Skip to content

P
Paddle

PaddlePaddle / Paddle
1 年多 前同步成功

通知 2301
Star 20931
Fork 5422
P
Paddle
提交 b94f9b16 编写于 作者: S sneaxiy
浏览文件
操作

merge release/1.1

上级 111959a0 c6571b3a
展开全部 隐藏空白更改
内联 并排
Showing with 3168 addition and 864 deletion
CMakeLists.txt
浏览文件 @ b94f9b16
......@@ -69,6 +69,7 @@ option(WITH_ANAKIN "Compile with Anakin library" OFF)
option(WITH_GRPC "Use grpc as the default rpc framework" ${WITH_DISTRIBUTE})
option(WITH_BRPC_RDMA "Use brpc rdma as the rpc protocal" OFF)
option(WITH_INFERENCE "Compile fluid inference library" ON)
option(ON_INFER "Turn on inference optimization." OFF)
option(WITH_INFERENCE_API_TEST "Test fluid inference high-level api interface" OFF)
option(WITH_SYSTEM_BLAS "Use system blas library" OFF)
option(PY_VERSION "Compile PaddlePaddle with python3 support" ${PY_VERSION})
......@@ -179,6 +180,7 @@ include(external/eigen) # download eigen3
include(external/pybind11) # download pybind11
include(external/cares)
include(external/cub)
include(external/xxhash) # download xxhash
if (NOT WIN32)
# there is no official support of snappystream, warpctc, nccl, cupti in windows
......@@ -301,3 +303,11 @@ if(WITH_DOC)
find_python_module(recommonmark REQUIRED)
add_subdirectory(doc)
endif()
if (ON_INFER)
message(STATUS "On inference mode, will take place some specific optimization.")
add_definitions(-DPADDLE_ON_INFERENCE)
else()
#TODO(luotao), combine this warning with `make inference_lib_dist` command.
message(WARNING "On inference mode, will take place some specific optimization. Turn on the ON_INFER flag when building inference_lib only.")
endif()
README.md
浏览文件 @ b94f9b16
......@@ -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://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)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.1/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/1.1/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 1.0.1](https://github.com/PaddlePaddle/Paddle/tree/release/1.0.0)
### Latest PaddlePaddle Release: [Fluid 1.1.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.1)
### 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==1.0.1.post87
pip install paddlepaddle-gpu==1.1.0.post87
# Linux GPU cuda8cudnn5
pip install paddlepaddle-gpu==1.0.1.post85
pip install paddlepaddle-gpu==1.1.0.post85
# For installation on other platform, refer to http://paddlepaddle.org/
```
......@@ -76,26 +76,26 @@ pip install paddlepaddle-gpu==1.0.1.post85
## Installation
It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html) on our website.
It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) on our website.
## Documentation
We provide [English](http://paddlepaddle.org/documentation/docs/en/1.0.0/getstarted/index_en.html) and
[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html) documentation.
We provide [English](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html) and
[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) documentation.
- [Deep Learning 101](https://github.com/PaddlePaddle/book)
You might want to start from this online interactive book that can run in a Jupyter Notebook.
- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.0/user_guides/howto/training/cluster_howto.html)
- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.1/user_guides/howto/training/cluster_howto.html)
You can run distributed training jobs on MPI clusters.
- [Python API](http://paddlepaddle.org/documentation/api/zh/1.0/fluid.html)
- [Python API](http://paddlepaddle.org/documentation/api/zh/1.1/fluid.html)
Our new API enables much shorter programs.
- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.0/advanced_usage/development/contribute_to_paddle.html)
- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.1/advanced_usage/development/contribute_to_paddle.html)
We appreciate your contributions!
......
benchmark/fluid/args.py
浏览文件 @ b94f9b16
......@@ -142,5 +142,10 @@ def parse_args():
choices=['reduce', 'all_reduce'],
default='all_reduce',
help='Specify the reduce strategy, can be reduce, all_reduce')
parser.add_argument(
'--fuse_broadcast_op',
action='store_true',
help='If set, would fuse multiple broadcast operators into one fused_broadcast operator.'
)
args = parser.parse_args()
return args
benchmark/fluid/fluid_benchmark.py
浏览文件 @ b94f9b16
......@@ -177,6 +177,7 @@ def train_parallel(train_args, test_args, args, train_prog, test_prog,
else:
build_strategy.reduce_strategy = fluid.BuildStrategy(
).ReduceStrategy.AllReduce
build_strategy.fuse_broadcast_op = args.fuse_broadcast_op
avg_loss = train_args[0]
......@@ -240,7 +241,6 @@ def train_parallel(train_args, test_args, args, train_prog, test_prog,
if args.use_fake_data or args.use_reader_op:
try:
fetch_ret = exe.run(fetch_list)
except fluid.core.EOFException as eof:
break
......
cmake/configure.cmake
浏览文件 @ b94f9b16
......@@ -50,11 +50,7 @@ if(NOT WITH_PROFILER)
endif(NOT WITH_PROFILER)
if(NOT CMAKE_CROSSCOMPILING)
if(WITH_AVX AND AVX512F_FOUND)
set(SIMD_FLAG ${AVX512F_FLAG})
elseif(WITH_AVX AND AVX2_FOUND)
set(SIMD_FLAG ${AVX2_FLAG})
elseif(WITH_AVX AND AVX_FOUND)
if(WITH_AVX AND AVX_FOUND)
set(SIMD_FLAG ${AVX_FLAG})
elseif(SSE3_FOUND)
set(SIMD_FLAG ${SSE3_FLAG})
......
cmake/external/xxhash.cmake 0 → 100644
浏览文件 @ b94f9b16
INCLUDE(ExternalProject)
set(XXHASH_SOURCE_DIR ${THIRD_PARTY_PATH}/xxhash)
set(XXHASH_INSTALL_DIR ${THIRD_PARTY_PATH}/install/xxhash)
set(XXHASH_INCLUDE_DIR "${XXHASH_INSTALL_DIR}/include")
IF(WITH_STATIC_LIB)
SET(BUILD_CMD make lib)
ELSE()
IF(APPLE)
SET(BUILD_CMD sed -i \"\" "s/-Wstrict-prototypes -Wundef/-Wstrict-prototypes -Wundef -fPIC/g" ${XXHASH_SOURCE_DIR}/src/extern_xxhash/Makefile && make lib)
ELSE(APPLE)
SET(BUILD_CMD sed -i "s/-Wstrict-prototypes -Wundef/-Wstrict-prototypes -Wundef -fPIC/g" ${XXHASH_SOURCE_DIR}/src/extern_xxhash/Makefile && make lib)
ENDIF(APPLE)
ENDIF()
ExternalProject_Add(
extern_xxhash
${EXTERNAL_PROJECT_LOG_ARGS}
GIT_REPOSITORY "https://github.com/Cyan4973/xxHash"
GIT_TAG "v0.6.5"
PREFIX ${XXHASH_SOURCE_DIR}
DOWNLOAD_NAME "xxhash"
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
BUILD_IN_SOURCE 1
PATCH_COMMAND
BUILD_COMMAND ${BUILD_CMD}
INSTALL_COMMAND export PREFIX=${XXHASH_INSTALL_DIR}/ && make install
TEST_COMMAND ""
)
set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/libxxhash.a")
INCLUDE_DIRECTORIES(${XXHASH_INCLUDE_DIR})
add_library(xxhash STATIC IMPORTED GLOBAL)
set_property(TARGET xxhash PROPERTY IMPORTED_LOCATION ${XXHASH_LIBRARIES})
include_directories(${XXHASH_INCLUDE_DIR})
add_dependencies(xxhash extern_xxhash)
LIST(APPEND external_project_dependencies xxhash)
IF(WITH_C_API)
INSTALL(DIRECTORY ${XXHASH_INCLUDE_DIR} DESTINATION third_party/xxhash)
IF(ANDROID)
INSTALL(FILES ${XXHASH_LIBRARIES} DESTINATION third_party/xxhash/lib/${ANDROID_ABI})
ELSE()
INSTALL(FILES ${XXHASH_LIBRARIES} DESTINATION third_party/xxhash/lib)
ENDIF()
ENDIF()
cmake/inference_lib.cmake
浏览文件 @ b94f9b16
......@@ -31,7 +31,7 @@ function(copy TARGET)
foreach(index RANGE ${len})
list(GET copy_lib_SRCS ${index} src)
list(GET copy_lib_DSTS ${index} dst)
add_custom_command(TARGET ${TARGET} PRE_BUILD
add_custom_command(TARGET ${TARGET} PRE_BUILD
COMMAND mkdir -p "${dst}"
COMMAND cp -r "${src}" "${dst}"
COMMENT "copying ${src} -> ${dst}")
......@@ -67,6 +67,13 @@ copy(boost_lib
DEPS boost
)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/xxhash")
copy(xxhash_lib
SRCS ${XXHASH_INCLUDE_DIR} ${XXHASH_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS xxhash
)
if(NOT PROTOBUF_FOUND)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/protobuf")
copy(protobuf_lib
......@@ -186,7 +193,7 @@ copy(cmake_cache
DSTS ${FLUID_INSTALL_DIR})
# This command generates a complete fluid library for both train and inference
add_custom_target(fluid_lib_dist DEPENDS ${fluid_lib_dist_dep})
add_custom_target(fluid_lib_dist DEPENDS ${fluid_lib_dist_dep})
# Following commands generate a inference-only fluid library
# third_party, version.txt and CMakeCache.txt are the same position with ${FLUID_INSTALL_DIR}
......
paddle/CMakeLists.txt
浏览文件 @ b94f9b16
......@@ -24,6 +24,7 @@ if(NOT WITH_FLUID_ONLY)
endif()
add_subdirectory(testing)
set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests CACHE INTERNAL "python tests directory")
if(NOT MOBILE_INFERENCE AND NOT RPI AND NOT WITH_C_API)
add_subdirectory(fluid)
endif()
paddle/fluid/API.spec
浏览文件 @ b94f9b16
......@@ -176,6 +176,7 @@ paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label'
paddle.fluid.layers.maxout ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.sequence_reverse ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.affine_channel ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None))
paddle.fluid.layers.hash ArgSpec(args=['input', 'hash_size', 'num_hash', 'name'], varargs=None, keywords=None, defaults=(1, None))
paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None)
......@@ -354,6 +355,8 @@ paddle.fluid.optimizer.ModelAverage.__init__ ArgSpec(args=['self', 'average_wind
paddle.fluid.optimizer.ModelAverage.apply ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
paddle.fluid.optimizer.ModelAverage.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.optimizer.ModelAverage.restore ArgSpec(args=['self', 'executor'], varargs=None, keywords=None, defaults=None)
paddle.fluid.optimizer.LarsMomentumOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'momentum', 'lars_coeff', 'lars_weight_decay', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.001, 0.0005, None, None))
paddle.fluid.optimizer.LarsMomentumOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.backward.append_backward ArgSpec(args=['loss', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.regularizer.L1DecayRegularizer.__init__ ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,))
paddle.fluid.regularizer.L2DecayRegularizer.__init__ ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,))
......
paddle/fluid/framework/attribute.cc
浏览文件 @ b94f9b16
......@@ -64,6 +64,13 @@ Attribute GetAttrValue(const proto::OpDesc::Attr& attr_desc) {
case proto::AttrType::LONG: {
return attr_desc.l();
}
case proto::AttrType::LONGS: {
std::vector<int64_t> val(attr_desc.longs_size());
for (int i = 0; i < attr_desc.longs_size(); ++i) {
val[i] = attr_desc.longs(i);
}
return val;
}
default:
PADDLE_THROW("Unsupport attr type %d", attr_desc.type());
}
......
paddle/fluid/framework/attribute.h
浏览文件 @ b94f9b16
......@@ -26,6 +26,113 @@ limitations under the License. */
namespace paddle {
namespace framework {
template <typename T>
struct ExtractAttribute {
explicit ExtractAttribute(const std::string& attr_name)
: attr_name_(attr_name) {}
T* operator()(Attribute& attr) const {
T* attr_value = nullptr;
try {
attr_value = &boost::get<T>(attr);
} catch (boost::bad_get& bad_get) {
PADDLE_THROW("Cannot get attribute %s by type %s, its type is %s",
attr_name_, paddle::platform::demangle(typeid(T).name()),
paddle::platform::demangle(attr.type().name()));
}
return attr_value;
}
const std::string& attr_name_;
};
// special handle bool
// FIXME(yuyang18): Currently we cast bool into int in python binding. It is
// hard to change the logic there. In another way, we should correct handle
// if the user set `some_flag=1`.
//
// FIX ME anytime if there is a better solution.
template <>
struct ExtractAttribute<bool> {
explicit ExtractAttribute(const std::string& attr_name)
: attr_name_(attr_name) {}
bool* operator()(Attribute& attr) const {
if (attr.type() == typeid(int)) { // NOLINT
int val = boost::get<int>(attr);
attr = static_cast<bool>(val);
} else if (attr.type() == typeid(float)) { // NOLINT
float val = boost::get<float>(attr);
attr = static_cast<bool>(val);
}
bool* attr_value = nullptr;
try {
attr_value = &boost::get<bool>(attr);
} catch (boost::bad_get& bad_get) {
PADDLE_THROW("Cannot get attribute %s by type bool, its type is %s",
attr_name_, paddle::platform::demangle(attr.type().name()));
}
return attr_value;
}
const std::string& attr_name_;
};
template <>
struct ExtractAttribute<int64_t> {
explicit ExtractAttribute(const std::string& attr_name)
: attr_name_(attr_name) {}
int64_t* operator()(Attribute& attr) const {
if (attr.type() == typeid(int)) { // NOLINT
int val = boost::get<int>(attr);
attr = static_cast<int64_t>(val);
} else if (attr.type() == typeid(float)) { // NOLINT
int val = boost::get<float>(attr);
attr = static_cast<int64_t>(val);
}
int64_t* attr_value = nullptr;
try {
attr_value = &boost::get<int64_t>(attr);
} catch (boost::bad_get& bad_get) {
PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s",
attr_name_, paddle::platform::demangle(attr.type().name()));
}
return attr_value;
}
const std::string& attr_name_;
};
template <>
struct ExtractAttribute<std::vector<int64_t>> {
explicit ExtractAttribute(const std::string& attr_name)
: attr_name_(attr_name) {}
std::vector<int64_t>* operator()(Attribute& attr) const {
if (attr.type() == typeid(std::vector<int>)) { // NOLINT
std::vector<int> val = boost::get<std::vector<int>>(attr);
std::vector<int64_t> vec(val.begin(), val.end());
attr = vec;
} else if (attr.type() == typeid(std::vector<float>)) { // NOLINT
std::vector<float> val = boost::get<std::vector<float>>(attr);
std::vector<int64_t> vec(val.begin(), val.end());
attr = vec;
}
std::vector<int64_t>* attr_value = nullptr;
try {
attr_value = &boost::get<std::vector<int64_t>>(attr);
} catch (boost::bad_get& bad_get) {
PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s",
attr_name_, paddle::platform::demangle(attr.type().name()));
}
return attr_value;
}
const std::string& attr_name_;
};
template <typename T>
inline proto::AttrType AttrTypeID() {
Attribute tmp = T();
......@@ -42,7 +149,11 @@ class AttrReader {
inline const T& Get(const std::string& name) const {
PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap",
name);
return boost::get<T>(attrs_.at(name));
Attribute& attr = const_cast<Attribute&>(attrs_.at(name));
ExtractAttribute<T> extract_attr(name);
T* attr_value = extract_attr(attr);
return *attr_value;
}
private:
......@@ -82,7 +193,7 @@ class DefaultValueSetter {
public:
explicit DefaultValueSetter(T default_value)
: default_value_(default_value) {}
void operator()(T& value) const { value = default_value_; }
void operator()(T& value) const { value = default_value_; } // NOLINT
private:
T default_value_;
......@@ -117,84 +228,6 @@ class EnumInContainer {
std::unordered_set<T> container_;
};
template <typename T>
struct ExtractAttribute {
explicit ExtractAttribute(const std::string& attr_name)
: attr_name_(attr_name) {}
T* operator()(Attribute& attr) const {
T* attr_value = nullptr;
try {
attr_value = &boost::get<T>(attr);
} catch (boost::bad_get& bad_get) {
PADDLE_THROW("Cannot get attribute %s by type %s, its type is %s",
attr_name_, paddle::platform::demangle(typeid(T).name()),
paddle::platform::demangle(attr.type().name()));
}
return attr_value;
}
const std::string& attr_name_;
};
// special handle bool
// FIXME(yuyang18): Currently we cast bool into int in python binding. It is
// hard to change the logic there. In another way, we should correct handle
// if the user set `some_flag=1`.
//
// FIX ME anytime if there is a better solution.
template <>
struct ExtractAttribute<bool> {
explicit ExtractAttribute(const std::string& attr_name)
: attr_name_(attr_name) {}
bool* operator()(Attribute& attr) const {
if (attr.type() == typeid(int)) { // NOLINT
int val = boost::get<int>(attr);
attr = static_cast<bool>(val);
} else if (attr.type() == typeid(float)) { // NOLINT
float val = boost::get<float>(attr);
attr = static_cast<bool>(val);
}
bool* attr_value = nullptr;
try {
attr_value = &boost::get<bool>(attr);
} catch (boost::bad_get& bad_get) {
PADDLE_THROW("Cannot get attribute %s by type bool, its type is %s",
attr_name_, paddle::platform::demangle(attr.type().name()));
}
return attr_value;
}
const std::string& attr_name_;
};
template <>
struct ExtractAttribute<int64_t> {
explicit ExtractAttribute(const std::string& attr_name)
: attr_name_(attr_name) {}
int64_t* operator()(Attribute& attr) const {
if (attr.type() == typeid(int)) { // NOLINT
int val = boost::get<int>(attr);
attr = static_cast<int64_t>(val);
} else if (attr.type() == typeid(float)) { // NOLINT
int val = boost::get<float>(attr);
attr = static_cast<int64_t>(val);
}
int64_t* attr_value = nullptr;
try {
attr_value = &boost::get<int64_t>(attr);
} catch (boost::bad_get& bad_get) {
PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s",
attr_name_, paddle::platform::demangle(attr.type().name()));
}
return attr_value;
}
const std::string& attr_name_;
};
// check whether a certain attribute fit its limits
// an attribute can have more than one limits
template <typename T>
......@@ -235,7 +268,7 @@ class TypedAttrChecker {
return *this;
}
void operator()(AttributeMap& attr_map) const {
void operator()(AttributeMap& attr_map) const { // NOLINT
if (!attr_map.count(attr_name_)) {
// user do not set this attr
PADDLE_ENFORCE(!default_value_setter_.empty(),
......@@ -271,7 +304,7 @@ class OpAttrChecker {
return *(checker.target<TypedAttrChecker<T>>());
}
void Check(AttributeMap& attr_map) const {
void Check(AttributeMap& attr_map) const { // NOLINT
for (const auto& checker : attr_checkers_) {
checker(attr_map);
}
......
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @ b94f9b16
......@@ -16,12 +16,14 @@ if(WITH_GPU)
dynload_cuda variable_visitor)
nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim dynload_cuda)
nv_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor dynload_cuda)
nv_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle)
else()
cc_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
variable_visitor)
cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim)
cc_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
cc_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle)
endif()
cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_base scope lod_tensor)
......@@ -34,7 +36,7 @@ if(WITH_GPU)
endif()
cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle
scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle)
scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle)
if(WITH_GPU)
cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto reference_count_pass)
......@@ -58,4 +60,4 @@ cc_library(fast_threaded_ssa_graph_executor SRCS fast_threaded_ssa_graph_executo
cc_library(build_strategy SRCS build_strategy.cc DEPS
graph_viz_pass multi_devices_graph_pass
multi_devices_graph_print_pass multi_devices_graph_check_pass
fuse_elewise_add_act_pass)
fuse_elewise_add_act_pass multi_batch_merge_pass)
paddle/fluid/framework/details/broadcast_op_handle.cc
浏览文件 @ b94f9b16
......@@ -48,16 +48,27 @@ void BroadcastOpHandle::RunImpl() {
var_scopes.emplace_back(s->FindVar(kLocalExecScopeName)->Get<Scope *>());
}
BroadcastOneVar(*in_var_handle, out_var_handles, var_scopes);
}
void BroadcastOpHandle::BroadcastOneVar(
const VarHandle &in_var_handle,
const std::vector<VarHandle *> &out_var_handles,
const std::vector<const Scope *> &var_scopes) {
auto *in_var =
var_scopes.at(in_var_handle->scope_idx_)->FindVar(in_var_handle->name_);
var_scopes.at(in_var_handle.scope_idx_)->FindVar(in_var_handle.name_);
PADDLE_ENFORCE_NOT_NULL(in_var);
Tensor &in_tensor = VariableVisitor::GetMutableTensor(in_var);
if (UNLIKELY(!in_tensor.IsInitialized())) {
VLOG(3) << "in var " << in_var_handle.name_ << "not inited, return!";
return;
}
InitOutputValue(*in_var_handle, out_var_handles);
InitOutputValue(in_var_handle, out_var_handles);
if (platform::is_cpu_place(in_tensor.place())) {
for (auto *out_var_handle : out_var_handles) {
if (out_var_handle->IsTheSameVar(*in_var_handle)) {
if (out_var_handle->IsTheSameVar(in_var_handle)) {
continue;
}
auto &out_p = out_var_handle->place_;
......@@ -114,12 +125,12 @@ void BroadcastOpHandle::RunImpl() {
}
}
if (!out_handle->IsTheSameVar(*in_var_handle)) {
auto out_var = var_scopes.at(in_var_handle->scope_idx_)
if (!out_handle->IsTheSameVar(in_var_handle)) {
auto out_var = var_scopes.at(in_var_handle.scope_idx_)
->FindVar(out_var_handles[0]->name_);
paddle::framework::TensorCopy(
in_tensor, in_var_handle->place_,
*(dev_ctxes_.at(in_var_handle->place_)),
in_tensor, in_var_handle.place_,
*(dev_ctxes_.at(in_var_handle.place_)),
&VariableVisitor::GetMutableTensor(out_var));
}
});
......
paddle/fluid/framework/details/broadcast_op_handle.h
浏览文件 @ b94f9b16
......@@ -61,7 +61,10 @@ struct BroadcastOpHandle : public OpHandleBase {
protected:
void RunImpl() override;
private:
void BroadcastOneVar(const VarHandle &in_var_handle,
const std::vector<VarHandle *> &out_var_handles,
const std::vector<const Scope *> &var_scopes);
std::vector<Scope *> local_scopes_;
std::vector<platform::Place> places_;
#ifdef PADDLE_WITH_CUDA
......
paddle/fluid/framework/details/build_strategy.cc
浏览文件 @ b94f9b16
......@@ -121,6 +121,7 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
USE_PASS(fuse_elewise_add_act_pass);
USE_PASS(graph_viz_pass);
USE_PASS(multi_batch_merge_pass);
USE_PASS(multi_devices_pass);
USE_PASS(multi_devices_check_pass);
USE_PASS(multi_devices_print_pass);
paddle/fluid/framework/details/build_strategy.h
浏览文件 @ b94f9b16
......@@ -69,6 +69,8 @@ struct BuildStrategy {
bool enable_data_balance_{false};
bool fuse_broadcast_op_{false};
// User normally doesn't need to call this API.
// The PassBuilder allows for more customized insert, remove of passes
// from python side.
......
paddle/fluid/framework/details/fused_broadcast_op_handle.cc 0 → 100644
浏览文件 @ b94f9b16
// 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/details/fused_broadcast_op_handle.h"
#include "paddle/fluid/framework/details/container_cast.h"
#include "paddle/fluid/framework/details/variable_visitor.h"
#include "paddle/fluid/platform/profiler.h"
namespace paddle {
namespace framework {
namespace details {
void FusedBroadcastOpHandle::RunImpl() {
platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second);
if (places_.size() == 1UL) return;
auto in_var_handles = DynamicCast<VarHandle>(inputs_);
auto out_var_handles = DynamicCast<VarHandle>(outputs_);
WaitInputVarGenerated();
std::vector<const Scope *> var_scopes;
for (auto *s : local_scopes_) {
var_scopes.emplace_back(s->FindVar(kLocalExecScopeName)->Get<Scope *>());
}
size_t place_num = places_.size();
PADDLE_ENFORCE_EQ(in_var_handles.size() * place_num, out_var_handles.size());
for (size_t i = 0; i < in_var_handles.size(); ++i) {
BroadcastOneVar(
*in_var_handles[i],
std::vector<VarHandle *>(out_var_handles.begin() + i * place_num,
out_var_handles.begin() + (i + 1) * place_num),
var_scopes);
}
}
std::string FusedBroadcastOpHandle::Name() const { return "fused_broadcast"; }
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/fused_broadcast_op_handle.h 0 → 100644
浏览文件 @ b94f9b16
// 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 <map>
#include <string>
#include <vector>
#include "paddle/fluid/framework/details/broadcast_op_handle.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/selected_rows.h"
#include "paddle/fluid/platform/device_context.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/nccl_helper.h"
#endif
namespace paddle {
namespace framework {
namespace details {
struct FusedBroadcastOpHandle : public BroadcastOpHandle {
public:
#ifdef PADDLE_WITH_CUDA
FusedBroadcastOpHandle(ir::Node *node,
const std::vector<Scope *> local_scopes,
const std::vector<platform::Place> &places,
const platform::NCCLContextMap *nccl_ctx)
: BroadcastOpHandle(node, local_scopes, places, nccl_ctx) {}
#else
FusedBroadcastOpHandle(ir::Node* node, const std::vector<Scope*> local_scopes,
const std::vector<platform::Place>& places)
: BroadcastOpHandle(node, local_scopes, places) {}
#endif
std::string Name() const override;
protected:
void RunImpl() override;
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/multi_devices_graph_pass.cc
浏览文件 @ b94f9b16
......@@ -21,6 +21,7 @@
#include "paddle/fluid/framework/details/broadcast_op_handle.h"
#include "paddle/fluid/framework/details/computation_op_handle.h"
#include "paddle/fluid/framework/details/data_balance_op_handle.h"
#include "paddle/fluid/framework/details/fused_broadcast_op_handle.h"
#include "paddle/fluid/framework/details/multi_devices_graph_pass.h"
#include "paddle/fluid/framework/details/reduce_op_handle.h"
#include "paddle/fluid/framework/details/rpc_op_handle.h"
......@@ -347,7 +348,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
BuildStrategy::GradientScaleStrategy::kCustomized) {
// TODO(paddle-dev): Why is there no input for this op_handle?
auto loss_grad_name = node->Op()->OutputArgumentNames()[0];
CreateScaleLossGradOp(&result, loss_grad_name);
CreateScaleLossGradOp(&result, loss_grad_name, node->outputs[0]);
}
// This assumes the backward generating code will ensure IsScaleLossOp
// is true only for the op that scale the final scalar loss.
......@@ -436,10 +437,14 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
if ((use_gpu &&
strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce) ||
is_dist_train) {
for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) {
auto &to_bcast_set = bcast_var_name_set[dev_id];
for (auto &bcast_name : to_bcast_set) {
CreateBroadcastOp(&result, bcast_name, dev_id);
if (strategy_.fuse_broadcast_op_) {
CreateFusedBroadcastOp(&result, bcast_var_name_set);
} else {
for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) {
auto &to_bcast_set = bcast_var_name_set[dev_id];
for (auto &bcast_name : to_bcast_set) {
CreateBroadcastOp(&result, bcast_name, dev_id);
}
}
}
}
......@@ -508,6 +513,44 @@ void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result,
}
}
void MultiDevSSAGraphBuilder::CreateFusedBroadcastOp(
ir::Graph *result,
const std::vector<std::unordered_set<std::string>> &bcast_varnames) const {
#ifdef PADDLE_WITH_CUDA
auto *op_handle = new FusedBroadcastOpHandle(
result->CreateEmptyNode("fused_broadcast", ir::Node::Type::kOperation),
local_scopes_, places_, nccl_ctxs_);
#else
auto *op_handle = new FusedBroadcastOpHandle(
result->CreateEmptyNode("fused_broadcast", ir::Node::Type::kOperation),
local_scopes_, places_);
#endif
result->Get<GraphOps>(kGraphOps).emplace_back(op_handle);
for (size_t i = 0; i < places_.size(); ++i) {
auto &p = places_[i];
SetCommunicationContext(op_handle, p);
}
for (size_t dev_id = 0; dev_id < bcast_varnames.size(); ++dev_id) {
for (auto &p_name : bcast_varnames[dev_id]) {
auto *in =
result->Get<GraphVars>(kGraphVars).at(dev_id).at(p_name).back().get();
op_handle->AddInput(in);
for (size_t out_dev_id = 0; out_dev_id < places_.size(); ++out_dev_id) {
auto &p = places_[out_dev_id];
auto &vars =
result->Get<GraphVars>(kGraphVars).at(out_dev_id).at(p_name);
auto *out_var = new VarHandle(
result->CreateEmptyNode(p_name, ir::Node::Type::kVariable),
vars.size(), out_dev_id, p_name, p);
vars.emplace_back(out_var);
op_handle->AddOutput(out_var);
}
}
}
}
void MultiDevSSAGraphBuilder::CreateComputationalOp(ir::Graph *result,
ir::Node *node,
int dev_id) const {
......@@ -602,7 +645,8 @@ int MultiDevSSAGraphBuilder::GetVarDeviceID(const ir::Graph &graph,
}
void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(
ir::Graph *result, const std::string &loss_grad_name) const {
ir::Graph *result, const std::string &loss_grad_name,
ir::Node *out_var_node) const {
for (size_t i = 0; i < places_.size(); ++i) {
// Insert ScaleCost OpHandle
auto *dev_ctx = platform::DeviceContextPool::Instance().Get(places_[i]);
......@@ -617,10 +661,8 @@ void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(
// loss->pending_ops_.emplace_back(op_handle);
// op_handle->inputs_.emplace_back(loss);
CreateOpOutput(
result, op_handle,
result->CreateEmptyNode(loss_grad_name, ir::Node::Type::kVariable),
places_[i], i);
CreateOpOutput(result, op_handle,
result->CreateVarNode(out_var_node->Var()), places_[i], i);
}
}
......@@ -680,7 +722,8 @@ int MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
}
if (node->Op()->Type() == "split_byref" ||
node->Op()->Type() == "split_selected_rows") {
node->Op()->Type() == "split_selected_rows" ||
node->Op()->Type() == "split_ids") {
// TODO(paddle-dev): getting the first var is not safe.
op_dev_id = GetVarDeviceID(*result, input_var_names[0]);
if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
......
paddle/fluid/framework/details/multi_devices_graph_pass.h
浏览文件 @ b94f9b16
......@@ -61,7 +61,8 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
size_t num_places) const;
void CreateScaleLossGradOp(ir::Graph *result,
const std::string &loss_grad_name) const;
const std::string &loss_grad_name,
ir::Node *out_var_node) const;
VarHandle *CreateReduceOp(ir::Graph *result, const std::string &og,
int dst_dev_id) const;
......@@ -78,6 +79,10 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
void CreateBroadcastOp(ir::Graph *result, const std::string &p_name,
size_t src_dev_id) const;
void CreateFusedBroadcastOp(
ir::Graph *result,
const std::vector<std::unordered_set<std::string>> &bcast_varnames) const;
bool IsSparseGradient(const std::string &og) const;
size_t GetAppropriateDeviceID(
......
paddle/fluid/framework/framework.proto
浏览文件 @ b94f9b16
......@@ -35,6 +35,7 @@ enum AttrType {
BLOCK = 8;
LONG = 9;
BLOCKS = 10;
LONGS = 11;
}
// OpDesc describes an instance of a C++ framework::OperatorBase
......@@ -55,6 +56,7 @@ message OpDesc {
optional int32 block_idx = 12;
optional int64 l = 13;
repeated int32 blocks_idx = 14;
repeated int64 longs = 15;
};
message Var {
......
paddle/fluid/framework/ir/CMakeLists.txt
浏览文件 @ b94f9b16
......@@ -36,6 +36,7 @@ pass_library(fc_lstm_fuse_pass inference)
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(multi_batch_merge_pass base)
pass_library(conv_bn_fuse_pass inference)
pass_library(seqconv_eltadd_relu_fuse_pass inference)
if(WITH_MKLDNN)
......
paddle/fluid/framework/ir/graph.cc
浏览文件 @ b94f9b16
......@@ -89,14 +89,20 @@ bool IsDistTrainOp(ir::Node *node, const std::vector<std::string> &send_vars,
Graph::Graph(const ProgramDesc &program) : program_(program) {
// Make the nodes id start from 0.
Node::ResetId();
auto var_nodes = InitFromProgram(program_);
ResolveHazard(var_nodes);
}
std::map<std::string, std::vector<ir::Node *>> Graph::InitFromProgram(
const ProgramDesc &program) {
VLOG(3) << "block in program:" << program_.Size();
std::unordered_map<std::string, VarDesc *> all_vars;
// var nodes for each var name, will have multiple versions in SSA
std::map<std::string, std::vector<ir::Node *>> var_nodes;
for (auto *var : program.Block(0).AllVars()) {
all_vars.emplace(var->Name(), var);
}
std::map<std::string, std::vector<ir::Node *>> var_nodes;
for (auto *op : program.Block(0).AllOps()) {
ir::Node *node = CreateOpNode(op);
// For input args, reuse the same var name if it was created before.
......@@ -134,7 +140,11 @@ Graph::Graph(const ProgramDesc &program) : program_(program) {
var->inputs.push_back(node);
}
}
return std::move(var_nodes);
}
void Graph::ResolveHazard(
const std::map<std::string, std::vector<ir::Node *>> &var_nodes) {
/**
* We should handle write after read(WAR) and write after write(WAW) here.
* Because some of the operators of the program can be executed parallelly.
......@@ -153,6 +163,7 @@ Graph::Graph(const ProgramDesc &program) : program_(program) {
auto it_old = versions.rbegin();
++it_old;
for (; it_old != versions.rend(); it_new = it_old, ++it_old) {
VLOG(3) << "deal with var: " << (*it_new)->Name();
ir::Node *write_op =
(*it_new)->inputs.empty() ? nullptr : (*it_new)->inputs[0];
const auto &read_ops = (*it_old)->outputs;
......
paddle/fluid/framework/ir/graph.h
浏览文件 @ b94f9b16
......@@ -160,6 +160,12 @@ class Graph {
return nullptr;
}
std::map<std::string, std::vector<ir::Node *>> InitFromProgram(
const ProgramDesc &program);
void ResolveHazard(
const std::map<std::string, std::vector<ir::Node *>> &var_nodes);
private:
// This method takes ownership of `node`.
ir::Node *AddNode(ir::Node *node) {
......
paddle/fluid/framework/ir/graph_helper.cc
浏览文件 @ b94f9b16
......@@ -120,19 +120,25 @@ size_t GraphNum(const Graph &graph) {
std::deque<ir::Node *> q_nodes;
std::vector<std::unordered_set<ir::Node *>> graph_nodes;
std::unordered_set<ir::Node *> g_nodes;
// q_set used to record records in the queue.
std::unordered_set<ir::Node *> q_set;
size_t graph_count = 0;
auto traverse_nodes = [&visited_nodes,
&q_nodes](const std::vector<ir::Node *> &nodes) {
std::copy_if(
nodes.begin(), nodes.end(), std::back_inserter(q_nodes),
[&visited_nodes](Node *node) { return !visited_nodes.count(node); });
auto traverse_nodes = [&visited_nodes, &q_nodes,
&q_set](const std::vector<ir::Node *> &nodes) {
for (auto n : nodes) {
if (visited_nodes.count(n) == 0 && q_set.count(n) == 0) {
q_nodes.push_back(n);
q_set.insert(n);
}
}
};
while (visited_nodes.size() != nodes.size()) {
if (!q_nodes.empty()) {
auto cur_node = q_nodes.front();
q_nodes.pop_front();
q_set.erase(cur_node);
visited_nodes.insert(cur_node);
g_nodes.insert(cur_node);
traverse_nodes(cur_node->inputs);
......@@ -146,6 +152,7 @@ size_t GraphNum(const Graph &graph) {
for (auto &n : nodes) {
if (visited_nodes.count(n) == 0) {
q_nodes.push_back(n);
q_set.insert(n);
break;
}
}
......
paddle/fluid/framework/ir/multi_batch_merge_pass.cc 0 → 100644
浏览文件 @ b94f9b16
// 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/multi_batch_merge_pass.h"
#include <map>
#include <string>
#include <vector>
#include "paddle/fluid/framework/ir/graph_helper.h"
#include "paddle/fluid/framework/op_proto_maker.h"
namespace paddle {
namespace framework {
namespace ir {
static const char kNumRepeats[] = "num_repeats";
typedef std::unordered_map<std::string, std::vector<ir::Node*>> SSAVarList;
ir::Node* SameNameVar(std::unordered_set<ir::Node*> all, ir::Node* target) {
for (auto n : all) {
if (target->IsVar() && target->Name() == n->Name()) {
return n;
}
}
return nullptr;
}
VarDesc CopyVarDesc(VarDesc* var_desc) {
VarDesc repeated_var(var_desc->Name());
// copy other variable attributes
if (var_desc->GetType() != proto::VarType::READER) {
repeated_var.SetType(var_desc->GetType());
repeated_var.SetShape(var_desc->GetShape());
repeated_var.SetDataType(var_desc->GetDataType());
repeated_var.SetLoDLevel(var_desc->GetLoDLevel());
repeated_var.SetPersistable(var_desc->Persistable());
} else {
// TODO(typhoonzero): copy reader var
}
return repeated_var;
}
VarDesc UpdateGradVarDesc(
VarDesc* var_desc, int repeat,
const std::unordered_set<std::string>& grad_names,
const std::unordered_set<std::string>& bn_vars_need_rename) {
if (grad_names.find(var_desc->Name()) != grad_names.end() ||
bn_vars_need_rename.find(var_desc->Name()) != bn_vars_need_rename.end()) {
std::string new_gname =
string::Sprintf("%s.repeat.%d", var_desc->Name(), repeat);
VarDesc repeated_var = CopyVarDesc(var_desc);
repeated_var.SetName(new_gname);
VLOG(3) << "update " << var_desc->Name() << " to repeat " << repeat;
return repeated_var;
}
return *var_desc;
}
std::unique_ptr<Graph> BatchMergePass::ApplyImpl(
std::unique_ptr<Graph> graph) const {
int num_repeats = Get<const int>(kNumRepeats);
std::vector<Node*> forward_backward_ops;
std::vector<Node*> optimize_ops;
std::vector<Node*> lr_ops; // ops other than forward/backward/optimize
std::unordered_set<std::string> grad_names;
std::vector<ir::Node*> nodes = TopologySortOperations(*graph);
auto origin_nodes = graph->ReleaseNodes();
VLOG(3) << "origin nodes count: " << origin_nodes.size();
ir::Graph& result = *graph;
// 1. record op nodes of different roles
for (auto node : nodes) {
if (node->IsVar()) continue;
int op_role = boost::get<int>(node->Op()->GetAttr(
framework::OpProtoAndCheckerMaker::OpRoleAttrName()));
if ((op_role == static_cast<int>(framework::OpRole::kForward)) ||
(op_role & static_cast<int>(framework::OpRole::kBackward)) ||
(op_role & static_cast<int>(framework::OpRole::kLoss))) {
forward_backward_ops.push_back(node);
} else if ((op_role & static_cast<int>(framework::OpRole::kOptimize)) ||
(op_role & static_cast<int>(framework::OpRole::kDist)) ||
(op_role & static_cast<int>(framework::OpRole::kRPC))) {
optimize_ops.push_back(node);
auto op_role_var = node->Op()->GetNullableAttr(
OpProtoAndCheckerMaker::OpRoleVarAttrName());
auto op_role_vars = boost::get<std::vector<std::string>>(op_role_var);
for (size_t i = 0; i < op_role_vars.size(); i += 2) {
grad_names.insert(op_role_vars[i + 1]);
}
} else if (op_role & static_cast<int>(framework::OpRole::kLRSched)) {
lr_ops.push_back(node);
} else { // NOLINT
PADDLE_THROW("Invalid op_role: %d", static_cast<int>(op_role));
}
}
// 2. copy forward backward
ir::Node* prev_repeat_last_op_node = nullptr;
// record origin_grad -> repeated grad list map.
std::map<ir::Node*, std::vector<ir::Node*>> grad_repeated_map;
std::map<std::string, std::vector<ir::Node*>> created;
std::unordered_set<std::string> bn_vars_need_rename;
for (int i = 0; i < num_repeats; ++i) {
std::unordered_set<ir::Node*> copied;
for (size_t node_idx = 0; node_idx < forward_backward_ops.size();
++node_idx) {
auto node = forward_backward_ops[node_idx];
OpDesc repeated_op(*(node->Op()), node->Op()->Block());
// 3. rename grad outputs to current repeat.
for (auto outname : repeated_op.OutputArgumentNames()) {
if (grad_names.find(outname) != grad_names.end()) {
std::string new_gname = string::Sprintf("%s.repeat.%d", outname, i);
repeated_op.RenameOutput(outname, new_gname);
}
}
// 3.5 let batch_norm ops use independent vars, note batch_norm_grad do
// not need this update
if (node->Name() == "batch_norm") {
// NOTE: assume bn op created by layers use save var as output mean and
// variance
std::string new_mean_name =
string::Sprintf("%s.repeat.%d", repeated_op.Input("Mean")[0], i);
std::string new_var_name = string::Sprintf(
"%s.repeat.%d", repeated_op.Input("Variance")[0], i);
bn_vars_need_rename.insert(repeated_op.Input("Mean")[0]);
bn_vars_need_rename.insert(repeated_op.Input("Variance")[0]);
VLOG(3) << "renaming " << repeated_op.Input("Mean")[0] << " to "
<< new_mean_name;
repeated_op.RenameInput(repeated_op.Input("Mean")[0], new_mean_name);
repeated_op.RenameInput(repeated_op.Input("Variance")[0], new_var_name);
repeated_op.RenameOutput(repeated_op.Output("MeanOut")[0],
new_mean_name);
repeated_op.RenameOutput(repeated_op.Output("VarianceOut")[0],
new_var_name);
}
// 3.9 do copy
auto repeated_node = result.CreateOpNode(&repeated_op);
copied.insert(node);
// 4. add deps between repeats
if (node_idx == forward_backward_ops.size() - 1) {
prev_repeat_last_op_node = repeated_node;
}
if (node_idx == 0 && prev_repeat_last_op_node) {
auto* depvar = result.CreateControlDepVar();
prev_repeat_last_op_node->outputs.push_back(depvar);
depvar->inputs.push_back(prev_repeat_last_op_node);
repeated_node->inputs.push_back(depvar);
depvar->outputs.push_back(repeated_node);
}
for (auto in_node : node->inputs) {
if (in_node->IsCtrlVar()) {
continue;
}
ir::Node* var = nullptr;
auto updated_var = UpdateGradVarDesc(in_node->Var(), i, grad_names,
bn_vars_need_rename);
// should be initialized by startup, how to initilize tensor in the
// scope?
if (node->Name() == "batch_norm" &&
bn_vars_need_rename.find(in_node->Name()) !=
bn_vars_need_rename.end()) {
// Create bn mean/variance for each repeat
var = result.CreateVarNode(&updated_var);
created[updated_var.Name()].push_back(var);
copied.insert(in_node);
repeated_node->inputs.push_back(var);
var->outputs.push_back(repeated_node);
continue;
}
// for other ops
if (in_node->inputs.empty() && i > 0) {
// do not copy head vars (inputs, params) in repeats > 0
var = created.at(in_node->Name()).back();
} else {
if (copied.find(in_node) == copied.end()) {
var = result.CreateVarNode(&updated_var);
if (grad_names.find(in_node->Var()->Name()) != grad_names.end()) {
grad_repeated_map[in_node].push_back(var);
}
copied.insert(in_node);
created[updated_var.Name()].push_back(var);
} else {
var = created.at(updated_var.Name()).back();
}
}
repeated_node->inputs.push_back(var);
var->outputs.push_back(repeated_node);
}
for (auto out_node : node->outputs) {
if (out_node->IsCtrlVar()) {
continue;
}
ir::Node* var = nullptr;
auto updated_var = UpdateGradVarDesc(out_node->Var(), i, grad_names,
bn_vars_need_rename);
if (copied.find(out_node) == copied.end()) {
var = result.CreateVarNode(&updated_var);
if (grad_names.find(out_node->Var()->Name()) != grad_names.end()) {
grad_repeated_map[out_node].push_back(var);
}
copied.insert(out_node);
created[updated_var.Name()].push_back(var);
} else {
var = created.at(updated_var.Name()).back();
}
repeated_node->outputs.push_back(var);
var->inputs.push_back(repeated_node);
}
}
}
// 5. create GRAD merge op node
for (auto kv : grad_repeated_map) {
OpDesc sum_op;
sum_op.SetType("sum");
std::vector<std::string> repeated_grad_names;
for (auto r : kv.second) {
repeated_grad_names.push_back(r->Var()->Name());
}
sum_op.SetInput("X", repeated_grad_names);
sum_op.SetOutput("Out", {kv.first->Var()->Name()});
sum_op.SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(),
static_cast<int>(OpRole::kBackward));
auto sum_op_node = result.CreateOpNode(&sum_op);
for (auto r : kv.second) {
sum_op_node->inputs.push_back(r);
r->outputs.push_back(sum_op_node);
}
auto sum_out_var_node = result.CreateVarNode(kv.first->Var());
sum_op_node->outputs.push_back(sum_out_var_node);
sum_out_var_node->inputs.push_back(sum_op_node);
created[sum_out_var_node->Name()].push_back(sum_out_var_node);
OpDesc scale_op;
scale_op.SetType("scale");
scale_op.SetInput("X", {sum_out_var_node->Var()->Name()});
// NOTE: inplace scale.
scale_op.SetOutput("Out", {sum_out_var_node->Var()->Name()});
scale_op.SetAttr("scale", static_cast<float>(1.0f / num_repeats));
scale_op.SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(),
static_cast<int>(OpRole::kBackward));
auto scale_op_node = result.CreateOpNode(&scale_op);
scale_op_node->inputs.push_back(sum_out_var_node);
sum_out_var_node->outputs.push_back(scale_op_node);
auto scale_out_var_node = result.CreateVarNode(sum_out_var_node->Var());
scale_op_node->outputs.push_back(scale_out_var_node);
scale_out_var_node->inputs.push_back(scale_op_node);
created[scale_out_var_node->Name()].push_back(scale_out_var_node);
}
// 6. add optimize ops
{
auto copy_node = [&result, &created](ir::Node* node) {
auto op_node = result.CreateOpNode(node->Op());
// copy op ins/outs
// NOTE: for send/recv ops, the OpDesc uses ctrldepvar to describe
// dependencies, so create those depvars if OpDesc have in/outs.
for (auto in_node : node->inputs) {
if (in_node->IsCtrlVar() && !in_node->Var()) {
continue;
}
ir::Node* var = nullptr;
if (created.find(in_node->Name()) == created.end()) {
var = result.CreateVarNode(in_node->Var());
created[in_node->Name()].push_back(var);
} else {
var = created.at(in_node->Name()).back();
}
op_node->inputs.push_back(var);
var->outputs.push_back(op_node);
}
for (auto out_node : node->outputs) {
if (out_node->IsCtrlVar() && !out_node->Var()) {
continue;
}
auto var = result.CreateVarNode(out_node->Var());
created[out_node->Name()].push_back(var);
op_node->outputs.push_back(var);
var->inputs.push_back(op_node);
}
};
for (auto node : lr_ops) {
copy_node(node);
}
for (auto node : optimize_ops) {
copy_node(node);
}
}
result.ResolveHazard(created);
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(multi_batch_merge_pass, paddle::framework::ir::BatchMergePass)
.RequirePassAttr(paddle::framework::ir::kNumRepeats);
paddle/fluid/framework/ir/multi_batch_merge_pass.h 0 → 100644
浏览文件 @ b94f9b16
// 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/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace ir {
// BatchMergePass is used to copy forward and backward ops for several
// times to run several batches to simulate large batch size training
// as if we have more than 1 GPUs.
// User can define how many batches to run, gradients will be merged
// through those repeats, and then do optimization using merged gradients.
// This pass is extremely useful when doing large batch-size distributed
// sync training, we can simulate even large batch size as if we have more
// GPUs.
class BatchMergePass : public Pass {
public:
virtual ~BatchMergePass() {}
protected:
std::unique_ptr<Graph> ApplyImpl(std::unique_ptr<Graph> graph) const override;
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/lod_tensor.cc
浏览文件 @ b94f9b16
......@@ -418,7 +418,7 @@ void LoDTensor::MergeLoDTensor(
PADDLE_ENFORCE_EQ(new_lod.size(), lod.size());
for (size_t j = 0; j < lod.size(); ++j) {
auto &sub_lod = new_lod[j];
auto &offset = sub_lod.back();
size_t offset = sub_lod.back();
for (size_t k = 1; k < lod[j].size(); ++k) {
sub_lod.push_back(lod[j][k] + offset);
}
......
paddle/fluid/framework/lod_tensor_array.h
浏览文件 @ b94f9b16
......@@ -18,6 +18,8 @@ limitations under the License. */
namespace paddle {
namespace framework {
using LoDTensorArray = std::vector<LoDTensor>;
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/op_desc.cc
浏览文件 @ b94f9b16
......@@ -419,8 +419,15 @@ struct SetAttrDescVisitor : public boost::static_visitor<void> {
}
VectorToRepeated(blocks_idx, attr_->mutable_blocks_idx());
}
void operator()(BlockDesc *desc) const { attr_->set_block_idx(desc->ID()); }
void operator()(int64_t v) const { attr_->set_l(v); }
void operator()(const std::vector<int64_t> &v) const {
VectorToRepeated(v, attr_->mutable_longs());
}
void operator()(boost::blank) const { PADDLE_THROW("Unexpected branch"); }
};
......
paddle/fluid/framework/op_proto_maker.h
浏览文件 @ b94f9b16
......@@ -28,12 +28,12 @@ enum class OpRole {
kBackward = 0x0001,
kOptimize = 0x0002,
// RPC role is for send/recv releated op
kRPC = 0x0003,
kRPC = 0x0004,
// Dist role is for split_byref/split_selected_rows/concat
// used for distributed training.
kDist = 0x0004,
kDist = 0x0008,
// Tag all learning rate scheduler operators.
kLRSched = 0x0005,
kLRSched = 0x0010,
kLoss = 0x0100,
// The default value of op's role. This should be only used for unittests and
......
paddle/fluid/framework/operator.cc
浏览文件 @ b94f9b16
......@@ -358,11 +358,11 @@ static bool VarIsTensor(const Variable* var) {
return var->IsType<LoDTensor>() || var->IsType<SelectedRows>();
}
static const Tensor* GetTensorFromVar(Variable* var) {
const Tensor* GetTensorFromVar(const Variable* var) {
if (var->IsType<LoDTensor>()) {
return var->GetMutable<LoDTensor>();
return static_cast<const Tensor*>(&(var->Get<LoDTensor>()));
} else if (var->IsType<SelectedRows>()) {
return var->GetMutable<SelectedRows>()->mutable_value();
return &(var->Get<SelectedRows>().value());
} else {
PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.",
var->Type().name());
......@@ -415,8 +415,7 @@ bool ExecutionContext::HasOutput(const std::string& name) const {
template <>
const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const {
auto* var = InputVar(name);
return var == nullptr ? nullptr
: GetTensorFromVar(const_cast<Variable*>(var));
return var == nullptr ? nullptr : GetTensorFromVar(var);
}
template <>
......
paddle/fluid/framework/operator.h
浏览文件 @ b94f9b16
......@@ -63,6 +63,7 @@ inline std::string GradVarName(const std::string& var_name) {
}
proto::VarType::Type GetDataTypeOfVar(const Variable* var);
const Tensor* GetTensorFromVar(const Variable* var);
class OperatorBase;
class ExecutionContext;
......
paddle/fluid/framework/parallel_executor.cc
浏览文件 @ b94f9b16
......@@ -109,18 +109,9 @@ ParallelExecutor::ParallelExecutor(
if (member_->local_scopes_.size() != 1 && local_scopes.empty()) {
BCastParamsToDevices(bcast_vars);
}
// Startup Program has been run. All local scopes has correct parameters.
// Startup Program has been run. All local scopes has correct parameters.
// Step 2. Create vars in each scope;
std::vector<details::VariableInfo> var_infos;
for (auto *var : main_program.Block(0).AllVars()) {
var_infos.emplace_back();
var_infos.back().name_ = var->Name();
var_infos.back().type_ = var->GetType();
var_infos.back().persistable_ = var->Persistable();
}
// Step 3. Convert main_program to SSA form and dependency graph. Also, insert
// Step 2. Convert main_program to SSA form and dependency graph. Also, insert
// ncclOp
#ifdef PADDLE_WITH_CUDA
std::unique_ptr<ir::Graph> graph = build_strategy.Apply(
......@@ -156,6 +147,26 @@ ParallelExecutor::ParallelExecutor(
params, member_->local_scopes_, member_->use_cuda_);
#endif
// Step 3. Create vars in each scope. Passes may also create new vars.
// skip control vars and empty vars
std::vector<details::VariableInfo> var_infos;
for (auto &node : graph->Nodes()) {
if (node->IsVar() && !node->IsCtrlVar() && node->Var()) {
var_infos.emplace_back();
var_infos.back().name_ = node->Var()->Name();
var_infos.back().type_ = node->Var()->GetType();
var_infos.back().persistable_ = node->Var()->Persistable();
}
}
if (VLOG_IS_ON(5)) {
// If the loss_var_name is given, the number of graph should be only one.
if (loss_var_name.size()) {
PADDLE_ENFORCE_EQ(ir::GraphNum(*graph), 1,
"The number of graph should be only one");
}
}
if (exec_strategy.type_ == ExecutionStrategy::kDefault) {
member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
exec_strategy, member_->local_scopes_, places, std::move(graph)));
......@@ -179,6 +190,10 @@ void ParallelExecutor::BCastParamsToDevices(
}
auto &main_tensor = main_var->Get<LoDTensor>();
if (!main_tensor.IsInitialized()) {
VLOG(3) << "one in var not inited, return!";
continue;
}
auto &dims = main_tensor.dims();
if (paddle::platform::is_gpu_place(main_tensor.place())) {
#ifdef PADDLE_WITH_CUDA
......
paddle/fluid/framework/scope.h
浏览文件 @ b94f9b16
......@@ -78,6 +78,8 @@ class Scope {
/// Drop all kids scopes belonged to this scope.
void DropKids();
std::list<Scope*>& kids() const { return kids_; }
/// Find if a scope exists in the kid scopes
bool HasKid(const Scope* scope) const;
......
paddle/fluid/framework/type_defs.h
浏览文件 @ b94f9b16
......@@ -36,7 +36,7 @@ using Attribute =
boost::variant<boost::blank, int, float, std::string, std::vector<int>,
std::vector<float>, std::vector<std::string>, bool,
std::vector<bool>, BlockDesc*, int64_t,
std::vector<BlockDesc*>>;
std::vector<BlockDesc*>, std::vector<int64_t>>;
using AttributeMap = std::unordered_map<std::string, Attribute>;
......
paddle/fluid/inference/CMakeLists.txt
浏览文件 @ b94f9b16
......@@ -30,7 +30,7 @@ if (WITH_GPU AND TENSORRT_FOUND)
endif()
# Create static library
cc_library(paddle_fluid DEPS ${fluid_modules} ${STATIC_INFERENCE_APIS} zero_copy_tensor)
cc_library(paddle_fluid DEPS ${fluid_modules} ${STATIC_INFERENCE_APIS} zero_copy_tensor reset_tensor_array)
if(NOT APPLE)
# TODO(liuyiqu: Temporarily disable the link flag because it is not support on Mac.
......@@ -40,7 +40,7 @@ endif()
# Create shared library
cc_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS}
DEPS ${fluid_modules} paddle_fluid_api)
DEPS ${fluid_modules} paddle_fluid_api reset_tensor_array)
set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)
if(NOT APPLE)
......
paddle/fluid/inference/analysis/analyzer.cc
浏览文件 @ b94f9b16
......@@ -107,6 +107,9 @@ void Analyzer::Run(Argument* argument) {
passes.push_back("mkldnn_placement_pass");
}
#endif
// infer_clean_graph_pass should be the first default pass
// after mkldnn_placement_pass.
passes.push_back("infer_clean_graph_pass");
for (auto& pass : ir_passes_) {
if (!disabled_ir_passes_.count(pass)) {
passes.push_back(pass);
......
paddle/fluid/inference/analysis/analyzer.h
浏览文件 @ b94f9b16
......@@ -67,7 +67,6 @@ class Analyzer : public OrderedRegistry<PassManager> {
// larger fusion.
const std::vector<std::string> all_ir_passes_{{
// Manual update the passes here.
"infer_clean_graph_pass", //
"attention_lstm_fuse_pass", //
"seqconv_eltadd_relu_fuse_pass", //
"embedding_fc_lstm_fuse_pass", //
......
paddle/fluid/inference/api/CMakeLists.txt
浏览文件 @ b94f9b16
......@@ -18,7 +18,8 @@ if(APPLE)
endif(APPLE)
set(inference_deps paddle_inference_api paddle_fluid_api analysis pass ir_pass_manager naive_executor ${GLOB_PASS_LIB})
set(inference_deps paddle_inference_api paddle_fluid_api analysis pass ir_pass_manager naive_executor ${GLOB_PASS_LIB}
)
if(WITH_GPU AND TENSORRT_FOUND)
set(inference_deps ${inference_deps} paddle_inference_tensorrt_subgraph_engine analysis_predictor)
......@@ -31,10 +32,17 @@ function(inference_api_test TARGET_NAME)
set(multiValueArgs ARGS)
cmake_parse_arguments(inference_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
cc_test(${TARGET_NAME}
SRCS ${inference_test_SRC}
DEPS "${inference_deps}"
ARGS --dirname=${PYTHON_TESTS_DIR}/book/)
if (WITH_GPU)
cc_test(${TARGET_NAME}
SRCS ${inference_test_SRC}
DEPS "${inference_deps}"
ARGS --dirname=${PYTHON_TESTS_DIR}/book/ --fraction_of_gpu_memory_to_use=0.15)
else()
cc_test(${TARGET_NAME}
SRCS ${inference_test_SRC}
DEPS "${inference_deps}"
ARGS --dirname=${PYTHON_TESTS_DIR}/book/)
endif()
if(inference_test_ARGS)
set_tests_properties(${TARGET_NAME}
PROPERTIES DEPENDS "${inference_test_ARGS}")
......@@ -42,7 +50,8 @@ function(inference_api_test TARGET_NAME)
endif(WITH_TESTING)
endfunction(inference_api_test)
cc_library(paddle_inference_api SRCS api.cc api_impl.cc helper.cc DEPS lod_tensor scope)
cc_library(reset_tensor_array SRCS details/reset_tensor_array.cc DEPS lod_tensor scope)
cc_library(paddle_inference_api SRCS api.cc api_impl.cc helper.cc DEPS reset_tensor_array lod_tensor scope)
cc_library(analysis_predictor SRCS analysis_predictor.cc DEPS paddle_inference_api analysis naive_executor zero_copy_tensor)
cc_library(zero_copy_tensor SRCS details/zero_copy_tensor.cc DEPS paddle_inference_api)
cc_library(zero_copy_tensor_dummy SRCS details/zero_copy_tensor_dummy.cc DEPS paddle_inference_api)
......@@ -52,8 +61,6 @@ cc_test(test_paddle_inference_api
inference_api_test(test_api_impl SRC api_impl_tester.cc
ARGS test_word2vec test_image_classification)
set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests)
cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS analysis_predictor ${inference_deps} paddle_inference_api
ARGS --dirname=${PYTHON_TESTS_DIR}/book)
......
paddle/fluid/inference/api/analysis_predictor.cc
浏览文件 @ b94f9b16
......@@ -82,6 +82,7 @@ bool AnalysisPredictor::Init(
// Get the feed_target_names and fetch_target_names
PrepareFeedFetch();
return true;
}
......@@ -109,6 +110,10 @@ bool AnalysisPredictor::Run(const std::vector<PaddleTensor> &inputs,
return false;
}
VLOG(3) << "predict cost: " << timer.toc() << "ms";
// Fix TensorArray reuse not cleaned bug.
tensor_array_batch_cleaner_.CollectTensorArrays(scope_.get());
tensor_array_batch_cleaner_.ResetTensorArray();
return true;
}
......@@ -322,6 +327,9 @@ std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetOutputTensor(
bool AnalysisPredictor::ZeroCopyRun() {
executor_->Run();
// Fix TensorArray reuse not cleaned bug.
tensor_array_batch_cleaner_.CollectTensorArrays(scope_.get());
tensor_array_batch_cleaner_.ResetTensorArray();
return true;
}
......
paddle/fluid/inference/api/analysis_predictor.h
浏览文件 @ b94f9b16
......@@ -18,6 +18,7 @@
#include "paddle/fluid/framework/naive_executor.h"
#include "paddle/fluid/inference/analysis/analyzer.h"
#include "paddle/fluid/inference/api/api_impl.h"
#include "paddle/fluid/inference/api/details/reset_tensor_array.h"
#include "paddle/fluid/inference/api/paddle_inference_api.h"
#include "paddle/fluid/string/printf.h"
......@@ -88,6 +89,7 @@ class AnalysisPredictor : public PaddlePredictor {
// Memory buffer for feed inputs. The temporary LoDTensor will cause serious
// concurrency problems, so cache them.
std::vector<framework::LoDTensor> feed_tensors_;
details::TensorArrayBatchCleaner tensor_array_batch_cleaner_;
};
} // namespace paddle
paddle/fluid/inference/api/api_impl.cc
浏览文件 @ b94f9b16
......@@ -22,6 +22,7 @@ limitations under the License. */
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/inference/api/api_impl.h"
#include "paddle/fluid/inference/api/details/reset_tensor_array.h"
#include "paddle/fluid/inference/api/helper.h"
#include "paddle/fluid/platform/cpu_helper.h"
#include "paddle/fluid/platform/profiler.h"
......@@ -157,6 +158,10 @@ bool NativePaddlePredictor::Run(const std::vector<PaddleTensor> &inputs,
return false;
}
VLOG(3) << "predict cost: " << timer.toc() << "ms";
// Fix TensorArray reuse not cleaned bug.
tensor_array_batch_cleaner_.CollectTensorArrays(scope_.get());
tensor_array_batch_cleaner_.ResetTensorArray();
return true;
}
......
paddle/fluid/inference/api/api_impl.h
浏览文件 @ b94f9b16
......@@ -26,11 +26,11 @@ limitations under the License. */
#include <string>
#include <vector>
#include "paddle/fluid/inference/api/paddle_inference_api.h"
#include "paddle/fluid/framework/ddim.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/naive_executor.h"
#include "paddle/fluid/inference/api/details/reset_tensor_array.h"
#include "paddle/fluid/inference/api/paddle_inference_api.h"
#include "paddle/fluid/inference/io.h"
#include "paddle/fluid/platform/init.h"
......@@ -77,6 +77,7 @@ class NativePaddlePredictor : public PaddlePredictor {
std::vector<framework::OpDesc *> fetchs_;
// Do not use unique_ptr, use parent scope to delete
framework::Scope *sub_scope_{nullptr};
details::TensorArrayBatchCleaner tensor_array_batch_cleaner_;
};
} // namespace paddle
paddle/fluid/inference/api/demo_ci/CMakeLists.txt
浏览文件 @ b94f9b16
......@@ -52,6 +52,7 @@ include_directories("${PADDLE_LIB}")
include_directories("${PADDLE_LIB}/third_party/install/protobuf/include")
include_directories("${PADDLE_LIB}/third_party/install/glog/include")
include_directories("${PADDLE_LIB}/third_party/install/gflags/include")
include_directories("${PADDLE_LIB}/third_party/install/xxhash/include")
if (NOT WIN32)
include_directories("${PADDLE_LIB}/third_party/install/snappy/include")
include_directories("${PADDLE_LIB}/third_party/install/snappystream/include")
......@@ -61,8 +62,8 @@ endif(NOT WIN32)
include_directories("${PADDLE_LIB}/third_party/boost")
include_directories("${PADDLE_LIB}/third_party/eigen3")
if (NOT WIN32)
if (USE_TENSORRT AND WITH_GPU)
if (NOT WIN32)
if (USE_TENSORRT AND WITH_GPU)
include_directories("${TENSORRT_INCLUDE_DIR}")
link_directories("${TENSORRT_LIB_DIR}")
endif()
......@@ -77,13 +78,14 @@ endif(NOT WIN32)
link_directories("${PADDLE_LIB}/third_party/install/protobuf/lib")
link_directories("${PADDLE_LIB}/third_party/install/glog/lib")
link_directories("${PADDLE_LIB}/third_party/install/gflags/lib")
link_directories("${PADDLE_LIB}/third_party/install/xxhash/lib")
link_directories("${PADDLE_LIB}/paddle/lib")
add_executable(${DEMO_NAME} ${DEMO_NAME}.cc)
if(WITH_MKL)
include_directories("${PADDLE_LIB}/third_party/install/mklml/include")
set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX}
set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX}
${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5${CMAKE_SHARED_LIBRARY_SUFFIX})
set(MKLDNN_PATH "${PADDLE_LIB}/third_party/install/mkldnn")
if(EXISTS ${MKLDNN_PATH})
......@@ -107,7 +109,7 @@ if (NOT WIN32)
set(EXTERNAL_LIB "-lrt -ldl -lpthread")
set(DEPS ${DEPS}
${MATH_LIB} ${MKLDNN_LIB}
glog gflags protobuf snappystream snappy z
glog gflags protobuf snappystream snappy z xxhash
${EXTERNAL_LIB})
else()
set(DEPS ${DEPS}
......@@ -120,7 +122,7 @@ endif(NOT WIN32)
if(WITH_GPU)
if(NOT WIN32)
if (USE_TENSORRT)
if (USE_TENSORRT)
set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/libnvinfer${CMAKE_STATIC_LIBRARY_SUFFIX})
set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/libnvinfer_plugin${CMAKE_STATIC_LIBRARY_SUFFIX})
endif()
......
paddle/fluid/inference/api/demo_ci/run.sh
浏览文件 @ b94f9b16
......@@ -16,7 +16,7 @@ if [ $2 == ON ]; then
fi
if [ $3 == ON ]; then
use_gpu_list='true false'
else
else
use_gpu_list='false'
fi
......@@ -83,7 +83,7 @@ for WITH_STATIC_LIB in ON OFF; do
-DWITH_STATIC_LIB=$WITH_STATIC_LIB
make -j
for use_gpu in $use_gpu_list; do
for vis_demo_name in $vis_demo_list; do
for vis_demo_name in $vis_demo_list; do
./vis_demo \
--modeldir=$DATA_DIR/$vis_demo_name/model \
--data=$DATA_DIR/$vis_demo_name/data.txt \
......@@ -95,7 +95,7 @@ for WITH_STATIC_LIB in ON OFF; do
fi
done
done
# --------tensorrt mobilenet------
if [ $USE_TENSORRT == ON -a $TEST_GPU_CPU == ON ]; then
rm -rf *
......@@ -107,7 +107,7 @@ for WITH_STATIC_LIB in ON OFF; do
-DUSE_TENSORRT=$USE_TENSORRT \
-DTENSORRT_INCLUDE_DIR=$TENSORRT_INCLUDE_DIR \
-DTENSORRT_LIB_DIR=$TENSORRT_LIB_DIR
make -j
make -j
./trt_mobilenet_demo \
--modeldir=$DATA_DIR/mobilenet/model \
--data=$DATA_DIR/mobilenet/data.txt \
......
paddle/fluid/inference/api/details/reset_tensor_array.cc 0 → 100644
浏览文件 @ b94f9b16
// 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/inference/api/details/reset_tensor_array.h"
namespace paddle {
namespace details {
// Should be called after the parameters are loaded.
void TensorArrayBatchCleaner::CollectTensorArrays(framework::Scope *scope) {
if (flag_) {
for (auto &var_name : scope->LocalVarNames()) {
auto *var = scope->FindVar(var_name);
// TODO(Superjomn) should avoid the case when a TensorArray is a
// parameter.
if (var_name == "feed" || var_name == "fetch") continue;
if (var->Type() == typeid(framework::LoDTensorArray)) {
VLOG(4) << "collect " << var_name;
arrays_.push_back(var->GetMutable<framework::LoDTensorArray>());
}
}
for (auto *kid : scope->kids()) {
CollectTensorArrays(kid);
}
VLOG(3) << "Collect " << arrays_.size() << " arrays";
flag_ = false;
}
}
// Should be called when `Run` finished.
void TensorArrayBatchCleaner::ResetTensorArray() {
for (auto *arr : arrays_) {
arr->clear();
}
}
} // namespace details
} // namespace paddle
paddle/fluid/inference/api/details/reset_tensor_array.h 0 → 100644
浏览文件 @ b94f9b16
// 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 <vector>
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/scope.h"
namespace paddle {
namespace details {
// Clean the TensorArray each batch to make the behavior the same with the
// training phase.
struct TensorArrayBatchCleaner {
// Fix the tensor array not clear in the inference scenarios.
void CollectTensorArrays(framework::Scope *scope);
void ResetTensorArray();
private:
bool flag_{true};
std::vector<framework::LoDTensorArray *> arrays_;
};
} // namespace details
} // namespace paddle
paddle/fluid/inference/api/paddle_inference_api.h
浏览文件 @ b94f9b16
......@@ -124,7 +124,7 @@ class ZeroCopyTensor {
std::vector<std::vector<size_t>> lod() const;
protected:
ZeroCopyTensor(void* scope) : scope_{scope} {}
explicit ZeroCopyTensor(void* scope) : scope_{scope} {}
void SetName(const std::string& name) { name_ = name; }
void* FindTensor() const;
......@@ -259,12 +259,6 @@ 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.
......
paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc
浏览文件 @ b94f9b16
......@@ -228,6 +228,7 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
TEST(Analyzer_rnn1, profile) {
contrib::AnalysisConfig cfg;
SetConfig(&cfg);
cfg.use_gpu = false;
std::vector<PaddleTensor> outputs;
std::vector<std::vector<PaddleTensor>> input_slots_all;
......
paddle/fluid/operators/CMakeLists.txt
浏览文件 @ b94f9b16
......@@ -268,6 +268,7 @@ if (WITH_GPU AND TENSORRT_FOUND)
else()
set(DEPS_OPS ${DEPS_OPS} tensorrt_engine_op)
endif()
op_library(hash_op DEPS xxhash)
op_library(clip_by_norm_op DEPS selected_rows_functor selected_rows)
op_library(sum_op DEPS selected_rows_functor)
op_library(sgd_op DEPS selected_rows_functor)
......
paddle/fluid/operators/beam_search_decode_op.cc
浏览文件 @ b94f9b16
......@@ -79,6 +79,9 @@ struct BeamSearchDecodeFunctor {
bool tensor_on_gpu_;
size_t beam_size_;
int end_id_;
// TODO(Superjomn) Here might result serious performance issue in the
// concurrency
// scenarios.
const LoDTensorArray& step_ids_origin_;
const LoDTensorArray& step_scores_origin_;
LoDTensorArray step_ids_ = LoDTensorArray();
......
paddle/fluid/operators/detection/generate_proposals_op.cc
浏览文件 @ b94f9b16
......@@ -284,7 +284,7 @@ static inline Tensor NMS(const platform::DeviceContext &ctx, Tensor *bbox,
selected_indices.push_back(idx);
++selected_num;
}
sorted_indices.erase(sorted_indices.end());
sorted_indices.erase(sorted_indices.end() - 1);
if (flag && eta < 1 && adaptive_threshold > 0.5) {
adaptive_threshold *= eta;
}
......
paddle/fluid/operators/fake_init_op.cc 0 → 100644
浏览文件 @ b94f9b16
/* 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/data_type.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
namespace paddle {
namespace operators {
class FakeInitInferShape : public framework::InferShapeBase {
public:
void operator()(framework::InferShapeContext *ctx) const override {
PADDLE_ENFORCE(ctx->HasOutput("Out"),
"Output(Out) of FakeInitOp should not be null.");
auto &shape = ctx->Attrs().Get<std::vector<int64_t>>("shape");
ctx->SetOutputDim("Out", framework::make_ddim(shape));
}
};
class FakeInitOp : public framework::OperatorBase {
public:
using framework::OperatorBase::OperatorBase;
private:
void RunImpl(const framework::Scope &scope,
const platform::Place &dev_place) const override {
framework::Tensor *tensor = nullptr;
auto &out_var = *scope.FindVar(Output("Out"));
if (out_var.IsType<framework::LoDTensor>()) {
tensor = out_var.GetMutable<framework::LoDTensor>();
tensor->Resize(framework::make_ddim(Attr<std::vector<int64_t>>("shape")));
} else if (out_var.IsType<framework::SelectedRows>()) {
tensor = out_var.GetMutable<framework::SelectedRows>()->mutable_value();
tensor->Resize(framework::make_ddim(Attr<std::vector<int64_t>>("shape")));
} else {
PADDLE_THROW(
"fake init op's output only"
"supports SelectedRows and LoDTensor");
}
}
};
class FakeInitOpVarTypeInference : public framework::VarTypeInference {
public:
void operator()(const framework::OpDesc &op_desc,
framework::BlockDesc *block) const override {}
};
class FakeInitOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddAttr<std::vector<int64_t>>("shape",
"(vector<int64_t>) The shape of the output");
AddOutput("Out",
"(Tensor) Tensor of specified shape will be filled "
"with the specified value");
AddComment(R"DOC(
FakeInit Operator.
Init an variable but not alloc memory for it, it is used for init the
table parameter at trainer side in distributed lookup table.
)DOC");
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(fake_init, ops::FakeInitOp, ops::FakeInitInferShape,
ops::FakeInitOpMaker, paddle::framework::EmptyGradOpMaker,
ops::FakeInitOpVarTypeInference);
paddle/fluid/operators/fill_constant_op.cc
浏览文件 @ b94f9b16
......@@ -24,7 +24,7 @@ class FillConstantInferShape : public framework::InferShapeBase {
void operator()(framework::InferShapeContext *ctx) const override {
PADDLE_ENFORCE(ctx->HasOutput("Out"),
"Output(Out) of FillConstantOp should not be null.");
auto &shape = ctx->Attrs().Get<std::vector<int>>("shape");
auto &shape = ctx->Attrs().Get<std::vector<int64_t>>("shape");
ctx->SetOutputDim("Out", framework::make_ddim(shape));
}
};
......@@ -47,10 +47,10 @@ class FillConstantOp : public framework::OperatorBase {
if (out_var.IsType<framework::LoDTensor>()) {
tensor = out_var.GetMutable<framework::LoDTensor>();
tensor->Resize(framework::make_ddim(Attr<std::vector<int>>("shape")));
tensor->Resize(framework::make_ddim(Attr<std::vector<int64_t>>("shape")));
} else if (out_var.IsType<framework::SelectedRows>()) {
tensor = out_var.GetMutable<framework::SelectedRows>()->mutable_value();
tensor->Resize(framework::make_ddim(Attr<std::vector<int>>("shape")));
tensor->Resize(framework::make_ddim(Attr<std::vector<int64_t>>("shape")));
} else {
PADDLE_THROW(
"fill constant op's output only"
......@@ -83,7 +83,8 @@ class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker {
"(int, default 5 (FP32)) "
"Output data type")
.SetDefault(framework::proto::VarType::FP32);
AddAttr<std::vector<int>>("shape", "(vector<int>) The shape of the output");
AddAttr<std::vector<int64_t>>("shape",
"(vector<int64_t>) The shape of the output");
AddAttr<float>("value", "(float, default 0) The value to be filled")
.SetDefault(0.0f);
AddAttr<bool>("force_cpu",
......
paddle/fluid/operators/gaussian_random_op.cc
浏览文件 @ b94f9b16
......@@ -52,7 +52,7 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasOutput("Out"),
"Output(Out) of GaussianRandomOp should not be null.");
auto shape = ctx->Attrs().Get<std::vector<int>>("shape");
auto shape = ctx->Attrs().Get<std::vector<int64_t>>("shape");
std::vector<int64_t> temp;
temp.reserve(shape.size());
for (auto dim : shape) {
......@@ -88,9 +88,9 @@ class GaussianRandomOpMaker : public framework::OpProtoAndCheckerMaker {
void Make() override {
AddOutput("Out", "Output matrix of gaussian random op");
AddAttr<std::vector<int>>("shape",
"(vector<int>) "
"The dimension of random tensor.");
AddAttr<std::vector<int64_t>>("shape",
"(vector<int64_t>) "
"The dimension of random tensor.");
AddAttr<float>("mean",
"(float, default 0.0) "
"mean of random tensor.")
......
paddle/fluid/operators/hash_op.cc 0 → 100644
浏览文件 @ b94f9b16
/* Copyright (c) 2016 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/operators/hash_op.h"
#include <string>
#include <vector>
namespace paddle {
namespace operators {
class HashOp : public framework::OperatorWithKernel {
public:
HashOp(const std::string &type, const framework::VariableNameMap &inputs,
const framework::VariableNameMap &outputs,
const framework::AttributeMap &attrs)
: OperatorWithKernel(type, inputs, outputs, attrs) {}
void InferShape(framework::InferShapeContext *ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"),
"Input(X) of HashOp should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("Out"),
"Output(Out) of HashOp should not be null.");
auto dims = ctx->GetInputDim("X");
PADDLE_ENFORCE_EQ(dims.size(), 2UL,
"The input of hash_op's dimensions must be 2");
std::vector<int64_t> out_dims;
out_dims.reserve(dims.size() + 1);
// copy all dims except the last one
for (size_t i = 0u; i != dims.size() - 1; ++i) {
out_dims.emplace_back(dims[i]);
}
int num_hash = ctx->Attrs().Get<int>("num_hash");
out_dims.emplace_back(num_hash);
// keep the last dim to 1
out_dims.emplace_back(1);
ctx->SetOutputDim("Out", framework::make_ddim(out_dims));
ctx->ShareLoD("X", /*->*/ "Out");
}
};
class HashOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("X", "(Tensor) Input tensor of scale operator.");
AddOutput("Out", "(Tensor) Output tensor of scale operator.");
AddComment(R"DOC(
**Hash Operator**
$$Out = scale * X$$
)DOC");
AddAttr<int>("num_hash", "").SetDefault(1);
AddAttr<int>("mod_by", "").SetDefault(100000);
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_WITHOUT_GRADIENT(hash, ops::HashOp, ops::HashOpMaker);
REGISTER_OP_CPU_KERNEL(hash, ops::HashKerel<int>, ops::HashKerel<int64_t>);
paddle/fluid/operators/hash_op.h 0 → 100644
浏览文件 @ b94f9b16
/* Copyright (c) 2016 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
extern "C" {
#include <xxhash.h>
}
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
namespace paddle {
namespace operators {
// template <typename DeviceContext, typename T>
template <typename T>
class HashKerel : public framework::OpKernel<T> {
public:
virtual void Compute(const framework::ExecutionContext& context) const {
auto* out_t = context.Output<framework::LoDTensor>("Out");
auto* in_t = context.Input<framework::LoDTensor>("X");
int mod_by = context.Attr<int>("mod_by");
int num_hash = context.Attr<int>("num_hash");
auto* output = out_t->mutable_data<T>(context.GetPlace());
auto in_dims = in_t->dims();
auto in_lod = in_t->lod();
PADDLE_ENFORCE_EQ(
static_cast<uint64_t>(in_dims[0]), in_lod[0].back(),
"The actual input data's size mismatched with LoD information.");
auto seq_length = in_dims[0];
auto last_dim = in_dims[in_dims.size() - 1];
auto* input = in_t->data<T>();
for (int idx = 0; idx < seq_length; ++idx) {
for (int ihash = 0; ihash != num_hash; ++ihash) {
output[idx * num_hash + ihash] =
XXH64(input, sizeof(int) * last_dim, ihash) % mod_by;
}
input += last_dim;
}
}
};
} // namespace operators
} // namespace paddle
paddle/fluid/operators/lars_momentum_op.cc 0 → 100644
浏览文件 @ b94f9b16
/* 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/operators/lars_momentum_op.h"
#include "paddle/fluid/operators/momentum_op.h"
namespace paddle {
namespace operators {
class LarsMomentumOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("Param",
"(LoDTensor, default LoDTensor<float>) "
"Input parameter that has to be updated");
AddInput("Grad",
"(LoDTensor, default LoDTensor<float>) "
"Input gradient of the parameter");
AddInput("Velocity",
"(LoDTensor, default LoDTensor<float>) "
"Input velocity (corresponding to the parameter) "
"that has to be updated");
AddInput("LearningRate",
"(LoDTensor, default LoDTensor<float>) "
"Input learning rate");
AddOutput("ParamOut",
"(LoDTensor) This output is updated parameter. "
"It shared memory with Input(Param).");
AddOutput("VelocityOut",
"(LoDTensor) This output is updated velocity. "
"It shared memory with Input(Velocity).");
AddAttr<float>("mu", "(float) Momentum coefficient");
AddAttr<float>("lars_coeff", "(float, default 0.001) LARS coefficient.")
.SetDefault(0.001);
AddAttr<float>("lars_weight_decay",
"(float, default 0.0005) LARS weight decay")
.SetDefault(0.0005);
AddComment(R"DOC(
Lars Momentum Optimizer.
This optimizer use LARS (https://arxiv.org/abs/1708.03888) to optimize each
weight using a local learning rate:
$$
local\_lr = \eta *
\frac{\left \| param \right \|}{\left \| grad \right \| + \beta *\left \| param \right \|} \\
velocity = mu * velocity +
local\_lr * (grad + \beta * param) \\
param = param - velocity. \\
$$
Note that we use lars_weight_decay here to decay weights, you may need not to
use L2 regularizers in case of using LARS.
)DOC");
}
};
class LarsMomentumOpVarTypeInference : public framework::VarTypeInference {
public:
void operator()(const framework::OpDesc &op_desc,
framework::BlockDesc *block) const override {}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(lars_momentum, ops::MomentumOp, ops::LarsMomentumOpMaker,
paddle::framework::EmptyGradOpMaker,
ops::LarsMomentumOpVarTypeInference);
REGISTER_OP_CPU_KERNEL(lars_momentum, ops::LarsMomentumOpKernel<float>,
ops::LarsMomentumOpKernel<double>);
paddle/fluid/operators/lars_momentum_op.cu 0 → 100644
浏览文件 @ b94f9b16
/* Copyright (c) 2016 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/op_registry.h"
#include "paddle/fluid/operators/lars_momentum_op.h"
namespace paddle {
namespace operators {
template <typename T>
__global__ void MomentumLarsKernel(const T* p, const T* g, const T* v,
const T* learning_rate, const T mu,
const int64_t num, const T lars_coeff,
const T lars_weight_decay, const T* p_norm,
const T* g_norm, T* p_out, T* v_out) {
T lr = learning_rate[0];
T local_lr = learning_rate[0];
for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num;
i += blockDim.x * gridDim.x) {
if (p_norm[0] > 0 && g_norm[0] > 0) {
local_lr = lr * lars_coeff * p_norm[0] /
(g_norm[0] + lars_weight_decay * p_norm[0]);
}
T v_new = v[i] * mu + local_lr * (g[i] + lars_weight_decay * p[i]);
v_out[i] = v_new;
p_out[i] = p[i] - v_new;
}
}
template <typename DeviceContext, typename T>
class LarsMomentumOpCUDAKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto param_out = ctx.Output<framework::LoDTensor>("ParamOut");
auto velocity_out = ctx.Output<framework::LoDTensor>("VelocityOut");
auto param = ctx.Input<framework::LoDTensor>("Param");
auto velocity = ctx.Input<framework::LoDTensor>("Velocity");
auto grad = ctx.Input<framework::LoDTensor>("Grad");
auto learning_rate = ctx.Input<framework::LoDTensor>("LearningRate");
T* p_out = param_out->mutable_data<T>(ctx.GetPlace());
T* v_out = velocity_out->mutable_data<T>(ctx.GetPlace());
T mu = static_cast<T>(ctx.Attr<float>("mu"));
T lars_coeff = ctx.Attr<float>("lars_coeff");
T lars_weight_decay = ctx.Attr<float>("lars_weight_decay");
auto* p = param->data<T>();
auto* v = velocity->data<T>();
auto* g = grad->data<T>();
auto* lr = learning_rate->data<T>();
int block = 512;
int grid = (param->numel() + block - 1) / block;
auto eigen_p = framework::EigenVector<T>::Flatten(*param);
auto eigen_g = framework::EigenVector<T>::Flatten(*grad);
// calculate norms using eigein and launch the kernel.
framework::Tensor p_norm_t, g_norm_t;
p_norm_t.Resize({1});
g_norm_t.Resize({1});
auto* p_norm_data = p_norm_t.mutable_data<T>(ctx.GetPlace());
auto* g_norm_data = g_norm_t.mutable_data<T>(ctx.GetPlace());
auto ep_norm = framework::EigenScalar<T>::From(p_norm_t);
auto eg_norm = framework::EigenScalar<T>::From(g_norm_t);
auto* place = ctx.template device_context<DeviceContext>().eigen_device();
ep_norm.device(*place) = eigen_p.square().sum().sqrt();
eg_norm.device(*place) = eigen_g.square().sum().sqrt();
MomentumLarsKernel<<<grid, block, 0, ctx.cuda_device_context().stream()>>>(
p, g, v, lr, mu, param->numel(), lars_coeff, lars_weight_decay,
p_norm_data, g_norm_data, p_out, v_out);
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(
lars_momentum,
ops::LarsMomentumOpCUDAKernel<paddle::platform::CUDADeviceContext, float>,
ops::LarsMomentumOpCUDAKernel<paddle::platform::CUDADeviceContext, double>);
paddle/fluid/operators/lars_momentum_op.h 0 → 100644
浏览文件 @ b94f9b16
/* Copyright (c) 2016 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/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
namespace paddle {
namespace operators {
template <typename T>
class LarsMomentumOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto param_out = ctx.Output<framework::LoDTensor>("ParamOut");
auto velocity_out = ctx.Output<framework::LoDTensor>("VelocityOut");
auto param = ctx.Input<framework::LoDTensor>("Param");
auto velocity = ctx.Input<framework::LoDTensor>("Velocity");
auto learning_rate = ctx.Input<framework::LoDTensor>("LearningRate");
auto* grad_var = ctx.InputVar("Grad");
// only support dense for now.
PADDLE_ENFORCE(grad_var->IsType<framework::LoDTensor>());
auto grad = ctx.Input<framework::LoDTensor>("Grad");
param_out->mutable_data<T>(ctx.GetPlace());
velocity_out->mutable_data<T>(ctx.GetPlace());
T mu = static_cast<T>(ctx.Attr<float>("mu"));
T lars_coeff = ctx.Attr<float>("lars_coeff");
T lars_weight_decay = ctx.Attr<float>("lars_weight_decay");
auto p_out = framework::EigenVector<T>::Flatten(*param_out);
auto v_out = framework::EigenVector<T>::Flatten(*velocity_out);
auto p = framework::EigenVector<T>::Flatten(*param);
auto v = framework::EigenVector<T>::Flatten(*velocity);
auto g = framework::EigenVector<T>::Flatten(*grad);
auto* lr = learning_rate->data<T>();
framework::Tensor p_norm_t, g_norm_t;
p_norm_t.Resize({1});
g_norm_t.Resize({1});
p_norm_t.mutable_data<T>(ctx.GetPlace());
g_norm_t.mutable_data<T>(ctx.GetPlace());
auto ep_norm = framework::EigenScalar<T>::From(p_norm_t);
auto eg_norm = framework::EigenScalar<T>::From(g_norm_t);
ep_norm = p.square().sum().sqrt();
eg_norm = g.square().sum().sqrt();
T local_lr = lr[0];
if (ep_norm(0) > 0 && eg_norm(0) > 0) {
local_lr = lr[0] * lars_coeff * ep_norm(0) /
(eg_norm(0) + lars_weight_decay * ep_norm(0));
}
v_out = v * mu + local_lr * (g + lars_weight_decay * p);
p_out = p - v_out;
}
};
} // namespace operators
} // namespace paddle
paddle/fluid/operators/listen_and_serv_op.cc
浏览文件 @ b94f9b16
......@@ -27,6 +27,10 @@ limitations under the License. */
#include "paddle/fluid/operators/distributed/request_handler_impl.h"
#include "paddle/fluid/operators/listen_and_serv_op.h"
DEFINE_int32(rpc_send_thread_num, 5, "number of threads for rpc send");
DEFINE_int32(rpc_get_thread_num, 5, "number of threads for rpc get");
DEFINE_int32(rpc_prefetch_thread_num, 5, "number of threads for rpc prefetch");
namespace paddle {
namespace operators {
......@@ -332,11 +336,14 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
sync_mode, checkpoint_block_id));
rpc_service_->RegisterRPC(distributed::kRequestSend,
request_send_handler_.get());
request_send_handler_.get(),
FLAGS_rpc_send_thread_num);
rpc_service_->RegisterRPC(distributed::kRequestGet,
request_get_handler_.get());
request_get_handler_.get(),
FLAGS_rpc_get_thread_num);
rpc_service_->RegisterRPC(distributed::kRequestPrefetch,
request_prefetch_handler_.get());
request_prefetch_handler_.get(),
FLAGS_rpc_prefetch_thread_num);
rpc_service_->RegisterRPC(distributed::kRequestCheckpoint,
request_checkpoint_handler_.get());
......
paddle/fluid/operators/lookup_table_op.cc
浏览文件 @ b94f9b16
......@@ -81,6 +81,12 @@ class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker {
"Otherwise the given value indicates padding the output "
"with zeros whenever lookup encounters it in Ids.")
.SetDefault(kNoPadding);
// NOTE(minqiyang): grad_inplace is an temporal attribute,
// please do NOT set this attribute in python layer.
AddAttr<bool>("grad_inplace",
"(boolean, default false) "
"If the grad op reuse the input's variable.")
.SetDefault(false);
AddComment(R"DOC(
Lookup Table Operator.
......@@ -115,7 +121,7 @@ class LookupTableOpGrad : public framework::OperatorWithKernel {
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("W"));
auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("Out"));
return framework::OpKernelType(data_type, ctx.device_context());
}
};
......
paddle/fluid/operators/lookup_table_op.h
浏览文件 @ b94f9b16
......@@ -21,6 +21,7 @@ limitations under the License. */
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/selected_rows.h"
#include "paddle/fluid/operators/math/blas.h"
namespace paddle {
namespace operators {
......@@ -68,6 +69,7 @@ class LookupTableKernel : public framework::OpKernel<T> {
const auto *table = table_t.value().data<T>();
auto *output = output_t->mutable_data<T>(context.GetPlace());
auto blas = math::GetBlas<platform::CPUDeviceContext, T>(context);
for (int64_t i = 0; i < ids_numel; ++i) {
if (padding_idx != kNoPadding && ids[i] == padding_idx) {
memset(output + i * row_width, 0, row_width * sizeof(T));
......@@ -75,8 +77,8 @@ class LookupTableKernel : public framework::OpKernel<T> {
PADDLE_ENFORCE_GE(ids[i], 0);
auto id_index = table_t.Index(ids[i]);
PADDLE_ENFORCE_GE(id_index, 0, "the input key should be exists.");
memcpy(output + i * row_width, table + id_index * row_width,
row_width * sizeof(T));
blas.VCOPY(row_width, table + id_index * row_width,
output + i * row_width);
}
}
}
......@@ -111,27 +113,37 @@ class LookupTableGradKernel : public framework::OpKernel<T> {
auto *ids_data = ids->data<int64_t>();
int64_t ids_num = ids->numel();
framework::Vector<int64_t> new_rows;
new_rows.reserve(ids_num);
for (int64_t i = 0; i < ids_num; i++) {
new_rows.push_back(ids_data[i]);
}
std::vector<int64_t> new_rows;
new_rows.resize(ids_num);
std::memcpy(&new_rows[0], ids_data, ids_num * sizeof(int64_t));
d_table->set_rows(new_rows);
auto *d_table_value = d_table->mutable_value();
d_table_value->Resize({ids_num, table_dim[1]});
d_table_value->mutable_data<T>(context.GetPlace());
d_table->set_height(table_dim[0]);
auto *d_output_data = d_output->data<T>();
auto *d_table_data = d_table_value->data<T>();
auto d_output_dims = d_output->dims();
PADDLE_ENFORCE_EQ(
d_table_value->dims(),
framework::flatten_to_2d(d_output_dims, d_output_dims.size() - 1));
memcpy(d_table_data, d_output_data, sizeof(T) * d_output->numel());
// FIXME(minqiyang):
// memory optimization will NOT reuse Tensor with SelectedRows
// so we could just share the tensor here directly.
// However, the InferVarType method will infer the output SelectedRows
// to Tensor sometimes, which is a bug, so we will add an attribute
// here to indicate the inplace and remove this attribute after
// the InferVarType's bug was fixed
bool grad_inplace = context.Attr<bool>("grad_inplace");
if (grad_inplace) {
d_table_value->ShareDataWith(*d_output);
} else {
d_table_value->mutable_data<T>(context.GetPlace());
d_table->set_height(table_dim[0]);
auto *d_output_data = d_output->data<T>();
auto *d_table_data = d_table_value->data<T>();
auto d_output_dims = d_output->dims();
PADDLE_ENFORCE_EQ(
d_table_value->dims(),
framework::flatten_to_2d(d_output_dims, d_output_dims.size() - 1));
memcpy(d_table_data, d_output_data, sizeof(T) * d_output->numel());
}
} else {
auto *ids = context.Input<LoDTensor>("Ids");
auto *d_output = context.Input<LoDTensor>(framework::GradVarName("Out"));
......
paddle/fluid/operators/math/selected_rows_functor.cc
浏览文件 @ b94f9b16
......@@ -12,9 +12,8 @@ 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 <map>
#include <set>
#include <vector>
#include <unordered_map>
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/selected_rows_functor.h"
......@@ -230,8 +229,24 @@ template struct SelectedRowsAddToTensor<platform::CPUDeviceContext, int64_t>;
// add or mul.
namespace scatter {
size_t FindPos(const std::vector<int64_t>& rows, int64_t value) {
return std::find(rows.begin(), rows.end(), value) - rows.begin();
template <typename DeviceContext, typename T>
typename std::enable_if<
std::is_floating_point<T>::value &&
std::is_same<DeviceContext, platform::CPUDeviceContext>::value>::type
elementwise_add_to(const DeviceContext& ctx, BlasT<DeviceContext, T>* blas,
size_t data_len, const T* in, T* out) {
blas->AXPY(data_len, 1., in, out);
}
template <typename DeviceContext, typename T>
typename std::enable_if<
!std::is_floating_point<T>::value &&
std::is_same<DeviceContext, platform::CPUDeviceContext>::value>::type
elementwise_add_to(const DeviceContext& ctx, BlasT<DeviceContext, T>* blas,
size_t data_len, const T* in, T* out) {
for (int64_t i = 0; i < data_len; i++) {
out[i] += in[i];
}
}
template <typename T>
......@@ -246,48 +261,84 @@ struct MergeAdd<platform::CPUDeviceContext, T> {
void operator()(const platform::CPUDeviceContext& context,
const framework::SelectedRows& input,
framework::SelectedRows* output) {
framework::SelectedRows& out = *output;
std::vector<int64_t> input_rows(input.rows());
std::vector<const framework::SelectedRows*> inputs;
inputs.push_back(&input);
(*this)(context, inputs, output);
}
std::map<int64_t, std::vector<int64_t>> merge_row_map;
for (size_t i = 0; i < input_rows.size(); ++i) {
merge_row_map[input_rows[i]].push_back(i);
void operator()(const platform::CPUDeviceContext& context,
const std::vector<const framework::SelectedRows*>& inputs,
framework::SelectedRows* output) {
if (inputs.size() == 0) {
VLOG(3) << "no input! return";
return;
}
std::vector<int64_t> merge_rows(merge_row_map.size());
size_t idx = 0;
int64_t input_width = input.value().dims()[1];
out.set_height(input.height());
T* out_data = out.mutable_value()->mutable_data<T>(
const framework::SelectedRows* has_value_input = nullptr;
for (auto* in : inputs) {
if (in->rows().size() > 0) {
has_value_input = in;
break;
}
}
if (has_value_input == nullptr) {
VLOG(3) << "no input has value! just return" << std::endl;
return;
}
auto input_width = has_value_input->value().dims()[1];
auto input_height = has_value_input->height();
framework::SelectedRows& out = *output;
std::set<int64_t> merged_row_set;
for (auto* input : inputs) {
if (input->rows().size() == 0) {
continue;
}
PADDLE_ENFORCE_EQ(input_width, input->value().dims()[1],
"all input should have same "
"dimension except for the first one");
PADDLE_ENFORCE_EQ(input_height, input->height(),
"all input should have same height");
merged_row_set.insert(input->rows().begin(), input->rows().end());
}
std::vector<int64_t> merge_rows(merged_row_set.begin(),
merged_row_set.end());
std::unordered_map<int64_t, size_t> rows_to_id;
for (size_t i = 0; i < merge_rows.size(); ++i) {
rows_to_id[merge_rows[i]] = i;
}
out.set_rows(merge_rows);
out.set_height(input_height);
out.mutable_value()->mutable_data<T>(
framework::make_ddim(
{static_cast<int64_t>(merge_rows.size()), input_width}),
context.GetPlace());
const T* in_data = input.value().data<T>();
for (auto& row_pair : merge_row_map) {
auto* out_ptr = out_data + idx * input_width;
auto& rows = row_pair.second;
merge_rows[idx] = row_pair.first;
++idx;
// rows.size() is always larger than 0
std::memcpy(out_ptr, in_data + rows[0] * input_width,
sizeof(T) * input_width);
for (size_t i = 1; i < rows.size(); ++i) {
auto* in_ptr = in_data + rows[i] * input_width;
for (int64_t j = 0; j < input_width; ++j) {
out_ptr[j] += in_ptr[j];
}
math::SetConstant<platform::CPUDeviceContext, T> constant_functor;
constant_functor(context, out.mutable_value(), 0.0);
auto* out_data = out.mutable_value()->data<T>();
auto blas = math::GetBlas<platform::CPUDeviceContext, T>(context);
for (auto* input : inputs) {
if (input->rows().size() == 0) {
continue;
}
auto* input_data = input->value().data<T>();
auto& input_rows = input->rows();
for (size_t i = 0; i < input_rows.size(); i++) {
size_t out_i = rows_to_id[input_rows[i]];
elementwise_add_to<platform::CPUDeviceContext, T>(
context, &blas, static_cast<size_t>(input_width),
&input_data[i * input_width], &out_data[out_i * input_width]);
}
}
out.set_rows(merge_rows);
}
};
template struct MergeAdd<platform::CPUDeviceContext, int>;
template struct MergeAdd<platform::CPUDeviceContext, int64_t>;
template struct MergeAdd<platform::CPUDeviceContext, float>;
template struct MergeAdd<platform::CPUDeviceContext, double>;
template <typename T>
struct UpdateToTensor<platform::CPUDeviceContext, T> {
......
paddle/fluid/operators/math/selected_rows_functor.cu
浏览文件 @ b94f9b16
......@@ -267,10 +267,15 @@ struct MergeAdd<platform::CUDADeviceContext, T> {
void operator()(const platform::CUDADeviceContext& context,
const framework::SelectedRows& input,
framework::SelectedRows* output) {
framework::SelectedRows& out = *output;
framework::Vector<int64_t> input_rows(input.rows());
if (input_rows.size() == 0) {
return;
}
framework::SelectedRows& out = *output;
std::set<int64_t> row_set(input_rows.begin(), input_rows.end());
std::vector<int64_t> merge_rows(row_set.begin(), row_set.end());
std::vector<int64_t> merge_rows_cpu(row_set.begin(), row_set.end());
framework::Vector<int64_t> merge_rows(merge_rows_cpu);
auto input_width = input.value().dims()[1];
......@@ -296,6 +301,73 @@ struct MergeAdd<platform::CUDADeviceContext, T> {
out.mutable_rows()->CUDAMutableData(context.GetPlace()),
out.rows().size(), input_width);
}
void operator()(const platform::CUDADeviceContext& context,
const std::vector<const framework::SelectedRows*>& inputs,
framework::SelectedRows* output) {
if (inputs.size() == 0) {
VLOG(3) << "no input! return";
return;
}
const framework::SelectedRows* has_value_input = nullptr;
for (auto* in : inputs) {
if (in->rows().size() > 0) {
has_value_input = in;
break;
}
}
if (has_value_input == nullptr) {
VLOG(3) << "no input has value! just return" << std::endl;
return;
}
auto input_width = has_value_input->value().dims()[1];
auto input_height = has_value_input->height();
framework::SelectedRows& out = *output;
std::set<int64_t> merged_row_set;
for (auto* input : inputs) {
if (input->rows().size() == 0) {
continue;
}
PADDLE_ENFORCE_EQ(input_width, input->value().dims()[1],
"all input should have same "
"dimension except for the first one");
PADDLE_ENFORCE_EQ(input_height, input->height(),
"all input should have same height");
merged_row_set.insert(input->rows().begin(), input->rows().end());
}
std::vector<int64_t> merge_rows_cpu(merged_row_set.begin(),
merged_row_set.end());
framework::Vector<int64_t> merge_rows(merge_rows_cpu);
out.set_rows(merge_rows);
out.set_height(input_height);
out.mutable_value()->mutable_data<T>(
framework::make_ddim(
{static_cast<int64_t>(merge_rows.size()), input_width}),
context.GetPlace());
math::SetConstant<platform::CUDADeviceContext, T> constant_functor;
constant_functor(context, out.mutable_value(), 0.0);
auto* out_data = out.mutable_value()->data<T>();
const int block_size = 256;
dim3 threads(block_size, 1);
for (auto* input : inputs) {
if (input->rows().size() == 0) {
continue;
}
auto* input_data = input->value().data<T>();
auto& input_rows = input->rows();
dim3 grid1(input_rows.size(), 1);
MergeAddKernel<T, 256><<<grid1, threads, 0, context.stream()>>>(
input_data, input_rows.CUDAData(context.GetPlace()), out_data,
out.mutable_rows()->CUDAMutableData(context.GetPlace()),
out.rows().size(), input_width);
}
}
};
template struct MergeAdd<platform::CUDADeviceContext, float>;
......
paddle/fluid/operators/math/selected_rows_functor.h
浏览文件 @ b94f9b16
......@@ -83,104 +83,9 @@ struct MergeAdd {
void operator()(const DeviceContext& context,
const framework::SelectedRows& input,
framework::SelectedRows* output);
};
template <>
struct MergeAdd<platform::CPUDeviceContext, float> {
framework::SelectedRows operator()(const platform::CPUDeviceContext& context,
const framework::SelectedRows& input) {
framework::SelectedRows out;
(*this)(context, input, &out);
return out;
}
void operator()(const platform::CPUDeviceContext& context,
const framework::SelectedRows& input,
framework::SelectedRows* output) {
framework::SelectedRows& out = *output;
std::vector<int64_t> input_rows(input.rows());
std::map<int64_t, std::vector<int64_t>> merge_row_map;
for (size_t i = 0; i < input_rows.size(); ++i) {
merge_row_map[input_rows[i]].push_back(i);
}
std::vector<int64_t> merge_rows(merge_row_map.size());
size_t idx = 0;
int64_t input_width = input.value().dims()[1];
out.set_height(input.height());
auto* out_data = out.mutable_value()->mutable_data<float>(
framework::make_ddim(
{static_cast<int64_t>(merge_rows.size()), input_width}),
context.GetPlace());
auto* in_data = input.value().data<float>();
auto blas = GetBlas<platform::CPUDeviceContext, float>(context);
for (auto& row_pair : merge_row_map) {
auto* out_ptr = out_data + idx * input_width;
auto& rows = row_pair.second;
merge_rows[idx] = row_pair.first;
++idx;
// rows.size() is always larger than 0
blas.VCOPY(input_width, in_data + rows[0] * input_width, out_ptr);
for (size_t i = 1; i < rows.size(); ++i) {
blas.AXPY(input_width, 1., in_data + rows[i] * input_width, out_ptr);
}
}
out.set_rows(merge_rows);
}
};
template <>
struct MergeAdd<platform::CPUDeviceContext, double> {
framework::SelectedRows operator()(const platform::CPUDeviceContext& context,
const framework::SelectedRows& input) {
framework::SelectedRows out;
(*this)(context, input, &out);
return out;
}
void operator()(const platform::CPUDeviceContext& context,
const framework::SelectedRows& input,
framework::SelectedRows* output) {
framework::SelectedRows& out = *output;
std::vector<int64_t> input_rows(input.rows());
std::map<int64_t, std::vector<int64_t>> merge_row_map;
for (size_t i = 0; i < input_rows.size(); ++i) {
merge_row_map[input_rows[i]].push_back(i);
}
std::vector<int64_t> merge_rows(merge_row_map.size());
size_t idx = 0;
int64_t input_width = input.value().dims()[1];
out.set_height(input.height());
auto* out_data = out.mutable_value()->mutable_data<double>(
framework::make_ddim(
{static_cast<int64_t>(merge_rows.size()), input_width}),
context.GetPlace());
auto* in_data = input.value().data<double>();
auto blas = GetBlas<platform::CPUDeviceContext, double>(context);
for (auto& row_pair : merge_row_map) {
auto* out_ptr = out_data + idx * input_width;
auto& rows = row_pair.second;
merge_rows[idx] = row_pair.first;
++idx;
// rows.size() is always larger than 0
blas.VCOPY(input_width, in_data + rows[0] * input_width, out_ptr);
for (size_t i = 1; i < rows.size(); ++i) {
blas.AXPY(input_width, 1., in_data + rows[i] * input_width, out_ptr);
}
}
out.set_rows(merge_rows);
}
void operator()(const DeviceContext& context,
const std::vector<const framework::SelectedRows*>& inputs,
framework::SelectedRows* output);
};
template <typename DeviceContext, typename T>
......
paddle/fluid/operators/math/selected_rows_functor_test.cc
浏览文件 @ b94f9b16
......@@ -302,6 +302,64 @@ TEST(selected_rows_functor, cpu_merge_add_int) {
EXPECT_EQ(out_data[1 * row_numel], 2);
EXPECT_EQ(out_data[2 * row_numel], 1);
}
TEST(selected_rows_functor, cpu_merge_add_multi) {
paddle::platform::CPUPlace cpu_place;
paddle::platform::CPUDeviceContext ctx(cpu_place);
paddle::operators::math::SetConstant<paddle::platform::CPUDeviceContext,
float>
set_const;
int64_t height = 10;
int64_t row_numel = 8;
std::vector<int64_t> rows1{5, 2, 5, 3, 5};
std::unique_ptr<paddle::framework::SelectedRows> selected_rows1{
new paddle::framework::SelectedRows(rows1, height)};
auto* in1_value = selected_rows1->mutable_value();
in1_value->mutable_data<float>(
paddle::framework::make_ddim(
{static_cast<int64_t>(rows1.size()), row_numel}),
cpu_place);
set_const(ctx, in1_value, 1.0);
std::vector<int64_t> rows2{2, 5, 3, 5, 3};
std::unique_ptr<paddle::framework::SelectedRows> selected_rows2{
new paddle::framework::SelectedRows(rows2, height)};
auto* in2_value = selected_rows2->mutable_value();
in2_value->mutable_data<float>(
paddle::framework::make_ddim(
{static_cast<int64_t>(rows2.size()), row_numel}),
cpu_place);
set_const(ctx, in2_value, 1.0);
std::unique_ptr<paddle::framework::SelectedRows> output{
new paddle::framework::SelectedRows()};
output->set_height(height);
paddle::operators::math::scatter::MergeAdd<paddle::platform::CPUDeviceContext,
float>
merge_add_functor;
std::vector<const paddle::framework::SelectedRows*> inputs;
inputs.push_back(selected_rows1.get());
inputs.push_back(selected_rows2.get());
merge_add_functor(ctx, inputs, output.get());
EXPECT_EQ(output->height(), height);
EXPECT_EQ(output->value().dims(),
paddle::framework::make_ddim({3, row_numel}));
std::vector<int64_t> ret_rows{2, 3, 5};
EXPECT_EQ(output->rows(), ret_rows);
auto* out_data = output->value().data<float>();
for (size_t i = 0; i < ret_rows.size(); ++i) {
for (size_t j = 0; j < row_numel; ++j) {
EXPECT_EQ(out_data[i * row_numel + j], ret_rows[i]);
}
}
}
TEST(selected_rows_functor, cpu_sum_to) {
paddle::platform::CPUPlace cpu_place;
paddle::platform::CPUDeviceContext ctx(cpu_place);
......@@ -318,6 +376,7 @@ TEST(selected_rows_functor, cpu_sum_to) {
paddle::framework::make_ddim(
{static_cast<int64_t>(rows1.size()), row_numel}),
cpu_place);
functor(ctx, in1_value, 1.0);
std::vector<int64_t> rows2{0, 5, 7, 9};
std::unique_ptr<paddle::framework::SelectedRows> selected_rows2{
......@@ -327,6 +386,7 @@ TEST(selected_rows_functor, cpu_sum_to) {
paddle::framework::make_ddim(
{static_cast<int64_t>(rows2.size()), row_numel}),
cpu_place);
functor(ctx, in2_value, 2.0);
std::unique_ptr<paddle::framework::SelectedRows> output{
new paddle::framework::SelectedRows()};
......
paddle/fluid/operators/math/selected_rows_functor_test.cu
浏览文件 @ b94f9b16
......@@ -241,3 +241,67 @@ TEST(selected_rows_functor, gpu_add_to) {
// row9: 2.0 + 3.0
EXPECT_EQ(tensor1_cpu_data[9 * row_numel + 6], 5.0);
}
TEST(selected_rows_functor, gpu_merge_add) {
paddle::platform::CUDAPlace gpu_place(0);
paddle::platform::CPUPlace cpu_place;
paddle::platform::CUDADeviceContext& ctx =
*reinterpret_cast<paddle::platform::CUDADeviceContext*>(
paddle::platform::DeviceContextPool::Instance().Get(gpu_place));
paddle::operators::math::SetConstant<paddle::platform::CUDADeviceContext,
float>
set_const;
int64_t height = 10;
int64_t row_numel = 8;
std::vector<int64_t> rows1{5, 2, 5, 3, 5};
std::unique_ptr<paddle::framework::SelectedRows> selected_rows1{
new paddle::framework::SelectedRows(rows1, height)};
auto* in1_value = selected_rows1->mutable_value();
in1_value->mutable_data<float>(
paddle::framework::make_ddim(
{static_cast<int64_t>(rows1.size()), row_numel}),
gpu_place);
set_const(ctx, in1_value, 1.0);
std::vector<int64_t> rows2{2, 5, 3, 5, 3};
std::unique_ptr<paddle::framework::SelectedRows> selected_rows2{
new paddle::framework::SelectedRows(rows2, height)};
auto* in2_value = selected_rows2->mutable_value();
in2_value->mutable_data<float>(
paddle::framework::make_ddim(
{static_cast<int64_t>(rows2.size()), row_numel}),
gpu_place);
set_const(ctx, in2_value, 1.0);
std::unique_ptr<paddle::framework::SelectedRows> output{
new paddle::framework::SelectedRows()};
output->set_height(height);
paddle::operators::math::scatter::MergeAdd<
paddle::platform::CUDADeviceContext, float>
merge_add_functor;
std::vector<const paddle::framework::SelectedRows*> inputs;
inputs.push_back(selected_rows1.get());
inputs.push_back(selected_rows2.get());
merge_add_functor(ctx, inputs, output.get());
paddle::framework::Tensor output_cpu;
paddle::framework::TensorCopy(output->value(), cpu_place, ctx, &output_cpu);
ctx.Wait();
EXPECT_EQ(output->height(), height);
EXPECT_EQ(output->value().dims(),
paddle::framework::make_ddim({3, row_numel}));
std::vector<int64_t> ret_rows{2, 3, 5};
EXPECT_EQ(output->rows(), ret_rows);
auto* out_data = output_cpu.data<float>();
for (size_t i = 0; i < ret_rows.size(); ++i) {
for (size_t j = 0; j < row_numel; ++j) {
EXPECT_EQ(out_data[i * row_numel + j], ret_rows[i]);
}
}
}
paddle/fluid/operators/merge_ids_op.cc
浏览文件 @ b94f9b16
......@@ -20,13 +20,16 @@ namespace operators {
class MergeIdsOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}");
AddInput(
"X",
"(LoDTensors) multi input tensor with shape{batch_num, N}, N is the "
"size of embedding table")
AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}")
.AsDuplicable();
AddInput("Rows", "(LoDTensor) the input ids with shape{row_size, 1}, ")
.AsDuplicable();
AddInput("X",
"(LoDTensors) multi input tensor with shape{Rows, N}, N is the "
"size of embedding table")
.AsDuplicable();
AddOutput("Out", "(LoDTensor) The merged outputs of the input tensors.")
.AsDuplicable();
AddOutput("Out", "(LoDTensor) The merged outputs of the input tensors.");
AddComment(R"DOC(
Merge multi LoDTensor's into one according to Ids's shard num.
......@@ -79,15 +82,19 @@ class MergeIdsOp : public framework::OperatorWithKernel {
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext *ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Ids"), "MergeIdsOp must has input Ids.");
PADDLE_ENFORCE(ctx->HasInputs("X"), "MergeIdsOp must has input X.");
PADDLE_ENFORCE(ctx->HasOutput("Out"), "MergeIdsOp must has output Out.");
PADDLE_ENFORCE(ctx->HasInputs("Ids"),
"MergeIdsOp must has multi input Ids.");
PADDLE_ENFORCE(ctx->HasInputs("Rows"),
"MergeIdsOp must has multi input Rows.");
PADDLE_ENFORCE(ctx->HasInputs("X"), "MergeIdsOp must has multi input X.");
PADDLE_ENFORCE(ctx->HasOutputs("Out"),
"MergeIdsOp must has multi output Out.");
auto ids_var_type = ctx->GetInputsVarType("Ids").front();
auto ids_dims = ctx->GetInputDim("Ids");
auto ids_dims = ctx->GetInputsDim("Ids");
if (ids_var_type == framework::proto::VarType::LOD_TENSOR) {
PADDLE_ENFORCE_EQ(ids_dims.size(), 2);
PADDLE_ENFORCE_EQ(ids_dims[1], 1);
PADDLE_ENFORCE_EQ(ids_dims[0].size(), 2);
PADDLE_ENFORCE_EQ(ids_dims[0][1], 1);
}
auto x_var_type = ctx->GetInputsVarType("X");
for (auto &var_type : x_var_type) {
......
paddle/fluid/operators/merge_ids_op.h
浏览文件 @ b94f9b16
......@@ -14,6 +14,8 @@ limitations under the License. */
#pragma once
#include <tuple>
#include <unordered_map>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/tensor_util.h"
......@@ -30,59 +32,70 @@ class MergeIdsOpKernel : public framework::OpKernel<T> {
if (!platform::is_cpu_place(place)) {
PADDLE_THROW("MergeIds do not support GPU kernel");
}
VLOG(3) << "run in MergeIdsOpKernel";
const auto *ids_var = ctx.InputVar("Ids");
PADDLE_ENFORCE(ids_var->IsType<framework::LoDTensor>(),
"only support to merge Ids of LoDTensor");
const auto ids = ctx.MultiInput<framework::LoDTensor>("Ids");
const auto row_ids = ctx.MultiInput<framework::LoDTensor>("Rows");
const auto x_tensors = ctx.MultiInput<framework::LoDTensor>("X");
auto outs = ctx.MultiOutput<framework::LoDTensor>("Out");
const auto &ids_tensor = ids_var->Get<framework::LoDTensor>();
const auto &ids_dims = ids_tensor.dims();
const int64_t *ids = ids_tensor.data<int64_t>();
PADDLE_ENFORCE_EQ(row_ids.size(), x_tensors.size(),
"the number of Rows and X should be the same");
PADDLE_ENFORCE_EQ(ids.size(), outs.size(),
"the number of Ids and Out should be the same");
auto x_tensors = ctx.MultiInput<framework::LoDTensor>("X");
int row_ids_size = 0;
int row_size = 0;
int embedding_size = 0;
auto *out = ctx.Output<framework::LoDTensor>("Out");
for (int i = 0; i < x_tensors.size(); ++i) {
const auto *x_tensor = x_tensors[i];
const auto *row_id = row_ids[i];
int batch_size = 0;
int embedding_size = 0;
for (auto &input : x_tensors) {
if (framework::product(input->dims()) != 0) {
if (embedding_size == 0) {
embedding_size = input->dims()[1];
}
PADDLE_ENFORCE_EQ(embedding_size, input->dims()[1],
"embedding size of all input should be the same");
batch_size += input->dims()[0];
if (embedding_size == 0) {
embedding_size = x_tensor->dims()[1];
}
PADDLE_ENFORCE_EQ(embedding_size, x_tensor->dims()[1],
"embedding size of all input should be the same");
row_size += x_tensor->dims()[0];
row_ids_size += row_id->dims()[0];
}
PADDLE_ENFORCE_EQ(
batch_size, ids_dims[0],
"the batch size of ids and merged embedding value should be the same");
row_size, row_ids_size,
"the merged X dim[0] and merged Rows dim[0] should be the same");
std::unordered_map<int64_t, std::tuple<int64_t, int64_t>>
selected_rows_idx_map;
for (int i = 0; i < x_tensors.size(); ++i) {
const auto *row_id = row_ids[i];
for (int j = 0; j < row_id->numel(); ++j) {
int64_t key = row_id->data<int64_t>()[j];
std::tuple<int64_t, int64_t> val = std::make_tuple(i, j);
selected_rows_idx_map.insert(std::make_pair(key, val));
}
}
PADDLE_ENFORCE_EQ(row_ids_size, selected_rows_idx_map.size(),
"the rows and tensor map size should be the same");
for (int i = 0; i < outs.size(); ++i) {
auto *out_ids = ids[i];
auto *out = outs[i];
const size_t shard_num = x_tensors.size();
out->set_lod(out_ids->lod());
if (shard_num == 1) {
VLOG(3) << "only one shard, we can copy the data directly";
TensorCopy(*x_tensors[0], place, out);
} else {
std::vector<int> in_indexs(shard_num, 0);
int nums = static_cast<int>(out_ids->dims()[0]);
auto *out_data = out->mutable_data<T>(
framework::make_ddim({batch_size, embedding_size}), place);
// copy data from ins[shard_num] to out.
for (int i = 0; i < ids_dims[0]; ++i) {
int64_t id = ids[i];
size_t shard_id = static_cast<size_t>(id) % shard_num;
int index = in_indexs[shard_id];
memcpy(out_data + embedding_size * i,
x_tensors[shard_id]->data<T>() + index * embedding_size,
framework::make_ddim({nums, embedding_size}), place);
for (int j = 0; j < nums; ++j) {
int id = out_ids->data<int64_t>()[j];
auto row_tuple = selected_rows_idx_map[id];
int64_t row_idx = std::get<1>(row_tuple);
const auto *x_tensor = x_tensors[std::get<0>(row_tuple)];
memcpy(out_data + embedding_size * j,
x_tensor->data<T>() + row_idx * embedding_size,
sizeof(T) * embedding_size);
in_indexs[shard_id] += 1;
}
for (size_t i = 0; i < shard_num; ++i) {
PADDLE_ENFORCE_EQ(in_indexs[i], x_tensors[i]->dims()[0],
"after merge, all data in x_tensor should be used");
}
}
}
......
paddle/fluid/operators/momentum_op.cc
浏览文件 @ b94f9b16
......@@ -19,54 +19,6 @@ namespace operators {
using Tensor = framework::Tensor;
class MomentumOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Param"),
"Input(param) of Momentum should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Grad"),
"Input(grad) of Momentum should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Velocity"),
"Input(velocity) of Momentum should not be null.");
PADDLE_ENFORCE(ctx->HasInput("LearningRate"),
"Input(LearningRate) of Momentum should not be null.");
PADDLE_ENFORCE(
ctx->GetInputsVarType("Param").front() ==
framework::proto::VarType::LOD_TENSOR,
"The input var's type should be LoDTensor, but the received is %s",
ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front());
PADDLE_ENFORCE(ctx->HasOutput("ParamOut"),
"Output(ParamOut) of Momentum should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("VelocityOut"),
"Output(VelocityOut) of Momentum should not be null.");
auto param_dim = ctx->GetInputDim("Param");
if (ctx->GetInputsVarType("Grad")[0] ==
framework::proto::VarType::LOD_TENSOR) {
PADDLE_ENFORCE_EQ(
param_dim, ctx->GetInputDim("Grad"),
"Param and Grad input of MomentumOp should have the same dimension.");
PADDLE_ENFORCE_EQ(
param_dim, ctx->GetInputDim("Velocity"),
"Param and Velocity of MomentumOp should have the same dimension.");
}
PADDLE_ENFORCE_EQ(framework::product(ctx->GetInputDim("LearningRate")), 1,
"Learning_rate should be a scalar");
ctx->SetOutputDim("ParamOut", param_dim);
ctx->SetOutputDim("VelocityOut", param_dim);
}
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
auto input_data_type = framework::GetDataTypeOfVar(ctx.InputVar("Param"));
return framework::OpKernelType(input_data_type, ctx.GetPlace());
}
};
class MomentumOpInferVarType : public framework::VarTypeInference {
public:
void operator()(const framework::OpDesc& op_desc,
......
paddle/fluid/operators/momentum_op.h
浏览文件 @ b94f9b16
......@@ -28,6 +28,54 @@ using framework::SelectedRows;
struct NoNesterov;
struct UseNesterov;
class MomentumOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Param"),
"Input(param) of Momentum should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Grad"),
"Input(grad) of Momentum should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Velocity"),
"Input(velocity) of Momentum should not be null.");
PADDLE_ENFORCE(ctx->HasInput("LearningRate"),
"Input(LearningRate) of Momentum should not be null.");
PADDLE_ENFORCE(
ctx->GetInputsVarType("Param").front() ==
framework::proto::VarType::LOD_TENSOR,
"The input var's type should be LoDTensor, but the received is %s",
ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front());
PADDLE_ENFORCE(ctx->HasOutput("ParamOut"),
"Output(ParamOut) of Momentum should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("VelocityOut"),
"Output(VelocityOut) of Momentum should not be null.");
auto param_dim = ctx->GetInputDim("Param");
if (ctx->GetInputsVarType("Grad")[0] ==
framework::proto::VarType::LOD_TENSOR) {
PADDLE_ENFORCE_EQ(
param_dim, ctx->GetInputDim("Grad"),
"Param and Grad input of MomentumOp should have the same dimension.");
PADDLE_ENFORCE_EQ(
param_dim, ctx->GetInputDim("Velocity"),
"Param and Velocity of MomentumOp should have the same dimension.");
}
PADDLE_ENFORCE_EQ(framework::product(ctx->GetInputDim("LearningRate")), 1,
"Learning_rate should be a scalar");
ctx->SetOutputDim("ParamOut", param_dim);
ctx->SetOutputDim("VelocityOut", param_dim);
}
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
auto input_data_type = framework::GetDataTypeOfVar(ctx.InputVar("Param"));
return framework::OpKernelType(input_data_type, ctx.GetPlace());
}
};
template <typename T>
class CPUDenseMomentumFunctor {
private:
......
paddle/fluid/operators/split_ids_op.cc
浏览文件 @ b94f9b16
......@@ -20,20 +20,27 @@ namespace operators {
class SplitIdsOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}");
AddOutput("Out", "(LoDTensor) The outputs of the input Ids.")
AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}")
.AsDuplicable();
AddOutput("Out", "(LoDTensors) The outputs of the input Ids.")
.AsDuplicable();
AddComment(R"DOC(
Split a LoDTensor of Ids into multi LoDTensors, the number is pserver's number
Example:
Input:
X = [1,2,3,4,5,6]
X = [[1,2,3,4,5,6],[2,3]]
Out(3 output):
out0 = [3, 6]
out1 = [1, 4]
out2 = [2, 5]
if compress is True:
out0 = [3, 3, 6]
out1 = [1, 4]
out2 = [2, 2, 5]
else:
out0 = [3, 6]
out1 = [1, 4]
out2 = [2, 5]
)DOC");
}
};
......@@ -43,16 +50,24 @@ class SplitIdsOp : public framework::OperatorWithKernel {
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext *ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Ids"), "SplitIdsOp must has input Ids.");
PADDLE_ENFORCE(ctx->HasInputs("Ids"), "SplitIdsOp must has input Ids.");
PADDLE_ENFORCE(ctx->HasOutputs("Out"), "SplitIdsOp must has output Out.");
auto ids_var_type = ctx->GetInputsVarType("Ids").front();
auto ids_dims = ctx->GetInputDim("Ids");
auto ids_dims = ctx->GetInputsDim("Ids");
if (ids_var_type == framework::proto::VarType::LOD_TENSOR) {
PADDLE_ENFORCE_EQ(ids_dims.size(), 2);
PADDLE_ENFORCE_EQ(ids_dims[1], 1);
PADDLE_ENFORCE_EQ(ids_dims[0].size(), 2);
}
}
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext &ctx) const override {
return framework::OpKernelType(
framework::ToDataType(
ctx.MultiInput<framework::Tensor>("Ids").front()->type()),
ctx.GetPlace());
}
};
class SplitIdsOpInferVarType : public framework::VarTypeInference {
......@@ -66,12 +81,28 @@ class SplitIdsOpInferVarType : public framework::VarTypeInference {
}
};
class SplitIdsOpGradMaker : public framework::SingleGradOpDescMaker {
public:
using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
protected:
std::unique_ptr<framework::OpDesc> Apply() const override {
auto grad = new framework::OpDesc();
grad->SetType("concat");
grad->SetInput("X", OutputGrad("Out"));
grad->SetOutput("Out", InputGrad("Ids"));
grad->SetAttr("axis", 0);
return std::unique_ptr<framework::OpDesc>(grad);
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(split_ids, ops::SplitIdsOp, ops::SplitIdsOpMaker,
ops::SplitIdsOpInferVarType);
ops::SplitIdsOpGradMaker, ops::SplitIdsOpInferVarType);
REGISTER_OP_CPU_KERNEL(
split_ids, ops::SplitIdsOpKernel<paddle::platform::CPUPlace, int64_t>,
ops::SplitIdsOpKernel<paddle::platform::CPUPlace, float>);
paddle/fluid/operators/split_ids_op.h
浏览文件 @ b94f9b16
......@@ -14,6 +14,8 @@ limitations under the License. */
#pragma once
#include <iterator>
#include <set>
#include <unordered_map>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
......@@ -31,19 +33,39 @@ class SplitIdsOpKernel : public framework::OpKernel<T> {
PADDLE_THROW("SplitIds do not support GPU kernel");
}
const auto *ids_var = ctx.InputVar("Ids");
const auto ids_vars = ctx.MultiInputVar("Ids");
PADDLE_ENFORCE_GT(ids_vars.size(), 0, "The number of Ids should > 0");
auto *ids_var = ids_vars[0];
if (ids_var->IsType<framework::LoDTensor>()) {
const auto &ids_dims = ctx.Input<framework::LoDTensor>("Ids")->dims();
const T *ids = ctx.Input<framework::LoDTensor>("Ids")->data<T>();
int batch_size = 0;
const auto ids_tensors = ctx.MultiInput<framework::LoDTensor>("Ids");
for (size_t i = 0; i < ids_tensors.size(); ++i) {
batch_size += ids_tensors[i]->dims()[0];
}
VLOG(4) << "Get Total BatchSize is: " << batch_size;
std::vector<T> all_ids(batch_size);
int offset = 0;
for (size_t i = 0; i < ids_tensors.size(); ++i) {
const auto *ids = ids_tensors[i];
std::memcpy(all_ids.data() + offset, ids->data<T>(),
ids->numel() * sizeof(T));
offset += ids->numel();
}
std::set<T> st(all_ids.begin(), all_ids.end());
all_ids.assign(st.begin(), st.end());
auto outs = ctx.MultiOutput<framework::LoDTensor>("Out");
const size_t shard_num = outs.size();
std::vector<std::vector<T>> out_ids;
out_ids.resize(outs.size());
// split id by their shard_num.
for (int i = 0; i < ids_dims[0]; ++i) {
T id = ids[i];
for (int i = 0; i < all_ids.size(); ++i) {
T id = all_ids[i];
size_t shard_id = static_cast<size_t>(id) % shard_num;
out_ids[shard_id].push_back(id);
}
......@@ -64,7 +86,7 @@ class SplitIdsOpKernel : public framework::OpKernel<T> {
PADDLE_ENFORCE_EQ(ids_dims[0],
static_cast<int64_t>(ids_selected_rows->rows().size()),
"");
const T *ids = ids_selected_rows->value().data<T>();
const T *ids_data = ids_selected_rows->value().data<T>();
const auto &ids_rows = ids_selected_rows->rows();
auto outs = ctx.MultiOutput<framework::SelectedRows>("Out");
const size_t shard_num = outs.size();
......@@ -87,7 +109,7 @@ class SplitIdsOpKernel : public framework::OpKernel<T> {
T *output = out->mutable_value()->mutable_data<T>(ddim, place);
for (int64_t i = 0; i < ddim[0]; ++i) {
memcpy(output + i * row_width,
ids + id_to_index[out->rows()[i]] * row_width,
ids_data + id_to_index[out->rows()[i]] * row_width,
row_width * sizeof(T));
}
}
......
paddle/fluid/operators/split_selected_rows_op.cc
浏览文件 @ b94f9b16
......@@ -22,9 +22,9 @@ class SplitSelectedRowsOpMaker : public framework::OpProtoAndCheckerMaker {
void Make() override {
AddInput("X", "The input SelectedRows.");
AddOutput("Out", "The outputs of the input SelectedRows.").AsDuplicable();
AddAttr<std::vector<int>>("height_sections",
"Height for each output SelectedRows.")
.SetDefault(std::vector<int>({}));
AddAttr<std::vector<int64_t>>("height_sections",
"Height for each output SelectedRows.")
.SetDefault(std::vector<int64_t>({}));
AddComment(R"DOC(
Split a SelectedRows with a specified rows section.
......
paddle/fluid/operators/split_selected_rows_op.h
浏览文件 @ b94f9b16
......@@ -21,7 +21,7 @@ limitations under the License. */
namespace paddle {
namespace operators {
static int FindOutIdx(int row, const std::vector<int>& abs_sections) {
static int FindOutIdx(int row, const std::vector<int64_t>& abs_sections) {
for (size_t i = 1; i < abs_sections.size(); ++i) {
if (row < abs_sections[i]) {
return i - 1;
......@@ -30,9 +30,9 @@ static int FindOutIdx(int row, const std::vector<int>& abs_sections) {
return abs_sections.size() - 1;
}
static std::vector<int> ToAbsoluteSection(
const std::vector<int>& height_sections) {
std::vector<int> abs_sections;
static std::vector<int64_t> ToAbsoluteSection(
const std::vector<int64_t>& height_sections) {
std::vector<int64_t> abs_sections;
abs_sections.resize(height_sections.size());
abs_sections[0] = 0;
for (size_t i = 1; i < height_sections.size(); ++i) {
......@@ -47,7 +47,7 @@ class SplitSelectedRowsOpKernel : public framework::OpKernel<T> {
void Compute(const framework::ExecutionContext& ctx) const override {
auto* x = ctx.Input<framework::SelectedRows>("X");
auto outs = ctx.MultiOutput<framework::SelectedRows>("Out");
auto height_sections = ctx.Attr<std::vector<int>>("height_sections");
auto height_sections = ctx.Attr<std::vector<int64_t>>("height_sections");
auto abs_sections = ToAbsoluteSection(height_sections);
......
paddle/fluid/operators/sum_op.cc
浏览文件 @ b94f9b16
......@@ -82,14 +82,15 @@ class SumOp : public framework::OperatorWithKernel {
if (x_vars[0]->IsType<framework::LoDTensor>()) {
int dtype = -1;
for (auto& x_var : x_vars) {
auto& lod_tensor = x_var->Get<framework::LoDTensor>();
if (lod_tensor.numel() == 0) {
// FIXME(zcd): The input x_var may be SelectedRows or LoDTensor.
auto tensor = framework::GetTensorFromVar(x_var);
if (tensor->numel() == 0) {
continue;
}
if (dtype == -1) {
dtype = framework::ToDataType(lod_tensor.type());
dtype = framework::ToDataType(tensor->type());
} else {
PADDLE_ENFORCE_EQ(dtype, framework::ToDataType(lod_tensor.type()));
PADDLE_ENFORCE_EQ(dtype, framework::ToDataType(tensor->type()));
}
}
PADDLE_ENFORCE_NE(dtype, -1,
......
paddle/fluid/operators/sum_op.h
浏览文件 @ b94f9b16
......@@ -83,79 +83,54 @@ class SumKernel : public framework::OpKernel<T> {
}
}
} else if (out_var->IsType<framework::SelectedRows>()) {
std::unique_ptr<framework::SelectedRows> in0;
if (in_place) {
// If is in_place, we store the input[0] to in0
auto &in_sel0 = in_vars[0]->Get<SelectedRows>();
auto &rows = in_sel0.rows();
#ifdef PADDLE_WITH_CUDA
std::vector<int64_t> rows_in_cpu;
rows_in_cpu.reserve(rows.size());
for (auto item : rows) {
rows_in_cpu.push_back(item);
}
in0.reset(new framework::SelectedRows(rows_in_cpu, in_sel0.height()));
#else
in0.reset(new framework::SelectedRows(rows, in_sel0.height()));
#endif
in0->mutable_value()->ShareDataWith(in_sel0.value());
if (in_place && in_vars.size() < 2) {
return;
}
auto get_selected_row = [&](size_t i) -> const SelectedRows & {
if (i == 0 && in0) {
return *in0.get();
} else {
return in_vars[i]->Get<SelectedRows>();
std::vector<const paddle::framework::SelectedRows *> inputs;
SelectedRows temp_in0;
if (in_place) {
auto &in0 = in_vars[0]->Get<SelectedRows>();
temp_in0.set_height(in0.height());
temp_in0.set_rows(in0.rows());
framework::TensorCopy(in0.value(), in0.place(),
context.device_context(),
temp_in0.mutable_value());
inputs.push_back(&temp_in0);
for (size_t i = 1; i < in_vars.size(); ++i) {
auto &in = in_vars[i]->Get<SelectedRows>();
if (in.rows().size() > 0) {
inputs.push_back(&in);
}
}
} else {
for (auto &in_var : in_vars) {
auto &in = in_var->Get<SelectedRows>();
if (in.rows().size() > 0) {
inputs.push_back(&in_var->Get<SelectedRows>());
}
}
};
}
auto *out = context.Output<SelectedRows>("Out");
out->mutable_rows()->clear();
auto *out_value = out->mutable_value();
// Runtime InferShape
size_t first_dim = 0;
for (size_t i = 0; i < in_num; i++) {
auto &sel_row = get_selected_row(i);
first_dim += sel_row.rows().size();
}
std::vector<int64_t> in_dim;
for (size_t i = 0; i < in_num; i++) {
auto &sel_row = get_selected_row(i);
if (sel_row.rows().size() > 0) {
in_dim = framework::vectorize(sel_row.value().dims());
bool has_data = false;
for (auto &in : inputs) {
if (in->rows().size() > 0) {
has_data = true;
break;
}
}
if (in_dim.empty()) {
VLOG(3) << "WARNING: all the inputs are empty";
in_dim =
framework::vectorize(get_selected_row(in_num - 1).value().dims());
if (has_data) {
math::scatter::MergeAdd<DeviceContext, T> merge_add;
merge_add(context.template device_context<DeviceContext>(), inputs,
out);
} else {
in_dim[0] = static_cast<int64_t>(first_dim);
}
out_value->Resize(framework::make_ddim(in_dim));
out_value->mutable_data<T>(context.GetPlace());
// if all the input sparse vars are empty, no need to
// merge these vars.
if (first_dim == 0UL) {
return;
}
math::SelectedRowsAddTo<DeviceContext, T> functor;
int64_t offset = 0;
for (size_t i = 0; i < in_num; i++) {
auto &sel_row = get_selected_row(i);
if (sel_row.rows().size() == 0) {
continue;
}
PADDLE_ENFORCE_EQ(out->height(), sel_row.height());
functor(context.template device_context<DeviceContext>(), sel_row,
offset, out);
offset += sel_row.value().numel();
// no data, just set a empty out tensor.
out->mutable_value()->mutable_data<T>(framework::make_ddim({0}),
context.GetPlace());
}
} else if (out_var->IsType<framework::LoDTensorArray>()) {
auto &out_array = *out_var->GetMutable<framework::LoDTensorArray>();
......
paddle/fluid/operators/uniform_random_op.cc
浏览文件 @ b94f9b16
......@@ -29,7 +29,7 @@ class CPUUniformRandomKernel : public framework::OpKernel<T> {
if (out_var->IsType<framework::LoDTensor>()) {
tensor = out_var->GetMutable<framework::LoDTensor>();
} else if (out_var->IsType<framework::SelectedRows>()) {
auto shape = ctx.Attr<std::vector<int>>("shape");
auto shape = ctx.Attr<std::vector<int64_t>>("shape");
auto *selected_rows = out_var->GetMutable<framework::SelectedRows>();
tensor = selected_rows->mutable_value();
tensor->Resize(framework::make_ddim(shape));
......@@ -67,7 +67,7 @@ class UniformRandomOp : public framework::OperatorWithKernel {
PADDLE_ENFORCE(
ctx->Attrs().Get<float>("min") < ctx->Attrs().Get<float>("max"),
"uniform_random's min must less then max");
auto &shape = ctx->Attrs().Get<std::vector<int>>("shape");
auto &shape = ctx->Attrs().Get<std::vector<int64_t>>("shape");
std::vector<int64_t> temp;
temp.reserve(shape.size());
for (auto dim : shape) {
......@@ -94,7 +94,7 @@ This operator initializes a tensor with random values sampled from a
uniform distribution. The random result is in set [min, max].
)DOC");
AddAttr<std::vector<int>>("shape", "The shape of the output tensor");
AddAttr<std::vector<int64_t>>("shape", "The shape of the output tensor");
AddAttr<float>("min", "Minimum value of uniform random. [default -1.0].")
.SetDefault(-1.0f);
AddAttr<float>("max", "Maximun value of uniform random. [default 1.0].")
......
paddle/fluid/operators/uniform_random_op.cu
浏览文件 @ b94f9b16
......@@ -48,7 +48,7 @@ class GPUUniformRandomKernel : public framework::OpKernel<T> {
if (out_var->IsType<framework::LoDTensor>()) {
tensor = out_var->GetMutable<framework::LoDTensor>();
} else if (out_var->IsType<framework::SelectedRows>()) {
auto shape = context.Attr<std::vector<int>>("shape");
auto shape = context.Attr<std::vector<int64_t>>("shape");
tensor = out_var->GetMutable<framework::SelectedRows>()->mutable_value();
tensor->Resize(framework::make_ddim(shape));
} else {
......
paddle/fluid/platform/init.cc
浏览文件 @ b94f9b16
......@@ -116,21 +116,51 @@ void InitDevices(bool init_p2p, const std::vector<int> devices) {
platform::SetNumThreads(FLAGS_paddle_num_threads);
#endif
if (platform::jit::MayIUse(platform::jit::avx512f)) {
#ifndef __AVX512F__
LOG(WARNING) << "AVX512F is available, Please re-compile on local machine";
#if !defined(_WIN32) && !defined(__APPLE__) && !defined(__OSX__)
if (platform::jit::MayIUse(platform::jit::avx)) {
#ifndef __AVX__
LOG(WARNING) << "AVX is available, Please re-compile on local machine";
#endif
}
if (platform::jit::MayIUse(platform::jit::avx2)) {
#ifndef __AVX2__
LOG(WARNING) << "AVX2 is available, Please re-compile on local machine";
// Throw some informations when CPU instructions mismatch.
#define AVX_GUIDE(compiletime, runtime) \
LOG(FATAL) \
<< "This version is compiled on higher instruction(" #compiletime \
") system, you may encounter illegal instruction error running on" \
" your local CPU machine. Please reinstall the " #runtime \
" version or compile from source code."
#ifdef __AVX512F__
if (!platform::jit::MayIUse(platform::jit::avx512f)) {
if (platform::jit::MayIUse(platform::jit::avx2)) {
AVX_GUIDE(AVX512, AVX2);
} else if (platform::jit::MayIUse(platform::jit::avx)) {
AVX_GUIDE(AVX512, AVX);
} else {
AVX_GUIDE(AVX512, NonAVX);
}
}
#endif
#ifdef __AVX2__
if (!platform::jit::MayIUse(platform::jit::avx2)) {
if (platform::jit::MayIUse(platform::jit::avx)) {
AVX_GUIDE(AVX2, AVX);
} else {
AVX_GUIDE(AVX2, NonAVX);
}
}
if (platform::jit::MayIUse(platform::jit::avx)) {
#ifndef __AVX__
LOG(WARNING) << "AVX is available, Please re-compile on local machine";
#endif
#ifdef __AVX__
if (!platform::jit::MayIUse(platform::jit::avx)) {
AVX_GUIDE(AVX, NonAVX);
}
#endif
#undef AVX_GUIDE
#endif
}
void InitGLOG(const std::string &prog_name) {
......
paddle/fluid/pybind/protobuf.cc
浏览文件 @ b94f9b16
......@@ -57,6 +57,18 @@ struct variant_caster<V<Ts...>> {
auto caster = make_caster<T>();
if (!load_success_ && caster.load(src, convert)) {
load_success_ = true;
if (std::is_same<T, std::vector<float>>::value) {
auto caster_ints = make_caster<std::vector<int64_t>>();
if (caster_ints.load(src, convert)) {
VLOG(4) << "This value are floats and int64_ts satisfy "
"simultaneously, will set it's type to "
"std::vector<int64_t>";
value = cast_op<std::vector<int64_t>>(caster_ints);
return true;
}
}
value = cast_op<T>(caster);
return true;
}
......@@ -259,6 +271,8 @@ void BindOpDesc(pybind11::module *m) {
pybind11::enum_<pd::proto::AttrType>(*m, "AttrType", "")
.value("INT", pd::proto::AttrType::INT)
.value("INTS", pd::proto::AttrType::INTS)
.value("LONG", pd::proto::AttrType::LONG)
.value("LONGS", pd::proto::AttrType::LONGS)
.value("FLOAT", pd::proto::AttrType::FLOAT)
.value("FLOATS", pd::proto::AttrType::FLOATS)
.value("STRING", pd::proto::AttrType::STRING)
......
paddle/fluid/pybind/pybind.cc
浏览文件 @ b94f9b16
......@@ -645,9 +645,13 @@ All parameter, weight, gradient are variables in Paddle.
py::class_<ir::Pass, std::shared_ptr<ir::Pass>> pass(m, "Pass");
pass.def(py::init())
.def("set_str", [](ir::Pass &self, const std::string &name,
const std::string &attr) {
self.Set<std::string>(name, new std::string(attr));
.def(
"set_str",
[](ir::Pass &self, const std::string &name, const std::string &attr) {
self.Set<std::string>(name, new std::string(attr));
})
.def("set_int", [](ir::Pass &self, const std::string &name, int val) {
self.Set<const int>(name, new int(val));
});
py::class_<ir::PassBuilder, std::shared_ptr<ir::PassBuilder>> pb(
......
paddle/fluid/train/demo/CMakeLists.txt
浏览文件 @ b94f9b16
......@@ -15,6 +15,7 @@ include_directories("${PADDLE_LIB}")
include_directories("${PADDLE_LIB}/third_party/install/protobuf/include")
include_directories("${PADDLE_LIB}/third_party/install/glog/include")
include_directories("${PADDLE_LIB}/third_party/install/gflags/include")
include_directories("${PADDLE_LIB}/third_party/install/xxhash/include")
include_directories("${PADDLE_LIB}/third_party/install/snappy/include")
include_directories("${PADDLE_LIB}/third_party/install/snappystream/include")
include_directories("${PADDLE_LIB}/third_party/install/zlib/include")
......@@ -27,6 +28,7 @@ link_directories("${PADDLE_LIB}/third_party/install/snappystream/lib")
link_directories("${PADDLE_LIB}/third_party/install/protobuf/lib")
link_directories("${PADDLE_LIB}/third_party/install/glog/lib")
link_directories("${PADDLE_LIB}/third_party/install/gflags/lib")
link_directories("${PADDLE_LIB}/third_party/install/xxhash/lib")
link_directories("${PADDLE_LIB}/third_party/install/zlib/lib")
add_executable(demo_trainer demo_trainer.cc)
......@@ -62,5 +64,5 @@ target_link_libraries(demo_trainer
${ARCHIVE_END}
${MATH_LIB}
${MKLDNN_LIB}
glog gflags protobuf snappystream snappy z
glog gflags protobuf snappystream snappy z xxhash
${EXTERNAL_LIB})
paddle/scripts/paddle_build.sh
浏览文件 @ b94f9b16
......@@ -95,9 +95,9 @@ function cmake_gen() {
exit 1
fi
fi
else
else
if [ "$1" != "" ]; then
echo "using python abi: $1"
echo "using python abi: $1"
if [ "$1" == "cp27-cp27m" ]; then
export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs2/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs4/lib:}
export PATH=/opt/python/cp27-cp27m/bin/:${PATH}
......@@ -119,7 +119,7 @@ function cmake_gen() {
fi
fi
fi
if [ "$SYSTEM" == "Darwin" ]; then
WITH_DISTRIBUTE=${WITH_DISTRIBUTE:-ON}
WITH_AVX=${WITH_AVX:-ON}
......@@ -127,7 +127,7 @@ function cmake_gen() {
else
INFERENCE_DEMO_INSTALL_DIR=${INFERENCE_DEMO_INSTALL_DIR:-/root/.cache/inference_demo}
fi
cat <<EOF
========================================
Configuring cmake in /paddle/build ...
......@@ -394,8 +394,8 @@ EOF
export http_proxy=
export https_proxy=
# TODO: jiabin need to refine this part when these tests fixed on mac
ctest --output-on-failure -j $1
# make install should also be test when unittest
ctest --output-on-failure -j $1
# make install should also be test when unittest
make install -j 8
pip install --user ${INSTALL_PREFIX:-/paddle/build}/opt/paddle/share/wheels/*.whl
if [[ ${WITH_FLUID_ONLY:-OFF} == "OFF" ]] ; then
......@@ -659,7 +659,7 @@ function gen_fluid_lib() {
Generating fluid library for train and inference ...
========================================
EOF
cmake .. -DWITH_DISTRIBUTE=OFF
cmake .. -DWITH_DISTRIBUTE=OFF -DON_INFER=ON
make -j `nproc` fluid_lib_dist
make -j `nproc` inference_lib_dist
fi
......
python/paddle/dataset/wmt16.py
浏览文件 @ b94f9b16
......@@ -78,7 +78,7 @@ def __build_dict(tar_file, dict_size, save_path, lang):
six.iteritems(word_dict), key=lambda x: x[1],
reverse=True)):
if idx + 3 == dict_size: break
fout.write("%s\n" % (word[0]))
fout.write("%s\n" % (cpt.to_bytes(word[0])))
def __load_dict(tar_file, dict_size, lang, reverse=False):
......
python/paddle/fluid/__init__.py
浏览文件 @ b94f9b16
......@@ -121,6 +121,9 @@ def __bootstrap__():
read_env_flags.append('rpc_server_profile_period')
read_env_flags.append('rpc_server_profile_path')
read_env_flags.append('enable_rpc_profiler')
read_env_flags.append('rpc_send_thread_num')
read_env_flags.append('rpc_get_thread_num')
read_env_flags.append('rpc_prefetch_thread_num')
if core.is_compiled_with_cuda():
read_env_flags += [
......
python/paddle/fluid/evaluator.py
浏览文件 @ b94f9b16
......@@ -316,7 +316,7 @@ class DetectionMAP(Evaluator):
gt_label (Variable): The ground truth label index, which is a LoDTensor
with shape [N, 1].
gt_box (Variable): The ground truth bounding box (bbox), which is a
LoDTensor with shape [N, 6]. The layout is [xmin, ymin, xmax, ymax].
LoDTensor with shape [N, 4]. The layout is [xmin, ymin, xmax, ymax].
gt_difficult (Variable|None): Whether this ground truth is a difficult
bounding bbox, which can be a LoDTensor [N, 1] or not set. If None,
it means all the ground truth labels are not difficult bbox.
......
python/paddle/fluid/layers/learning_rate_scheduler.py
浏览文件 @ b94f9b16
......@@ -27,7 +27,7 @@ from . import nn
from . import ops
from . import tensor
from ..initializer import init_on_cpu
from ..framework import default_main_program, Parameter, unique_name
from ..framework import default_main_program, Parameter, unique_name, name_scope
__all__ = [
'exponential_decay', 'natural_exp_decay', 'inverse_time_decay',
......@@ -332,14 +332,16 @@ def append_LARS(params_grads, learning_rate, weight_decay):
return grad_norm + weight_decay * param_norm
for param, grad in params_grads:
param_lr = param.optimize_attr['learning_rate']
param_norm = ops.sqrt(nn.reduce_sum(input=ops.square(param)))
grad_norm = ops.sqrt(nn.reduce_sum(input=ops.square(grad)))
if type(param_lr) == float and param_lr == 1.0:
decayed_lr = learning_rate * param_norm \
/ _balanced_weight(param_norm, grad_norm)
else:
decayed_lr = learning_rate * param_lr * param_norm \
/ _balanced_weight(param_norm, grad_norm)
# set back param local learning rate
param.optimize_attr['learning_rate'] = decayed_lr
with param.block.program.optimized_guard(
[param, grad]), name_scope("optimizer"):
param_lr = param.optimize_attr['learning_rate']
param_norm = ops.sqrt(nn.reduce_sum(input=ops.square(param)))
grad_norm = ops.sqrt(nn.reduce_sum(input=ops.square(grad)))
if type(param_lr) == float and param_lr == 1.0:
decayed_lr = learning_rate * param_norm \
/ _balanced_weight(param_norm, grad_norm)
else:
decayed_lr = learning_rate * param_lr * param_norm \
/ _balanced_weight(param_norm, grad_norm)
# set back param local learning rate
param.optimize_attr['learning_rate'] = decayed_lr
python/paddle/fluid/layers/nn.py
浏览文件 @ b94f9b16
......@@ -156,6 +156,7 @@ __all__ = [
'maxout',
'sequence_reverse',
'affine_channel',
'hash',
]
......@@ -1991,17 +1992,17 @@ def sequence_slice(input, offset, length, name=None):
"""
**Sequence Slice Layer**
The layer crops a subsequence from given sequence with given start
The layer crops a subsequence from given sequence with given start
offset and subsequence length.
It only supports sequence data (LoDTensor with lod_level equal to 1).
.. code-block:: text
- Case:
Given the input Variable **input**:
input.data = [[a1, a2], [b1, b2], [c1, c2], [d1, d2], [e1, e2]],
input.lod = [[3, 2]],
input.dims = (5, 2),
......@@ -2009,16 +2010,16 @@ def sequence_slice(input, offset, length, name=None):
with offset.data = [[0], [1]] and length.data = [[2], [1]],
the output Variable will be
out.data = [[a1, a2], [b1, b2], [e1, e2]],
out.lod = [[2, 1]],
out.dims = (3, 2).
NOTE: The first dimension size of **input**, **offset** and **length**
NOTE: The first dimension size of **input**, **offset** and **length**
should be equal. The **offset** should start from 0.
Args:
input(Variable): The input Variable which consists of the complete
input(Variable): The input Variable which consists of the complete
sequences.
offset(Variable): The offset to slice each sequence.
length(Variable): The length of each subsequence.
......@@ -2037,7 +2038,7 @@ def sequence_slice(input, offset, length, name=None):
dtype='float32', lod_level=1)
offset = fluid.layers.assign(input=np.array([[0, 1]]).astype("int32"))
length = fluid.layers.assign(input=np.array([[2, 1]]).astype("int32"))
subseqs = fluid.layers.sequence_slice(input=seqs, offset=offset,
subseqs = fluid.layers.sequence_slice(input=seqs, offset=offset,
length=length)
"""
helper = LayerHelper("sequence_slice", **locals())
......@@ -2420,12 +2421,12 @@ def layer_norm(input,
param_attr(ParamAttr|None): The parameter attribute for the learnable
gain :math:`g`. If :attr:`scale` is False, :attr:`param_attr` is
omitted. If :attr:`scale` is True and :attr:`param_attr` is None,
a default :code:`ParamAttr` would be added as scale. The
:attr:`param_attr` is initialized as 1 if it is added. Default None.
a default :code:`ParamAttr` would be added as scale. The
:attr:`param_attr` is initialized as 1 if it is added. Default None.
bias_attr(ParamAttr|None): The parameter attribute for the learnable
bias :math:`b`. If :attr:`shift` is False, :attr:`bias_attr` is
omitted. If :attr:`shift` is True and :attr:`param_attr` is None,
a default :code:`ParamAttr` would be added as bias. The
a default :code:`ParamAttr` would be added as bias. The
:attr:`bias_attr` is initialized as 0 if it is added. Default None.
act(str): Activation to be applied to the output of layer normalizaiton.
Default None.
......@@ -3043,8 +3044,8 @@ def sequence_unpad(x, length, name=None):
"""
**Sequence Unpad Layer**
This layer removes the padding data in the input sequences and convert
them into sequences with actual length as output, identitied by lod
This layer removes the padding data in the input sequences and convert
them into sequences with actual length as output, identitied by lod
information.
.. code-block:: text
......@@ -3054,9 +3055,9 @@ def sequence_unpad(x, length, name=None):
Given input Variable **x**:
x.data = [[ 1.0, 2.0, 3.0, 4.0, 5.0],
[ 6.0, 7.0, 8.0, 9.0, 10.0],
[11.0, 12.0, 13.0, 14.0, 15.0]],
in which there are 3 sequences padded to length 5, and the acutal length
[11.0, 12.0, 13.0, 14.0, 15.0]],
in which there are 3 sequences padded to length 5, and the acutal length
specified by input Variable **length**:
length.data = [[2], [3], [4]],
......@@ -3064,7 +3065,7 @@ def sequence_unpad(x, length, name=None):
after unpadding, the output Variable will be:
out.data = [[1.0, 2.0, 6.0, 7.0, 8.0, 11.0, 12.0, 13.0, 14.0]]
out.lod = [[2, 3, 4]]
out.lod = [[2, 3, 4]]
Args:
x(Variable): Input Variable which contains the padded sequences with
......@@ -5499,9 +5500,9 @@ def roi_align(input,
Examples:
.. code-block:: python
align_out = fluid.layers.roi_align(input=x,
rois=rois,
pooled_height=7,
align_out = fluid.layers.roi_align(input=x,
rois=rois,
pooled_height=7,
pooled_width=7,
spatial_scale=0.5,
sampling_ratio=-1)
......@@ -7498,7 +7499,6 @@ def sequence_reverse(x, name=None):
out(${y_type}): ${y_comment}
"""
helper = LayerHelper("sequence_reverse", **locals())
if name is None:
out = helper.create_variable_for_type_inference(dtype=x.dtype)
else:
......@@ -7519,7 +7519,7 @@ def affine_channel(x, scale=None, bias=None, data_layout='NCHW', name=None):
Useful for replacing spatial batch norm with its equivalent fixed
transformation. The input also can be 2D tensor and applies a affine
transformation in second dimension.
Args:
x (Variable): Feature map input can be a 4D tensor with order NCHW
or NHWC. It also can be a 2D tensor and the affine transformation
......@@ -7552,3 +7552,31 @@ def affine_channel(x, scale=None, bias=None, data_layout='NCHW', name=None):
attrs={"data_layout": data_layout},
outputs={"Out": out})
return out
def hash(input, hash_size, num_hash=1, name=None):
"""
hash the input
Args:
input (Variable): The input variable which is a one-hot word.
hash_size (int): The space size for hash algorithm.
num_hash (int): The times of hash, default 1.
name (str, default None): The name of this layer.
Returns:
Variable: The hash result variable which is a LoDTensor.
Examples:
.. code-block:: python
word_dict = paddle.dataset.imdb.word_dict()
x = fluid.layers.data(shape[1], dtype='int32', lod_level=1)
out = fluid.layers.hash(input=x, len(word_dict))
"""
helper = LayerHelper('hash', **locals())
out = helper.create_variable_for_type_inference(
helper.input_dtype(), stop_gradient=True)
helper.append_op(
type='hash',
inputs={'X': input},
outputs={'Out': out},
attrs={'num_hash': num_hash,
'mod_by': hash_size})
return out
python/paddle/fluid/metrics.py
浏览文件 @ b94f9b16
......@@ -13,8 +13,6 @@
# limitations under the License.
"""
Fluid Metrics
The metrics are accomplished via Python natively.
"""
from __future__ import print_function
......@@ -24,6 +22,12 @@ import copy
import warnings
import six
from .layer_helper import LayerHelper
from .initializer import Constant
from . import unique_name
from .framework import Program, Variable, program_guard
from . import layers
__all__ = [
'MetricBase',
'CompositeMetric',
......@@ -478,67 +482,6 @@ class EditDistance(MetricBase):
return avg_distance, avg_instance_error
class DetectionMAP(MetricBase):
"""
Calculate the detection mean average precision (mAP).
mAP is the metric to measure the accuracy of object detectors
like Faster R-CNN, SSD, etc.
It is the average of the maximum precisions at different recall values.
Please get more information from the following articles:
https://sanchom.wordpress.com/tag/average-precision/
https://arxiv.org/abs/1512.02325
The general steps are as follows:
1. calculate the true positive and false positive according to the input
of detection and labels.
2. calculate mAP value, support two versions: '11 point' and 'integral'.
Examples:
.. code-block:: python
pred = fluid.layers.fc(input=data, size=1000, act="tanh")
batch_map = layers.detection_map(
input,
label,
class_num,
background_label,
overlap_threshold=overlap_threshold,
evaluate_difficult=evaluate_difficult,
ap_version=ap_version)
metric = fluid.metrics.DetectionMAP()
for data in train_reader():
loss, preds, labels = exe.run(fetch_list=[cost, batch_map])
batch_size = data[0]
metric.update(value=batch_map, weight=batch_size)
numpy_map = metric.eval()
"""
def __init__(self, name=None):
super(DetectionMAP, self).__init__(name)
# the current map value
self.value = .0
self.weight = .0
def update(self, value, weight):
if not _is_number_or_matrix_(value):
raise ValueError(
"The 'value' must be a number(int, float) or a numpy ndarray.")
if not _is_number_(weight):
raise ValueError("The 'weight' must be a number(int, float).")
self.value += value
self.weight += weight
def eval(self):
if self.weight == 0:
raise ValueError(
"There is no data in DetectionMAP Metrics. "
"Please check layers.detection_map output has added to DetectionMAP."
)
return self.value / self.weight
class Auc(MetricBase):
"""
Auc metric adapts to the binary classification.
......@@ -616,3 +559,179 @@ class Auc(MetricBase):
idx -= 1
return auc / tot_pos / tot_neg if tot_pos > 0.0 and tot_neg > 0.0 else 0.0
class DetectionMAP(object):
"""
Calculate the detection mean average precision (mAP).
The general steps are as follows:
1. calculate the true positive and false positive according to the input
of detection and labels.
2. calculate mAP value, support two versions: '11 point' and 'integral'.
Please get more information from the following articles:
https://sanchom.wordpress.com/tag/average-precision/
https://arxiv.org/abs/1512.02325
Args:
input (Variable): The detection results, which is a LoDTensor with shape
[M, 6]. The layout is [label, confidence, xmin, ymin, xmax, ymax].
gt_label (Variable): The ground truth label index, which is a LoDTensor
with shape [N, 1].
gt_box (Variable): The ground truth bounding box (bbox), which is a
LoDTensor with shape [N, 4]. The layout is [xmin, ymin, xmax, ymax].
gt_difficult (Variable|None): Whether this ground truth is a difficult
bounding bbox, which can be a LoDTensor [N, 1] or not set. If None,
it means all the ground truth labels are not difficult bbox.
class_num (int): The class number.
background_label (int): The index of background label, the background
label will be ignored. If set to -1, then all categories will be
considered, 0 by defalut.
overlap_threshold (float): The threshold for deciding true/false
positive, 0.5 by defalut.
evaluate_difficult (bool): Whether to consider difficult ground truth
for evaluation, True by defalut. This argument does not work when
gt_difficult is None.
ap_version (string): The average precision calculation ways, it must be
'integral' or '11point'. Please check
https://sanchom.wordpress.com/tag/average-precision/ for details.
- 11point: the 11-point interpolated average precision.
- integral: the natural integral of the precision-recall curve.
Examples:
.. code-block:: python
exe = fluid.Executor(place)
map_evaluator = fluid.Evaluator.DetectionMAP(input,
gt_label, gt_box, gt_difficult)
cur_map, accum_map = map_evaluator.get_map_var()
fetch = [cost, cur_map, accum_map]
for epoch in PASS_NUM:
map_evaluator.reset(exe)
for data in batches:
loss, cur_map_v, accum_map_v = exe.run(fetch_list=fetch)
In the above example:
'cur_map_v' is the mAP of current mini-batch.
'accum_map_v' is the accumulative mAP of one pass.
"""
def __init__(self,
input,
gt_label,
gt_box,
gt_difficult=None,
class_num=None,
background_label=0,
overlap_threshold=0.5,
evaluate_difficult=True,
ap_version='integral'):
self.helper = LayerHelper('map_eval')
gt_label = layers.cast(x=gt_label, dtype=gt_box.dtype)
if gt_difficult:
gt_difficult = layers.cast(x=gt_difficult, dtype=gt_box.dtype)
label = layers.concat([gt_label, gt_difficult, gt_box], axis=1)
else:
label = layers.concat([gt_label, gt_box], axis=1)
# calculate mean average precision (mAP) of current mini-batch
map = layers.detection_map(
input,
label,
class_num,
background_label,
overlap_threshold=overlap_threshold,
evaluate_difficult=evaluate_difficult,
ap_version=ap_version)
states = []
states.append(
self._create_state(
dtype='int32', shape=None, suffix='accum_pos_count'))
states.append(
self._create_state(
dtype='float32', shape=None, suffix='accum_true_pos'))
states.append(
self._create_state(
dtype='float32', shape=None, suffix='accum_false_pos'))
var = self._create_state(dtype='int32', shape=[1], suffix='has_state')
self.helper.set_variable_initializer(
var, initializer=Constant(value=int(0)))
self.has_state = var
# calculate accumulative mAP
accum_map = layers.detection_map(
input,
label,
class_num,
background_label,
overlap_threshold=overlap_threshold,
evaluate_difficult=evaluate_difficult,
has_state=self.has_state,
input_states=states,
out_states=states,
ap_version=ap_version)
layers.fill_constant(
shape=self.has_state.shape,
value=1,
dtype=self.has_state.dtype,
out=self.has_state)
self.cur_map = map
self.accum_map = accum_map
def _create_state(self, suffix, dtype, shape):
"""
Create state variable.
Args:
suffix(str): the state suffix.
dtype(str|core.VarDesc.VarType): the state data type
shape(tuple|list): the shape of state
Returns: State variable
"""
state = self.helper.create_variable(
name="_".join([unique_name.generate(self.helper.name), suffix]),
persistable=True,
dtype=dtype,
shape=shape)
return state
def get_map_var(self):
"""
Returns: mAP variable of current mini-batch and
accumulative mAP variable cross mini-batches.
"""
return self.cur_map, self.accum_map
def reset(self, executor, reset_program=None):
"""
Reset metric states at the begin of each pass/user specified batch.
Args:
executor(Executor): a executor for executing
the reset_program.
reset_program(Program|None): a single Program for reset process.
If None, will create a Program.
"""
def _clone_var_(block, var):
assert isinstance(var, Variable)
return block.create_var(
name=var.name,
shape=var.shape,
dtype=var.dtype,
type=var.type,
lod_level=var.lod_level,
persistable=var.persistable)
if reset_program is None:
reset_program = Program()
with program_guard(main_program=reset_program):
var = _clone_var_(reset_program.current_block(), self.has_state)
layers.fill_constant(
shape=var.shape, value=0, dtype=var.dtype, out=var)
executor.run(reset_program)
python/paddle/fluid/op.py
浏览文件 @ b94f9b16
......@@ -120,6 +120,8 @@ class OpDescCreationMethod(object):
new_attr.strings.extend(user_defined_attr)
elif attr.type == framework_pb2.BOOLEANS:
new_attr.bools.extend(user_defined_attr)
elif attr.type == framework_pb2.LONGS:
new_attr.longs.extend(user_defined_attr)
elif attr.type == framework_pb2.INT_PAIRS:
for p in user_defined_attr:
pair = new_attr.int_pairs.add()
......
python/paddle/fluid/optimizer.py
浏览文件 @ b94f9b16
......@@ -14,6 +14,7 @@
from __future__ import print_function
import re
import sys
from collections import defaultdict
from paddle.fluid.framework import Program, Variable, name_scope, default_main_program
from . import framework
......@@ -32,7 +33,8 @@ __all__ = [
'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad', 'Ftrl',
'SGDOptimizer', 'MomentumOptimizer', 'AdagradOptimizer', 'AdamOptimizer',
'AdamaxOptimizer', 'DecayedAdagradOptimizer', 'RMSPropOptimizer',
'FtrlOptimizer', 'Adadelta', 'ModelAverage', 'RMSPropOptimizer'
'FtrlOptimizer', 'Adadelta', 'ModelAverage', 'LarsMomentum',
'LarsMomentumOptimizer'
]
......@@ -105,7 +107,6 @@ class Optimizer(object):
param = param_and_grad[0]
param_lr = param.optimize_attr['learning_rate']
if type(param_lr) == Variable:
print("returns updated param lr ", param_lr)
return param_lr
else:
if param_lr == 1.0:
......@@ -400,6 +401,91 @@ class MomentumOptimizer(Optimizer):
return momentum_op
class LarsMomentumOptimizer(Optimizer):
"""
Momentum optimizer with LARS support
The update equations are as follows:
.. math::
& local\_learning\_rate = learning\_rate * lars\_coeff * \\
\\frac{||param||}{||gradient|| + lars\_weight\_decay * ||param||}
& velocity = mu * velocity + local\_learning\_rate * (gradient + lars\_weight\_decay * param)
& param = param - velocity
Args:
learning_rate (float|Variable): the learning rate used to update parameters. \
Can be a float value or a Variable with one float value as data element.
momentum (float): momentum factor
lars_coeff (float): defines how much we trust the layer to change its weights.
lars_weight_decay (float): weight decay coefficient for decaying using LARS.
regularization: A Regularizer, such as
fluid.regularizer.L2DecayRegularizer.
name: A optional name prefix.
Examples:
.. code-block:: python
optimizer = fluid.optimizer.LarsMomentum(learning_rate=0.2, momentum=0.1, lars_weight_decay=0.001)
optimizer.minimize(cost)
"""
_velocity_acc_str = "velocity"
def __init__(self,
learning_rate,
momentum,
lars_coeff=0.001,
lars_weight_decay=0.0005,
regularization=None,
name=None):
assert learning_rate is not None
assert momentum is not None
super(LarsMomentumOptimizer, self).__init__(
learning_rate=learning_rate,
regularization=regularization,
name=name)
self.type = "lars_momentum"
self._momentum = momentum
self._lars_coeff = float(lars_coeff)
self._lars_weight_decay = float(lars_weight_decay)
def _create_accumulators(self, block, parameters):
assert isinstance(block, framework.Block)
for p in parameters:
self._add_accumulator(self._velocity_acc_str, p)
def _append_optimize_op(self, block, param_and_grad):
assert isinstance(block, framework.Block)
velocity_acc = self._get_accumulator(self._velocity_acc_str,
param_and_grad[0])
# create the momentum optimize op
momentum_op = block.append_op(
type=self.type,
inputs={
"Param": param_and_grad[0],
"Grad": param_and_grad[1],
"Velocity": velocity_acc,
"LearningRate": self._create_param_lr(param_and_grad)
},
outputs={
"ParamOut": param_and_grad[0],
"VelocityOut": velocity_acc
},
attrs={
"mu": self._momentum,
"lars_coeff": self._lars_coeff,
"lars_weight_decay": self._lars_weight_decay
})
return momentum_op
class AdagradOptimizer(Optimizer):
"""
**Adaptive Gradient Algorithm (Adagrad)**
......@@ -1221,6 +1307,7 @@ DecayedAdagrad = DecayedAdagradOptimizer
Adadelta = AdadeltaOptimizer
RMSProp = RMSPropOptimizer
Ftrl = FtrlOptimizer
LarsMomentum = LarsMomentumOptimizer
class ModelAverage(Optimizer):
......
python/paddle/fluid/tests/CMakeLists.txt
浏览文件 @ b94f9b16
set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests CACHE INTERNAL "python tests directory")
file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
......
python/paddle/fluid/tests/book/high-level-api/image_classification/CMakeLists.txt
浏览文件 @ b94f9b16
file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
# default test
foreach(src ${TEST_OPS})
py_test(${src} SRCS ${src}.py)
endforeach()
if(NOT APPLE)
# default test
foreach(src ${TEST_OPS})
py_test(${src} SRCS ${src}.py)
endforeach()
else()
foreach(src ${TEST_OPS})
if(${src} STREQUAL "test_image_classification_vgg")
message(WARNING "These tests has been disabled in OSX for random fail: \n" ${src})
elseif(${src} STREQUAL "test_image_classification_resnet")
message(WARNING "These tests has been disabled in OSX for random fail: \n" ${src})
elseif()
py_test(${src} SRCS ${src}.py)
endif()
endforeach()
endif()
python/paddle/fluid/tests/unittests/CMakeLists.txt
浏览文件 @ b94f9b16
......@@ -17,6 +17,10 @@ if(NOT WITH_DISTRIBUTE)
list(REMOVE_ITEM TEST_OPS test_listen_and_serv_op)
LIST(REMOVE_ITEM TEST_OPS test_dist_mnist)
LIST(REMOVE_ITEM TEST_OPS test_dist_word2vec)
LIST(REMOVE_ITEM TEST_OPS test_dist_ctr)
LIST(REMOVE_ITEM TEST_OPS test_dist_simnet_bow)
LIST(REMOVE_ITEM TEST_OPS test_dist_mnist_batch_merge)
LIST(REMOVE_ITEM TEST_OPS test_dist_text_classification)
endif(NOT WITH_DISTRIBUTE)
list(REMOVE_ITEM TEST_OPS test_seq_concat_op) # FIXME(helin): https://github.com/PaddlePaddle/Paddle/issues/8290
......@@ -88,4 +92,6 @@ py_test_modules(test_parallel_executor_crf MODULES test_parallel_executor_crf SE
py_test_modules(test_parallel_executor_fetch_feed MODULES test_parallel_executor_fetch_feed SERIAL)
set_tests_properties(test_parallel_executor_fetch_feed PROPERTIES TIMEOUT 150)
py_test_modules(test_parallel_executor_transformer MODULES test_parallel_executor_transformer SERIAL)
py_test_modules(test_image_classification_resnet MODULES test_image_classification_resnet SERIAL)
if(NOT APPLE)
py_test_modules(test_image_classification_resnet MODULES test_image_classification_resnet SERIAL)
endif()
python/paddle/fluid/tests/unittests/dist_mnist.py
浏览文件 @ b94f9b16
......@@ -95,7 +95,7 @@ class TestDistMnist2x2(TestDistRunnerBase):
# Reader
train_reader = paddle.batch(
paddle.dataset.mnist.train(), batch_size=batch_size)
paddle.dataset.mnist.test(), batch_size=batch_size)
test_reader = paddle.batch(
paddle.dataset.mnist.test(), batch_size=batch_size)
opt.minimize(avg_cost)
......
python/paddle/fluid/tests/unittests/dist_mnist_batch_merge.py 0 → 100644
浏览文件 @ b94f9b16
# 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.
from __future__ import print_function
import numpy as np
import argparse
import time
import math
import paddle
import paddle.fluid as fluid
import paddle.fluid.profiler as profiler
from paddle.fluid import core
import unittest
from multiprocessing import Process
import os
import signal
from functools import reduce
from test_dist_base import TestDistRunnerBase, runtime_main
from dist_mnist import cnn_model
DTYPE = "float32"
def test_merge_reader(repeat_batch_size=8):
orig_reader = paddle.dataset.mnist.test()
record_batch = []
b = 0
for d in orig_reader():
if b >= repeat_batch_size:
break
record_batch.append(d)
b += 1
while True:
for d in record_batch:
yield d
class TestDistMnist2x2(TestDistRunnerBase):
def get_model(self, batch_size=2):
# Input data
images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE)
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
# Train program
predict = cnn_model(images)
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(x=cost)
# Evaluator
batch_size_tensor = fluid.layers.create_tensor(dtype='int64')
batch_acc = fluid.layers.accuracy(
input=predict, label=label, total=batch_size_tensor)
inference_program = fluid.default_main_program().clone()
# Optimization
opt = fluid.optimizer.Momentum(learning_rate=0.001, momentum=0.9)
# Reader
train_reader = paddle.batch(test_merge_reader, batch_size=batch_size)
test_reader = paddle.batch(
paddle.dataset.mnist.test(), batch_size=batch_size)
opt.minimize(avg_cost)
return inference_program, avg_cost, train_reader, test_reader, batch_acc, predict
if __name__ == "__main__":
runtime_main(TestDistMnist2x2)
python/paddle/fluid/tests/unittests/dist_mnist_lars.py 0 → 100644
浏览文件 @ b94f9b16
# 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.
from __future__ import print_function
import numpy as np
import argparse
import time
import math
import paddle
import paddle.fluid as fluid
import paddle.fluid.profiler as profiler
from paddle.fluid import core
import unittest
from multiprocessing import Process
import os
import signal
from functools import reduce
from test_dist_base import TestDistRunnerBase, runtime_main
from dist_mnist import cnn_model
DTYPE = "float32"
paddle.dataset.mnist.fetch()
# Fix seed for test
fluid.default_startup_program().random_seed = 1
fluid.default_main_program().random_seed = 1
class TestDistMnist2x2(TestDistRunnerBase):
def get_model(self, batch_size=2):
# Input data
images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE)
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
# Train program
predict = cnn_model(images)
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(x=cost)
# Evaluator
batch_size_tensor = fluid.layers.create_tensor(dtype='int64')
batch_acc = fluid.layers.accuracy(
input=predict, label=label, total=batch_size_tensor)
inference_program = fluid.default_main_program().clone()
# Optimization
opt = fluid.optimizer.LarsMomentumOptimizer(
learning_rate=0.001, momentum=0.9)
# Reader
train_reader = paddle.batch(
paddle.dataset.mnist.test(), batch_size=batch_size)
test_reader = paddle.batch(
paddle.dataset.mnist.test(), batch_size=batch_size)
opt.minimize(avg_cost)
return inference_program, avg_cost, train_reader, test_reader, batch_acc, predict
if __name__ == "__main__":
runtime_main(TestDistMnist2x2)
python/paddle/fluid/tests/unittests/test_hash_op.py 0 → 100644
浏览文件 @ b94f9b16
此差异已折叠。
python/paddle/fluid/tests/unittests/test_metrics.py 0 → 100644
浏览文件 @ b94f9b16
此差异已折叠。
python/paddle/utils/__init__.py
浏览文件 @ b94f9b16
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或