Skip to content

P
Paddle

BaiXuePrincess / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 5a389306 编写于 作者: S sneaxiy
浏览文件
操作

test=develop

上级 ac2eba44 9517a453
显示空白变更内容
内联 并排
Showing with 553 addition and 172 deletion
README.md
浏览文件 @ 5a389306
......@@ -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 @ 5a389306
......@@ -311,6 +311,8 @@ function(cc_test TARGET_NAME)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cpu_deterministic=true)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true)
# No unit test should exceed 10 minutes.
set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 600)
endif()
endfunction(cc_test)
......@@ -629,6 +631,8 @@ function(py_test TARGET_NAME)
PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
# No unit test should exceed 10 minutes.
set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 600)
endif()
endfunction()
......
paddle/fluid/API.spec
浏览文件 @ 5a389306
......@@ -61,12 +61,12 @@ paddle.fluid.layers.cos_sim ArgSpec(args=['X', 'Y'], varargs=None, keywords=None
paddle.fluid.layers.cross_entropy ArgSpec(args=['input', 'label', 'soft_label', 'ignore_index'], varargs=None, keywords=None, defaults=(False, -100))
paddle.fluid.layers.square_error_cost ArgSpec(args=['input', 'label'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.chunk_eval ArgSpec(args=['input', 'label', 'chunk_scheme', 'num_chunk_types', 'excluded_chunk_types'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None))
paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None, None))
paddle.fluid.layers.conv2d ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, None, None))
paddle.fluid.layers.conv3d ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, None, None))
paddle.fluid.layers.sequence_pool ArgSpec(args=['input', 'pool_type'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'param_attr', 'bias_attr', 'use_cudnn'], varargs=None, keywords=None, defaults=(None, None, False))
paddle.fluid.layers.softmax ArgSpec(args=['input', 'param_attr', 'bias_attr', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(None, None, True, None))
paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(False, None))
paddle.fluid.layers.softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(True, None))
paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None))
paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None))
paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False))
......@@ -97,8 +97,8 @@ paddle.fluid.layers.warpctc ArgSpec(args=['input', 'label', 'blank', 'norm_by_ti
paddle.fluid.layers.sequence_reshape ArgSpec(args=['input', 'new_dim'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.transpose ArgSpec(args=['x', 'perm', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.im2sequence ArgSpec(args=['input', 'filter_size', 'stride', 'padding', 'input_image_size', 'out_stride', 'name'], varargs=None, keywords=None, defaults=(1, 1, 0, None, 1, None))
paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples'], varargs=None, keywords=None, defaults=(None, None, None, None))
paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, None))
paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'name'], varargs=None, keywords=None, defaults=(0, None))
paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None)
......
paddle/fluid/framework/ir/CMakeLists.txt
浏览文件 @ 5a389306
......@@ -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 +37,10 @@ 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)
if(WITH_MKLDNN)
pass_library(mkldnn_placement_pass base)
pass_library(conv_relu_mkldnn_fuse_pass inference)
endif()
cc_library(fuse_elewise_add_act_pass SRCS fuse_elewise_add_act_pass.cc DEPS pass graph_pattern_detector )
......
paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc
浏览文件 @ 5a389306
......@@ -262,7 +262,7 @@ std::unique_ptr<ir::Graph> AttentionLSTMFusePass::ApplyImpl(
std::unordered_set<std::string> specified_vars({"data_lod_attention",
"cell_init", "hidden_init",
"data", "week", "minute"});
int count = 0;
size_t count = 0;
for (auto* node : graph->Nodes()) {
if (node->IsVar() && specified_vars.count(node->Name())) {
++count;
......
paddle/fluid/framework/ir/conv_bn_fuse_pass.cc
浏览文件 @ 5a389306
......@@ -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
// 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);
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});
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++;
}
};
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_relu_mkldnn_fuse_pass.cc
浏览文件 @ 5a389306
......@@ -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 @ 5a389306
......@@ -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;
}
......@@ -88,11 +101,17 @@ TEST(ConvReLUFusePass, basic) {
auto* op = node->Op();
ASSERT_TRUE(op->HasAttr("use_mkldnn"));
EXPECT_TRUE(boost::get<bool>(op->GetAttr("use_mkldnn")));
// 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"));
}
}
}
EXPECT_EQ(conv_relu_count, 1);
......
paddle/fluid/framework/ir/fuse_pass_base.cc 0 → 100644
浏览文件 @ 5a389306
// 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 @ 5a389306
......@@ -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_pattern_detector.cc
浏览文件 @ 5a389306
......@@ -259,6 +259,8 @@ GraphPatternDetector::DetectPatterns() {
return result;
}
// TODO(Superjomn) enhance the function as it marks unique unique as duplicates
// see https://github.com/PaddlePaddle/Paddle/issues/13550
void GraphPatternDetector::UniquePatterns(
std::vector<GraphPatternDetector::subgraph_t> *subgraphs) {
if (subgraphs->empty()) return;
......
paddle/fluid/framework/ir/mkldnn_placement_pass.cc 0 → 100644
浏览文件 @ 5a389306
/* 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 @ 5a389306
/* 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/op_desc.cc
浏览文件 @ 5a389306
......@@ -85,10 +85,6 @@ class CompileTimeInferShapeContext : public InferShapeContext {
VLOG(3) << "input " << in << " is not LodTensor";
return;
}
PADDLE_ENFORCE_EQ(in_var->GetType(), proto::VarType::LOD_TENSOR,
"The %d-th output of Output(%s) must be LoDTensor.", j,
out);
out_var->SetLoDLevel(in_var->GetLoDLevel());
}
......
paddle/fluid/framework/program_desc.cc
浏览文件 @ 5a389306
......@@ -126,7 +126,7 @@ const std::vector<std::string> ProgramDesc::GetFeedTargetNames() {
std::vector<std::string> feed_target_names;
for (auto *op : global_block.AllOps()) {
if (op->Type() == kFeedOpType) {
int col = boost::get<int>(op->GetAttr("col"));
size_t col = boost::get<int>(op->GetAttr("col"));
if (col >= feed_target_names.size()) {
feed_target_names.resize(col + 1);
}
......@@ -143,7 +143,7 @@ const std::vector<std::string> ProgramDesc::GetFetchTargetNames() {
std::vector<std::string> fetch_target_names;
for (auto *op : global_block.AllOps()) {
if (op->Type() == kFetchOpType) {
int col = boost::get<int>(op->GetAttr("col"));
size_t col = boost::get<int>(op->GetAttr("col"));
if (col >= fetch_target_names.size()) {
fetch_target_names.resize(col + 1);
}
......
paddle/fluid/framework/reader_test.cc
浏览文件 @ 5a389306
......@@ -39,7 +39,7 @@ TEST(READER, decorate_chain) {
{
auto endpoints = root->GetEndPoints();
ASSERT_EQ(endpoints.size(), 2U);
ASSERT_NE(endpoints.count(end_point1.get()), 0);
ASSERT_NE(endpoints.count(end_point1.get()), 0UL);
ASSERT_NE(endpoints.count(end_point2.get()), 0);
}
......
paddle/fluid/framework/selected_rows_test.cc
浏览文件 @ 5a389306
......@@ -91,7 +91,7 @@ TEST(SelectedRows, SparseTable) {
ASSERT_TRUE(table.HasKey(10));
ASSERT_TRUE(table.HasKey(8));
ASSERT_TRUE(table.HasKey(6));
ASSERT_EQ(table.rows().size(), 3);
ASSERT_EQ(table.rows().size(), 3UL);
framework::Tensor ids;
ids.Resize(framework::make_ddim({4}));
......
paddle/fluid/inference/analysis/analyzer.cc
浏览文件 @ 5a389306
......@@ -101,7 +101,11 @@ Analyzer::Analyzer() { Register("manager1", new DfgPassManagerImpl); }
void Analyzer::Run(Argument* argument) {
std::vector<std::string> passes;
for (auto& pass : all_ir_passes_) {
if (use_mkldnn_) {
VLOG(3) << "Adding MKL-DNN placement pass";
passes.push_back("mkldnn_placement_pass");
}
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 +121,26 @@ 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 @ 5a389306
......@@ -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);
......@@ -81,6 +84,9 @@ class Analyzer : public OrderedRegistry<PassManager> {
}};
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 @ 5a389306
......@@ -225,10 +225,24 @@ 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 @ 5a389306
......@@ -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 @ 5a389306
......@@ -52,9 +52,10 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
}
// Easy for profiling independently.
TEST(Analyzer_resnet50, profile) {
void profile(bool use_mkldnn = false) {
AnalysisConfig cfg;
SetConfig(&cfg);
cfg._use_mkldnn = use_mkldnn;
std::vector<PaddleTensor> outputs;
std::vector<std::vector<PaddleTensor>> input_slots_all;
......@@ -69,6 +70,11 @@ TEST(Analyzer_resnet50, profile) {
}
}
TEST(Analyzer_resnet50, profile) { profile(); }
#ifndef PADDLE_WITH_MKLDNN
TEST(Analyzer_resnet50, profile_mkldnn) { profile(true /* use_mkldnn */); }
#endif
// Check the fuse status
TEST(Analyzer_resnet50, fuse_statis) {
AnalysisConfig cfg;
......@@ -82,15 +88,21 @@ TEST(Analyzer_resnet50, fuse_statis) {
}
// Compare result of NativeConfig and AnalysisConfig
TEST(Analyzer_resnet50, compare) {
void compare(bool use_mkldnn = false) {
AnalysisConfig cfg;
SetConfig(&cfg);
cfg._use_mkldnn = use_mkldnn;
std::vector<std::vector<PaddleTensor>> input_slots_all;
SetInput(&input_slots_all);
CompareNativeAndAnalysis(cfg, input_slots_all);
}
TEST(Analyzer_resnet50, compare) { compare(); }
#ifdef PADDLE_WITH_MKLDNN
TEST(Analyzer_resnet50, compare_mkldnn) { compare(true /* use_mkldnn */); }
#endif
} // namespace analysis
} // namespace inference
} // namespace paddle
paddle/fluid/inference/tests/api/analyzer_vis_tester.cc
浏览文件 @ 5a389306
......@@ -59,9 +59,6 @@ void SetConfig(AnalysisConfig *cfg) {
cfg->specify_input_name = true;
// TODO(TJ): fix fusion gru
cfg->ir_passes.push_back("fc_gru_fuse_pass");
#ifdef PADDLE_WITH_MKLDNN
cfg->_use_mkldnn = true;
#endif
}
void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
......@@ -84,9 +81,10 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
// Easy for profiling independently.
// ocr, mobilenet and se_resnext50
TEST(Analyzer_vis, profile) {
void profile(bool use_mkldnn = false) {
AnalysisConfig cfg;
SetConfig(&cfg);
cfg._use_mkldnn = use_mkldnn;
std::vector<PaddleTensor> outputs;
std::vector<std::vector<PaddleTensor>> input_slots_all;
......@@ -108,6 +106,12 @@ TEST(Analyzer_vis, profile) {
}
}
TEST(Analyzer_vis, profile) { profile(); }
#ifdef PADDLE_WITH_MKLDNN
TEST(Analyzer_vis, profile_mkldnn) { profile(true /* use_mkldnn */); }
#endif
// Check the fuse status
TEST(Analyzer_vis, fuse_statis) {
AnalysisConfig cfg;
......@@ -118,15 +122,21 @@ TEST(Analyzer_vis, fuse_statis) {
}
// Compare result of NativeConfig and AnalysisConfig
TEST(Analyzer_vis, compare) {
void compare(bool use_mkldnn = false) {
AnalysisConfig cfg;
SetConfig(&cfg);
cfg._use_mkldnn = use_mkldnn;
std::vector<std::vector<PaddleTensor>> input_slots_all;
SetInput(&input_slots_all);
CompareNativeAndAnalysis(cfg, input_slots_all);
}
TEST(Analyzer_vis, compare) { compare(); }
#ifdef PADDLE_WITH_MKLDNN
TEST(Analyzer_vis, compare_mkldnn) { compare(true /* use_mkldnn */); }
#endif
} // namespace analysis
} // namespace inference
} // namespace paddle
paddle/fluid/inference/tests/api/tester_helper.h
浏览文件 @ 5a389306
......@@ -163,7 +163,8 @@ void TestPrediction(const AnalysisConfig &config,
const std::vector<std::vector<PaddleTensor>> &inputs,
std::vector<PaddleTensor> *outputs, int num_threads,
bool use_analysis = FLAGS_use_analysis) {
LOG(INFO) << "use_analysis: " << use_analysis;
LOG(INFO) << "use_analysis: " << use_analysis
<< ", use_mkldnn: " << config._use_mkldnn;
if (num_threads == 1) {
TestOneThreadPrediction(config, inputs, outputs, use_analysis);
} else {
......@@ -175,6 +176,7 @@ void TestPrediction(const AnalysisConfig &config,
void CompareNativeAndAnalysis(
const AnalysisConfig &config,
const std::vector<std::vector<PaddleTensor>> &inputs) {
LOG(INFO) << "use_mkldnn: " << config._use_mkldnn;
std::vector<PaddleTensor> native_outputs, analysis_outputs;
TestOneThreadPrediction(config, inputs, &native_outputs, false);
TestOneThreadPrediction(config, inputs, &analysis_outputs, true);
......
paddle/fluid/operators/reader/reader_blocking_queue_test.cc
浏览文件 @ 5a389306
......@@ -229,7 +229,7 @@ TEST(BlockingQueue, speed_test_mode) {
q1.Receive(&b);
EXPECT_EQ(b, i);
}
EXPECT_EQ(q1.Size(), 0);
EXPECT_EQ(q1.Size(), 0UL);
BlockingQueue<size_t> q2(queue_size, true);
for (size_t i = 0; i < queue_size; ++i) {
......
paddle/fluid/operators/sequence_unpad_op.cc
浏览文件 @ 5a389306
......@@ -50,7 +50,7 @@ class SequenceUnpadOp : public framework::OperatorWithKernel {
if (x_dims.size() == 2) {
out_dims_vec.push_back(1);
} else {
for (size_t i = 2; i < x_dims.size(); ++i) {
for (int i = 2; i < x_dims.size(); ++i) {
out_dims_vec.push_back(x_dims[i]);
}
}
......
paddle/fluid/operators/sequence_unpad_op.h
浏览文件 @ 5a389306
......@@ -61,7 +61,7 @@ class SequenceUnpadOpKernel : public framework::OpKernel<T> {
if (x_t->dims().size() == 2) {
out_dims_vec.push_back(1);
} else {
for (size_t i = 2; i < x_t->dims().size(); ++i) {
for (int i = 2; i < x_t->dims().size(); ++i) {
out_dims_vec.push_back(x_t->dims()[i]);
}
}
......
python/paddle/fluid/layers/nn.py
浏览文件 @ 5a389306
......@@ -356,7 +356,6 @@ def dynamic_lstm(input,
c_0(Variable): The initial cell state is an optional input, default is zero.
This is a tensor with shape (N x D), where N is the
batch size. `h_0` and `c_0` can be NULL but only at the same time.
param_attr(ParamAttr|None): The parameter attribute for the learnable
hidden-hidden weights.
......@@ -364,6 +363,11 @@ def dynamic_lstm(input,
W_{fh}, W_{oh}`}
- The shape is (D x 4D), where D is the hidden
size.
If it is set to None or one attribute of ParamAttr,
dynamic_lstm will create ParamAttr as param_attr.
If the Initializer of the param_attr is not set, the
parameter is initialized with Xavier. Default: None.
bias_attr (ParamAttr|None): The bias attribute for the learnable bias
weights, which contains two parts, input-hidden
bias weights and peephole connections weights if
......@@ -376,6 +380,11 @@ def dynamic_lstm(input,
- Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
W_{fc}, W_{oc}`}.
- The shape is (1 x 7D).
If it is set to None or one attribute of ParamAttr,
dynamic_lstm will create ParamAttr as bias_attr.
If the Initializer of the bias_attr is not set,
the bias is initialized zero. Default: None.
use_peepholes (bool): ${use_peepholes_comment}
is_reverse (bool): ${is_reverse_comment}
gate_activation (str): ${gate_activation_comment}
......@@ -394,11 +403,11 @@ def dynamic_lstm(input,
hidden_dim = 512
forward_proj = fluid.layers.fc(input=input_seq, size=hidden_dim * 4,
act=None, bias_attr=None)
bias_attr=False)
forward, _ = fluid.layers.dynamic_lstm(
input=forward_proj, size=hidden_dim * 4, use_peepholes=False)
"""
assert bias_attr is not False, "bias_attr should not be False in dynamic_lstmp."
helper = LayerHelper('lstm', **locals())
size = size // 4
weight = helper.create_parameter(
......@@ -533,6 +542,11 @@ def dynamic_lstmp(input,
size.
- Projection weight = {:math:`W_{rh}`}.
- The shape of projection weight is (D x P).
If it is set to None or one attribute of ParamAttr,
dynamic_lstm will create ParamAttr as param_attr.
If the Initializer of the param_attr is not set, the
parameter is initialized with Xavier. Default: None.
bias_attr(ParamAttr|None): The bias attribute for the learnable bias
weights, which contains two parts, input-hidden
bias weights and peephole connections weights if
......@@ -545,6 +559,11 @@ def dynamic_lstmp(input,
- Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
W_{fc}, W_{oc}`}.
- The shape is (1 x 7D).
If it is set to None or one attribute of ParamAttr,
dynamic_lstm will create ParamAttr as bias_attr.
If the Initializer of the bias_attr is not set,
the bias is initialized zero. Default: None.
use_peepholes(bool): Whether to enable diagonal/peephole connections,
default `True`.
is_reverse(bool): Whether to compute reversed LSTM, default `False`.
......@@ -589,6 +608,7 @@ def dynamic_lstmp(input,
proj_activation="tanh")
"""
assert bias_attr is not False, "bias_attr should not be False in dynamic_lstmp."
helper = LayerHelper('lstmp', **locals())
size = size // 4
weight = helper.create_parameter(
......@@ -1270,7 +1290,8 @@ def sequence_conv(input,
padding=None,
bias_attr=None,
param_attr=None,
act=None):
act=None,
name=None):
"""
This function creates the op for sequence_conv, using the inputs and
other convolutional configurations for the filters and stride as given
......@@ -1282,9 +1303,19 @@ def sequence_conv(input,
filter_size (int): the filter size (H and W).
filter_stride (int): stride of the filter.
padding (bool): if True, add paddings.
bias_attr (ParamAttr|None): attributes for bias
param_attr (ParamAttr|None): attributes for parameter
act (str): the activation type
bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of sequence_conv.
If it is set to False, no bias will be added to the output units.
If it is set to None or one attribute of ParamAttr, sequence_conv
will create ParamAttr as bias_attr. If the Initializer of the bias_attr
is not set, the bias is initialized zero. Default: None.
param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
of sequence_conv. If it is set to None or one attribute of ParamAttr, sequence_conv
will create ParamAttr as param_attr. If the Initializer of the param_attr
is not set, the parameter is initialized with Xavier. Default: None.
act (str): Activation type, if it is set to None, activation is not appended.
Default: None.
name (str|None): A name for this layer(optional). If set None, the layer
will be named automatically. Default: None.
Returns:
Variable: output of sequence_conv
......@@ -1313,7 +1344,7 @@ def sequence_conv(input,
return helper.append_activation(pre_act)
def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=False):
def sequence_softmax(input, use_cudnn=False, name=None):
"""
This function computes the softmax activation among all time-steps for each
sequence. The dimension of each time-step should be 1. Thus, the shape of
......@@ -1333,10 +1364,10 @@ def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=False):
Args:
input (Variable): The input variable which is a LoDTensor.
bias_attr (ParamAttr|None): attributes for bias
param_attr (ParamAttr|None): attributes for parameter
use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn \
library is installed. Default: False
library is installed. Default: False.
name (str|None): A name for this layer(optional). If set None, the layer
will be named automatically. Default: None.
Returns:
Variable: output of sequence_softmax
......@@ -1360,7 +1391,7 @@ def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=False):
return softmax_out
def softmax(input, param_attr=None, bias_attr=None, use_cudnn=True, name=None):
def softmax(input, use_cudnn=True, name=None):
"""
The input of the softmax operator is a tensor of any rank. The output tensor
has the same shape as the input.
......@@ -1387,10 +1418,10 @@ def softmax(input, param_attr=None, bias_attr=None, use_cudnn=True, name=None):
Args:
input (Variable): The input variable.
bias_attr (ParamAttr): attributes for bias
param_attr (ParamAttr): attributes for parameter
use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn \
library is installed.
name (str|None): A name for this layer(optional). If set None, the layer
will be named automatically. Default: None.
Returns:
Variable: output of softmax
......@@ -1496,14 +1527,23 @@ def conv2d(input,
convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
the first half of the filters is only connected to the first half
of the input channels, while the second half of the filters is only
connected to the second half of the input channels. Default: groups=1
param_attr (ParamAttr): The parameters to the Conv2d Layer. Default: None
bias_attr (ParamAttr): Bias parameter for the Conv2d layer. Default: None
connected to the second half of the input channels. Default: groups=1.
param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
of conv2d. If it is set to None or one attribute of ParamAttr, conv2d
will create ParamAttr as param_attr. If the Initializer of the param_attr
is not set, the parameter is initialized with :math:`Normal(0.0, std)`,
and the :math:`std` is :math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`. Default: None.
bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv2d.
If it is set to False, no bias will be added to the output units.
If it is set to None or one attribute of ParamAttr, conv2d
will create ParamAttr as bias_attr. If the Initializer of the bias_attr
is not set, the bias is initialized zero. Default: None.
use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn
library is installed. Default: True
act (str): Activation type. Default: None
act (str): Activation type, if it is set to None, activation is not appended.
Default: None
name (str|None): A name for this layer(optional). If set None, the layer
will be named automatically.
will be named automatically. Default: None
Returns:
Variable: The tensor variable storing the convolution and \
......@@ -1521,7 +1561,7 @@ def conv2d(input,
"""
num_channels = input.shape[1]
assert param_attr is not False, "param_attr should not be False here."
l_type = 'conv2d'
if (num_channels == groups and num_filters % num_channels == 0 and
not use_cudnn):
......@@ -1549,7 +1589,8 @@ def conv2d(input,
filter_shape = [num_filters, int(num_filter_channels)] + filter_size
def _get_default_param_initializer():
std = (2.0 / (filter_size[0]**2 * num_channels))**0.5
filter_elem_num = filter_size[0] * filter_size[1] * num_channels
std = (2.0 / filter_elem_num)**0.5
return Normal(0.0, std, 0)
filter_param = helper.create_parameter(
......@@ -1660,13 +1701,22 @@ def conv3d(input,
the first half of the filters is only connected to the first half
of the input channels, while the second half of the filters is only
connected to the second half of the input channels. Default: groups=1
param_attr (ParamAttr): The parameters to the Conv3d Layer. Default: None
bias_attr (ParamAttr): Bias parameter for the Conv3d layer. Default: None
param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
of conv3d. If it is set to None or one attribute of ParamAttr, conv3d
will create ParamAttr as param_attr. If it is set to None, the parameter
is initialized with :math:`Normal(0.0, std)`, and the :math:`std` is
:math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`. Default: None.
bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv3d.
If it is set to False, no bias will be added to the output units.
If it is set to None or one attribute of ParamAttr, conv3d
will create ParamAttr as bias_attr. If the Initializer of the bias_attr
is not set, the bias is initialized zero. Default: None.
use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn
library is installed. Default: True
act (str): Activation type. Default: None
act (str): Activation type, if it is set to None, activation is not appended.
Default: None.
name (str|None): A name for this layer(optional). If set None, the layer
will be named automatically.
will be named automatically. Default: None.
Returns:
Variable: The tensor variable storing the convolution and \
......@@ -1684,7 +1734,7 @@ def conv3d(input,
"""
l_type = 'conv3d'
assert param_attr is not False, "param_attr should not be False here."
helper = LayerHelper(l_type, **locals())
dtype = helper.input_dtype()
......@@ -1709,7 +1759,9 @@ def conv3d(input,
filter_shape = [num_filters, num_filter_channels] + filter_size
def _get_default_param_initializer():
std = (2.0 / (filter_size[0]**3 * num_channels))**0.5
filter_elem_num = filter_size[0] * filter_size[1] * filter_size[
2] * num_channels
std = (2.0 / filter_elem_num)**0.5
return Normal(0.0, std, 0)
filter_param = helper.create_parameter(
......@@ -2181,8 +2233,14 @@ def batch_norm(input,
is_test(bool, Default False): Used for training or training.
momentum(float, Default 0.9):
epsilon(float, Default 1e-05):
param_attr(ParamAttr): The parameter attribute for Parameter `scale`.
bias_attr(ParamAttr): The parameter attribute for Parameter `bias`.
param_attr(ParamAttr|None): The parameter attribute for Parameter `scale`
of batch_norm. If it is set to None or one attribute of ParamAttr, batch_norm
will create ParamAttr as param_attr. If the Initializer of the param_attr
is not set, the parameter is initialized with Xavier. Default: None.
bias_attr(ParamAttr|None): The parameter attribute for the bias of batch_norm.
If it is set to None or one attribute of ParamAttr, batch_norm
will create ParamAttr as bias_attr. If the Initializer of the bias_attr
is not set, the bias is initialized zero. Default: None.
data_layout(string, default NCHW): NCHW|NHWC
in_place(bool, Default False): Make the input and output of batch norm reuse memory.
name(string, Default None): A name for this layer(optional). If set None, the layer
......@@ -2202,6 +2260,7 @@ def batch_norm(input,
hidden1 = fluid.layers.fc(input=x, size=200, param_attr='fc1.w')
hidden2 = fluid.layers.batch_norm(input=hidden1)
"""
assert bias_attr is not False, "bias_attr should not be False in batch_norm."
helper = LayerHelper('batch_norm', **locals())
dtype = helper.input_dtype()
......@@ -2480,15 +2539,22 @@ def conv2d_transpose(input,
when group=2, the first half of the filters is only connected to the
first half of the input channels, while the second half of the
filters is only connected to the second half of the input channels.
Default: groups=1
param_attr(ParamAttr): The parameters to the Conv2d_transpose Layer.
Default: None
bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None
Default: groups = 1.
param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
of conv2d_transpose. If it is set to None or one attribute of ParamAttr, conv2d_transpose
will create ParamAttr as param_attr. If the Initializer of the param_attr
is not set, the parameter is initialized with Xavier. Default: None.
bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv2d_transpose.
If it is set to False, no bias will be added to the output units.
If it is set to None or one attribute of ParamAttr, conv2d_transpose
will create ParamAttr as bias_attr. If the Initializer of the bias_attr
is not set, the bias is initialized zero. Default: None.
use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
library is installed. Default: True
act(str): Activation type. Default: None
library is installed. Default: True.
act (str): Activation type, if it is set to None, activation is not appended.
Default: None.
name(str|None): A name for this layer(optional). If set None, the layer
will be named automatically.
will be named automatically. Default: True.
Returns:
Variable: The tensor variable storing the convolution transpose result.
......@@ -2503,7 +2569,7 @@ def conv2d_transpose(input,
data = fluid.layers.data(name='data', shape=[3, 32, 32], dtype='float32')
conv2d_transpose = fluid.layers.conv2d_transpose(input=data, num_filters=2, filter_size=3)
"""
assert param_attr is not False, "param_attr should not be False in conv2d_transpose."
input_channel = input.shape[1]
op_type = 'conv2d_transpose'
......@@ -2539,6 +2605,7 @@ def conv2d_transpose(input,
else:
filter_size = utils.convert_to_list(filter_size, 2,
'conv2d_transpose.filter_size')
if output_size is None:
output_size = []
elif isinstance(output_size, list) or isinstance(output_size, int):
......@@ -2548,6 +2615,7 @@ def conv2d_transpose(input,
padding = utils.convert_to_list(padding, 2, 'padding')
groups = 1 if groups is None else groups
filter_shape = [input_channel, num_filters // groups] + filter_size
img_filter = helper.create_parameter(
dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)
......@@ -2660,12 +2728,19 @@ def conv3d_transpose(input,
first half of the input channels, while the second half of the
filters is only connected to the second half of the input channels.
Default: groups=1
param_attr(ParamAttr): The parameters to the Conv3d_transpose Layer.
Default: None
bias_attr(ParamAttr): Bias parameter for the Conv3d layer. Default: None
param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
of conv3d_transpose. If it is set to None or one attribute of ParamAttr, conv3d_transpose
will create ParamAttr as param_attr. If the Initializer of the param_attr
is not set, the parameter is initialized with Xavier. Default: None.
bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv3d_transpose.
If it is set to False, no bias will be added to the output units.
If it is set to None or one attribute of ParamAttr, conv3d_transpose
will create ParamAttr as bias_attr. If the Initializer of the bias_attr
is not set, the bias is initialized zero. Default: None.
use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
library is installed. Default: True
act(str): Activation type. Default: None
act (str): Activation type, if it is set to None, activation is not appended.
Default: None.
name(str|None): A name for this layer(optional). If set None, the layer
will be named automatically.
......@@ -2682,6 +2757,7 @@ def conv3d_transpose(input,
data = fluid.layers.data(name='data', shape=[3, 12, 32, 32], dtype='float32')
conv3d_transpose = fluid.layers.conv3d_transpose(input=data, num_filters=2, filter_size=3)
"""
assert param_attr is not False, "param_attr should not be False in conv3d_transpose."
l_type = "conv3d_transpose"
helper = LayerHelper(l_type, **locals())
if not isinstance(input, Variable):
......@@ -3200,10 +3276,18 @@ def lstm_unit(x_t,
cell_t_prev (Variable): The cell value of lstm unit, a 2-D tensor with
shape M x S, M for batch size and S for size of lstm unit.
forget_bias (float): The forget bias of lstm unit.
param_attr (ParamAttr): The attributes of parameter weights, used to set
initializer, name etc.
bias_attr (ParamAttr): The attributes of bias weights, if not False,
bias weights will be created and be set to default value.
param_attr(ParamAttr|None): The parameter attribute for the learnable
hidden-hidden weights.
If it is set to None or one attribute of ParamAttr,
lstm_unit will create ParamAttr as param_attr.
If the Initializer of the param_attr is not set, the
parameter is initialized with Xavier. Default: None.
bias_attr (ParamAttr|None): The bias attribute for the learnable bias
weights. If it is set to False, no bias will be added
to the output units. If it is set to None or one attribute of ParamAttr,
lstm_unit will create ParamAttr as bias_attr.
If the Initializer of the bias_attr is not set,
the bias is initialized zero. Default: None.
name(str|None): A name for this layer(optional). If set None, the layer
will be named automatically.
......@@ -4117,7 +4201,8 @@ def nce(input,
sample_weight=None,
param_attr=None,
bias_attr=None,
num_neg_samples=None):
num_neg_samples=None,
name=None):
"""
${comment}
......@@ -4128,9 +4213,18 @@ def nce(input,
sample_weight (Variable|None): A Variable of shape [batch_size, 1]
storing a weight for each sample. The default weight for each
sample is 1.0.
param_attr (ParamAttr|None): attributes for parameter
bias_attr (ParamAttr|None): attributes for bias
param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
of nce. If it is set to None or one attribute of ParamAttr, nce
will create ParamAttr as param_attr. If the Initializer of the param_attr
is not set, the parameter is initialized with Xavier. Default: None.
bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of nce.
If it is set to False, no bias will be added to the output units.
If it is set to None or one attribute of ParamAttr, nce
will create ParamAttr as bias_attr. If the Initializer of the bias_attr
is not set, the bias is initialized zero. Default: None.
num_neg_samples (int): ${num_neg_samples_comment}
name (str|None): A name for this layer(optional). If set None, the layer
will be named automatically. Default: None.
Returns:
Variable: The output nce loss.
......@@ -4163,19 +4257,28 @@ def nce(input,
"""
helper = LayerHelper('nce', **locals())
assert isinstance(input, Variable)
dim = input.shape[1]
assert isinstance(label, Variable)
dim = input.shape[1]
num_true_class = label.shape[1]
w = helper.create_parameter(
attr=helper.param_attr,
shape=[num_total_classes, dim],
is_bias=False,
dtype=input.dtype)
inputs = {
'Input': input,
'Label': label,
'Weight': w,
'SampleWeight': sample_weight if sample_weight is not None else []
}
if helper.bias_attr:
b = helper.create_parameter(
attr=helper.bias_attr,
shape=[num_total_classes, 1],
is_bias=True,
dtype=input.dtype)
inputs['Bias'] = b
cost = helper.create_tmp_variable(dtype=input.dtype)
sample_logits = helper.create_tmp_variable(dtype=input.dtype)
sample_labels = helper.create_tmp_variable(dtype=label.dtype)
......@@ -4192,13 +4295,7 @@ def nce(input,
helper.append_op(
type='nce',
inputs={
'Input': input,
'Label': label,
'Weight': w,
'Bias': b,
'SampleWeight': sample_weight if sample_weight is not None else []
},
inputs=inputs,
outputs={
'Cost': cost,
'SampleLogits': sample_logits,
......@@ -4208,7 +4305,12 @@ def nce(input,
return cost / (num_neg_samples + 1)
def hsigmoid(input, label, num_classes, param_attr=None, bias_attr=None):
def hsigmoid(input,
label,
num_classes,
param_attr=None,
bias_attr=None,
name=None):
"""
The hierarchical sigmoid operator is used to accelerate the training
process of language model. This operator organizes the classes into a
......@@ -4229,11 +4331,17 @@ def hsigmoid(input, label, num_classes, param_attr=None, bias_attr=None):
label (Variable): The tensor variable contains labels of training data.
It's a tensor with shape is :math:`[N \\times 1]`.
num_classes: (int), The number of classes, must not be less than 2.
param_attr (ParamAttr|list of ParamAttr, default None): The parameter
attribute for learnable parameters/weights of this layer.
bias_attr (ParamAttr|list of ParamAttr, default None): The parameter
attribute for the bias of this layer. If it is set to False, no
bias will be applied.
param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
of hsigmoid. If it is set to None or one attribute of ParamAttr, hsigmoid
will create ParamAttr as param_attr. If the Initializer of the param_attr
is not set, the parameter is initialized with Xavier. Default: None.
bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of hsigmoid.
If it is set to False, no bias will be added to the output units.
If it is set to None or one attribute of ParamAttr, hsigmoid
will create ParamAttr as bias_attr. If the Initializer of the bias_attr
is not set, the bias is initialized zero. Default: None.
name (str|None): A name for this layer(optional). If set None, the layer
will be named automatically. Default: None.
Returns:
Out: (Tensor) The cost of hierarchical sigmoid operator. the shape is [N, 1]
......@@ -7327,6 +7435,7 @@ def affine_channel(x, scale=None, bias=None, data_layout='NCHW', name=None):
data_layout (string, default NCHW): NCHW or NHWC. If input is 2D
tensor, you can ignore data_layout.
name (str, default None): The name of this layer.
Returns:
out (Variable): A tensor of the same shape and data layout with x.
"""
......
python/paddle/fluid/nets.py
浏览文件 @ 5a389306
......@@ -64,23 +64,33 @@ def simple_img_conv_pool(input,
average-pooling. Default :math:`max`.
global_pooling (bool): Whether to use the global pooling. If global_pooling = true,
pool_size and pool_padding while be ignored. Default False
conv_stride (int|list|tuple): The stride size of the Conv2d Layer. If stride is a
conv_stride (int|list|tuple): The stride size of the conv2d Layer. If stride is a
list or tuple, it must contain two integers, (conv_stride_H, conv_stride_W). Otherwise,
the conv_stride_H = conv_stride_W = conv_stride. Default: conv_stride = 1.
conv_padding (int|list|tuple): The padding size of the Conv2d Layer. If padding is
conv_padding (int|list|tuple): The padding size of the conv2d Layer. If padding is
a list or tuple, it must contain two integers, (conv_padding_H, conv_padding_W).
Otherwise, the conv_padding_H = conv_padding_W = conv_padding. Default: conv_padding = 0.
conv_dilation (int|list|tuple): The dilation size of the Conv2d Layer. If dilation is
conv_dilation (int|list|tuple): The dilation size of the conv2d Layer. If dilation is
a list or tuple, it must contain two integers, (conv_dilation_H, conv_dilation_W).
Otherwise, the conv_dilation_H = conv_dilation_W = conv_dilation. Default: conv_dilation = 1.
conv_groups (int): The groups number of the Conv2d Layer. According to grouped
conv_groups (int): The groups number of the conv2d Layer. According to grouped
convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
the first half of the filters is only connected to the first half
of the input channels, while the second half of the filters is only
connected to the second half of the input channels. Default: groups=1
param_attr (ParamAttr): The parameters to the Conv2d Layer. Default: None
bias_attr (ParamAttr): Bias parameter for the Conv2d layer. Default: None
act (str): Activation type for Conv2d. Default: None
connected to the second half of the input channels. Default: groups=1.
param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
of conv2d. If it is set to None or one attribute of ParamAttr, conv2d
will create ParamAttr as param_attr. If the Initializer of the param_attr
is not set, the parameter is initialized with :math:`Normal(0.0, std)`,
and the :math:`std` is :math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`.
Default: None.
bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv2d.
If it is set to False, no bias will be added to the output units.
If it is set to None or one attribute of ParamAttr, conv2d
will create ParamAttr as bias_attr. If the Initializer of the bias_attr
is not set, the bias is initialized zero. Default: None.
act (str): Activation type for conv2d, if it is set to None, activation is not
appended. Default: None.
use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn
library is installed. Default: True
......
python/paddle/fluid/regularizer.py
浏览文件 @ 5a389306
......@@ -237,6 +237,7 @@ class L1DecayRegularizer(WeightDecayRegularizer):
'Ids': idx},
outputs={'Out': decay},
attrs={'is_sparse': True})
param = decay
# Append sign op
block.append_op(
......
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