m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

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

Browse Source
revert-13637-optimize-opyreader
Dang Qingqing 7 years ago
parent e3c7348f52 225ecee5ea
commit a121c89830
79 changed files with 2024 additions and 978 deletions
Show Stats Download Patch File Download Diff File

2
Dockerfile
Unescape View File

@ -53,7 +53,7 @@ RUN curl -s -q https://glide.sh/get | sh
# and its size is only one-third of the official one.
# 2. Manually add ~IPluginFactory() in IPluginFactory class of NvInfer.h, otherwise, it couldn't work in paddle.
# See https://github.com/PaddlePaddle/Paddle/issues/10129 for details.
RUN wget -qO- http://paddlepaddledeps.bj.bcebos.com/TensorRT-4.0.0.3.Ubuntu-16.04.4.x86_64-gnu.cuda-8.0.cudnn7.0.tar.gz | \
RUN wget -qO- http://paddlepaddledeps.cdn.bcebos.com/TensorRT-4.0.0.3.Ubuntu-16.04.4.x86_64-gnu.cuda-8.0.cudnn7.0.tar.gz | \
tar -xz -C /usr/local && \
cp -rf /usr/local/TensorRT/include /usr && \
cp -rf /usr/local/TensorRT/lib /usr

16
cmake/inference_lib.cmake
Unescape View File

@ -128,16 +128,13 @@ set(src_dir "${PADDLE_SOURCE_DIR}/paddle/fluid")
set(dst_dir "${FLUID_INSTALL_DIR}/paddle/fluid")
set(module "framework")
if (NOT WIN32)
copy(framework_lib DEPS framework_py_proto
SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/details/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/framework/framework.pb.h
DSTS ${dst_dir}/${module} ${dst_dir}/${module}/details ${dst_dir}/${module}
)
else()
copy(framework_lib
set(framework_lib_deps framework_py_proto)
endif(NOT WIN32)
copy(framework_lib DEPS ${framework_lib_deps}
SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/details/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/framework/framework.pb.h
DSTS ${dst_dir}/${module} ${dst_dir}/${module}/details ${dst_dir}/${module}
${src_dir}/${module}/ir/*.h
DSTS ${dst_dir}/${module} ${dst_dir}/${module}/details ${dst_dir}/${module} ${dst_dir}/${module}/ir
)
endif(NOT WIN32)
set(module "memory")
copy(memory_lib
@ -161,7 +158,8 @@ set(module "inference")
copy(inference_lib DEPS ${inference_deps}
SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
${src_dir}/${module}/api/paddle_inference_api.h ${src_dir}/${module}/api/demo_ci
DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/paddle_inference_pass.h
DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
)
set(module "platform")

4
doc/fluid/new_docs/beginners_guide/basics/machine_translation/README.cn.md
Unescape View File

@ -60,6 +60,7 @@
图3. 编码器-解码器框架
</div>
<a name="编码器"></a>
#### 编码器
编码阶段分为三步：
@ -81,7 +82,7 @@
机器翻译任务的训练过程中，解码阶段的目标是最大化下一个正确的目标语言词的概率。思路是：
1. 每一个时刻，根据源语言句子的编码信息（又叫上下文向量，context vector）`$c$`、真实目标语言序列的第`$i$`个词`$u_i$`和`$i$`时刻RNN的隐层状态`$z_i$`，计算出下一个隐层状态`$z_{i+1}$`。计算公式如下：
$$z_{i+1}=\phi_{\theta '} \left ( c,u_i,z_i \right )$$
其中`$\phi _{\theta '}$`是一个非线性激活函数；`$c=q\mathbf{h}$`是源语言句子的上下文向量，在不使用[注意力机制](#注意力机制)时，如果[编码器](#编码器)的输出是源语言句子编码后的最后一个元素，则可以定义`$c=h_T$`；`$u_i$`是目标语言序列的第`$i$`个单词，`$u_0$`是目标语言序列的开始标记`<s>`，表示解码开始；`$z_i$`是`$i$`时刻解码RNN的隐层状态，`$z_0$`是一个全零的向量。
其中`$\phi _{\theta '}$`是一个非线性激活函数；`$c=q\mathbf{h}$`是源语言句子的上下文向量，在不使用注意力机制时，如果[编码器](#编码器)的输出是源语言句子编码后的最后一个元素，则可以定义`$c=h_T$`；`$u_i$`是目标语言序列的第`$i$`个单词，`$u_0$`是目标语言序列的开始标记`<s>`，表示解码开始；`$z_i$`是`$i$`时刻解码RNN的隐层状态，`$z_0$`是一个全零的向量。
2. 将`$z_{i+1}$`通过`softmax`归一化，得到目标语言序列的第`$i+1$`个单词的概率分布`$p_{i+1}$`。概率分布公式如下：
$$p\left ( u_{i+1}|u_{&lt;i+1},\mathbf{x} \right )=softmax(W_sz_{i+1}+b_z)$$
@ -93,6 +94,7 @@ $$p\left ( u_{i+1}|u_{&lt;i+1},\mathbf{x} \right )=softmax(W_sz_{i+1}+b_z)$$
机器翻译任务的生成过程，通俗来讲就是根据预先训练的模型来翻译源语言句子。生成过程中的解码阶段和上述训练过程的有所差异，具体介绍请见[柱搜索算法](#柱搜索算法)。
<a name="柱搜索算法"></a>
### 柱搜索算法
柱搜索（[beam search](http://en.wikipedia.org/wiki/Beam_search)）是一种启发式图搜索算法，用于在图或树中搜索有限集合中的最优扩展节点，通常用在解空间非常大的系统（如机器翻译、语音识别）中，原因是内存无法装下图或树中所有展开的解。如在机器翻译任务中希望翻译“`<s>你好<e>`”，就算目标语言字典中只有3个词（`<s>`, `<e>`, `hello`），也可能生成无限句话（`hello`循环出现的次数不定），为了找到其中较好的翻译结果，我们可采用柱搜索算法。

2
doc/fluid/new_docs/beginners_guide/basics/understand_sentiment/README.cn.md
Unescape View File

@ -149,6 +149,8 @@ def convolution_net(data, input_dim, class_dim, emb_dim, hid_dim):
网络的输入`input_dim`表示的是词典的大小，`class_dim`表示类别数。这里，我们使用[`sequence_conv_pool`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/trainer_config_helpers/networks.py) API实现了卷积和池化操作。
<a name="栈值双向LSTM"></a>
### 栈式双向LSTM
栈式双向神经网络`stacked_lstm_net`的代码片段如下：

4
doc/fluid/new_docs/beginners_guide/basics/word2vec/README.cn.md
Unescape View File

@ -50,7 +50,7 @@ similarity: -0.0997506977351
```
以上结果可以通过运行`calculate_dis.py`, 加载字典里的单词和对应训练特征结果得到，我们将在[应用模型](#应用模型)中详细描述用法。
以上结果可以通过运行`calculate_dis.py`, 加载字典里的单词和对应训练特征结果得到，我们将在[模型应用](#模型应用)中详细描述用法。
## 模型概览
@ -189,6 +189,7 @@ dream that one day <e>
最后，每个输入会按其单词次在字典里的位置，转化成整数的索引序列，作为PaddlePaddle的输入。
<a name="训练模型"></a>
## 编程实现
本配置的模型结构如下图所示：
@ -349,6 +350,7 @@ Step 20: Average Cost 5.766995
...
```
<a name="模型应用"></a>
## 模型应用
在模型训练后，我们可以用它做一些预测。

2
doc/fluid/new_docs/beginners_guide/quick_start/recognize_digits/README.cn.md
Unescape View File

@ -102,7 +102,7 @@ Softmax回归模型采用了最简单的两层神经网络，即只有输入层
池化是非线性下采样的一种形式，主要作用是通过减少网络的参数来减小计算量，并且能够在一定程度上控制过拟合。通常在卷积层的后面会加上一个池化层。池化包括最大池化、平均池化等。其中最大池化是用不重叠的矩形框将输入层分成不同的区域，对于每个矩形框的数取最大值作为输出层，如图6所示。
更详细的关于卷积神经网络的具体知识可以参考[斯坦福大学公开课]( http://cs231n.github.io/convolutional-networks/ )和[图像分类](https://github.com/PaddlePaddle/book/blob/develop/image_classification/README.md)教程。
更详细的关于卷积神经网络的具体知识可以参考[斯坦福大学公开课]( http://cs231n.github.io/convolutional-networks/ )和[图像分类]( https://github.com/PaddlePaddle/book/tree/develop/03.image_classification )教程。
### 常见激活函数介绍
- sigmoid激活函数： $ f(x) = sigmoid(x) = \frac{1}{1+e^{-x}} $

3
doc/fluid/new_docs/user_guides/howto/debug/visualdl.md
Unescape View File

@ -104,6 +104,7 @@ visualDL --logdir=scratch_log --port=8080
# 访问 http://127.0.0.1:8080
```
如果出现`TypeError: __init__() got an unexpected keyword argument 'file'`, 是因为protobuf不是3.5以上，运行`pip install --upgrade protobuf`就能解决。
如果在虚拟环境下仍然遇到安装问题，请尝试以下方法。
@ -149,7 +150,7 @@ python setup.py bdist_wheel
pip install --upgrade dist/visualdl-*.whl
```
如果打包和安装遇到其他问题，不安装只想运行Visual DL可以看[这里](https://github.com/PaddlePaddle/VisualDL/blob/develop/docs/how_to_dev_frontend_en.md)
如果打包和安装遇到其他问题，不安装只想运行Visual DL可以看[这里](https://github.com/PaddlePaddle/VisualDL/blob/develop/docs/develop/how_to_dev_frontend_cn.md)
## SDK

8
doc/fluid/new_docs/user_guides/howto/inference/native_infer.rst
Unescape View File

@ -4,13 +4,12 @@ Paddle 预测 API
为了更简单方便的预测部署，Fluid 提供了一套高层 API
用来隐藏底层不同的优化实现。
`预测库相关代码 <https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/contrib/inference>`__
`预测库相关代码 <https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/inference/api>`_
包括
- 头文件 ``paddle_inference_api.h`` 定义了所有的接口
- 库文件\ ``libpaddle_fluid.so`` 或 ``libpaddle_fluid.a``
- 库文件 ``libpaddle_inference_api.so`` 或
``libpaddle_inference_api.a``
编译和依赖可以参考 :ref:`install_or_build_cpp_inference_lib` 。
@ -97,8 +96,7 @@ engine
CHECK(predictor->Run(slots, &outputs));
// 获取 outputs ...
编译时，联编 ``libpaddle_fluid.a/.so`` 和
``libpaddle_inference_api.a/.so`` 便可。
编译时，联编 ``libpaddle_fluid.a/.so`` 便可。
详细代码参考
------------

5
paddle/fluid/API.spec
Unescape View File

@ -43,6 +43,7 @@ paddle.fluid.Executor.run ArgSpec(args=['self', 'program', 'feed', 'fetch_list',
paddle.fluid.global_scope ArgSpec(args=[], varargs=None, keywords=None, defaults=None)
paddle.fluid.scope_guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
paddle.fluid.Trainer.__init__ ArgSpec(args=['self', 'train_func', 'optimizer_func', 'param_path', 'place', 'parallel', 'checkpoint_config'], varargs=None, keywords=None, defaults=(None, None, False, None))
paddle.fluid.Trainer.save_inference_model ArgSpec(args=['self', 'param_path', 'feeded_var_names', 'target_var_indexes'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Trainer.save_params ArgSpec(args=['self', 'param_path'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Trainer.stop ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Trainer.test ArgSpec(args=['self', 'reader', 'feed_order'], varargs=None, keywords=None, defaults=None)
@ -312,7 +313,7 @@ paddle.fluid.layers.iou_similarity ArgSpec(args=[], varargs='args', keywords='kw
paddle.fluid.layers.box_coder ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
paddle.fluid.layers.polygon_box_transform ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
paddle.fluid.layers.accuracy ArgSpec(args=['input', 'label', 'k', 'correct', 'total'], varargs=None, keywords=None, defaults=(1, None, None))
paddle.fluid.layers.auc ArgSpec(args=['input', 'label', 'curve', 'num_thresholds', 'topk'], varargs=None, keywords=None, defaults=('ROC', 200, 1))
paddle.fluid.layers.auc ArgSpec(args=['input', 'label', 'curve', 'num_thresholds', 'topk'], varargs=None, keywords=None, defaults=('ROC', 4095, 1))
paddle.fluid.layers.exponential_decay ArgSpec(args=['learning_rate', 'decay_steps', 'decay_rate', 'staircase'], varargs=None, keywords=None, defaults=(False,))
paddle.fluid.layers.natural_exp_decay ArgSpec(args=['learning_rate', 'decay_steps', 'decay_rate', 'staircase'], varargs=None, keywords=None, defaults=(False,))
paddle.fluid.layers.inverse_time_decay ArgSpec(args=['learning_rate', 'decay_steps', 'decay_rate', 'staircase'], varargs=None, keywords=None, defaults=(False,))
@ -376,7 +377,7 @@ paddle.fluid.optimizer.DecayedAdagradOptimizer.__init__ ArgSpec(args=['self', 'l
paddle.fluid.optimizer.DecayedAdagradOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.optimizer.FtrlOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'l1', 'l2', 'lr_power'], varargs=None, keywords='kwargs', defaults=(0.0, 0.0, -0.5))
paddle.fluid.optimizer.FtrlOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.optimizer.RMSPropOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'rho', 'epsilon', 'momentum'], varargs=None, keywords='kwargs', defaults=(0.95, 1e-06, 0.0))
paddle.fluid.optimizer.RMSPropOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'rho', 'epsilon', 'momentum', 'centered'], varargs=None, keywords='kwargs', defaults=(0.95, 1e-06, 0.0, False))
paddle.fluid.optimizer.RMSPropOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.optimizer.AdadeltaOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'epsilon', 'rho'], varargs=None, keywords='kwargs', defaults=(1e-06, 0.95))
paddle.fluid.optimizer.AdadeltaOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))

28
paddle/fluid/framework/details/multi_devices_graph_pass.cc
Unescape View File

@ -326,7 +326,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
ir::Graph &result = *graph;
for (auto &node : nodes) {
if (node->NodeType() == ir::Node::Type::kVariable && node->Var()) {
if (node->IsVar() && node->Var()) {
all_vars_.emplace(node->Name(), node->Var());
}
}
@ -583,18 +583,6 @@ void MultiDevSSAGraphBuilder::InsertDataBalanceOp(
}
}
bool MultiDevSSAGraphBuilder::IsParameterGradientOnce(
const std::string &og,
std::unordered_set<std::string> *og_has_been_broadcast) const {
bool is_pg_once =
grad_names_.count(og) != 0 && og_has_been_broadcast->count(og) == 0;
if (is_pg_once) {
// Insert NCCL AllReduce Op
og_has_been_broadcast->insert(og);
}
return is_pg_once;
}
int MultiDevSSAGraphBuilder::GetOpDeviceID(const ir::Graph &graph,
ir::Node *node) const {
if (strategy_.reduce_ != BuildStrategy::ReduceStrategy::kReduce) {
@ -688,20 +676,6 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(ir::Graph *result,
return var;
}
// Find the first occurence of `prev_op_name` and make current `op` depend
// on it.
void MultiDevSSAGraphBuilder::ConnectOp(ir::Graph *result, OpHandleBase *op,
const std::string &prev_op_name) const {
for (auto &prev_op : result->Get<GraphOps>(kGraphOps)) {
if (prev_op->Name() == prev_op_name) {
auto *dep_var = new DummyVarHandle(result->CreateControlDepVar());
prev_op->AddOutput(dep_var);
result->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
op->AddInput(dep_var);
}
}
}
void MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
ir::Node *node) const {
int op_dev_id = -1;

7
paddle/fluid/framework/details/multi_devices_graph_pass.h
Unescape View File

@ -69,9 +69,6 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
std::vector<std::string> FindDistTrainRecvVars(
const std::vector<ir::Node *> &nodes) const;
void ConnectOp(ir::Graph *result, OpHandleBase *op,
const std::string &prev_op_name) const;
void CreateComputationalOps(ir::Graph *result, ir::Node *node,
size_t num_places) const;
@ -83,10 +80,6 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
void CreateComputationalOp(ir::Graph *result, ir::Node *node,
int dev_id) const;
bool IsParameterGradientOnce(
const std::string &og,
std::unordered_set<std::string> *og_has_been_broadcast) const;
int GetOpDeviceID(const ir::Graph &graph, ir::Node *node) const;
void InsertAllReduceOp(ir::Graph *result, const std::string &og) const;

31
paddle/fluid/framework/ir/CMakeLists.txt
Unescape View File

@ -1,20 +1,35 @@
set(pass_file ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/paddle_inference_pass.h)
file(WRITE ${pass_file} "// Generated by the paddle/fluid/framework/ir/CMakeLists.txt. DO NOT EDIT!\n\n")
file(APPEND ${pass_file} "\#include \"paddle/fluid/framework/ir/pass.h\"\n")
function(pass_library TARGET)
set(options "")
set(oneValueArgs "")
set(multiValueArgs SRCS DEPS)
cmake_parse_arguments(op_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
cc_library(${TARGET} SRCS ${TARGET}.cc DEPS graph_pattern_detector pass)
file(APPEND ${pass_file} "USE_PASS(${TARGET});\n")
set(PASS_LIBRARY ${TARGET} ${PASS_LIBRARY} PARENT_SCOPE)
endfunction()
cc_library(node SRCS node.cc DEPS proto_desc)
cc_library(graph SRCS graph.cc DEPS node)
cc_library(graph_helper SRCS graph_helper.cc DEPS graph)
cc_library(pass SRCS pass.cc DEPS graph node graph_helper)
cc_library(graph_viz_pass SRCS graph_viz_pass.cc DEPS graph pass graph_helper)
cc_library(graph_to_program_pass SRCS graph_to_program_pass.cc DEPS graph pass 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(fc_fuse_pass SRCS fc_fuse_pass.cc DEPS graph graph_pattern_detector)
cc_library(attention_lstm_fuse_pass SRCS attention_lstm_fuse_pass.cc DEPS graph graph_pattern_detector)
cc_library(infer_clean_graph_pass SRCS infer_clean_graph_pass.cc DEPS graph pass)
cc_library(fc_lstm_fuse_pass SRCS fc_lstm_fuse_pass.cc DEPS graph graph_pattern_detector)
cc_library(seq_concat_fc_fuse_pass SRCS seq_concat_fc_fuse_pass.cc DEPS graph graph_pattern_detector)
pass_library(graph_to_program_pass)
pass_library(graph_viz_pass)
pass_library(fc_fuse_pass)
pass_library(attention_lstm_fuse_pass)
pass_library(infer_clean_graph_pass)
pass_library(fc_lstm_fuse_pass)
pass_library(seq_concat_fc_fuse_pass)
set(GLOB_PASS_LIB ${PASS_LIBRARY} CACHE INTERNAL "Global PASS library")
cc_test(pass_test SRCS pass_test.cc DEPS graph pass graph_helper)
cc_test(graph_test SRCS graph_test.cc DEPS graph graph_helper op_registry)
cc_test(graph_helper_test SRCS graph_helper_test.cc DEPS graph graph_helper op_registry)
cc_test(graph_to_program_pass_test SRCS graph_to_program_pass_test.cc DEPS graph_to_program_pass)
cc_test(test_graph_pattern_detector SRCS graph_pattern_detector_tester.cc DEPS graph_pattern_detector)
cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass graph_pattern_detector graph pass graph_traits framework_proto)
cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framework_proto)

15
paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc
Unescape View File

@ -13,13 +13,10 @@
// limitations under the License.
#include "paddle/fluid/framework/ir/attention_lstm_fuse_pass.h"
#include <string>
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
#include "paddle/fluid/framework/ir/graph_viz_pass.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/inference/api/helper.h"
namespace paddle {
namespace framework {
@ -99,17 +96,13 @@ void FindWhileOp(Graph* graph) {
auto* cell_init = graph->RetriveNode(6);
auto* hidden_init = graph->RetriveNode(8);
#define LINK_TO(node0, node1) \
node0->outputs.push_back(node1); \
node1->inputs.push_back(node0);
auto* lstm_op = graph->CreateOpNode(&op_desc);
PrepareParameters(graph, param);
LINK_TO(X, lstm_op);
LINK_TO(cell_init, lstm_op);
LINK_TO(hidden_init, lstm_op);
LINK_TO(lstm_op, LSTMOUT);
IR_NODE_LINK_TO(X, lstm_op);
IR_NODE_LINK_TO(cell_init, lstm_op);
IR_NODE_LINK_TO(hidden_init, lstm_op);
IR_NODE_LINK_TO(lstm_op, LSTMOUT);
GraphSafeRemoveNodes(graph, marked_nodes);
}

118
paddle/fluid/framework/ir/fc_fuse_pass.cc
Unescape View File

@ -21,74 +21,26 @@ namespace paddle {
namespace framework {
namespace ir {
bool VarOutLinksToOp(Node* node, const std::string& op_type) {
for (auto* out : node->outputs) {
if (out->IsOp() && out->Op()->Type() == op_type) {
return true;
}
}
return false;
}
void BuildFCPattern(PDPattern* pattern) {
// Create Operators
auto* mul_op = pattern->NewNode("mul")->assert_is_op("mul");
auto* elementwise_add_op =
pattern->NewNode("elementwise_add")->assert_is_op("elementwise_add");
// Create variables
// w
auto* mul_weight_var = pattern->NewNode("mul_weight")
->AsInput()
->assert_is_op_nth_input("mul", "Y", 0);
// x
auto* mul_tmp_var = pattern->NewNode("mul_tmp_var")
->AsInput()
->assert_is_op_nth_input("mul", "X", 0);
// intermediate variable, will be removed in the IR after fuse.
auto* mul_out_var = pattern->NewNode("mul_out")
->AsIntermediate()
->assert_is_only_output_of_op("mul")
->assert_is_op_input("elementwise_add");
// bias
auto* elementwise_add_tmp_var = pattern->NewNode("elementwise_add_tmpvar")
->assert_is_op_input("elementwise_add")
->AsInput();
// output
auto* elementwise_add_out_var = pattern->NewNode("elementwise_add_out")
->AsOutput()
->assert_is_op_output("elementwise_add");
mul_op->LinksFrom({mul_weight_var, mul_tmp_var}).LinksTo({mul_out_var});
elementwise_add_op->LinksFrom({mul_out_var, elementwise_add_tmp_var})
.LinksTo({elementwise_add_out_var});
}
// Replace the node `from` in the links to `to`
bool LinksReplace(std::vector<Node*>* links, Node* from, Node* to) {
for (auto*& n : *links) {
if (n == from) {
n = to;
return true;
}
}
return false;
}
std::unique_ptr<ir::Graph> FCFusePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
PADDLE_ENFORCE(graph.get());
FusePassBase::Init("fc", graph.get());
FusePassBase::Init("fc_fuse", graph.get());
std::unordered_set<Node*> nodes2delete;
GraphPatternDetector gpd;
BuildFCPattern(gpd.mutable_pattern());
#define GET_NODE(id) \
PADDLE_ENFORCE(subgraph.count(gpd.pattern().RetrieveNode(#id)), \
"pattern has no Node called %s", #id); \
auto* id = subgraph.at(gpd.pattern().RetrieveNode(#id)); \
PADDLE_ENFORCE_NOT_NULL(id, "subgraph has no node %s", #id);
// BuildFCPattern(gpd.mutable_pattern());
auto* x = gpd.mutable_pattern()
->NewNode("fc_fuse/x")
->AsInput()
->assert_is_op_input("mul", "X");
patterns::FC(gpd.mutable_pattern(), "fc_fuse", x, true /*with bias*/);
#define GET_NODE(id) \
PADDLE_ENFORCE(subgraph.count(gpd.pattern().RetrieveNode("fc_fuse/" #id)), \
"pattern has no Node called %s", #id); \
auto* id = subgraph.at(gpd.pattern().RetrieveNode("fc_fuse/" #id)); \
PADDLE_ENFORCE_NOT_NULL(id, "subgraph has no node %s", "fc_fuse/" #id);
int found_fc_count = 0;
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
@ -98,43 +50,33 @@ std::unique_ptr<ir::Graph> FCFusePass::ApplyImpl(
// scenerio.
// FC's fusion is simple, just op fuse, no need to process the
// parameters.
GET_NODE(mul_tmp_var); // x
GET_NODE(mul_weight); // Y
GET_NODE(elementwise_add_tmpvar); // bias
GET_NODE(elementwise_add_out); // Out
GET_NODE(mul); // MUL op
GET_NODE(elementwise_add); // ELEMENT_ADD op
GET_NODE(mul_out); // tmp
GET_NODE(x); // x
GET_NODE(w); // Y
GET_NODE(fc_bias); // bias
GET_NODE(fc_out); // Out
GET_NODE(mul); // MUL op
GET_NODE(elementwise_add); // ELEMENT_ADD op
GET_NODE(mul_out); // tmp
#undef GET_NODE
// Create an FC Node.
OpDesc desc;
std::string fc_x_in = mul_tmp_var->Name();
std::string fc_Y_in = mul_weight->Name();
std::string fc_bias_in = elementwise_add_tmpvar->Name();
std::string fc_out = elementwise_add_out->Name();
std::string fc_x_in = x->Name();
std::string fc_Y_in = w->Name();
std::string fc_bias_in = fc_bias->Name();
std::string fc_out_out = fc_out->Name();
desc.SetInput("Input", std::vector<std::string>({fc_x_in}));
desc.SetInput("W", std::vector<std::string>({fc_Y_in}));
desc.SetInput("Bias", std::vector<std::string>({fc_bias_in}));
desc.SetOutput("Out", std::vector<std::string>({fc_out}));
desc.SetOutput("Out", std::vector<std::string>({fc_out_out}));
desc.SetType("fc");
auto fc_node = g->CreateOpNode(&desc); // OpDesc will be copied.
fc_node->inputs =
std::vector<Node*>({mul_tmp_var, mul_weight, elementwise_add_tmpvar});
fc_node->outputs.push_back(elementwise_add_out);
// Update link relatons
PADDLE_ENFORCE(LinksReplace(&mul_tmp_var->outputs, mul, fc_node));
PADDLE_ENFORCE(LinksReplace(&mul_weight->outputs, mul, fc_node));
PADDLE_ENFORCE(LinksReplace(&elementwise_add_tmpvar->outputs,
elementwise_add, fc_node));
PADDLE_ENFORCE(
LinksReplace(&elementwise_add_out->inputs, elementwise_add, fc_node));
GraphSafeRemoveNodes(graph.get(), {mul, elementwise_add, mul_out});
// Drop old nodes
graph->RemoveNode(mul);
graph->RemoveNode(elementwise_add);
graph->RemoveNode(mul_out); // tmp variable
IR_NODE_LINK_TO(x, fc_node);
IR_NODE_LINK_TO(w, fc_node);
IR_NODE_LINK_TO(fc_bias, fc_node);
IR_NODE_LINK_TO(fc_node, fc_out);
found_fc_count++;
};

47
paddle/fluid/framework/ir/fc_lstm_fuse_pass.cc
Unescape View File

@ -11,7 +11,6 @@
// 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/fc_lstm_fuse_pass.h"
#include <string>
#include "paddle/fluid/framework/lod_tensor.h"
@ -87,15 +86,24 @@ int BuildFusion(Graph* graph, const std::string& name_scope, Scope* scope,
}
op_desc.SetInput("Bias", {new_bias_var});
}
#undef GET_NODE
// Create temp variables.
scope->Var(name_scope + "/BatchedInput.new")
->GetMutable<framework::LoDTensor>();
scope->Var(name_scope + "/BatchCellPreAct.new")
->GetMutable<framework::LoDTensor>();
scope->Var(name_scope + "/BatchedGate.new")
->GetMutable<framework::LoDTensor>();
op_desc.SetInput("H0", {});
op_desc.SetInput("C0", {});
op_desc.SetOutput("Hidden", {hidden_n->Name()});
op_desc.SetOutput("Cell", {cell_n->Name()});
op_desc.SetOutput("XX", {xx_n->Name()});
op_desc.SetOutput("BatchedInput", {"blstm_0.tmp_2"});
op_desc.SetOutput("BatchedGate", {name_scope + "/BatchedGate.new"});
op_desc.SetOutput("BatchCellPreAct", {name_scope + "/BatchCellPreAct.new"});
op_desc.SetOutput("BatchedInput", {name_scope + "/BatchedInput.new"});
op_desc.SetAttr("is_reverse", lstm_n->Op()->GetAttr("is_reverse"));
op_desc.SetAttr("use_peepholes", lstm_n->Op()->GetAttr("use_peepholes"));
// TODO(TJ): get from attr
@ -121,22 +129,18 @@ int BuildFusion(Graph* graph, const std::string& name_scope, Scope* scope,
#undef TMP_NEW
#undef TMP_NAME
#define LINK_TO(a, b) \
a->outputs.push_back(b); \
b->inputs.push_back(a);
LINK_TO(input_n, op);
LINK_TO(weight_x_n, op);
LINK_TO(weight_h_n, op);
LINK_TO(bias_n, op);
LINK_TO(op, hidden_n);
#undef LINK_TO
IR_NODE_LINK_TO(input_n, op);
IR_NODE_LINK_TO(weight_x_n, op);
IR_NODE_LINK_TO(weight_h_n, op);
IR_NODE_LINK_TO(bias_n, op);
IR_NODE_LINK_TO(op, hidden_n);
return op;
};
int fusion_count{0};
auto fc_no_bias_handler = [&](
const GraphPatternDetector::subgraph_t& subgraph, Graph* g) {
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
#define GET_NODE(name__) \
std::string name__##key = name_scope + "/" + #name__; \
auto* name__##n = pattern->RetrieveNode(name__##key); \
@ -157,21 +161,24 @@ int BuildFusion(Graph* graph, const std::string& name_scope, Scope* scope,
if (with_fc_bias) {
GET_NODE(fc_bias);
GET_NODE(elementwise_add);
lstm_creator(lstm, x, w, Weight, Bias, Hidden, Cell, fc_out, fc_bias);
// Remove unneeded nodes.
std::unordered_set<const Node*> marked_nodes(
{mul_n, lstm_n, elementwise_add_n});
GraphSafeRemoveNodes(graph, marked_nodes);
} else {
lstm_creator(lstm, x, w, Weight, Bias, Hidden, Cell, fc_out, -1);
// Remove unneeded nodes.
std::unordered_set<const Node*> marked_nodes({mul_n, lstm_n});
GraphSafeRemoveNodes(graph, marked_nodes);
}
#undef GET_NODE
// Remove unneeded nodes.
std::unordered_set<const Node*> marked_nodes({mul_n, lstm_n});
GraphSafeRemoveNodes(graph, marked_nodes);
++fusion_count;
};
gpd(graph, fc_no_bias_handler);
gpd(graph, handler);
return fusion_count;
}

2
paddle/fluid/framework/ir/fc_lstm_fuse_pass.h
Unescape View File

@ -12,6 +12,8 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#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"

24
paddle/fluid/framework/ir/graph_pattern_detector.cc
Unescape View File

@ -73,7 +73,6 @@ void PDPattern::AddEdge(PDNode* a, PDNode* b) {
void GraphPatternDetector::operator()(Graph* graph,
GraphPatternDetector::handle_t handler) {
if (!MarkPDNodesInGraph(*graph)) {
LOG(INFO) << "Mark failed";
return;
}
@ -86,7 +85,7 @@ void GraphPatternDetector::operator()(Graph* graph,
LOG(INFO) << "detect " << subgraphs.size() << " subgraph matches the pattern";
int id = 0;
for (auto& g : subgraphs) {
LOG(INFO) << "optimizing #" << id++ << " subgraph";
VLOG(3) << "optimizing #" << id++ << " subgraph";
handler(g, graph);
}
}
@ -111,6 +110,11 @@ bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph& graph) {
return false;
}
}
for (auto& item : pdnodes2nodes_) {
for (auto& n : item.second) {
GetMarkedNodes(const_cast<Graph*>(&graph)).insert(n);
}
}
VLOG(3) << pdnodes2nodes_.size() << " nodes marked";
return !pdnodes2nodes_.empty();
@ -278,7 +282,7 @@ void GraphPatternDetector::RemoveOverlappedMatch(
for (const auto& subgraph : *subgraphs) {
bool valid = true;
for (auto& item : subgraph) {
if (node_set.count(item.second)) {
if (item.first->IsIntermediate() && node_set.count(item.second)) {
valid = false;
break;
}
@ -334,22 +338,22 @@ PDNode& PDNode::LinksFrom(const std::vector<PDNode*>& others) {
}
PDNode* PDNode::assert_is_op() {
asserts_.emplace_back([this](Node* x) { return x && x->IsOp(); });
asserts_.emplace_back([](Node* x) { return x && x->IsOp(); });
return this;
}
PDNode* PDNode::assert_is_op(const std::string& op_type) {
asserts_.emplace_back([this, op_type](Node* x) {
asserts_.emplace_back([op_type](Node* x) {
return x && x->IsOp() && x->Op()->Type() == op_type;
});
return this;
}
PDNode* PDNode::assert_is_var() {
asserts_.emplace_back([this](Node* x) { return x && x->IsVar(); });
asserts_.emplace_back([](Node* x) { return x && x->IsVar(); });
return this;
}
PDNode* PDNode::assert_var_not_persistable() {
assert_is_var();
asserts_.emplace_back([this](Node* x) { return !x->Var()->Persistable(); });
asserts_.emplace_back([](Node* x) { return !x->Var()->Persistable(); });
return this;
}
PDNode* PDNode::assert_is_persistable_var() {
@ -491,14 +495,16 @@ void GraphSafeRemoveNodes(Graph* graph,
for (auto it = node->inputs.begin(); it != node->inputs.end();) {
if (nodes.count(*it)) {
it = const_cast<Node*>(node)->inputs.erase(it);
} else
} else {
it++;
}
}
for (auto it = node->outputs.begin(); it != node->outputs.end();) {
if (nodes.count(*it)) {
it = const_cast<Node*>(node)->outputs.erase(it);
} else
} else {
it++;
}
}
}
}

9
paddle/fluid/framework/ir/graph_pattern_detector.h
Unescape View File

@ -19,6 +19,9 @@
#endif
#include <numeric>
#include <string>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/node.h"
#include "paddle/fluid/inference/analysis/dot.h"
@ -245,6 +248,8 @@ class GraphPatternDetector {
void UniquePatterns(std::vector<subgraph_t>* subgraphs);
// Remove overlapped match subgraphs, when overlapped, keep the previous one.
// The intermediate PDNodes will be removed, so can't shared by multiple
// patterns.
void RemoveOverlappedMatch(std::vector<subgraph_t>* subgraphs);
// Validate whether the intermediate nodes are linked by external nodes.
@ -295,6 +300,10 @@ PDNode* LSTM(PDPattern* pattern, const std::string& name_scope, PDNode* x);
} // namespace patterns
#define IR_NODE_LINK_TO(a, b) \
a->outputs.push_back(b); \
b->inputs.push_back(a);
} // namespace ir
} // namespace framework
} // namespace paddle

5
paddle/fluid/framework/ir/graph_pattern_detector_tester.cc
Unescape View File

@ -140,8 +140,9 @@ TEST(GraphPatternDetecter, MultiSubgraph) {
return node->IsOp() && (node->Name() == "op2" || node->Name() == "op3");
},
"OP0");
auto* any_var = x.mutable_pattern()->NewNode(
[](Node* node) { return node->IsVar(); }, "VAR");
auto* any_var = x.mutable_pattern()
->NewNode([](Node* node) { return node->IsVar(); }, "VAR")
->AsIntermediate();
auto* any_op1 = x.mutable_pattern()->NewNode(
[](Node* node) { return node->IsOp(); }, "OP1");

59
paddle/fluid/framework/ir/graph_viz_pass.cc
Unescape View File

@ -50,20 +50,37 @@ std::unique_ptr<ir::Graph> GraphVizPass::ApplyImpl(
Dot dot;
std::vector<Dot::Attr> op_attrs({Dot::Attr("style", "filled"),
Dot::Attr("shape", "box"),
Dot::Attr("fillcolor", "red")});
std::vector<Dot::Attr> var_attrs({Dot::Attr("style", "filled,rounded"),
// Dot::Attr("shape", "diamond"),
Dot::Attr("fillcolor", "yellow")});
std::vector<Dot::Attr> marked_op_attrs({Dot::Attr("style", "filled"),
Dot::Attr("shape", "box"),
Dot::Attr("fillcolor", "lightgray")});
std::vector<Dot::Attr> marked_var_attrs(
{Dot::Attr("style", "filled,rounded"),
// Dot::Attr("shape", "diamond"),
Dot::Attr("fillcolor", "lightgray")});
const std::vector<Dot::Attr> op_attrs({
Dot::Attr("style", "rounded,filled,bold"), //
Dot::Attr("shape", "box"), //
Dot::Attr("color", "#303A3A"), //
Dot::Attr("fontcolor", "#ffffff"), //
Dot::Attr("width", "1.3"), //
Dot::Attr("height", "0.84"), //
Dot::Attr("fontname", "Arial"), //
});
const std::vector<Dot::Attr> arg_attrs({
Dot::Attr("shape", "box"), //
Dot::Attr("style", "rounded,filled,bold"), //
Dot::Attr("fontname", "Arial"), //
Dot::Attr("fillcolor", "#999999"), //
Dot::Attr("color", "#dddddd"), //
});
const std::vector<Dot::Attr> param_attrs({
Dot::Attr("shape", "box"), //
Dot::Attr("style", "rounded,filled,bold"), //
Dot::Attr("fontname", "Arial"), //
Dot::Attr("color", "#148b97"), //
Dot::Attr("fontcolor", "#ffffff"), //
});
const std::vector<Dot::Attr> marked_op_attrs(
{Dot::Attr("style", "rounded,filled,bold"), Dot::Attr("shape", "box"),
Dot::Attr("fillcolor", "yellow")});
const std::vector<Dot::Attr> marked_var_attrs(
{Dot::Attr("style", "filled,rounded"), Dot::Attr("shape", "box"),
Dot::Attr("fillcolor", "yellow")});
auto marked_nodes = ConsumeMarkedNodes(graph.get());
// Create nodes
@ -74,9 +91,17 @@ std::unique_ptr<ir::Graph> GraphVizPass::ApplyImpl(
marked_nodes.count(n) ? marked_op_attrs : op_attrs;
dot.AddNode(node_id, attr, node_id);
} else if (n->IsVar()) {
decltype(op_attrs) attr =
marked_nodes.count(n) ? marked_var_attrs : var_attrs;
dot.AddNode(node_id, attr, node_id);
decltype(op_attrs)* attr;
if (marked_nodes.count(n)) {
attr = &marked_var_attrs;
} else if (const_cast<Node*>(n)->Var() &&
const_cast<Node*>(n)->Var()->Persistable()) {
attr = &param_attrs;
} else {
attr = &arg_attrs;
}
dot.AddNode(node_id, *attr, node_id);
}
node2dot[n] = node_id;
}

23
paddle/fluid/framework/ir/infer_clean_graph_pass.cc
Unescape View File

@ -13,42 +13,41 @@
// limitations under the License.
#include <algorithm>
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
namespace paddle {
namespace framework {
namespace ir {
class InferCleanGraphPass : public Pass {
class InferCleanGraphPass : public FusePassBase {
public:
virtual ~InferCleanGraphPass() {}
protected:
std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const {
FusePassBase::Init("original_graph", graph.get());
PADDLE_ENFORCE(graph.get());
auto is_valid_node = [](Node* x) {
return x && IsControlDepVar(*x) && x->IsVar() && !x->Var();
};
std::unordered_set<Node*> invalid_nodes;
std::unordered_set<const Node*> invalid_nodes;
int valid_op = 0;
for (auto* node : graph->Nodes()) {
if (is_valid_node(node)) {
invalid_nodes.insert(node);
} else if (node->IsOp()) {
// Collect all the operators to help tracking number of operators.
++valid_op;
}
}
// remove nodes from the graph.
for (auto* node : invalid_nodes) {
graph->RemoveNode(node);
}
GraphSafeRemoveNodes(graph.get(), invalid_nodes);
// clean edges.
for (auto* node : graph->Nodes()) {
CleanEdges(&node->inputs, invalid_nodes);
CleanEdges(&node->outputs, invalid_nodes);
}
AddStatis(valid_op);
return graph;
}

18
paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.cc
Unescape View File

@ -219,16 +219,13 @@ std::unique_ptr<ir::Graph> SeqConcatFcFusePass::ApplyImpl(
op_desc.SetAttr("fc_activation", act->Op()->Type());
auto* op_node = graph->CreateOpNode(&op_desc);
// Add links
#define NODE_LINKS(a, b) \
a->outputs.push_back(b); \
b->inputs.push_back(a);
NODE_LINKS(fc_w, op_node);
NODE_LINKS(fc_bias, op_node);
NODE_LINKS(concat_in0, op_node);
NODE_LINKS(sequence_expand0_in, op_node);
NODE_LINKS(sequence_expand1_in, op_node);
NODE_LINKS(op_node, fc_out);
// Add links
IR_NODE_LINK_TO(fc_w, op_node);
IR_NODE_LINK_TO(fc_bias, op_node);
IR_NODE_LINK_TO(concat_in0, op_node);
IR_NODE_LINK_TO(sequence_expand0_in, op_node);
IR_NODE_LINK_TO(sequence_expand1_in, op_node);
IR_NODE_LINK_TO(op_node, fc_out);
// Clean nodes.
std::unordered_set<const Node*> marked_nodes;
@ -241,7 +238,6 @@ std::unique_ptr<ir::Graph> SeqConcatFcFusePass::ApplyImpl(
marked_nodes.erase(sequence_expand0_in);
marked_nodes.erase(sequence_expand1_in);
marked_nodes.erase(fc_out);
GraphSafeRemoveNodes(graph, marked_nodes);
});

5
paddle/fluid/inference/CMakeLists.txt
Unescape View File

@ -10,7 +10,7 @@ set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor)
# TODO(panyx0718): Should this be called paddle_fluid_inference_api_internal?
cc_library(paddle_fluid_api
SRCS io.cc
DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB} graph_to_program_pass)
DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
@ -22,7 +22,7 @@ cc_library(paddle_fluid_origin DEPS ${fluid_modules} paddle_fluid_api)
#endif()
# Create static library
cc_library(paddle_fluid DEPS ${fluid_modules} paddle_fluid_api paddle_inference_api)
cc_library(paddle_fluid DEPS ${fluid_modules} paddle_fluid_api paddle_inference_api analysis_predictor)
if(NOT APPLE)
# TODO(liuyiqu: Temporarily disable the link flag because it is not support on Mac.
set(LINK_FLAGS "-Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.sym")
@ -32,6 +32,7 @@ endif()
# Create shared library
cc_library(paddle_fluid_shared SHARED
SRCS io.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api_impl.cc
${CMAKE_CURRENT_SOURCE_DIR}/api/analysis_predictor.cc
DEPS ${fluid_modules} paddle_fluid_api)
set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)

38
paddle/fluid/inference/analysis/CMakeLists.txt
Unescape View File

@ -33,7 +33,7 @@ function (inference_analysis_test TARGET)
endif()
cc_test(${TARGET}
SRCS "${analysis_test_SRCS}"
DEPS analysis graph fc_fuse_pass graph_viz_pass infer_clean_graph_pass graph_pattern_detector pass ${analysis_test_EXTRA_DEPS}
DEPS analysis pass ${GLOB_PASS_LIB} ${analysis_test_EXTRA_DEPS}
ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model ${mem_opt} ${analysis_test_ARGS})
set_tests_properties(${TARGET} PROPERTIES DEPENDS test_word2vec)
endif(WITH_TESTING)
@ -56,25 +56,13 @@ if (NOT EXISTS ${DITU_INSTALL_DIR} AND WITH_TESTING)
endif()
inference_analysis_test(test_analyzer SRCS analyzer_tester.cc
EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis
analysis_predictor
# ir
fc_fuse_pass
fc_lstm_fuse_pass
seq_concat_fc_fuse_pass
graph_viz_pass
infer_clean_graph_pass
graph_pattern_detector
infer_clean_graph_pass
attention_lstm_fuse_pass
paddle_inference_api
pass
EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor
ARGS --infer_ditu_rnn_model=${DITU_INSTALL_DIR}/model
--infer_ditu_rnn_data=${DITU_INSTALL_DIR}/data.txt)
--infer_ditu_rnn_data=${DITU_INSTALL_DIR}/data.txt)
inference_analysis_test(test_data_flow_graph SRCS data_flow_graph_tester.cc)
inference_analysis_test(test_data_flow_graph_to_fluid_pass SRCS data_flow_graph_to_fluid_pass_tester.cc EXTRA_DEPS paddle_inference_api)
inference_analysis_test(test_fluid_to_ir_pass SRCS fluid_to_ir_pass_tester.cc EXTRA_DEPS paddle_fluid)
inference_analysis_test(test_data_flow_graph_to_fluid_pass SRCS data_flow_graph_to_fluid_pass_tester.cc)
inference_analysis_test(test_fluid_to_ir_pass SRCS fluid_to_ir_pass_tester.cc)
inference_analysis_test(test_fluid_to_data_flow_graph_pass SRCS fluid_to_data_flow_graph_pass_tester.cc)
inference_analysis_test(test_subgraph_splitter SRCS subgraph_splitter_tester.cc)
inference_analysis_test(test_dfg_graphviz_draw_pass SRCS dfg_graphviz_draw_pass_tester.cc)
@ -86,7 +74,7 @@ inference_analysis_test(test_model_store_pass SRCS model_store_pass_tester.cc)
set(CHINESE_NER_MODEL_URL "http://paddle-inference-dist.bj.bcebos.com/chinese_ner_model.tar.gz")
set(CHINESE_NER_DATA_URL "http://paddle-inference-dist.bj.bcebos.com/chinese_ner-data.txt.tar.gz")
set(CHINESE_NER_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo/chinese_ner" CACHE PATH "Chinese ner model and data root." FORCE)
if (NOT EXISTS ${CHINESE_NER_INSTALL_DIR} AND WITH_TESTING)
if (NOT EXISTS ${CHINESE_NER_INSTALL_DIR} AND WITH_TESTING AND WITH_INFERENCE)
inference_download_and_uncompress(${CHINESE_NER_INSTALL_DIR} ${CHINESE_NER_MODEL_URL} "chinese_ner_model.tar.gz")
inference_download_and_uncompress(${CHINESE_NER_INSTALL_DIR} ${CHINESE_NER_DATA_URL} "chinese_ner-data.txt.tar.gz")
endif()
@ -99,7 +87,7 @@ inference_analysis_test(test_analyzer_ner SRCS analyzer_ner_tester.cc
set(LAC_MODEL_URL "http://paddle-inference-dist.bj.bcebos.com/lac_model.tar.gz")
set(LAC_DATA_URL "http://paddle-inference-dist.bj.bcebos.com/lac_data.txt.tar.gz")
set(LAC_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo/lac" CACHE PATH "LAC model and data root." FORCE)
if (NOT EXISTS ${LAC_INSTALL_DIR} AND WITH_TESTING)
if (NOT EXISTS ${LAC_INSTALL_DIR} AND WITH_TESTING AND WITH_INFERENCE)
inference_download_and_uncompress(${LAC_INSTALL_DIR} ${LAC_MODEL_URL} "lac_model.tar.gz")
inference_download_and_uncompress(${LAC_INSTALL_DIR} ${LAC_DATA_URL} "lac_data.txt.tar.gz")
endif()
@ -108,3 +96,15 @@ inference_analysis_test(test_analyzer_lac SRCS analyzer_lac_tester.cc
EXTRA_DEPS paddle_inference_api paddle_fluid_api
ARGS --infer_model=${LAC_INSTALL_DIR}/model
--infer_data=${LAC_INSTALL_DIR}/data.txt)
set(TEXT_CLASSIFICATION_MODEL_URL "http://paddle-inference-dist.bj.bcebos.com/text-classification-Senta.tar.gz")
set(TEXT_CLASSIFICATION_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo/text_classification" CACHE PATH "Text Classification model and data root." FORCE)
if (NOT EXISTS ${TEXT_CLASSIFICATION_INSTALL_DIR} AND WITH_TESTING AND WITH_INFERENCE)
inference_download_and_uncompress(${TEXT_CLASSIFICATION_INSTALL_DIR} ${TEXT_CLASSIFICATION_MODEL_URL} "text-classification-Senta.tar.gz")
endif()
inference_analysis_test(test_text_classification SRCS analyzer_text_classification_tester.cc
EXTRA_DEPS paddle_inference_api paddle_fluid_api analysis_predictor
ARGS --infer_model=${TEXT_CLASSIFICATION_INSTALL_DIR}/text-classification-Senta)

38
paddle/fluid/inference/analysis/analyzer.cc
Unescape View File

@ -14,6 +14,7 @@
#include "paddle/fluid/inference/analysis/analyzer.h"
#include <string>
#include <vector>
#include "paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h"
#include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h"
#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h"
@ -41,20 +42,16 @@ class DfgPassManagerImpl final : public DfgPassManager {
public:
DfgPassManagerImpl() {
// TODO(Superjomn) set the key with pass reprs.
LOG(INFO)
<< "-----------------------------------------------------------------";
if (FLAGS_IA_enable_ir) {
AddPass("fluid-to-ir-pass", new FluidToIrPass);
} else {
if (!FLAGS_IA_enable_ir) {
AddPass("fluid-to-data-flow-graph", new FluidToDataFlowGraphPass);
} else {
AddPass("fluid-to-ir-pass", new FluidToIrPass);
}
TryAddTensorRtPass();
AddPass("data-flow-graph-to-fluid", new DataFlowGraphToFluidPass);
if (!FLAGS_IA_output_storage_path.empty()) {
AddPass("model-store-pass", new ModelStorePass);
}
LOG(INFO)
<< "-----------------------------------------------------------------";
}
std::string repr() const override { return "dfg-pass-manager"; }
@ -101,19 +98,15 @@ class DfgPassManagerImpl final : public DfgPassManager {
Analyzer::Analyzer() { Register("manager1", new DfgPassManagerImpl); }
void Analyzer::Run(Argument* argument) {
// Ugly support fluid-to-ir-pass
argument->Set(kFluidToIrPassesAttr,
new std::vector<std::string>({
// Manual update the passes here.
"graph_viz_pass", //
"infer_clean_graph_pass", "graph_viz_pass", //
"attention_lstm_fuse_pass", "graph_viz_pass", //
"fc_lstm_fuse_pass", "graph_viz_pass", //
"mul_lstm_fuse_pass", "graph_viz_pass", //
"seq_concat_fc_fuse_pass", "graph_viz_pass", //
"fc_fuse_pass", "graph_viz_pass" //
}));
std::vector<std::string> passes;
for (auto& pass : all_ir_passes_) {
if (!disabled_ir_passes_.count(pass)) {
passes.push_back(pass);
passes.push_back("graph_viz_pass"); // add graphviz for debug.
}
}
passes.push_back("graph_viz_pass");
argument->Set(kFluidToIrPassesAttr, new std::vector<std::string>(passes));
for (auto& x : data_) {
PADDLE_ENFORCE(x->Initialize(argument));
@ -122,6 +115,11 @@ void Analyzer::Run(Argument* argument) {
}
}
Analyzer& Analyzer::DisableIrPasses(const std::vector<std::string>& passes) {
disabled_ir_passes_.insert(passes.begin(), passes.end());
return *this;
}
} // namespace analysis
} // namespace inference
} // namespace paddle

Some files were not shown because too many files have changed in this diff Show More

Write Preview
Loading…
Cancel
Save