indices) {
static inline std::vector<std::pair<T, int>> GetSortedScoreIndex(
const std::vector<T> &scores) {
std::vector<std::pair<T, int>> sorted_indices;
sorted_indices.reserve(scores.size());
for (size_t i = 0; i < scores.size(); ++i) {
sorted_indices->push_back(std::make_pair(scores[i], i));
sorted_indices.emplace_back(scores[i], i);
}
// Sort the score pair according to the scores in descending order
std::stable_sort(sorted_indices->begin(), sorted_indices->end(),
SortScorePairDescend);
std::stable_sort(sorted_indices.begin(), sorted_indices.end(),
[](const std::pair<T, int> &a, const std::pair<T, int> &b) {
return a.first < b.first;
});
return sorted_indices;
}
template <class T>
T BBoxArea(const T *box, const bool normalized) {
static inline T BBoxArea(const T *box, bool normalized) {
if (box[2] < box[0] || box[3] < box[1]) {
// If coordinate values are is invalid
// (e.g. xmax < xmin or ymax < ymin), return 0.
......@@ -220,7 +221,7 @@ T BBoxArea(const T *box, const bool normalized) {
}
template <class T>
T JaccardOverlap(const T *box1, const T *box2, const bool normalized) {
static inline T JaccardOverlap(const T *box1, const T *box2, bool normalized) {
if (box2[0] > box1[2] || box2[2] < box1[0] || box2[1] > box1[3] ||
box2[3] < box1[1]) {
return static_cast<T>(0.);
......@@ -229,8 +230,8 @@ T JaccardOverlap(const T *box1, const T *box2, const bool normalized) {
const T inter_ymin = std::max(box1[1], box2[1]);
const T inter_xmax = std::min(box1[2], box2[2]);
const T inter_ymax = std::min(box1[3], box2[3]);
const T inter_w = std::max(0.0f, inter_xmax - inter_xmin + 1);
const T inter_h = std::max(0.0f, inter_ymax - inter_ymin + 1);
const T inter_w = std::max(T(0), inter_xmax - inter_xmin + 1);
const T inter_h = std::max(T(0), inter_ymax - inter_ymin + 1);
const T inter_area = inter_w * inter_h;
const T bbox1_area = BBoxArea<T>(box1, normalized);
const T bbox2_area = BBoxArea<T>(box2, normalized);
......@@ -238,9 +239,21 @@ T JaccardOverlap(const T *box1, const T *box2, const bool normalized) {
}
}
template <typename T>
static inline Tensor VectorToTensor(const std::vector<T> &selected_indices,
int selected_num) {
Tensor keep_nms;
keep_nms.Resize({selected_num});
auto *keep_data = keep_nms.mutable_data<T>(platform::CPUPlace());
for (int i = 0; i < selected_num; ++i) {
keep_data[i] = selected_indices[i];
}
return keep_nms;
}
template <class T>
Tensor NMS(const platform::DeviceContext &ctx, Tensor *bbox, Tensor *scores,
const T nms_threshold, const float eta) {
static inline Tensor NMS(const platform::DeviceContext &ctx, Tensor *bbox,
Tensor *scores, T nms_threshold, float eta) {
PADDLE_ENFORCE_NOT_NULL(bbox);
int64_t num_boxes = bbox->dims()[0];
// 4: [xmin ymin xmax ymax]
......@@ -248,20 +261,18 @@ Tensor NMS(const platform::DeviceContext &ctx, Tensor *bbox, Tensor *scores,
std::vector<T> scores_data(num_boxes);
std::copy_n(scores->data<T>(), num_boxes, scores_data.begin());
std::vector<std::pair<T, int>> sorted_indices;
GetMaxScoreIndex<T>(scores_data, &sorted_indices);
std::vector<std::pair<T, int>> sorted_indices =
GetSortedScoreIndex<T>(scores_data);
std::vector<int> selected_indices;
int selected_num = 0;
T adaptive_threshold = nms_threshold;
const T *bbox_data = bbox->data<T>();
bool flag;
while (sorted_indices.size() != 0) {
int idx = sorted_indices.front().second;
flag = true;
for (size_t k = 0; k < selected_indices.size(); ++k) {
int idx = sorted_indices.back().second;
bool flag = true;
for (int kept_idx : selected_indices) {
if (flag) {
const int kept_idx = selected_indices[k];
T overlap = JaccardOverlap<T>(bbox_data + idx * box_size,
bbox_data + kept_idx * box_size, false);
flag = (overlap <= adaptive_threshold);
......@@ -271,32 +282,29 @@ Tensor NMS(const platform::DeviceContext &ctx, Tensor *bbox, Tensor *scores,
}
if (flag) {
selected_indices.push_back(idx);
selected_num++;
++selected_num;
}
sorted_indices.erase(sorted_indices.begin());
sorted_indices.erase(sorted_indices.end());
if (flag && eta < 1 && adaptive_threshold > 0.5) {
adaptive_threshold *= eta;
}
}
Tensor keep_nms;
keep_nms.Resize({selected_num});
int *keep_data = keep_nms.mutable_data<int>(ctx.GetPlace());
for (int i = 0; i < selected_num; ++i) {
keep_data[i] = selected_indices[i];
}
return keep_nms;
return VectorToTensor(selected_indices, selected_num);
}
template <typename DeviceContext, typename T>
template <typename T>
class GenerateProposalsKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &context) const override {
auto *scores = context.Input<Tensor>("Scores");
auto *bbox_deltas = context.Input<Tensor>("BboxDeltas");
auto *im_info = context.Input<Tensor>("ImInfo");
auto *anchors = context.Input<Tensor>("Anchors");
auto *variances = context.Input<Tensor>("Variances");
auto anchors = detail::Ref(context.Input<Tensor>("Anchors"),
"Cannot find input Anchors(%s) in scope",
context.Inputs("Anchors")[0]);
auto variances = detail::Ref(context.Input<Tensor>("Variances"),
"Cannot find input Variances(%s) in scope",
context.Inputs("Variances")[0]);
auto *rpn_rois = context.Output<LoDTensor>("RpnRois");
auto *rpn_roi_probs = context.Output<LoDTensor>("RpnRoiProbs");
......@@ -307,15 +315,16 @@ class GenerateProposalsKernel : public framework::OpKernel<T> {
float min_size = context.Attr<float>("min_size");
float eta = context.Attr<float>("eta");
auto &dev_ctx = context.template device_context<DeviceContext>();
auto &dev_ctx =
context.template device_context<platform::CPUDeviceContext>();
auto scores_dim = scores->dims();
auto &scores_dim = scores->dims();
int64_t num = scores_dim[0];
int64_t c_score = scores_dim[1];
int64_t h_score = scores_dim[2];
int64_t w_score = scores_dim[3];
auto bbox_dim = bbox_deltas->dims();
auto &bbox_dim = bbox_deltas->dims();
int64_t c_bbox = bbox_dim[1];
int64_t h_bbox = bbox_dim[2];
int64_t w_bbox = bbox_dim[3];
......@@ -330,17 +339,17 @@ class GenerateProposalsKernel : public framework::OpKernel<T> {
scores_swap.mutable_data<T>({num, h_score, w_score, c_score},
dev_ctx.GetPlace());
math::Transpose<DeviceContext, T, 4> trans;
math::Transpose<platform::CPUDeviceContext, T, 4> trans;
std::vector<int> axis = {0, 2, 3, 1};
trans(dev_ctx, *bbox_deltas, &bbox_deltas_swap, axis);
trans(dev_ctx, *scores, &scores_swap, axis);
framework::LoD lod;
std::vector<size_t> lod0(1, 0);
Tensor *anchor = const_cast<framework::Tensor *>(anchors);
anchor->Resize({anchors->numel() / 4, 4});
Tensor *var = const_cast<framework::Tensor *>(variances);
var->Resize({var->numel() / 4, 4});
lod.resize(1);
auto &lod0 = lod[0];
lod0.push_back(0);
anchors.Resize({anchors.numel() / 4, 4});
variances.Resize({variances.numel() / 4, 4});
int64_t num_proposals = 0;
for (int64_t i = 0; i < num; ++i) {
......@@ -352,24 +361,17 @@ class GenerateProposalsKernel : public framework::OpKernel<T> {
scores_slice.Resize({h_score * w_score * c_score, 1});
std::pair<Tensor, Tensor> tensor_pair =
ProposalForOneImage(dev_ctx, im_info_slice, *anchor, *var,
ProposalForOneImage(dev_ctx, im_info_slice, anchors, variances,
bbox_deltas_slice, scores_slice, pre_nms_top_n,
post_nms_top_n, nms_thresh, min_size, eta);
Tensor proposals = tensor_pair.first;
Tensor scores = tensor_pair.second;
framework::VisitDataType(
framework::ToDataType(rpn_rois->type()),
AppendProposalsFunctor(rpn_rois, 4 * num_proposals, &proposals));
framework::VisitDataType(
framework::ToDataType(rpn_roi_probs->type()),
AppendProposalsFunctor(rpn_roi_probs, num_proposals, &scores));
Tensor &proposals = tensor_pair.first;
Tensor &scores = tensor_pair.second;
AppendProposals(rpn_rois, 4 * num_proposals, proposals);
AppendProposals(rpn_roi_probs, num_proposals, scores);
num_proposals += proposals.dims()[0];
lod0.emplace_back(num_proposals);
lod0.push_back(num_proposals);
}
lod.emplace_back(lod0);
rpn_rois->set_lod(lod);
rpn_roi_probs->set_lod(lod);
rpn_rois->Resize({num_proposals, 4});
......@@ -377,7 +379,7 @@ class GenerateProposalsKernel : public framework::OpKernel<T> {
}
std::pair<Tensor, Tensor> ProposalForOneImage(
const DeviceContext &ctx, const Tensor &im_info_slice,
const platform::CPUDeviceContext &ctx, const Tensor &im_info_slice,
const Tensor &anchors, const Tensor &variances,
const Tensor &bbox_deltas_slice, // [M, 4]
const Tensor &scores_slice, // [N, 1]
......@@ -392,10 +394,9 @@ class GenerateProposalsKernel : public framework::OpKernel<T> {
for (int i = 0; i < scores_slice.numel(); ++i) {
index[i] = i;
}
std::function<bool(const int64_t &, const int64_t &)> compare =
[scores_data](const int64_t &i, const int64_t &j) {
return scores_data[i] > scores_data[j];
};
auto compare = [scores_data](const int64_t &i, const int64_t &j) {
return scores_data[i] > scores_data[j];
};
if (pre_nms_top_n <= 0 || pre_nms_top_n >= scores_slice.numel()) {
std::sort(index, index + scores_slice.numel(), compare);
......@@ -452,33 +453,45 @@ class GenerateProposalsKernel : public framework::OpKernel<T> {
class GenerateProposalsOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("Scores", "The scores of anchors should be foreground.");
AddInput("BboxDeltas", "bbox_deltas.");
AddInput("ImInfo", "Information for image reshape.");
AddInput("Anchors", "All anchors.");
AddInput("Variances", " variances");
AddOutput("RpnRois", "Anchors.");
AddOutput("RpnRoiProbs", "Anchors.");
AddAttr<int>("pre_nms_topN", "pre_nms_topN");
AddAttr<int>("post_nms_topN", "post_nms_topN");
AddAttr<float>("nms_thresh", "nms_thres");
AddAttr<float>("min_size", "min size");
AddInput("Scores",
"(Tensor) The scores from conv is in shape (N, A, H, W), "
"N is batch size, A is number of anchors, "
"H and W are height and width of the feature map");
AddInput("BboxDeltas",
"(Tensor) Bounding box deltas from conv is in "
"shape (N, 4*A, H, W).");
AddInput("ImInfo",
"(Tensor) Information for image reshape is in shape (N, 3), "
"in format (height, width, scale)");
AddInput("Anchors",
"(Tensor) Bounding box anchors from anchor_generator_op "
"is in shape (A, H, W, 4).");
AddInput("Variances",
"(Tensor) Bounding box variances with same shape as `Anchors`.");
AddOutput("RpnRois",
"(LoDTensor), Output proposals with shape (rois_num, 4).");
AddOutput("RpnRoiProbs",
"(LoDTensor) Scores of proposals with shape (rois_num, 1).");
AddAttr<int>("pre_nms_topN",
"Number of top scoring RPN proposals to keep before "
"applying NMS.");
AddAttr<int>("post_nms_topN",
"Number of top scoring RPN proposals to keep after "
"applying NMS");
AddAttr<float>("nms_thresh", "NMS threshold used on RPN proposals.");
AddAttr<float>("min_size",
"Proposal height and width both need to be greater "
"than this min_size.");
AddAttr<float>("eta", "The parameter for adaptive NMS.");
AddComment(R"DOC(
Generate Proposals OP
This operator proposes rois according to each box with their probability to be a foreground object and
the box can be calculated by anchors. Bbox_deltais and scores are the output of RPN. Final proposals
could be used to train detection net.
Scores is the probability for each box to be an object. In format of (N, A, H, W) where N is batch size, A is number
of anchors, H and W are height and width of the feature map.
BboxDeltas is the differece between predicted box locatoin and anchor location. In format of (N, 4*A, H, W)
This operator Generate bounding box proposals for Faster RCNN.
The propoasls are generated for a list of images based on image
score 'Scores', bounding box regression result 'BboxDeltas' as
well as predefined bounding box shapes 'anchors'. Greedy
non-maximum suppression is applied to generate the final bounding
boxes.
For generating proposals, this operator transposes and resizes scores and bbox_deltas in size of (H*W*A, 1) and (H*W*A, 4) and
calculate box locations as proposals candidates. Then clip boxes to image and remove predicted boxes with small area.
Finally, apply nms to get final proposals as output.
)DOC");
}
};
......@@ -490,6 +503,5 @@ namespace ops = paddle::operators;
REGISTER_OPERATOR(generate_proposals, ops::GenerateProposalsOp,
ops::GenerateProposalsOpMaker,
paddle::framework::EmptyGradOpMaker);
REGISTER_OP_CPU_KERNEL(
generate_proposals,
ops::GenerateProposalsKernel<paddle::platform::CPUDeviceContext, float>);
REGISTER_OP_CPU_KERNEL(generate_proposals, ops::GenerateProposalsKernel<float>,
ops::GenerateProposalsKernel<double>);
paddle/fluid/operators/detection/generate_proposals_op.cu
浏览文件 @ 049c9c7d
......@@ -16,10 +16,13 @@ limitations under the License. */
#include <string>
#include <vector>
#include "cub/cub.cuh"
#include "paddle/fluid/framework/mixed_vector.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/memory/memory.h"
#include "paddle/fluid/operators/detail/safe_ref.h"
#include "paddle/fluid/operators/gather.cu.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/platform/for_range.h"
namespace paddle {
namespace operators {
......@@ -36,36 +39,38 @@ namespace {
int const kThreadsPerBlock = sizeof(uint64_t) * 8;
template <typename T>
__global__ void RangeInitKernel(const T start, const T delta, const int size,
T *out) {
CUDA_1D_KERNEL_LOOP(i, size) { out[i] = start + i * delta; }
}
static const double kBBoxClipDefault = std::log(1000.0 / 16.0);
struct RangeInitFunctor {
int start_;
int delta_;
int *out_;
__device__ void operator()(size_t i) { out_[i] = start_ + i * delta_; }
};
template <typename T>
void SortDescending(const platform::CUDADeviceContext &ctx, const Tensor &value,
Tensor *value_out, Tensor *index_out) {
int num = value.numel();
static void SortDescending(const platform::CUDADeviceContext &ctx,
const Tensor &value, Tensor *value_out,
Tensor *index_out) {
int num = static_cast<int>(value.numel());
Tensor index_in_t;
int *idx_in = index_in_t.mutable_data<int>({num}, ctx.GetPlace());
int block = 512;
auto stream = ctx.stream();
RangeInitKernel<<<DIVUP(num, block), block, 0, stream>>>(0, 1, num, idx_in);
platform::ForRange<platform::CUDADeviceContext> for_range(ctx, num);
for_range(RangeInitFunctor{0, 1, idx_in});
int *idx_out = index_out->mutable_data<int>({num}, ctx.GetPlace());
const T *keys_in = value.data<T>();
T *keys_out = value_out->mutable_data<T>({num}, ctx.GetPlace());
// Determine temporary device storage requirements
void *d_temp_storage = NULL;
size_t temp_storage_bytes = 0;
cub::DeviceRadixSort::SortPairsDescending<T, int>(
d_temp_storage, temp_storage_bytes, keys_in, keys_out, idx_in, idx_out,
num);
nullptr, temp_storage_bytes, keys_in, keys_out, idx_in, idx_out, num);
// Allocate temporary storage
auto place = boost::get<platform::CUDAPlace>(ctx.GetPlace());
d_temp_storage = memory::Alloc(place, temp_storage_bytes);
void *d_temp_storage = memory::Alloc(place, temp_storage_bytes);
// Run sorting operation
cub::DeviceRadixSort::SortPairsDescending<T, int>(
......@@ -76,22 +81,27 @@ void SortDescending(const platform::CUDADeviceContext &ctx, const Tensor &value,
}
template <typename T>
__device__ __forceinline__ T Min(T x, T y) {
return x < y ? x : y;
}
template <typename T>
__device__ __forceinline__ T Max(T x, T y) {
return x > y ? x : y;
}
template <typename T>
__global__ void BoxDecodeAndClipKernel(const T *anchor, const T *deltas,
const T *var, const int *index,
const T *im_info, const int num,
T *proposals) {
T kBBoxClipDefault = log(1000.0 / 16.0);
CUDA_1D_KERNEL_LOOP(i, num) {
struct BoxDecodeAndClipFunctor {
const T *anchor;
const T *deltas;
const T *var;
const int *index;
const T *im_info;
T *proposals;
BoxDecodeAndClipFunctor(const T *anchor, const T *deltas, const T *var,
const int *index, const T *im_info, T *proposals)
: anchor(anchor),
deltas(deltas),
var(var),
index(index),
im_info(im_info),
proposals(proposals) {}
T bbox_clip_default{static_cast<T>(kBBoxClipDefault)};
__device__ void operator()(size_t i) {
int k = index[i] * 4;
T axmin = anchor[k];
T aymin = anchor[k + 1];
......@@ -108,17 +118,17 @@ __global__ void BoxDecodeAndClipKernel(const T *anchor, const T *deltas,
T dxmax = deltas[k + 2];
T dymax = deltas[k + 3];
T d_cx = 0., d_cy = 0., d_w = 0., d_h = 0.;
T d_cx, d_cy, d_w, d_h;
if (var) {
d_cx = cx + dxmin * w * var[k];
d_cy = cy + dymin * h * var[k + 1];
d_w = exp(Min<T>(dxmax * var[k + 2], kBBoxClipDefault)) * w;
d_h = exp(Min<T>(dymax * var[k + 3], kBBoxClipDefault)) * h;
d_w = exp(Min(dxmax * var[k + 2], bbox_clip_default)) * w;
d_h = exp(Min(dymax * var[k + 3], bbox_clip_default)) * h;
} else {
d_cx = cx + dxmin * w;
d_cy = cy + dymin * h;
d_w = exp(Min<T>(dxmax, kBBoxClipDefault)) * w;
d_h = exp(Min<T>(dymax, kBBoxClipDefault)) * h;
d_w = exp(Min(dxmax, bbox_clip_default)) * w;
d_h = exp(Min(dymax, bbox_clip_default)) * h;
}
T oxmin = d_cx - d_w * 0.5;
......@@ -126,17 +136,21 @@ __global__ void BoxDecodeAndClipKernel(const T *anchor, const T *deltas,
T oxmax = d_cx + d_w * 0.5 - 1.;
T oymax = d_cy + d_h * 0.5 - 1.;
proposals[i * 4] = Max<T>(Min<T>(oxmin, im_info[1] - 1.), 0.);
proposals[i * 4 + 1] = Max<T>(Min<T>(oymin, im_info[0] - 1.), 0.);
proposals[i * 4 + 2] = Max<T>(Min<T>(oxmax, im_info[1] - 1.), 0.);
proposals[i * 4 + 3] = Max<T>(Min<T>(oymax, im_info[0] - 1.), 0.);
proposals[i * 4] = Max(Min(oxmin, im_info[1] - 1.), 0.);
proposals[i * 4 + 1] = Max(Min(oymin, im_info[0] - 1.), 0.);
proposals[i * 4 + 2] = Max(Min(oxmax, im_info[1] - 1.), 0.);
proposals[i * 4 + 3] = Max(Min(oymax, im_info[0] - 1.), 0.);
}
}
__device__ __forceinline__ T Min(T a, T b) const { return a > b ? b : a; }
__device__ __forceinline__ T Max(T a, T b) const { return a > b ? a : b; }
};
template <typename T, int BlockSize>
__global__ void FilterBBoxes(const T *bboxes, const T *im_info,
const T min_size, const int num, int *keep_num,
int *keep) {
static __global__ void FilterBBoxes(const T *bboxes, const T *im_info,
const T min_size, const int num,
int *keep_num, int *keep) {
T im_h = im_info[0];
T im_w = im_info[1];
T im_scale = im_info[2];
......@@ -181,7 +195,7 @@ __global__ void FilterBBoxes(const T *bboxes, const T *im_info,
}
}
__device__ inline float IoU(const float *a, const float *b) {
static __device__ inline float IoU(const float *a, const float *b) {
float left = max(a[0], b[0]), right = min(a[2], b[2]);
float top = max(a[1], b[1]), bottom = min(a[3], b[3]);
float width = max(right - left + 1, 0.f), height = max(bottom - top + 1, 0.f);
......@@ -191,8 +205,9 @@ __device__ inline float IoU(const float *a, const float *b) {
return inter_s / (s_a + s_b - inter_s);
}
__global__ void NMSKernel(const int n_boxes, const float nms_overlap_thresh,
const float *dev_boxes, uint64_t *dev_mask) {
static __global__ void NMSKernel(const int n_boxes,
const float nms_overlap_thresh,
const float *dev_boxes, uint64_t *dev_mask) {
const int row_start = blockIdx.y;
const int col_start = blockIdx.x;
......@@ -234,9 +249,9 @@ __global__ void NMSKernel(const int n_boxes, const float nms_overlap_thresh,
}
template <typename T>
void NMS(const platform::CUDADeviceContext &ctx, const Tensor &proposals,
const Tensor &sorted_indices, const T nms_threshold,
Tensor *keep_out) {
static void NMS(const platform::CUDADeviceContext &ctx, const Tensor &proposals,
const Tensor &sorted_indices, const T nms_threshold,
Tensor *keep_out) {
int boxes_num = proposals.dims()[0];
PADDLE_ENFORCE_EQ(boxes_num, sorted_indices.dims()[0]);
......@@ -247,13 +262,10 @@ void NMS(const platform::CUDADeviceContext &ctx, const Tensor &proposals,
const T *boxes = proposals.data<T>();
auto place = boost::get<platform::CUDAPlace>(ctx.GetPlace());
int size_bytes = boxes_num * col_blocks * sizeof(uint64_t);
uint64_t *d_mask =
reinterpret_cast<uint64_t *>(memory::Alloc(place, size_bytes));
NMSKernel<<<blocks, threads>>>(boxes_num, nms_threshold, boxes, d_mask);
uint64_t *h_mask = reinterpret_cast<uint64_t *>(
memory::Alloc(platform::CPUPlace(), size_bytes));
memory::Copy(platform::CPUPlace(), h_mask, place, d_mask, size_bytes, 0);
framework::Vector<uint64_t> mask(boxes_num * col_blocks);
NMSKernel<<<blocks, threads>>>(
boxes_num, nms_threshold, boxes,
mask.CUDAMutableData(boost::get<platform::CUDAPlace>(ctx.GetPlace())));
std::vector<uint64_t> remv(col_blocks);
memset(&remv[0], 0, sizeof(uint64_t) * col_blocks);
......@@ -267,7 +279,7 @@ void NMS(const platform::CUDADeviceContext &ctx, const Tensor &proposals,
if (!(remv[nblock] & (1ULL << inblock))) {
++num_to_keep;
keep_vec.push_back(i);
uint64_t *p = &h_mask[0] + i * col_blocks;
uint64_t *p = &mask[0] + i * col_blocks;
for (int j = nblock; j < col_blocks; j++) {
remv[j] |= p[j];
}
......@@ -276,12 +288,10 @@ void NMS(const platform::CUDADeviceContext &ctx, const Tensor &proposals,
int *keep = keep_out->mutable_data<int>({num_to_keep}, ctx.GetPlace());
memory::Copy(place, keep, platform::CPUPlace(), keep_vec.data(),
sizeof(int) * num_to_keep, 0);
memory::Free(place, d_mask);
memory::Free(platform::CPUPlace(), h_mask);
}
template <typename T>
std::pair<Tensor, Tensor> ProposalForOneImage(
static std::pair<Tensor, Tensor> ProposalForOneImage(
const platform::CUDADeviceContext &ctx, const Tensor &im_info,
const Tensor &anchors, const Tensor &variances,
const Tensor &bbox_deltas, // [M, 4]
......@@ -300,18 +310,20 @@ std::pair<Tensor, Tensor> ProposalForOneImage(
// 2. box decode and clipping
Tensor proposals;
proposals.mutable_data<T>({pre_nms_num, 4}, ctx.GetPlace());
int block = 512;
auto stream = ctx.stream();
BoxDecodeAndClipKernel<T><<<DIVUP(pre_nms_num, block), block, 0, stream>>>(
anchors.data<T>(), bbox_deltas.data<T>(), variances.data<T>(),
index_sort.data<int>(), im_info.data<T>(), pre_nms_num,
proposals.data<T>());
{
platform::ForRange<platform::CUDADeviceContext> for_range(ctx, pre_nms_num);
for_range(BoxDecodeAndClipFunctor<T>{
anchors.data<T>(), bbox_deltas.data<T>(), variances.data<T>(),
index_sort.data<int>(), im_info.data<T>(), proposals.data<T>()});
}
// 3. filter
Tensor keep_index, keep_num_t;
keep_index.mutable_data<int>({pre_nms_num}, ctx.GetPlace());
keep_num_t.mutable_data<int>({1}, ctx.GetPlace());
min_size = std::max(min_size, 1.0f);
auto stream = ctx.stream();
FilterBBoxes<T, 512><<<1, 512, 0, stream>>>(
proposals.data<T>(), im_info.data<T>(), min_size, pre_nms_num,
keep_num_t.data<int>(), keep_index.data<int>());
......@@ -355,8 +367,12 @@ class CUDAGenerateProposalsKernel : public framework::OpKernel<T> {
auto *scores = context.Input<Tensor>("Scores");
auto *bbox_deltas = context.Input<Tensor>("BboxDeltas");
auto *im_info = context.Input<Tensor>("ImInfo");
auto *anchors = context.Input<Tensor>("Anchors");
auto *variances = context.Input<Tensor>("Variances");
auto anchors = detail::Ref(context.Input<Tensor>("Anchors"),
"Cannot find input Anchors(%s) in scope",
context.Inputs("Anchors")[0]);
auto variances = detail::Ref(context.Input<Tensor>("Variances"),
"Cannot find input Variances(%s) in scope",
context.Inputs("Variances")[0]);
auto *rpn_rois = context.Output<LoDTensor>("RpnRois");
auto *rpn_roi_probs = context.Output<LoDTensor>("RpnRoiProbs");
......@@ -392,10 +408,8 @@ class CUDAGenerateProposalsKernel : public framework::OpKernel<T> {
trans(dev_ctx, *bbox_deltas, &bbox_deltas_swap, axis);
trans(dev_ctx, *scores, &scores_swap, axis);
Tensor *anchor = const_cast<framework::Tensor *>(anchors);
anchor->Resize({anchors->numel() / 4, 4});
Tensor *var = const_cast<framework::Tensor *>(variances);
var->Resize({var->numel() / 4, 4});
anchors.Resize({anchors.numel() / 4, 4});
variances.Resize({variances.numel() / 4, 4});
rpn_rois->mutable_data<T>({bbox_deltas->numel() / 4, 4},
context.GetPlace());
......@@ -417,12 +431,12 @@ class CUDAGenerateProposalsKernel : public framework::OpKernel<T> {
scores_slice.Resize({h_score * w_score * c_score, 1});
std::pair<Tensor, Tensor> box_score_pair =
ProposalForOneImage<T>(dev_ctx, im_info_slice, *anchor, *var,
ProposalForOneImage<T>(dev_ctx, im_info_slice, anchors, variances,
bbox_deltas_slice, scores_slice, pre_nms_top_n,
post_nms_top_n, nms_thresh, min_size, eta);
Tensor proposals = box_score_pair.first;
Tensor scores = box_score_pair.second;
Tensor &proposals = box_score_pair.first;
Tensor &scores = box_score_pair.second;
memory::Copy(place, rpn_rois_data + num_proposals * 4, place,
proposals.data<T>(), sizeof(T) * proposals.numel(), 0);
......
paddle/fluid/operators/detection/gpc.cc 0 → 100644
浏览文件 @ 049c9c7d
此差异已折叠。
paddle/fluid/operators/detection/gpc.h 0 → 100644
浏览文件 @ 049c9c7d
此差异已折叠。
paddle/fluid/operators/detection/poly_util.cc 0 → 100644
浏览文件 @ 049c9c7d
此差异已折叠。
paddle/fluid/operators/detection/poly_util.h 0 → 100644
浏览文件 @ 049c9c7d
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#ifndef POLY_UTIL_H_
#define POLY_UTIL_H_
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/detection/gpc.h"
namespace paddle {
namespace operators {
template <class T>
class Point_ {
public:
// default constructor
Point_() {}
Point_(T _x, T _y) {}
Point_(const Point_& pt) {}
Point_& operator=(const Point_& pt);
// conversion to another data type
// template<typename _T> operator Point_<_T>() const;
// conversion to the old-style C structures
// operator Vec<T, 2>() const;
// checks whether the point is inside the specified rectangle
// bool inside(const Rect_<T>& r) const;
T x; //!< x coordinate of the point
T y; //!< y coordinate of the point
};
template <class T>
void Array2PointVec(const T*& box, const size_t box_size,
std::vector<Point_<T>>& vec);
template <class T>
void Array2Poly(const T*& box, const size_t box_size, gpc::gpc_polygon& poly);
template <class T>
void PointVec2Poly(const std::vector<Point_<T>>& vec, gpc::gpc_polygon& poly);
template <class T>
void Poly2PointVec(const gpc::gpc_vertex_list& contour,
std::vector<Point_<T>>& vec);
template <class T>
T GetContourArea(std::vector<Point_<T>>& vec);
template <class T>
T PolyArea(const T* box, const size_t box_size, const bool normalized);
template <class T>
T PolyOverlapArea(const T* box1, const T* box2, const size_t box_size,
const bool normalized);
} // namespace operators
} // namespace paddle
#include "paddle/fluid/operators/detection/poly_util.cc"
#endif // POLY_UTIL_H_
paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.cc 0 → 100644
浏览文件 @ 049c9c7d
此差异已折叠。
paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.h 0 → 100644
浏览文件 @ 049c9c7d
此差异已折叠。
paddle/fluid/operators/gather.h
浏览文件 @ 049c9c7d
此差异已折叠。
paddle/fluid/operators/math/concat.h → paddle/fluid/operators/math/concat_and_split.h
浏览文件 @ 049c9c7d
......@@ -54,7 +54,7 @@ class ConcatFunctor {
* Output[1] = [[5,6]]
*/
template <typename DeviceContext, typename T>
class ConcatGradFunctor {
class SplitFunctor {
public:
void operator()(const DeviceContext& context, const framework::Tensor& input,
const std::vector<const framework::Tensor*>& ref_inputs,
......
paddle/fluid/operators/roi_align_op.cc 0 → 100644
浏览文件 @ 049c9c7d
此差异已折叠。
paddle/fluid/operators/roi_align_op.cu 0 → 100644
浏览文件 @ 049c9c7d
此差异已折叠。
paddle/fluid/operators/roi_align_op.h 0 → 100644
浏览文件 @ 049c9c7d
此差异已折叠。
paddle/fluid/operators/split_op.cc
浏览文件 @ 049c9c7d
此差异已折叠。
paddle/fluid/operators/split_op.h
浏览文件 @ 049c9c7d
此差异已折叠。
paddle/fluid/platform/profiler.cc
浏览文件 @ 049c9c7d
此差异已折叠。
paddle/fluid/platform/profiler.h
浏览文件 @ 049c9c7d
此差异已折叠。
python/paddle/fluid/__init__.py
浏览文件 @ 049c9c7d
此差异已折叠。
python/paddle/fluid/layer_helper.py
浏览文件 @ 049c9c7d
此差异已折叠。
python/paddle/fluid/layers/io.py
浏览文件 @ 049c9c7d
此差异已折叠。
python/paddle/fluid/layers/nn.py
浏览文件 @ 049c9c7d
此差异已折叠。
python/paddle/fluid/regularizer.py
浏览文件 @ 049c9c7d
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或