Skip to content

P
Paddle

PaddlePaddle / Paddle
10 个月 前同步成功

通知 2286
Star 20931
Fork 5422
P
Paddle

体验新版 GitCode，发现更多精彩内容 >>

提交 a1bf4c28 编写于 作者: D dangqingqing
浏览文件
操作

Update code and small fix.

上级 5e7e90ce 21053c16
展开全部 隐藏空白更改
内联 并排
Showing with 2602 addition and 811 deletion
.clang-format
浏览文件 @ a1bf4c28
......@@ -25,4 +25,3 @@ AllowAllParametersOfDeclarationOnNextLine: true
BinPackParameters: false
BinPackArguments: false
...
.travis.yml
浏览文件 @ a1bf4c28
......@@ -42,7 +42,7 @@ before_install:
script:
- |
timeout 2580 paddle/scripts/travis/${JOB}.sh # 43min timeout
RESULT=$?; if [ $RESULT -eq 0 ] || [ $RESULT -eq 142 ]; then true; else false; fi;
RESULT=$?; if [ $RESULT -eq 0 ] || [ $RESULT -eq 142 ]; then true ;else exit 1; fi;
- |
if [[ "$JOB" != "build_doc" ]]; then exit 0; fi;
if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi;
......
CMakeLists.txt
浏览文件 @ a1bf4c28
......@@ -133,6 +133,8 @@ include(external/any) # download libn::any
include(external/eigen) # download eigen3
include(external/pybind11) # download pybind11
include(external/nccl)
include(external/cares)
include(external/grpc)
include(cudnn) # set cudnn libraries, must before configure
include(configure) # add paddle env configuration
......
Dockerfile
浏览文件 @ a1bf4c28
......@@ -29,7 +29,7 @@ RUN apt-get update && \
automake locales clang-format swig doxygen cmake \
liblapack-dev liblapacke-dev libboost-dev \
clang-3.8 llvm-3.8 libclang-3.8-dev \
net-tools && \
net-tools libtool && \
apt-get clean -y
# Install Go and glide
......
cmake/external/cares.cmake 0 → 100644
浏览文件 @ a1bf4c28
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
#
# 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.
#
IF(MOBILE_INFERENCE)
return()
ENDIF()
include (ExternalProject)
# NOTE: c-ares is needed when linking with grpc.
SET(CARES_SOURCES_DIR ${THIRD_PARTY_PATH}/cares)
SET(CARES_INSTALL_DIR ${THIRD_PARTY_PATH}/install/cares)
SET(CARES_INCLUDE_DIR "${CARES_INSTALL_DIR}/include/" CACHE PATH "cares include directory." FORCE)
ExternalProject_Add(
extern_cares
GIT_REPOSITORY "https://github.com/c-ares/c-ares.git"
GIT_TAG "cares-1_13_0"
PREFIX ${CARES_SOURCES_DIR}
UPDATE_COMMAND ""
CONFIGURE_COMMAND ./buildconf && ./configure --disable-shared --prefix=${CARES_INSTALL_DIR}
BUILD_IN_SOURCE 1
BUILD_COMMAND make
INSTALL_COMMAND make install
)
ADD_LIBRARY(cares STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET cares PROPERTY IMPORTED_LOCATION
"${CARES_INSTALL_DIR}/lib/libcares.a")
include_directories(${CARES_INCLUDE_DIR})
ADD_DEPENDENCIES(cares extern_cares)
cmake/external/grpc.cmake 0 → 100644
浏览文件 @ a1bf4c28
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
#
# 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.
#
IF(MOBILE_INFERENCE)
return()
ENDIF()
include (ExternalProject)
SET(GRPC_SOURCES_DIR ${THIRD_PARTY_PATH}/grpc)
SET(GRPC_INSTALL_DIR ${THIRD_PARTY_PATH}/install/grpc)
SET(GRPC_INCLUDE_DIR "${GRPC_INSTALL_DIR}/include/" CACHE PATH "grpc include directory." FORCE)
SET(GRPC_CPP_PLUGIN "${GRPC_INSTALL_DIR}/bin/grpc_cpp_plugin" CACHE FILEPATH "GRPC_CPP_PLUGIN" FORCE)
IF(APPLE)
SET(BUILD_CMD make -n | sed "s/-Werror//g" | sh)
ELSE()
SET(BUILD_CMD make)
ENDIF()
ExternalProject_Add(
extern_grpc
DEPENDS protobuf zlib
GIT_REPOSITORY "https://github.com/grpc/grpc.git"
GIT_TAG "v1.7.x"
PREFIX ${GRPC_SOURCES_DIR}
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
BUILD_IN_SOURCE 1
# NOTE(yuyang18):
# Disable -Werror, otherwise the compile will fail in MacOS.
# It seems that we cannot configure that by make command.
# Just dry run make command and remove `-Werror`, then use a shell to run make commands
BUILD_COMMAND ${BUILD_CMD}
INSTALL_COMMAND make prefix=${GRPC_INSTALL_DIR} install
)
# FIXME(typhoonzero): hack to get static lib path, try a better way like merge them.
ADD_LIBRARY(grpc++_unsecure STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET grpc++_unsecure PROPERTY IMPORTED_LOCATION
"${GRPC_INSTALL_DIR}/lib/libgrpc++_unsecure.a")
ADD_LIBRARY(grpc++ STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET grpc++ PROPERTY IMPORTED_LOCATION
"${GRPC_INSTALL_DIR}/lib/libgrpc++.a")
ADD_LIBRARY(gpr STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET gpr PROPERTY IMPORTED_LOCATION
"${GRPC_INSTALL_DIR}/lib/libgpr.a")
ADD_LIBRARY(grpc_unsecure STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET grpc_unsecure PROPERTY IMPORTED_LOCATION
"${GRPC_INSTALL_DIR}/lib/libgrpc_unsecure.a")
include_directories(${GRPC_INCLUDE_DIR})
ADD_DEPENDENCIES(grpc++_unsecure extern_grpc)
cmake/external/protobuf.cmake
浏览文件 @ a1bf4c28
......@@ -15,7 +15,18 @@
INCLUDE(ExternalProject)
# Always invoke `FIND_PACKAGE(Protobuf)` for importing function protobuf_generate_cpp
FIND_PACKAGE(Protobuf QUIET)
SET(PROTOBUF_FOUND "OFF")
macro(UNSET_VAR VAR_NAME)
UNSET(${VAR_NAME} CACHE)
UNSET(${VAR_NAME})
endmacro()
UNSET_VAR(PROTOBUF_INCLUDE_DIR)
UNSET_VAR(PROTOBUF_FOUND)
UNSET_VAR(PROTOBUF_PROTOC_EXECUTABLE)
UNSET_VAR(PROTOBUF_PROTOC_LIBRARY)
UNSET_VAR(PROTOBUF_LITE_LIBRARY)
UNSET_VAR(PROTOBUF_LIBRARY)
UNSET_VAR(PROTOBUF_INCLUDE_DIR)
UNSET_VAR(Protobuf_PROTOC_EXECUTABLE)
if(NOT COMMAND protobuf_generate_python) # before cmake 3.4, protobuf_genrerate_python is not defined.
function(protobuf_generate_python SRCS)
......@@ -110,7 +121,6 @@ macro(PROMPT_PROTOBUF_LIB)
# FIND_Protobuf.cmake uses `Protobuf_PROTOC_EXECUTABLE`.
# make `protobuf_generate_cpp` happy.
SET(Protobuf_PROTOC_EXECUTABLE ${PROTOBUF_PROTOC_EXECUTABLE})
FOREACH(dep ${protobuf_DEPS})
ADD_DEPENDENCIES(protobuf ${dep})
ADD_DEPENDENCIES(protobuf_lite ${dep})
......@@ -128,11 +138,11 @@ endmacro()
set(PROTOBUF_ROOT "" CACHE PATH "Folder contains protobuf")
if (NOT "${PROTOBUF_ROOT}" STREQUAL "")
find_path(PROTOBUF_INCLUDE_DIR google/protobuf/message.h PATHS ${PROTOBUF_ROOT}/include)
find_library(PROTOBUF_LIBRARY protobuf PATHS ${PROTOBUF_ROOT}/lib)
find_library(PROTOBUF_LITE_LIBRARY protobuf-lite PATHS ${PROTOBUF_ROOT}/lib)
find_library(PROTOBUF_PROTOC_LIBRARY protoc PATHS ${PROTOBUF_ROOT}/lib)
find_program(PROTOBUF_PROTOC_EXECUTABLE protoc PATHS ${PROTOBUF_ROOT}/bin)
find_path(PROTOBUF_INCLUDE_DIR google/protobuf/message.h PATHS ${PROTOBUF_ROOT}/include NO_DEFAULT_PATH)
find_library(PROTOBUF_LIBRARY protobuf PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
find_library(PROTOBUF_LITE_LIBRARY protobuf-lite PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
find_library(PROTOBUF_PROTOC_LIBRARY protoc PATHS ${PROTOBUF_ROOT}/lib NO_DEFAULT_PATH)
find_program(PROTOBUF_PROTOC_EXECUTABLE protoc PATHS ${PROTOBUF_ROOT}/bin NO_DEFAULT_PATH)
if (PROTOBUF_INCLUDE_DIR AND PROTOBUF_LIBRARY AND PROTOBUF_LITE_LIBRARY AND PROTOBUF_PROTOC_LIBRARY AND PROTOBUF_PROTOC_EXECUTABLE)
message(STATUS "Using custom protobuf library in ${PROTOBUF_ROOT}.")
SET_PROTOBUF_VERSION()
......
cmake/external/zlib.cmake
浏览文件 @ a1bf4c28
......@@ -50,6 +50,8 @@ ExternalProject_Add(
)
LIST(APPEND external_project_dependencies zlib)
ADD_LIBRARY(zlib_target STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET zlib_target PROPERTY IMPORTED_LOCATION ${ZLIB_LIBRARIES})
IF(WITH_C_API)
INSTALL(DIRECTORY ${ZLIB_INCLUDE_DIR} DESTINATION third_party/zlib)
......
cmake/generic.cmake
浏览文件 @ a1bf4c28
......@@ -459,11 +459,58 @@ function(py_test TARGET_NAME)
if(WITH_TESTING)
set(options STATIC static SHARED shared)
set(oneValueArgs "")
set(multiValueArgs SRCS DEPS)
cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
set(multiValueArgs SRCS DEPS ARGS)
cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
add_test(NAME ${TARGET_NAME}
COMMAND env PYTHONPATH=${PADDLE_PYTHON_BUILD_DIR}/lib-python
${PYTHON_EXECUTABLE} ${py_test_SRCS}
${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}
WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
endif()
endfunction()
# grpc_library generate grpc code using grpc_cpp_plugin and protoc
# then build the generated protobuf code and grpc code with your
# implementation source codes together. Use SRCS argument for your
# implementation source files and PROTO argument for your .proto
# files.
#
# Usage: grpc_library(my_target SRCS my_client.cc PROTO my_target.proto DEPS my_dep)
function(grpc_library TARGET_NAME)
set(oneValueArgs PROTO)
set(multiValueArgs SRCS DEPS)
set(options "")
cmake_parse_arguments(grpc_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
message(STATUS "generating grpc ${grpc_library_PROTO}")
get_filename_component(ABS_PROTO ${grpc_library_PROTO} ABSOLUTE)
get_filename_component(PROTO_WE ${grpc_library_PROTO} NAME_WE)
get_filename_component(PROTO_PATH ${ABS_PROTO} PATH)
protobuf_generate_cpp(grpc_proto_srcs grpc_proto_hdrs "${ABS_PROTO}")
set(grpc_grpc_srcs "${CMAKE_CURRENT_BINARY_DIR}/${PROTO_WE}.grpc.pb.cc")
set(grpc_grpc_hdrs "${CMAKE_CURRENT_BINARY_DIR}/${PROTO_WE}.grpc.pb.h")
cc_library("${TARGET_NAME}_proto" SRCS "${grpc_proto_srcs}")
add_custom_command(
OUTPUT "${grpc_grpc_srcs}" "${grpc_grpc_hdrs}"
COMMAND ${PROTOBUF_PROTOC_EXECUTABLE}
ARGS --grpc_out "${CMAKE_CURRENT_BINARY_DIR}" -I "${PROTO_PATH}"
--plugin=protoc-gen-grpc="${GRPC_CPP_PLUGIN}" "${ABS_PROTO}"
DEPENDS "${ABS_PROTO}" ${PROTOBUF_PROTOC_EXECUTABLE} extern_grpc)
# FIXME(typhoonzero): grpc generated code do not generate virtual-dtor, mark it
# as compiler warnings instead of error. Should try remove the warnings also.
set_source_files_properties(
${grpc_grpc_srcs}
PROPERTIES
COMPILE_FLAGS "-Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
cc_library("${TARGET_NAME}_grpc" SRCS "${grpc_grpc_srcs}")
set_source_files_properties(
${grpc_library_SRCS}
PROPERTIES
COMPILE_FLAGS "-Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
cc_library("${TARGET_NAME}" SRCS "${grpc_library_SRCS}" DEPS "${TARGET_NAME}_grpc" "${TARGET_NAME}_proto" "${grpc_library_DEPS}")
endfunction()
doc/api/v2/config/layer.rst
浏览文件 @ a1bf4c28
......@@ -54,7 +54,7 @@ img_conv
.. _api_v2.layer_context_projection:
context_projection
context_projection
------------------
.. autoclass:: paddle.v2.layer.context_projection
:noindex:
......@@ -70,7 +70,7 @@ Image Pooling Layer
img_pool
--------
.. autoclass:: paddle.v2.layer.img_pool
:noindex:
:noindex:
spp
---
......@@ -104,7 +104,7 @@ sum_to_one_norm
---------------
.. autoclass:: paddle.v2.layer.sum_to_one_norm
:noindex:
cross_channel_norm
------------------
.. autoclass:: paddle.v2.layer.cross_channel_norm
......@@ -114,7 +114,7 @@ row_l2_norm
-----------
.. autoclass:: paddle.v2.layer.row_l2_norm
:noindex:
Recurrent Layers
================
......@@ -415,6 +415,13 @@ multiplex
.. autoclass:: paddle.v2.layer.multiplex
:noindex:
Factorization Machine Layer
============================
factorization_machine
---------------------
.. autoclass:: paddle.v2.layer.factorization_machine
:noindex:
Slicing and Joining Layers
==========================
......
doc/design/refactor/distributed_architecture.md
浏览文件 @ a1bf4c28
......@@ -2,106 +2,70 @@
## Abstract
PaddlePaddle v0.10.0 uses the "trainer-parameter server"
architecture. We run multiple replicated instances of trainers (runs
the same code written by the user) and parameter servers for
distributed training. This architecture served us well, but has some
limitations:
PaddlePaddle version 0.10.0 uses the "trainer-parameter server" architecture. We run multiple instances of trainers (where each trainer runs the same model) and parameter servers for distributed training. This architecture serves well, but has few limitations:
1. Need to write special code to handle tasks which should only be run
by a single trainer. E.g., initializing model and saving model.
1. There is a need to write special code that handles tasks which should only be run on a single trainer. E.g., initializing the model, saving the model etc.
2. Model parallelism is hard: need to write if-else branches conditioned
on the trainer ID to partition model onto each trainer, and manually
write the inter-model-shard communication code.
2. Model parallelism is hard: It would need all the if-else branches conditioned on the trainer ID to partition the model onto the trainers, and eventually manually writing out the inter-model-shard communication code to communicate between different trainers.
3. The user can not directly specify the parameter update rule: need
to modify the parameter server C++ code and compile a new
binary. This adds complication for researchers: A lot of extra
effort is required. Besides, the training job submission program
may not allow running arbitrary binaries.
3. The user can not directly specify the parameter update rule: This would need to modify the parameter server code and compile a new binary. This makes things more complicated for researchers: A lot of extra effort is required to make this work. Besides, the training job submission program may not allow running arbitrary binaries.
This design doc discusses PaddlePaddle's new distributed training
architecture that addresses the above limitations.
This design doc discusses PaddlePaddle's new distributed training architecture that addresses the above mentioned limitations.
## Analysis
We will assume the user writes the trainer program by Python, the same
analysis holds if the trainer program is written in C++.
The assumption is that the user writes the trainer program in either Python or C++.
### Limitation 1
If we look at the Python code that the user writes, there are two
kinds of functionalities:
There are two basic functionalities in the trainer program:
- The training logic such as load / save model and print log.
- The neural network definition such as the definition of the data
layer, the fully connected layer, the cost function and the
1. The training logic such as loading / saving the model and printing out the logs.
2. The neural network definition such as the definition of the data layer, the fully connected layer, the cost function and the
optimizer.
When we training with PaddlePaddle v0.10.0 distributedly, multiple
replicated Python instances are running on different nodes: both the
training logic and the neural network computation is replicated.
When we train using PaddlePaddle v0.10.0 in a distributed fashion, multiple instances of the same Python code are run on different nodes, hence both: the
training logic as well as the neural network computation logic, is replicated.
The tasks that should only run once all belong to the training logic,
if we only replicate the neural network computation, but do **not**
replicate the training logic, the limitation could be solved.
The tasks that only need to be run once belong to the training logic. Hence if we only replicate the neural network computation part, and do **not**
replicate the training logic, the limitation mentioned above can be avoided.
### Limitation 2
Model parallelism means running a single model on multiple nodes by
partitioning the model onto different nodes and managing the
inter-model-shard communications.
Model parallelism means that a single model is partitioned into different components and each node runs one of the component separately. This comes at the extra cost of managing the
inter-model-shard communication between nodes.
PaddlePaddle should be able to modify the nerual network computation
definition to support model parallelism automatically. However, the
computation is only specified in Python code, and PaddlePaddle can not
modify Python code.
PaddlePaddle should ideally be able to modify the neural network computation and figure out the support for model parallelism automatically. However, the
computation is only specified in Python code which sits outside of PaddlePaddle, hence PaddlePaddle can not support the feature in this setup.
Just like compiler uses a intermediate representation (IR) so that
programmer does not need to manually optimize their code in most of
the cases - the compiler will optimize the IR:
Similar to how a compiler uses an intermediate representation (IR) so that the programmer does not need to manually optimize their code for most of the cases, we can have an intermediate representation in PaddlePaddle as well. The compiler optimizes the IR as follows:
<img src="src/compiler.png"/>
We can have our own IR too: PaddlePaddle can support model parallel by
converting the IR so the user no longer need to manually do it in
Python:
PaddlePaddle can support model parallelism by converting the IR so that the user no longer needs to manually perform the computation and operations in the Python component:
<img src="src/paddle-compile.png"/>
The IR for PaddlePaddle after refactor is called `Block`, it specifies
the computation dependency graph and the variables used in the
computation.
The IR for PaddlePaddle after refactoring is called a `Block`, it specifies the computation dependency graph and the variables used in the computation.
### Limitation 3
The user can not directly specify the parameter update rule for the
parameter server because the parameter server does not use the same
computation definition as the trainer. Instead, the update rule is
baked in the parameter server. The user can not specify the update
rule in the same way of specifying the trainer computation.
The user can not directly specify the parameter update rule for the parameter server in the Python module, since the parameter server does not use the same computation definition as the trainer. Instead, the update rule is baked inside the parameter server. The user can not specify the update rule explicitly.
This could be fixed by making the parameter server run the same
computation definition as the trainer. For a detailed explanation,
please
see
This could be fixed by making the parameter server run the same computation definition as the trainer (the user's Python module). For a detailed explanation, refer to this document -
[Design Doc: Operation Graph Based Parameter Server](./dist_train.md)
## Distributed Training Architecture
The new distributed training architecture can address the above
limitations. Below is the illustration:
The revamped distributed training architecture can address the above discussed limitations. Below is the illustration of how it does so:
<img src="src/distributed_architecture.png"/>
The architecture includes major components: *PaddlePaddle Python*,
*PaddlePaddle converter* and *PaddlePaddle runtime*:
The major components in the architecture are: *PaddlePaddle Python*, *PaddlePaddle converter* and *PaddlePaddle runtime*.
### PaddlePaddle Python
PaddlePaddle Python is the Python library that user's Python trainer
invoke to build the neural network topology, start training, etc.
PaddlePaddle Python is the Python library that user's Python code invokes, to read the data. build the neural network topology, start training, etc.
```Python
paddle.init()
......@@ -117,102 +81,60 @@ for i in range(1000):
print cost_val
```
The code above is a typical Python trainer code, the neural network
topology is built using helper functions such as
`paddle.layer.fc`. The training is done by calling `session.eval`
iteratively.
The above code is what a typical Python trainer code is, the neural network topology is built using the helper functions such as `paddle.layer.fc`. Training is done by calling `session.eval` iteratively.
#### session.eval
As shown in the graph, `session.eval` sends the IR and the evaluation
inputs/targets to the PaddlePaddle cluster for evaluation. The
targets can be any variable in the computation graph. When the target
is the `optimizer` variable, the neural network will be optimized
once. When the target is the `cost` variable, `session.eval` returns
the cost value.
As shown in the graph, `session.eval` sends the IR and the evaluation inputs or targets to the PaddlePaddle cluster for evaluation.
The targets can be any variable in the computation graph. When the target is say, the `optimizer` variable, the neural network will be optimized once. When the target is the `cost` variable, `session.eval` returns the cost value. Based on what the target is, an appropriate action is taken.
The Python `session` is a wrapper of the C++ `Session` class. For more
information about `Session`, please
see [Design Doc: Session](./session.md).
The Python `session` is a wrapper of the C++ `Session` class. For more information about `Session`, refer to this document - [Design Doc: Session](./session.md).
### PaddlePaddle Converter
PaddlePaddle converter automatically converts the IR in the request
(IR and evaluation inputs/targets) from PaddlePaddle Python to new
partitioned IRs and dispatch the new IRs and evaluation inputs/targets
to different PaddlePaddle runtimes. Below are the steps:
The PaddlePaddle converter automatically converts the IR in the request (IR and evaluation inputs/targets) from PaddlePaddle Python to partitioned IRs and dispatches the new IRs and evaluation inputs/targets to different PaddlePaddle runtimes. Below are the steps that are followed :
1. Add `feed` OP that feeds the eval inputs, and `fetch` OP that
fetches the eval targets to the IR.
1. Add a `feed` OP that feeds the eval inputs, and a `fetch` OP that fetches the eval targets to the IR.
1. Extract a new computation (sub)graph with `feed` and `fetch` OP as
the boundary. The runtime does not need to run the OP that is not
dependent by the `fetch` OP.
2. Extract a new computation (sub)graph with the `feed` and `fetch` OPs as the boundary. The runtime does not need to run the OP that is not dependent on the `fetch` OP.
1. Optimizes the computation graph.
3. Optimize the computation graph.
1. Place the OPs in the graph onto different devices on different
PaddlePaddle runtime according to a placement algorithm and device
constraint specified by the user.
4. Place the OPs in the graph onto different devices on different PaddlePaddle runtime according to a placement algorithm and the device constraints specified by the user.
1. Partition the graph according to runtime boundaries and add `send` /
`recv` OP pair on the runtime boundaries.
5. Partition the graph according to runtime boundaries and add `send` / `recv` OP pair on the runtime boundaries.
1. Dispatch the partitioned graph to different PaddlePaddle runtimes.
6. Dispatch the partitioned graph to different PaddlePaddle runtimes.
7. PaddlePaddle runtimes with the `fetch` OP reports evaluation results back to the converter, the converter reports the evaluation results back to the PaddlePaddle Python.
1. PaddlePaddle runtimes with the `fetch` OP reports evaluation
results back to the converter, the convert reports the evaluation
results back to the PaddlePaddle Python.
The output IRs will be cached to optimize the conversion latency.
#### Placement Algorithm
Our first implementation will only support "trainer-parameter server"
placement: the parameters, initializers, and optimizers are placed on
the PaddlePaddle runtimes with the parameter server role. And
everything else will be placed on the PaddlePaddle runtimes with the
trainer role. This has the same functionality of our
"trainer-parameter server" architecture of PaddlePaddle v0.10.0, but
is more general and flexible.
Our first implementation will only support "trainer-parameter server" placement: the parameters, initializers, and optimizers are all placed on the PaddlePaddle runtimes with the parameter server role. Everything else will be placed on the PaddlePaddle runtimes with the trainer role. This has the same functionality as the "trainer-parameter server" architecture of PaddlePaddle v0.10.0, but is more generic and flexible.
In the future, we will implement the general placement algorithm,
which makes placements according to the input IR, and a model of
device computation time and device communication time. Model
parallelism requires the general placement algorithm.
In the future, a more general placement algorithm should be implemented, which makes placements according to the input IR, and a model of device computation time and device communication time. Model parallelism requires the generic placement algorithm.
### PaddlePaddle Runtime
The PaddlePaddle runtime owns multiple devices (e.g., CPUs, GPUs) and
runs the IR. The runtime does not need to do OP placement since it's
already done by the converter.
The PaddlePaddle runtime owns multiple devices (e.g., CPUs, GPUs) and runs the IR. The runtime does not need to do OP placement since it is already done by the converter.
### Local Training Architecture
The local training architecture will be the same as the distributed
training architecture, the differences are everything runs locally,
and there is just one PaddlePaddle runtime:
The local training architecture will be the same as the distributed training architecture, the difference is that everything runs locally, and there is just one PaddlePaddle runtime:
<img src="src/local_architecture.png"/>
### Training Data
In PaddlePaddle v0.10.0, training data is typically read
with [data reader](../reader/README.md) from Python. This approach is
no longer efficient when training distributedly since the Python
process no longer runs on the same node with the trainer processes,
the Python reader will need to read from the distributed filesystem
(assuming it has the access) and send to the trainers, doubling the
network traffic.
When doing distributed training, the user can still use Python data
reader: the training data are sent with `session.eval`. However should
be used for debugging purpose only. The users are encouraged to use
the read data OPs.
In PaddlePaddle v0.10.0, training data is typically read with a [data reader](../reader/README.md) from Python. This approach is no longer efficient when training in a distributed fashion since the Python process no longer runs on the same node with the trainer processes. The Python reader will need to read from the distributed filesystem (assuming it has the required access) and send to the trainers, doubling the network traffic.
When doing distributed training, the user can still use Python data reader: the training data are sent with `session.eval`. However this should be used for debugging purpose only. The users are encouraged to use the read data OPs.
## References:
......
paddle/capi/Matrix.cpp
浏览文件 @ a1bf4c28
......@@ -55,7 +55,7 @@ paddle_error paddle_matrix_set_row(paddle_matrix mat,
}
PD_API paddle_error paddle_matrix_set_value(paddle_matrix mat,
paddle_real* value) {
paddle_real* value) {
if (mat == nullptr || value == nullptr) return kPD_NULLPTR;
auto ptr = cast(mat);
if (ptr->mat == nullptr) return kPD_NULLPTR;
......@@ -75,7 +75,7 @@ PD_API paddle_error paddle_matrix_set_value(paddle_matrix mat,
}
PD_API paddle_error paddle_matrix_get_value(paddle_matrix mat,
paddle_real* result) {
paddle_real* result) {
if (mat == nullptr || result == nullptr) return kPD_NULLPTR;
auto ptr = cast(mat);
if (ptr->mat == nullptr) return kPD_NULLPTR;
......
paddle/capi/matrix.h
浏览文件 @ a1bf4c28
......@@ -79,7 +79,7 @@ PD_API paddle_error paddle_matrix_set_row(paddle_matrix mat,
* @note value should contain enough element of data to init the mat
*/
PD_API paddle_error paddle_matrix_set_value(paddle_matrix mat,
paddle_real* value);
paddle_real* value);
/**
* @brief PDMatGetRow Get raw row buffer from matrix
......@@ -93,14 +93,14 @@ PD_API paddle_error paddle_matrix_get_row(paddle_matrix mat,
paddle_real** rawRowBuffer);
/**
* @brief copy data from the matrix
* @brief copy data from the matrix
* @param [in] mat Target matrix
* @param [out] result pointer to store the matrix data
* @param [out] result pointer to store the matrix data
* @return paddle_error
* @note the space of the result should allocated before invoke this API
*/
PD_API paddle_error paddle_matrix_get_value(paddle_matrix mat,
paddle_real* result);
paddle_real* result);
/**
* @brief PDMatCreateNone Create None Matrix
* @return
......
paddle/framework/lod_tensor.cc
浏览文件 @ a1bf4c28
......@@ -13,6 +13,8 @@
limitations under the License. */
#include "paddle/framework/lod_tensor.h"
#include "paddle/framework/data_type.h"
#include "paddle/framework/framework.pb.h"
#include "paddle/memory/memcpy.h"
#include "paddle/memory/memory.h"
......@@ -27,11 +29,11 @@
namespace paddle {
namespace framework {
std::ostream& operator<<(std::ostream& os, const LoD& lod) {
std::ostream &operator<<(std::ostream &os, const LoD &lod) {
os << "{";
for (auto& v : lod) {
for (auto &v : lod) {
os << "{";
for (auto& i : v) {
for (auto &i : v) {
os << i << ",";
}
os << "}";
......@@ -41,7 +43,7 @@ std::ostream& operator<<(std::ostream& os, const LoD& lod) {
return os;
}
LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end) {
LoD SliceLevels(const LoD &in, size_t level_begin, size_t level_end) {
LoD new_lod;
new_lod.reserve(level_end - level_begin);
for (size_t i = level_begin; i < level_end; i++) {
......@@ -53,7 +55,7 @@ LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end) {
return new_lod;
}
LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin,
LoD SliceInLevel(const LoD &in, size_t level, size_t elem_begin,
size_t elem_end) {
PADDLE_ENFORCE_LT(level, in.size());
PADDLE_ENFORCE_LT(elem_end, in[level].size());
......@@ -64,9 +66,9 @@ LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin,
res[0].assign(in[level].begin() + elem_begin,
in[level].begin() + elem_end + 1);
for (size_t lvl = 1; lvl < res.size(); lvl++) {
const auto& in_level = in[level + lvl];
const auto& above_level = res[lvl - 1];
auto& out_level = res[lvl];
const auto &in_level = in[level + lvl];
const auto &above_level = res[lvl - 1];
auto &out_level = res[lvl];
out_level.assign(in_level.begin() + above_level.front(),
in_level.begin() + above_level.back() + 1);
}
......@@ -74,33 +76,33 @@ LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin,
// to make the first offset equals 0, all the elements minus the first
// element
size_t front = res[lvl].front();
for (auto& ele : res[lvl]) {
for (auto &ele : res[lvl]) {
ele -= front;
}
}
return res;
}
LoD ToAbsOffset(const LoD& in) {
LoD ToAbsOffset(const LoD &in) {
// the lowest level stores relative offsets
if (in.empty() || in.size() == 1) return in;
LoD result = in;
for (int level = result.size() - 2; level >= 0; level--) {
for (auto& ele : result[level]) {
for (auto &ele : result[level]) {
ele = result[level + 1][ele];
}
}
return result;
}
bool operator==(const LoD& a, const LoD& b) {
bool operator==(const LoD &a, const LoD &b) {
if (a.size() != b.size()) {
return false;
}
for (size_t i = 0; i < a.size(); i++) {
const auto& a_level = a[i];
const auto& b_level = b[i];
const auto &a_level = a[i];
const auto &b_level = b[i];
if (a_level.size() != b_level.size()) {
return false;
}
......@@ -151,7 +153,7 @@ void LoDTensor::ShrinkInLevel(size_t level, size_t elem_begin,
}
using LoDAndOffset = std::pair<LoD, std::pair<size_t, size_t>>;
LoDAndOffset GetSubLoDAndAbsoluteOffset(const LoD& lod, size_t start_idx,
LoDAndOffset GetSubLoDAndAbsoluteOffset(const LoD &lod, size_t start_idx,
size_t end_idx, size_t start_level) {
LoD sub_lod;
......@@ -170,7 +172,7 @@ LoDAndOffset GetSubLoDAndAbsoluteOffset(const LoD& lod, size_t start_idx,
return LoDAndOffset{sub_lod, {start_idx, end_idx}};
}
void AppendLoD(LoD* lod, const LoD& lod_length) {
void AppendLoD(LoD *lod, const LoD &lod_length) {
PADDLE_ENFORCE(
lod->empty() || lod->size() == lod_length.size(),
"The lod_length should has the same size with the appended lod.");
......@@ -178,12 +180,139 @@ void AppendLoD(LoD* lod, const LoD& lod_length) {
*lod = LoD(lod_length.size(), std::vector<size_t>({0}));
}
for (size_t i = 0; i < lod->size(); ++i) {
auto& level = (*lod)[i];
auto &level = (*lod)[i];
for (size_t len : lod_length[i]) {
level.push_back(level.back() + len);
}
}
}
void SerializeToStream(std::ostream &os, const LoDTensor &tensor,
const platform::DeviceContext &dev_ctx) {
// TODO(typhoonzero): serialize to ostream
{ // the 1st field, uint32_t version
constexpr uint32_t version = 0;
os.write(reinterpret_cast<const char *>(&version), sizeof(version));
}
{ // the 2nd field, tensor description
// int32_t size
// void* protobuf message
framework::TensorDesc desc;
desc.set_data_type(framework::ToDataType(tensor.type()));
auto dims = framework::vectorize(tensor.dims());
auto *pb_dims = desc.mutable_dims();
pb_dims->Resize(static_cast<int>(dims.size()), 0);
std::copy(dims.begin(), dims.end(), pb_dims->begin());
int32_t size = desc.ByteSize();
os.write(reinterpret_cast<const char *>(&size), sizeof(size));
auto out = desc.SerializeAsString();
os.write(out.data(), size);
}
{ // the 3rd field, tensor data
uint64_t size = tensor.memory_size();
auto *data_ptr = tensor.data<void>();
PADDLE_ENFORCE(size < std::numeric_limits<std::streamsize>::max(),
"Index overflow when writing tensor");
if (platform::is_gpu_place(tensor.place())) {
#ifdef PADDLE_WITH_CUDA
constexpr size_t kBufSize = 1024 * 1024 * 64; // 64MB
std::unique_ptr<char[]> buf(new char[kBufSize]);
auto &gpu_dev_ctx =
static_cast<const platform::CUDADeviceContext &>(dev_ctx);
platform::CPUPlace cpu;
uintptr_t data = reinterpret_cast<uintptr_t>(data_ptr);
while (size != 0) {
size_t size_to_write = std::min(kBufSize, static_cast<size_t>(size));
memory::Copy(cpu, buf.get(),
boost::get<platform::GPUPlace>(tensor.place()),
reinterpret_cast<const void *>(data), size_to_write,
gpu_dev_ctx.stream());
gpu_dev_ctx.Wait();
os.write(buf.get(), size_to_write);
data += size_to_write;
size -= size_to_write;
}
#else
PADDLE_THROW("Unexpected branch");
#endif
} else {
os.write(static_cast<const char *>(data_ptr),
static_cast<std::streamsize>(size));
}
}
{ // the 4th field, lod information
// uint64_t lod_level
// uint64_t lod_level_1 size in byte.
// int* lod_level_1 data
// ...
auto lod = tensor.lod();
uint64_t size = lod.size();
os.write(reinterpret_cast<const char *>(&size), sizeof(size));
for (auto &each : lod) {
size = each.size() * sizeof(framework::LoD::value_type::value_type);
os.write(reinterpret_cast<const char *>(&size), sizeof(size));
os.write(reinterpret_cast<const char *>(each.data()),
static_cast<std::streamsize>(size));
}
}
}
void DeserializeFromStream(std::istream &is, LoDTensor *tensor) {
uint32_t version;
is.read(reinterpret_cast<char *>(&version), sizeof(version));
PADDLE_ENFORCE_EQ(version, 0U, "Only version 0 is supported");
framework::TensorDesc desc;
{ // int32_t size
// proto buffer
int32_t size;
is.read(reinterpret_cast<char *>(&size), sizeof(size));
std::unique_ptr<char[]> buf(new char[size]);
is.read(reinterpret_cast<char *>(buf.get()), size);
PADDLE_ENFORCE(desc.ParseFromArray(buf.get(), size),
"Cannot parse tensor desc");
}
{ // read tensor
std::vector<int64_t> dims;
dims.reserve(static_cast<size_t>(desc.dims().size()));
std::copy(desc.dims().begin(), desc.dims().end(), std::back_inserter(dims));
tensor->Resize(framework::make_ddim(dims));
void *buf;
platform::Place cpu = platform::CPUPlace();
switch (desc.data_type()) {
case framework::FP32:
buf = tensor->mutable_data<float>(cpu);
break;
case framework::FP64:
buf = tensor->mutable_data<double>(cpu);
break;
case framework::INT32:
buf = tensor->mutable_data<int>(cpu);
break;
case framework::INT64:
buf = tensor->mutable_data<int64_t>(cpu);
break;
default:
PADDLE_THROW("DataType %d not supported", desc.data_type());
}
is.read(static_cast<char *>(buf), tensor->memory_size());
}
{ // read lod
uint64_t lod_level;
is.read(reinterpret_cast<char *>(&lod_level), sizeof(lod_level));
auto &lod = *tensor->mutable_lod();
lod.resize(lod_level);
for (uint64_t i = 0; i < lod_level; ++i) {
uint64_t size;
is.read(reinterpret_cast<char *>(&size), sizeof(size));
std::vector<size_t> tmp(size / sizeof(size_t));
is.read(reinterpret_cast<char *>(tmp.data()),
static_cast<std::streamsize>(size));
lod[i] = tmp;
}
}
}
} // namespace framework
} // namespace paddle
paddle/framework/lod_tensor.h
浏览文件 @ a1bf4c28
......@@ -189,5 +189,14 @@ std::pair<LoD, std::pair<size_t, size_t>> GetSubLoDAndAbsoluteOffset(
void AppendLoD(LoD* lod, const LoD& lod_length);
/*
* Serialize/Desiralize LoDTensor to std::ostream
* You can pass ofstream or ostringstream to serilize to file
* or to a in memory string. GPU tensor will be copied to CPU.
*/
void SerializeToStream(std::ostream& os, const LoDTensor& tensor,
const platform::DeviceContext& dev_ctx);
void DeserializeFromStream(std::istream& is, LoDTensor* tensor);
} // namespace framework
} // namespace paddle
paddle/framework/tensor_util.h
浏览文件 @ a1bf4c28
......@@ -135,18 +135,17 @@ inline void CopyToVector(const Tensor& src, const platform::DeviceContext& ctx,
auto dst_ptr = static_cast<void*>(dst->data());
if (platform::is_cpu_place(src.place())) {
memory::Copy(dst_place, dst_ptr, boost::get<platform::CPUPlace>(src.place()),
src_ptr, size);
memory::Copy(dst_place, dst_ptr,
boost::get<platform::CPUPlace>(src.place()), src_ptr, size);
}
#ifdef PADDLE_WITH_CUDA
else if (platform::is_gpu_place(src.place())) { // NOLINT
memory::Copy(
dst_place, dst_ptr, boost::get<platform::GPUPlace>(src.place()), src_ptr,
size,
dst_place, dst_ptr, boost::get<platform::GPUPlace>(src.place()),
src_ptr, size,
reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
}
#endif
}
} // namespace framework
......
paddle/gserver/layers/FactorizationMachineLayer.cpp 0 → 100644
浏览文件 @ a1bf4c28
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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 "FactorizationMachineLayer.h"
#include <algorithm>
#include <vector>
#include "paddle/math/SparseMatrix.h"
#include "paddle/utils/Logging.h"
#include "paddle/utils/Stat.h"
namespace paddle {
REGISTER_LAYER(factorization_machine, FactorizationMachineLayer);
bool FactorizationMachineLayer::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
/* Initialize the basic parent class */
Layer::init(layerMap, parameterMap);
factorSize_ = config_.factor_size();
/* initialize the latentVectors_ */
CHECK_EQ(inputLayers_.size(), 1UL);
size_t inputSize = inputLayers_[0]->getSize();
CHECK_EQ(parameters_[0]->getSize(), inputSize * factorSize_);
latentVectors_ = std::unique_ptr<Weight>(
new Weight(inputSize, factorSize_, parameters_[0]));
return true;
}
void FactorizationMachineLayer::forward(PassType passType) {
Layer::forward(passType);
const MatrixPtr& inputV = getInputValue(0);
size_t batchSize = inputV->getHeight();
size_t outputSize = getSize();
size_t inputSize = inputLayers_[0]->getSize();
reserveOutput(batchSize, outputSize);
MatrixPtr outV = getOutputValue();
Matrix::resizeOrCreate(
latentVectorsSquare_, inputSize, factorSize_, false, useGpu_);
Matrix::resizeOrCreate(
inputMulFactor_, batchSize, factorSize_, false, useGpu_);
Matrix::resizeOrCreate(tmpOut_, batchSize, factorSize_, false, useGpu_);
REGISTER_TIMER_INFO("FmInputMulFactorTimer", getName().c_str());
inputMulFactor_->mul(*inputV, *latentVectors_->getW());
inputMulFactor_->square2(*tmpOut_);
outV->sumRows(*tmpOut_, 0.5, 0);
if (dynamic_cast<CpuSparseMatrix*>(inputV.get())) {
Matrix::resizeOrCreateSparseMatrix(inputSquare_,
inputV->getHeight(),
inputV->getWidth(),
inputV->getElementCnt(),
inputV->getValueType());
inputSquare_->copyFrom(*inputV);
(dynamic_cast<CpuSparseMatrix*>(inputSquare_.get()))->square2();
} else {
Matrix::resizeOrCreate(
inputSquare_, inputV->getHeight(), inputV->getWidth(), false, useGpu_);
inputV->square2(*inputSquare_);
}
latentVectors_->getW()->square2(*latentVectorsSquare_);
tmpOut_->mul(*inputSquare_, *latentVectorsSquare_);
outV->sumRows(*tmpOut_, -0.5, 1.0);
/* activation */ {
REGISTER_TIMER_INFO("FmFwAtvTimer", getName().c_str());
forwardActivation();
}
}
void FactorizationMachineLayer::backward(const UpdateCallback& callback) {
/* Do derivation */ { backwardActivation(); }
const MatrixPtr& inputV = getInputValue(0);
const MatrixPtr& oGrad = getOutputGrad();
Matrix::resizeOrCreate(
tmpSum_, 1, latentVectors_->getW()->getHeight(), false, useGpu_);
MatrixPtr tmpSumTrans = Matrix::create(tmpSum_->getRowBuf(0),
latentVectors_->getW()->getHeight(),
1,
false,
useGpu_);
/* Calculate the gradients of the latentVectors_ matrix */
if (latentVectors_->getWGrad()) {
if (dynamic_cast<CpuSparseMatrix*>(inputV.get())) {
Matrix::resizeOrCreateSparseMatrix(tmpInput_,
inputV->getHeight(),
inputV->getWidth(),
inputV->getElementCnt());
CpuSparseMatrix* sparseInputV =
dynamic_cast<CpuSparseMatrix*>(inputV.get());
CpuSparseMatrix* sparseInputSquare =
dynamic_cast<CpuSparseMatrix*>(inputSquare_.get());
CpuSparseMatrix* sparseTmpInput =
dynamic_cast<CpuSparseMatrix*>(tmpInput_.get());
sparseTmpInput->copyFrom(*sparseInputV);
sparseTmpInput->rowScale(0, *sparseInputV, *oGrad);
latentVectors_->getWGrad()->mul(
*sparseTmpInput->getTranspose(), *inputMulFactor_, 1, 1);
sparseTmpInput->rowScale(0, *sparseInputSquare, *oGrad);
Matrix::resizeOrCreate(negOnes_, 1, inputV->getHeight(), false, useGpu_);
negOnes_->zeroMem();
negOnes_->add(-1);
tmpSum_->mul(*negOnes_, *sparseTmpInput, 1, 0);
} else {
Matrix::resizeOrCreate(
tmpInput_, inputV->getHeight(), inputV->getWidth(), false, useGpu_);
tmpInput_->rowScale(0, *inputV, *oGrad);
latentVectors_->getWGrad()->mul(
*tmpInput_->getTranspose(), *inputMulFactor_, 1, 1);
tmpInput_->rowScale(0, *inputSquare_, *oGrad);
tmpSum_->sumCols(*tmpInput_, -1, 0);
}
latentVectors_->getWGrad()->addRowScale(
0, *latentVectors_->getW(), *tmpSumTrans);
/* Increasing the number of gradient */
latentVectors_->getParameterPtr()->incUpdate(callback);
}
/* Calculate the input layers gradient */
MatrixPtr inGrad = getInputGrad(0);
if (inGrad != NULL) {
inGrad->mul(
*inputMulFactor_, *latentVectors_->getW()->getTranspose(), 1, 1);
tmpSumTrans->sumRows(*latentVectorsSquare_, -1, 0);
inGrad->addColScale(0, *inputV, *tmpSum_);
inGrad->rowScale(0, *inGrad, *oGrad);
}
}
} // namespace paddle
paddle/gserver/layers/FactorizationMachineLayer.h 0 → 100644
浏览文件 @ a1bf4c28
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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 "Layer.h"
#include "paddle/math/Matrix.h"
#include "paddle/utils/ThreadLocal.h"
namespace paddle {
/**
* @brief The Factorization Machine models pairwise (order-2) feature
* interactions as inner product of the learned latent vectors corresponding
* to each input feature.
*
* The Factorization Machine can effectively capture feature interactions
* especially when the input is sparse. While in principle FM can model higher
* order feature interaction, in practice usually only order-2 feature
* interactions are considered. The Factorization Machine Layer here only
* computes the order-2 interations with the formula:
*
* \f[
* y = \sum_{i=1}^{n-1}\sum_{j=i+1}^n\langle v_i, v_j \rangle x_i x_j
* \f]
*
* The detailed calculation for forward and backward can be found at this paper:
*
* Factorization machines.
*
* The config file api is factorization_machine.
*/
class FactorizationMachineLayer : public Layer {
protected:
// The latent vectors, shape: (size, factorSize_)
// Each row of the latentVectors_ matrix is the latent vector
// corresponding to one input feature dimension
std::unique_ptr<Weight> latentVectors_;
// The hyperparameter that defines the dimensionality of the factorization
size_t factorSize_;
private:
// Store the square values of the letent vectors matrix
MatrixPtr latentVectorsSquare_;
// Store the square values of input matrix
MatrixPtr inputSquare_;
// The result of input matrix * latent vector matrix that will be used in
// both forward and backward step
MatrixPtr inputMulFactor_;
// Store temporary calculation result
MatrixPtr tmpOut_;
MatrixPtr tmpSum_;
MatrixPtr tmpInput_;
// Negative identity matrix
MatrixPtr negOnes_;
public:
explicit FactorizationMachineLayer(const LayerConfig& config)
: Layer(config) {}
~FactorizationMachineLayer() {}
bool init(const LayerMap& layerMap,
const ParameterMap& parameterMap) override;
void forward(PassType passType) override;
void backward(const UpdateCallback& callback = nullptr) override;
};
} // namespace paddle
paddle/gserver/layers/HierarchicalSigmoidLayer.cpp
浏览文件 @ a1bf4c28
......@@ -64,49 +64,111 @@ void HierarchicalSigmoidLayer::forward(PassType passType) {
batchSize,
codeLength_,
/* trans */ false,
useGpu(deviceId_));
false);
Matrix::resizeOrCreate(preOutput_.grad,
batchSize,
codeLength_,
/* trans */ false,
useGpu(deviceId_));
false);
IVectorPtr label = getInput(*getLabelLayer()).ids;
preOutput_.value->zeroMem();
if (useGpu_) {
Matrix::resizeOrCreate(cpuOutput_,
output_.value->getHeight(),
output_.value->getWidth(),
/* trans */ false,
false);
IVector::resizeOrCreate(cpuLabel_, label->getSize(), false);
cpuLabel_->copyFrom(*label);
cpuOutput_->copyFrom(*output_.value);
} else {
cpuOutput_ = output_.value;
cpuLabel_ = label;
}
/* add the bias-vector */
if (biases_.get() != NULL) {
preOutput_.value->addByBitCode(numClasses_, *label, *biases_->getW());
if (useGpu_) {
Matrix::resizeOrCreate(cpuBias_,
1,
numClasses_ - 1,
/* trans */ false,
false);
cpuBias_->copyFrom(*biases_->getW());
} else {
cpuBias_ = biases_->getW();
}
preOutput_.value->addByBitCode(numClasses_, *cpuLabel_, *cpuBias_);
}
for (size_t i = 0; i < inputLayers_.size() - 1; ++i) {
MatrixPtr input = getInputValue(i);
if (useGpu_) {
Matrix::resizeOrCreate(cpuInput_,
input->getHeight(),
input->getWidth(),
/* trans */ false,
false);
Matrix::resizeOrCreate(cpuWeight_,
weights_[i]->getW()->getHeight(),
weights_[i]->getW()->getWidth(),
/* trans */ false,
false);
cpuInput_->copyFrom(*input);
cpuWeight_->copyFrom(*weights_[i]->getW());
} else {
cpuInput_ = input;
cpuWeight_ = weights_[i]->getW();
}
preOutput_.value->mulByBitCode(
numClasses_, *label, *weights_[i]->getW(), *input);
numClasses_, *cpuLabel_, *cpuWeight_, *cpuInput_);
}
// keep consistent with the clipping in the following softrelu
preOutput_.value->clip(-40.0, 40.0);
preOutput_.value->sumByBitCode(numClasses_,
*label,
*output_.value,
*cpuLabel_,
*cpuOutput_,
-1); // scaleSum
preOutput_.value->softrelu(*preOutput_.value);
MatrixPtr sum =
Matrix::create(batchSize, 1, /* trans= */ false, useGpu(deviceId_));
MatrixPtr sum = Matrix::create(batchSize, 1, /* trans= */ false, false);
preOutput_.value->rowSum(*sum);
output_.value->add(*sum);
cpuOutput_->add(*sum);
if (useGpu_) {
output_.value->copyFrom(*cpuOutput_);
} else {
output_.value = cpuOutput_;
}
}
void HierarchicalSigmoidLayer::backward(const UpdateCallback& callback) {
IVectorPtr label = getInput(*getLabelLayer()).ids;
if (useGpu_) {
IVector::resizeOrCreate(cpuLabel_, label->getSize(), false);
cpuLabel_->copyFrom(*label);
} else {
cpuLabel_ = label;
}
preOutput_.grad->one();
preOutput_.grad->softreluDerivative(*preOutput_.value);
preOutput_.grad->subByBitCode(numClasses_, *label);
preOutput_.grad->subByBitCode(numClasses_, *cpuLabel_);
if (biases_ && biases_->getWGrad()) {
preOutput_.grad->addByBitCodeBackward(
numClasses_, *label, *biases_->getWGrad());
MatrixPtr biases_grad = biases_->getWGrad();
if (useGpu_) {
Matrix::resizeOrCreate(cpuBias_,
1,
numClasses_ - 1,
/* trans */ false,
false);
cpuBias_->copyFrom(*biases_grad);
} else {
cpuBias_ = biases_grad;
}
preOutput_.grad->addByBitCodeBackward(numClasses_, *cpuLabel_, *cpuBias_);
if (useGpu_) {
biases_grad->copyFrom(*cpuBias_);
} else {
biases_grad = cpuBias_;
}
/* Increasing the number of gradient */
biases_->getParameterPtr()->incUpdate(callback);
}
......@@ -115,9 +177,31 @@ void HierarchicalSigmoidLayer::backward(const UpdateCallback& callback) {
/* Calculate the W-gradient for the current layer */
MatrixPtr input = getInputValue(i);
if (weights_[i]->getWGrad()) {
MatrixPtr weights_grad = weights_[i]->getWGrad();
if (useGpu_) {
Matrix::resizeOrCreate(cpuInput_,
input->getHeight(),
input->getWidth(),
/* trans */ false,
false);
Matrix::resizeOrCreate(cpuWeightGrad_,
weights_grad->getHeight(),
weights_grad->getWidth(),
/* trans */ false,
false);
cpuInput_->copyFrom(*input);
cpuWeightGrad_->copyFrom(*weights_grad);
} else {
cpuInput_ = input;
cpuWeightGrad_ = weights_grad;
}
preOutput_.grad->mulByBitCodeBackwardWeight(
numClasses_, *label, *weights_[i]->getWGrad(), *input);
numClasses_, *cpuLabel_, *cpuWeightGrad_, *cpuInput_);
if (useGpu_) {
weights_grad->copyFrom(*cpuWeightGrad_);
} else {
weights_grad = cpuWeightGrad_;
}
/* Increasing the number of gradient */
weights_[i]->getParameterPtr()->incUpdate(callback);
}
......@@ -125,8 +209,30 @@ void HierarchicalSigmoidLayer::backward(const UpdateCallback& callback) {
/* Calculate the input layers error */
MatrixPtr inputGrad = getInputGrad(i);
if (inputGrad) {
if (useGpu_) {
Matrix::resizeOrCreate(cpuInputGrad_,
inputGrad->getHeight(),
inputGrad->getWidth(),
/* trans */ false,
false);
Matrix::resizeOrCreate(cpuWeight_,
weights_[i]->getW()->getHeight(),
weights_[i]->getW()->getWidth(),
/* trans */ false,
false);
cpuInputGrad_->copyFrom(*inputGrad);
cpuWeight_->copyFrom(*weights_[i]->getW());
} else {
cpuInputGrad_ = inputGrad;
cpuWeight_ = weights_[i]->getW();
}
preOutput_.grad->mulByBitCodeBackwardError(
numClasses_, *label, *weights_[i]->getW(), *inputGrad);
numClasses_, *cpuLabel_, *cpuWeight_, *cpuInputGrad_);
if (useGpu_) {
inputGrad->copyFrom(*cpuInputGrad_);
} else {
inputGrad = cpuInputGrad_;
}
}
}
}
......
paddle/gserver/layers/HierarchicalSigmoidLayer.h
浏览文件 @ a1bf4c28
......@@ -80,6 +80,15 @@ protected:
int codeLength_;
/// temporary result of output_
Argument preOutput_;
/// The temporary variables in CPU memory.
MatrixPtr cpuWeight_;
MatrixPtr cpuWeightGrad_;
MatrixPtr cpuInput_;
MatrixPtr cpuInputGrad_;
MatrixPtr cpuBias_;
MatrixPtr cpuOutput_;
IVectorPtr cpuLabel_;
};
} // namespace paddle
paddle/gserver/tests/test_LayerGrad.cpp
浏览文件 @ a1bf4c28
......@@ -681,12 +681,13 @@ TEST(Layer, hsigmoidLayer) {
config.layerConfig.add_inputs();
config.layerConfig.add_inputs();
// Not support GPU now
testLayerGrad(config,
"hsigmoid",
100,
/* trans */ false, /* useGpu */
false);
for (auto useGpu : {false, true}) {
testLayerGrad(config,
"hsigmoid",
100,
/* trans */ false,
/* useGpu */ useGpu);
}
}
TEST(Layer, multi_cross) {
......@@ -2464,6 +2465,25 @@ TEST(Layer, L2DistanceLayer) {
}
}
void testFactorizationMachineLayer(InputType type, bool useGpu) {
const int FACTOR_SIZE = 10;
TestConfig config;
config.layerConfig.set_type("factorization_machine");
config.layerConfig.set_factor_size(FACTOR_SIZE);
config.layerConfig.set_size(1);
config.biasSize = 0;
config.inputDefs.push_back({type, "layer_0", 128, 1280});
config.layerConfig.add_inputs();
testLayerGrad(config, "factorization_machine", 16, false, useGpu, false);
}
TEST(Layer, FactorizationMachineLayer) {
for (auto useGpu : {false, true}) {
testFactorizationMachineLayer(INPUT_DATA, useGpu);
}
testFactorizationMachineLayer(INPUT_SPARSE_FLOAT_VALUE_DATA, false);
}
int main(int argc, char** argv) {
testing::InitGoogleTest(&argc, argv);
initMain(argc, argv);
......
paddle/math/CpuSparseMatrix.cpp
浏览文件 @ a1bf4c28
......@@ -260,6 +260,35 @@ void CpuSparseMatrix::printOneRow(std::ostream& os, size_t idx) const {
os << ";";
}
void CpuSparseMatrix::rowScale(size_t cCol, CpuSparseMatrix& b, Matrix& c) {
CHECK(getFormat() != SPARSE_CSC) << "Not supported";
CHECK_EQ(height_, b.getHeight());
CHECK_EQ(width_, b.getWidth());
real* A = getValue();
real* B = b.getValue();
if (b.getValueType() == FLOAT_VALUE) {
for (size_t i = 0; i < height_; i++) {
size_t start = getRowStartIdx(i);
size_t end = getRowStartIdx(i + 1);
CHECK_EQ(start, b.getRowStartIdx(i));
CHECK_EQ(end, b.getRowStartIdx(i + 1));
for (size_t j = start; j < end; j++) {
A[j] = B[j] * c.getElement(i, cCol);
}
}
} else if (b.getValueType() == NO_VALUE) {
for (size_t i = 0; i < height_; i++) {
size_t start = getRowStartIdx(i);
size_t end = getRowStartIdx(i + 1);
CHECK_EQ(start, b.getRowStartIdx(i));
CHECK_EQ(end, b.getRowStartIdx(i + 1));
for (size_t j = start; j < end; j++) {
A[j] = c.getElement(i, cCol);
}
}
}
}
void CpuSparseMatrix::randomizeUniform() {
CHECK_LE(elementCnt_, height_ * width_);
if (valueType_ == FLOAT_VALUE) {
......
paddle/math/CpuSparseMatrix.h
浏览文件 @ a1bf4c28
......@@ -239,6 +239,15 @@ public:
const unsigned int* cols,
const real* values);
/**
* @brief this_row = b_row * c_row[cCol]
*
* @param[in] cCol the column of matrix c used to scale each row of b
* @param[in] b CpuSparseMatrix
* @param[in] c Matrix
*/
void rowScale(size_t cCol, CpuSparseMatrix& b, Matrix& c);
void randomizeUniform();
void copyFrom(const GpuSparseMatrix& src, hl_stream_t stream);
......
paddle/operators/CMakeLists.txt
浏览文件 @ a1bf4c28
......@@ -205,8 +205,24 @@ set(DEPS_OPS
tensor_array_read_write_op
gru_op
adagrad_op
sgd_op)
sgd_op
save_op
load_op
send_op
recv_op)
add_subdirectory(detail)
op_library(send_op SRCS send_op.cc DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib_target protobuf)
set_source_files_properties(
send_op.cc
PROPERTIES
COMPILE_FLAGS "-Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
op_library(recv_op SRCS recv_op.cc DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib_target protobuf)
set_source_files_properties(
recv_op.cc
PROPERTIES
COMPILE_FLAGS "-Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
op_library(cond_op SRCS cond_op.cc DEPS framework_proto tensor operator net_op)
op_library(cross_entropy_op DEPS cross_entropy)
......@@ -235,6 +251,10 @@ op_library(conv_transpose_op DEPS vol2col)
op_library(gru_op DEPS sequence2batch gru_compute)
op_library(recurrent_op SRCS recurrent_op.cc DEPS executor)
# FIXME(typhoonzero): save/load depends lodtensor serialization functions
op_library(save_op DEPS lod_tensor)
op_library(load_op DEPS lod_tensor)
list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
foreach(src ${GENERAL_OPS})
op_library(${src})
......@@ -242,6 +262,8 @@ endforeach()
set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")
cc_test(gather_test SRCS gather_test.cc DEPS tensor)
cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)
......@@ -251,3 +273,4 @@ if(WITH_GPU)
cc_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
endif()
cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op)
cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS send_op recv_op sum_op executor)
paddle/operators/batch_norm_op.cc
浏览文件 @ a1bf4c28
......@@ -62,13 +62,14 @@ class BatchNormOp : public framework::OperatorWithKernel {
const auto x_dims = ctx->GetInputDim("X");
const TensorFormat tensor_format =
StringToTensorFormat(ctx->Attrs().Get<std::string>("tensor_format"));
PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5,
"Input X must have 2 to 5 dimensions.");
const int C =
(tensor_format == TensorFormat::NCHW ? x_dims[1]
: x_dims[x_dims.size() - 1]);
PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5,
"Input X must have 3 to 5 dimensions.");
PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale").size(), 1UL);
PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale")[0], C);
PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias").size(), 1UL);
......@@ -146,8 +147,8 @@ class BatchNormKernel<platform::CPUPlace, T> : public framework::OpKernel<T> {
const auto *x = ctx.Input<Tensor>("X");
const auto &x_dims = x->dims();
PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5,
"The Input dim size should be between 3 and 5");
PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5,
"The Input dim size should be between 2 and 5");
const int N = x_dims[0];
const int C =
(tensor_format == TensorFormat::NCHW ? x_dims[1]
......@@ -339,8 +340,8 @@ class BatchNormGradKernel<platform::CPUPlace, T>
// Get the size for each dimension.
// NCHW [batch_size, in_channels, in_height, in_width]
const auto &x_dims = x->dims();
PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5,
"The Input dim size should be between 3 and 5");
PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5,
"The Input dim size should be between 2 and 5");
const int N = x_dims[0];
const int C =
(tensor_format == TensorFormat::NCHW ? x_dims[1]
......
paddle/operators/batch_norm_op.cu.cc
浏览文件 @ a1bf4c28
......@@ -29,14 +29,21 @@ void ExtractNCWHD(const framework::DDim &dims,
const TensorFormat &tensor_format, int *N, int *C, int *H,
int *W, int *D) {
*N = dims[0];
*C = tensor_format == TensorFormat::NCHW ? dims[1] : dims[dims.size() - 1];
*H = tensor_format == TensorFormat::NCHW ? dims[2] : dims[1];
*W = dims.size() > 3
? (tensor_format == TensorFormat::NCHW ? dims[3] : dims[2])
: 1;
*D = dims.size() > 4
? (tensor_format == TensorFormat::NCHW ? dims[4] : dims[3])
: 1;
if (dims.size() == 2) {
*C = dims[1];
*H = 1;
*W = 1;
*D = 1;
} else {
*C = tensor_format == TensorFormat::NCHW ? dims[1] : dims[dims.size() - 1];
*H = tensor_format == TensorFormat::NCHW ? dims[2] : dims[1];
*W = dims.size() > 3
? (tensor_format == TensorFormat::NCHW ? dims[3] : dims[2])
: 1;
*D = dims.size() > 4
? (tensor_format == TensorFormat::NCHW ? dims[4] : dims[3])
: 1;
}
}
template <typename T>
......@@ -56,8 +63,8 @@ class BatchNormKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
// NCHW [batch_size, in_channels, in_height, in_width]
const auto *x = ctx.Input<Tensor>("X");
const auto &x_dims = x->dims();
PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5,
"The Input dim size should be between 3 and 5");
PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5,
"The Input dim size should be between 2 and 5");
int N, C, H, W, D;
ExtractNCWHD(x_dims, tensor_format, &N, &C, &H, &W, &D);
......@@ -180,8 +187,8 @@ class BatchNormGradKernel<platform::GPUPlace, T>
const auto &x_dims = x->dims();
PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5,
"The Input dim size should be between 3 and 5");
PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5,
"The Input dim size should be between 2 and 5");
int N, C, H, W, D;
ExtractNCWHD(x_dims, tensor_format, &N, &C, &H, &W, &D);
......
paddle/operators/conv_transpose_op.cc
浏览文件 @ a1bf4c28
......@@ -74,12 +74,12 @@ Conv2DTransposeOpMaker::Conv2DTransposeOpMaker(
"The format of output tensor is also NCHW.");
AddAttr<std::vector<int>>(
"strides",
"(vector<int> defalut:{1, 1}), the strides(h_stride, w_stride) of "
"(vector<int> default:{1, 1}), the strides(h_stride, w_stride) of "
"convolution transpose operator.")
.SetDefault({1, 1});
AddAttr<std::vector<int>>(
"paddings",
"(vector<int> defalut:{0, 0}), the paddings(h_pad, w_pad) of convolution "
"(vector<int> default:{0, 0}), the paddings(h_pad, w_pad) of convolution "
"transpose operator.")
.SetDefault({0, 0});
AddComment(R"DOC(
......@@ -101,8 +101,8 @@ Example:
Output:
Output shape: (N, C_out, H_out, W_out)
where
H_out = (H_in - 1) * strides[0] - 2 * paddings[0] + filter_size[0];
W_out = (W_in - 1) * strides[1] - 2 * paddings[1] + filter_size[1];
H_out = (H_in - 1) * strides[0] - 2 * paddings[0] + H_f;
W_out = (W_in - 1) * strides[1] - 2 * paddings[1] + W_f;
)DOC");
}
......@@ -130,12 +130,12 @@ Conv3DTransposeOpMaker::Conv3DTransposeOpMaker(
"the number of channels, D is the depth of the feature, H is the "
"height of the feature, and W is the width of the feature.");
AddAttr<std::vector<int>>("strides",
"(vector<int> defalut:{1, 1, 1}), the "
"(vector<int> default:{1, 1, 1}), the "
"strides{d_stride, h_stride, w_stride} of "
"convolution transpose operator.")
.SetDefault({1, 1, 1});
AddAttr<std::vector<int>>("paddings",
"(vector<int> defalut:{0, 0, 0}), paddings(d_pad, "
"(vector<int> default:{0, 0, 0}), paddings(d_pad, "
"h_pad, w_pad) of convolution transpose operator.")
.SetDefault({0, 0, 0});
AddComment(R"DOC(
......@@ -158,9 +158,9 @@ Example:
Output:
Output shape: (N, C_out, D_out, H_out, W_out)
where
D_out = (D_in - 1) * strides[0] - 2 * paddings[0] + filter_size[0];
H_out = (H_in - 1) * strides[1] - 2 * paddings[1] + filter_size[1];
W_out = (W_in - 1) * strides[2] - 2 * paddings[2] + filter_size[2];
D_out = (D_in - 1) * strides[0] - 2 * paddings[0] + D_f;
H_out = (H_in - 1) * strides[1] - 2 * paddings[1] + H_f;
W_out = (W_in - 1) * strides[2] - 2 * paddings[2] + W_f;
)DOC");
}
......
paddle/operators/detail/CMakeLists.txt 0 → 100644
浏览文件 @ a1bf4c28
grpc_library(sendrecvop_grpc SRCS recv_impl.cc send_impl.cc PROTO send_recv.proto DEPS lod_tensor selected_rows)
paddle/operators/detail/recv_impl.cc 0 → 100644
浏览文件 @ a1bf4c28
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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 "send_recv_impl.h"
namespace paddle {
namespace operators {
namespace detail {
Status SendRecvServerImpl::SendVariable(ServerContext *context,
const VariableMessage *in_var,
VariableMessage *out_var) {
framework::LoDTensor t;
// TODO(typhoonzero): desirealize in_tensor and run pserver network.
std::istringstream iss(in_var->serialized());
framework::DeserializeFromStream(iss, &t);
lodtensor_queue_.Push(std::move(t));
// Block util the sub graph is done.
t = lodtensor_return_queue_.Pop();
std::ostringstream oss;
// FIXME(typhoonzero): get context from op.
framework::SerializeToStream(oss, t, platform::CPUDeviceContext());
std::string *varname = out_var->mutable_varname();
*varname = in_var->varname();
std::string *serialized = out_var->mutable_serialized();
*serialized = oss.str();
return Status::OK;
}
} // namespace detail
} // namespace operators
} // namespace paddle
paddle/operators/detail/send_impl.cc 0 → 100644
浏览文件 @ a1bf4c28
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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 "send_recv_impl.h"
namespace paddle {
namespace operators {
namespace detail {
bool RPCClient::SendVariable(const framework::Scope& scope,
const std::string& inname,
const std::string& outname) {
ClientContext context;
VariableMessage msg, out_msg;
// FIXME(typhoonzero): pass device context to here.
auto ctx = platform::CPUDeviceContext();
auto* var = scope.FindVar(inname);
PADDLE_ENFORCE(var);
// TODO(typhoonzero): support SelectedRows
PADDLE_ENFORCE(var->IsType<framework::LoDTensor>(),
"Only support LoDTensor, %s has wrong type", inname);
const framework::LoDTensor& tensor = var->Get<framework::LoDTensor>();
std::ostringstream oss;
framework::SerializeToStream(oss, tensor, ctx);
msg.set_varname(inname);
msg.set_serialized(oss.str());
Status status = stub_->SendVariable(&context, msg, &out_msg);
if (!status.ok()) {
return false;
}
std::istringstream iss(out_msg.serialized());
framework::LoDTensor ret_tensor;
framework::DeserializeFromStream(iss, &ret_tensor);
auto* outvar = scope.FindVar(outname);
framework::LoDTensor* out_tensor = outvar->GetMutable<framework::LoDTensor>();
// FIXME(typhoonzero): do not copy.
framework::CopyFrom(ret_tensor, ctx.GetPlace(), ctx, out_tensor);
return true;
}
} // namespace detail
} // namespace operators
} // namespace paddle
paddle/operators/detail/send_recv.proto 0 → 100644
浏览文件 @ a1bf4c28
/* 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. */
syntax = "proto3";
package sendrecv;
service SendRecvService {
// For parameter server round-robin like hashing, do not split tensors.
// Send and recv only one tensor
rpc SendVariable(VariableMessage) returns (VariableMessage) {}
}
// VariableMessage is serialized paddle variable message.
// It can be:
// Tensor
// LoDTensor
// SelectedRows
message VariableMessage {
string varname = 1;
bytes serialized = 2;
}
message VoidMessage {}
\ No newline at end of file
paddle/operators/detail/send_recv_impl.h 0 → 100644
浏览文件 @ a1bf4c28
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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/framework/data_type.h"
#include "paddle/framework/lod_tensor.h"
#include "paddle/framework/scope.h"
#include "paddle/framework/selected_rows.h"
#include "paddle/operators/detail/simple_block_queue.h"
// #include <grpc++/channel.h>
// #include <grpc++/client_context.h>
// #include <grpc++/create_channel.h>
// #include <grpc++/security/credentials.h>
#include "paddle/operators/detail/send_recv.grpc.pb.h"
#include "paddle/operators/detail/send_recv.pb.h"
#include <grpc++/grpc++.h>
using grpc::Channel;
using grpc::Server;
using grpc::ServerContext;
using grpc::ServerReader;
using grpc::ServerBuilder;
using grpc::ClientContext;
using grpc::ClientReader;
using grpc::ClientReaderWriter;
using grpc::ClientWriter;
using grpc::Status;
using sendrecv::SendRecvService;
using sendrecv::VariableMessage;
using sendrecv::VoidMessage;
namespace paddle {
namespace operators {
namespace detail {
class SendRecvServerImpl final : public SendRecvService::Service {
public:
explicit SendRecvServerImpl() {}
Status SendVariable(ServerContext *context, const VariableMessage *in_var,
VariableMessage *out_var) override;
const framework::LoDTensor Get() { return this->lodtensor_queue_.Pop(); }
void Push(const framework::LoDTensor &tensor) {
this->lodtensor_return_queue_.Push(tensor);
}
private:
SimpleBlockQueue<framework::LoDTensor> lodtensor_queue_;
SimpleBlockQueue<framework::LoDTensor> lodtensor_return_queue_;
SimpleBlockQueue<framework::SelectedRows> selected_rows_queue_;
SimpleBlockQueue<framework::SelectedRows> selected_rows_return_queue_;
};
// RPCClient is a class to send tensors to pserver sub-network
// using different hashing methods.
class RPCClient {
public:
RPCClient(std::shared_ptr<Channel> channel)
: stub_(SendRecvService::NewStub(channel)) {}
bool SendVariable(const framework::Scope &scope, const std::string &inname,
const std::string &outname);
private:
std::unique_ptr<SendRecvService::Stub> stub_;
};
} // namespace detail
} // namespace operators
} // namespace paddle
paddle/operators/detail/simple_block_queue.h 0 → 100644
浏览文件 @ a1bf4c28
/* 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 <condition_variable>
#include <deque>
#include <mutex>
namespace paddle {
namespace operators {
namespace detail {
template <typename T>
class SimpleBlockQueue {
private:
std::mutex mutex_;
std::condition_variable condition_;
std::deque<T> queue_;
public:
void Push(T const& value) {
{
std::unique_lock<std::mutex> lock(this->mutex_);
queue_.push_front(value);
}
this->condition_.notify_one();
}
T Pop() {
std::unique_lock<std::mutex> lock(this->mutex_);
this->condition_.wait(lock, [=] { return !this->queue_.empty(); });
T rc(std::move(this->queue_.back()));
this->queue_.pop_back();
return rc;
}
};
} // namespace detail
} // namespace operators
} // namespace paddle
paddle/operators/load_op.cc
浏览文件 @ a1bf4c28
......@@ -38,61 +38,7 @@ class LoadOp : public framework::OperatorBase {
out_var_name);
auto *tensor = out_var->GetMutable<framework::LoDTensor>();
uint32_t version;
fin.read(reinterpret_cast<char *>(&version), sizeof(version));
PADDLE_ENFORCE_EQ(version, 0U, "Only version 0 is supported");
framework::TensorDesc desc;
{ // int32_t size
// proto buffer
int32_t size;
fin.read(reinterpret_cast<char *>(&size), sizeof(size));
std::unique_ptr<char[]> buf(new char[size]);
fin.read(reinterpret_cast<char *>(buf.get()), size);
PADDLE_ENFORCE(desc.ParseFromArray(buf.get(), size),
"Cannot parse tensor desc");
}
{ // read tensor
std::vector<int64_t> dims;
dims.reserve(static_cast<size_t>(desc.dims().size()));
std::copy(desc.dims().begin(), desc.dims().end(),
std::back_inserter(dims));
tensor->Resize(framework::make_ddim(dims));
void *buf;
platform::Place cpu = platform::CPUPlace();
switch (desc.data_type()) {
case framework::FP32:
buf = tensor->mutable_data<float>(cpu);
break;
case framework::FP64:
buf = tensor->mutable_data<double>(cpu);
break;
case framework::INT32:
buf = tensor->mutable_data<int>(cpu);
break;
case framework::INT64:
buf = tensor->mutable_data<int64_t>(cpu);
break;
default:
PADDLE_THROW("DataType %d not supported", desc.data_type());
}
fin.read(static_cast<char *>(buf), tensor->memory_size());
}
{ // read lod
uint64_t lod_level;
fin.read(reinterpret_cast<char *>(&lod_level), sizeof(lod_level));
auto &lod = *tensor->mutable_lod();
lod.resize(lod_level);
for (uint64_t i = 0; i < lod_level; ++i) {
uint64_t size;
fin.read(reinterpret_cast<char *>(&size), sizeof(size));
std::vector<size_t> tmp(size / sizeof(size_t));
fin.read(reinterpret_cast<char *>(tmp.data()),
static_cast<std::streamsize>(size));
lod[i] = tmp;
}
}
framework::DeserializeFromStream(fin, tensor);
auto place = dev_ctx.GetPlace();
if (platform::is_gpu_place(place)) {
......
paddle/operators/log_loss_op.cc 0 → 100644
浏览文件 @ a1bf4c28
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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/operators/log_loss_op.h"
namespace paddle {
namespace operators {
class LogLossOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Predicted"),
"Input(Predicted) must be initialized.");
PADDLE_ENFORCE(ctx->HasInput("Labels"),
"Input(Labels) must be initialized.");
auto pred_dims = ctx->GetInputDim("Predicted");
auto label_dims = ctx->GetInputDim("Labels");
PADDLE_ENFORCE_EQ(pred_dims, label_dims);
PADDLE_ENFORCE_EQ(pred_dims.size(), 2,
"The rank of Input(Predicted) must be 2 and the shape is "
"[batch_size, 1].");
PADDLE_ENFORCE_EQ(pred_dims[1], 1,
"Each row of Input(Predicted) contains a real value, "
"so the 2nd dimension of Input(X) must be 1.");
ctx->SetOutputDim("Loss", {pred_dims[0], 1});
ctx->ShareLoD("Predicted", "Loss");
}
};
template <typename AttrType>
class LogLossOpMaker : public framework::OpProtoAndCheckerMaker {
public:
LogLossOpMaker(framework::OpProto* proto,
framework::OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("Predicted",
"The input value (Predicted) of Log loss op."
"Predicted is a 2-D tensor with shape [batch_size, 1].");
AddInput("Labels",
"The target value (Labels) of Log loss op."
"Labels is a 2-D tensor with shape [batch_size, 1].");
AddOutput("Loss",
"The output tensor with shape [batch_size, 1] "
"which represents the log loss.");
AddAttr<AttrType>("epsilon", "Epsilon in log loss.");
AddComment(R"DOC(
LogLoss Operator.
Log loss is a loss function used for binary classification. Log Loss quantifies
the accuracy of a classifier by penalising false classifications. Minimising the
Log Loss is equivalent to maximising the accuracy of the classifier. We define
Predicted as the values predicted by our model and Labels as the target ground
truth value. Log loss can evaluate how close the predicted values are to the
target. The shapes of Predicted and Labels are both [batch_size, 1].
The equation is:
$$
Loss = - Labels * log(Predicted + \epsilon) -
(1 - Labels) * log(1 - Predicted + \epsilon)
$$
)DOC");
}
};
class LogLossGradOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Predicted"),
"Input(Predicted) should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Labels"),
"Input(Labels) should not be null.");
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Loss")),
"Input(Loss@GRAD) should not be null.");
PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Predicted")),
"Output(Predicted@GRAD) should not be null.");
auto pred_dims = ctx->GetInputDim("Predicted");
auto label_dims = ctx->GetInputDim("Labels");
auto loss_grad_dims = ctx->GetInputDim(framework::GradVarName("Loss"));
PADDLE_ENFORCE_EQ(loss_grad_dims, pred_dims);
auto pred_grad_name = framework::GradVarName("Predicted");
ctx->SetOutputDim(pred_grad_name, pred_dims);
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(log_loss, ops::LogLossOp, ops::LogLossOpMaker<float>, log_loss_grad,
ops::LogLossGradOp);
REGISTER_OP_CPU_KERNEL(log_loss,
ops::LogLossKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(
log_loss_grad, ops::LogLossGradKernel<paddle::platform::CPUPlace, float>);
paddle/operators/log_loss_op.cu 0 → 100644
浏览文件 @ a1bf4c28
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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. */
#define EIGEN_USE_GPU
#include "paddle/operators/log_loss_op.h"
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(log_loss,
ops::LogLossKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(
log_loss_grad, ops::LogLossGradKernel<paddle::platform::GPUPlace, float>);
paddle/operators/log_loss_op.h 0 → 100644
浏览文件 @ a1bf4c28
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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/framework/eigen.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
template <typename Place, typename T, typename AttrType = T>
class LogLossKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* loss_out = ctx.Output<Tensor>("Loss");
loss_out->mutable_data<T>(ctx.GetPlace());
auto epsilon = static_cast<T>(ctx.Attr<AttrType>("epsilon"));
auto prediction = EigenVector<T>::Flatten(*ctx.Input<Tensor>("Predicted"));
auto label = EigenVector<T>::Flatten(*ctx.Input<Tensor>("Labels"));
auto loss = EigenVector<T>::Flatten(*loss_out);
auto place = ctx.GetEigenDevice<Place>();
loss.device(place) = (-(label * (prediction + epsilon).log()) -
((static_cast<T>(1) - label) *
(static_cast<T>(1) - prediction + epsilon).log()));
}
};
template <typename Place, typename T, typename AttrType = T>
class LogLossGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto epsilon = static_cast<T>(ctx.Attr<AttrType>("epsilon"));
auto prediction = EigenVector<T>::Flatten(*ctx.Input<Tensor>("Predicted"));
auto label = EigenVector<T>::Flatten(*ctx.Input<Tensor>("Labels"));
auto* dloss = ctx.Input<Tensor>(framework::GradVarName("Loss"));
auto* dpred = ctx.Output<Tensor>(framework::GradVarName("Predicted"));
auto dl = EigenVector<T>::Flatten(*dloss);
auto place = ctx.GetEigenDevice<Place>();
if (dpred) {
dpred->mutable_data<T>(ctx.GetPlace());
auto dx = framework::EigenVector<T>::Flatten(*dpred);
dx.device(place) = dl * (-(label / (prediction + epsilon)) +
((static_cast<T>(1) - label) /
(static_cast<T>(1) - prediction + epsilon)));
}
}
};
} // namespace operators
} // namespace paddle
paddle/operators/math/maxouting.cc
浏览文件 @ a1bf4c28
......@@ -23,8 +23,7 @@ template <typename T>
class MaxOutFunctor<platform::CPUPlace, T> {
public:
void operator()(const platform::DeviceContext& context,
const framework::Tensor& input,
framework::Tensor * output,
const framework::Tensor& input, framework::Tensor* output,
int groups) {
const int batch_size = input.dims()[0];
const int input_height = input.dims()[2];
......@@ -37,34 +36,30 @@ class MaxOutFunctor<platform::CPUPlace, T> {
T* output_data = output->mutable_data<T>(context.GetPlace());
for (int i = 0; i < batch_size; ++i) {
int new_bindex = c_size * i;
int new_bindex = c_size * i;
for (int c = 0; c < output_channels; ++c) {
int new_cindex = fea_size * c;
for (int f = 0; f < fea_size; ++f) {
T ele = static_cast<T>(-FLT_MAX);
for (int ph = 0; ph < groups; ++ph) {
T x = input_data[(new_bindex + new_cindex) * groups
+ ph * fea_size + f];
T x = input_data[(new_bindex + new_cindex) * groups +
ph * fea_size + f];
ele = ele > x ? ele : x;
}
output_data[(new_bindex+new_cindex+f)] = ele;
output_data[(new_bindex + new_cindex + f)] = ele;
}
}
}
}
};
template <class T>
class MaxOutGradFunctor<platform::CPUPlace, T> {
public:
public:
void operator()(const platform::DeviceContext& context,
const framework::Tensor& input,
framework::Tensor * input_grad,
const framework::Tensor& input, framework::Tensor* input_grad,
const framework::Tensor& output,
const framework::Tensor& output_grad,
int groups) {
const framework::Tensor& output_grad, int groups) {
const int batch_size = input.dims()[0];
const int input_height = input.dims()[2];
const int input_width = input.dims()[3];
......@@ -84,11 +79,11 @@ public:
bool continue_match = true;
int output_idx = blen + clen + f;
for (int g = 0; g < groups && continue_match; ++g) {
int input_idx = input_idx0 + fea_size * g;
if (input_data[input_idx] == output_data[output_idx]) {
input_grad_data[input_idx] += output_grad_data[output_idx];
continue_match = false;
}
int input_idx = input_idx0 + fea_size * g;
if (input_data[input_idx] == output_data[output_idx]) {
input_grad_data[input_idx] += output_grad_data[output_idx];
continue_match = false;
}
}
}
}
......
paddle/operators/math/maxouting.cu
浏览文件 @ a1bf4c28
......@@ -21,9 +21,9 @@ namespace math {
template <typename T>
__global__ void KernelMaxOut(const int nthreads, const T* input_data,
const int channels,
const int input_height, const int input_width,
int groups, T* output_data ) {
const int channels, const int input_height,
const int input_width, int groups,
T* output_data) {
const int size = input_height * input_width * channels / groups;
const int feat_len = input_height * input_width;
int index = blockIdx.x * blockDim.x + threadIdx.x;
......@@ -34,7 +34,7 @@ __global__ void KernelMaxOut(const int nthreads, const T* input_data,
int channel_idx = batch_offset / feat_len;
int feat_idx = batch_offset % feat_len;
int data_idx =
(batch_idx * size + channel_idx * feat_len) * groups + feat_idx;
(batch_idx * size + channel_idx * feat_len) * groups + feat_idx;
T ele = static_cast<T>(-FLT_MAX);
for (int g = 0; g < groups; ++g) {
T x = input_data[data_idx + g * feat_len];
......@@ -44,34 +44,35 @@ __global__ void KernelMaxOut(const int nthreads, const T* input_data,
}
}
template <typename T>
__global__ void KernelMaxoutGrad(
const int nthreads, const T* input_data, const T* output_data,
const T* output_grad, T* input_grad, const int channels,
const int input_height, const int input_width, int groups) {
const int size = input_height * input_width * channels / groups;
const int feat_len = input_height * input_width;
int index = blockIdx.x * blockDim.x + threadIdx.x;
int offset = blockDim.x * gridDim.x;
for (int i = index; i < nthreads; i += offset) {
int batch_idx = i / size;
int batch_offset = i % size;
int channel_idx = batch_offset / feat_len;
int feat_idx = batch_offset % feat_len;
int data_idx =
__global__ void KernelMaxoutGrad(const int nthreads, const T* input_data,
const T* output_data, const T* output_grad,
T* input_grad, const int channels,
const int input_height, const int input_width,
int groups) {
const int size = input_height * input_width * channels / groups;
const int feat_len = input_height * input_width;
int index = blockIdx.x * blockDim.x + threadIdx.x;
int offset = blockDim.x * gridDim.x;
for (int i = index; i < nthreads; i += offset) {
int batch_idx = i / size;
int batch_offset = i % size;
int channel_idx = batch_offset / feat_len;
int feat_idx = batch_offset % feat_len;
int data_idx =
(batch_idx * size + channel_idx * feat_len) * groups + feat_idx;
int max_index = -1;
bool continue_match = true;
for (int g = 0; g < groups && continue_match; ++g) {
if (input_data[data_idx + g * feat_len] == output_data[i]) {
max_index = data_idx + g * feat_len;
continue_match = false;
break;
}
}
if (max_index != -1) {
input_grad[max_index] += output_grad[index];
int max_index = -1;
bool continue_match = true;
for (int g = 0; g < groups && continue_match; ++g) {
if (input_data[data_idx + g * feat_len] == output_data[i]) {
max_index = data_idx + g * feat_len;
continue_match = false;
break;
}
}
if (max_index != -1) {
input_grad[max_index] += output_grad[index];
}
}
}
/*
* All tensors are in NCHW format.
......@@ -80,7 +81,7 @@ template <typename T>
class MaxOutFunctor<platform::GPUPlace, T> {
public:
void operator()(const platform::DeviceContext& context,
const framework::Tensor& input, framework::Tensor * output,
const framework::Tensor& input, framework::Tensor* output,
int groups) {
const int batch_size = input.dims()[0];
const int input_channels = input.dims()[1];
......@@ -92,7 +93,7 @@ class MaxOutFunctor<platform::GPUPlace, T> {
const T* input_data = input.data<T>();
T* output_data = output->mutable_data<T>(context.GetPlace());
int nthreads = output->numel();
int nthreads = output->numel();
int blocks = (nthreads + 1024 - 1) / 1024;
dim3 threads(1024, 1);
dim3 grid(blocks, 1);
......@@ -101,8 +102,7 @@ class MaxOutFunctor<platform::GPUPlace, T> {
T><<<grid, threads, 0,
reinterpret_cast<const platform::CUDADeviceContext&>(context)
.stream()>>>(nthreads, input_data, input_channels,
input_height, input_width, groups,
output_data);
input_height, input_width, groups, output_data);
}
};
/*
......@@ -112,11 +112,9 @@ template <typename T>
class MaxOutGradFunctor<platform::GPUPlace, T> {
public:
void operator()(const platform::DeviceContext& context,
const framework::Tensor& input,
framework::Tensor * input_grad,
const framework::Tensor& input, framework::Tensor* input_grad,
const framework::Tensor& output,
const framework::Tensor& output_grad,
int groups) {
const framework::Tensor& output_grad, int groups) {
const int batch_size = input.dims()[0];
const int input_channels = input.dims()[1];
const int input_height = input.dims()[2];
......@@ -129,7 +127,7 @@ class MaxOutGradFunctor<platform::GPUPlace, T> {
const T* output_data = output.data<T>();
const T* output_grad_data = output_grad.data<T>();
T* input_grad_data = input_grad->mutable_data<T>(context.GetPlace());
int nthreads = output.numel();
int nthreads = output.numel();
int blocks = (nthreads + 1024 - 1) / 1024;
dim3 threads(1024, 1);
dim3 grid(blocks, 1);
......@@ -137,9 +135,9 @@ class MaxOutGradFunctor<platform::GPUPlace, T> {
KernelMaxoutGrad<
T><<<grid, threads, 0,
reinterpret_cast<const platform::CUDADeviceContext&>(context)
.stream()>>>(
nthreads, input_data, output_data, output_grad_data, input_grad_data,
input_channels, input_height, input_width, groups);
.stream()>>>(nthreads, input_data, output_data,
output_grad_data, input_grad_data, input_channels,
input_height, input_width, groups);
}
};
......
paddle/operators/math/maxouting.h
浏览文件 @ a1bf4c28
......@@ -21,15 +21,14 @@ namespace paddle {
namespace operators {
namespace math {
#define FLT_MAX \
__FLT_MAX__
#define FLT_MAX __FLT_MAX__
template <typename Place, typename T>
class MaxOutFunctor {
public:
void operator()(const platform::DeviceContext& context,
const framework::Tensor& input, framework::Tensor * output,
const framework::Tensor& input, framework::Tensor* output,
int groups);
};
......@@ -37,8 +36,7 @@ template <typename Place, class T>
class MaxOutGradFunctor {
public:
void operator()(const platform::DeviceContext& context,
const framework::Tensor& input,
framework::Tensor * input_grad,
const framework::Tensor& input, framework::Tensor* input_grad,
const framework::Tensor& output,
const framework::Tensor& output_grad, int groups);
};
......
paddle/operators/maxout_op.cc
浏览文件 @ a1bf4c28
......@@ -22,16 +22,17 @@ class MaxOutOpMaker : public framework::OpProtoAndCheckerMaker {
public:
MaxOutOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X",
AddInput(
"X",
"(Tensor) The input tensor of maxout operator. "
"The format of input tensor is NCHW. Where N is batch size, C is the "
"number of channels, H and W is the height and width of feature.");
AddOutput("Out",
"(Tensor) The output tensor of maxout operator."
"The format of output tensor is also NCHW."
"Where N is batch size, C is "
"the number of channels, H and W is the height and "
"width of feature.");
"(Tensor) The output tensor of maxout operator."
"The format of output tensor is also NCHW."
"Where N is batch size, C is "
"the number of channels, H and W is the height and "
"width of feature.");
AddAttr<int>(
"groups",
R"DOC("Specifies how many groups the input tensor will be split"
......@@ -59,21 +60,19 @@ class MaxOutOpMaker : public framework::OpProtoAndCheckerMaker {
}
};
class MaxOutOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of MaxoutOp"
PADDLE_ENFORCE(ctx->HasInput("X"),
"Input(X) of MaxoutOp"
"should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("Out"),
"Output(Out) of MaxoutOp should not be null.");
auto in_x_dims = ctx->GetInputDim("X");
int groups = ctx->Attrs().Get<int>("groups");
// check groups > 1
PADDLE_ENFORCE_GT(
groups, 1,
"groups should be larger than 1 in maxoutop");
PADDLE_ENFORCE_GT(groups, 1, "groups should be larger than 1 in maxoutop");
std::vector<int64_t> output_shape({in_x_dims[0], in_x_dims[1] / groups});
output_shape.push_back(in_x_dims[2]);
output_shape.push_back(in_x_dims[3]);
......@@ -87,18 +86,17 @@ class MaxOutOpGrad : public framework::OperatorWithKernel {
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null.");
PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
"Input(X@GRAD) should not be null.");
"Input(X@GRAD) should not be null.");
ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
}
};
} // namespace operators
} // namespace paddle
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(maxout, ops::MaxOutOp, ops::MaxOutOpMaker, maxout_grad,
ops::MaxOutOpGrad);
REGISTER_OP_CPU_KERNEL(maxout, ops::MaxOutKernel<paddle::platform::CPUPlace,
float>);
REGISTER_OP_CPU_KERNEL(maxout_grad,
ops::MaxOutGradKernel<paddle::platform::CPUPlace,
float>);
ops::MaxOutOpGrad);
REGISTER_OP_CPU_KERNEL(maxout,
ops::MaxOutKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(
maxout_grad, ops::MaxOutGradKernel<paddle::platform::CPUPlace, float>);
paddle/operators/maxout_op.cu.cc
浏览文件 @ a1bf4c28
......@@ -18,8 +18,6 @@ namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(maxout,
ops::MaxOutKernel<paddle::platform::GPUPlace, float>,
ops::MaxOutKernel<paddle::platform::GPUPlace, double>);
REGISTER_OP_GPU_KERNEL(maxout_grad,
ops::MaxOutGradKernel<paddle::platform::GPUPlace,
float>,
ops::MaxOutGradKernel<paddle::platform::GPUPlace,
double>);
REGISTER_OP_GPU_KERNEL(
maxout_grad, ops::MaxOutGradKernel<paddle::platform::GPUPlace, float>,
ops::MaxOutGradKernel<paddle::platform::GPUPlace, double>);
paddle/operators/maxout_op.h
浏览文件 @ a1bf4c28
......@@ -53,7 +53,7 @@ class MaxOutGradKernel : public framework::OpKernel<T> {
zero(device_ctx, in_x_grad, static_cast<T>(0.0));
math::MaxOutGradFunctor<Place, T> maxout_backward;
maxout_backward(context.device_context(), *in_x, in_x_grad, *out,
*out_grad, groups);
*out_grad, groups);
}
}
};
......
paddle/operators/recv_op.cc 0 → 100644
浏览文件 @ a1bf4c28
/* 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 <stdint.h>
#include <sys/stat.h>
#include <ostream>
#include <thread>
#include <unistd.h>
#include "paddle/framework/data_type.h"
#include "paddle/framework/executor.h"
#include "paddle/framework/framework.pb.h"
#include "paddle/framework/lod_tensor.h"
#include "paddle/framework/op_registry.h"
#include "paddle/operators/detail/send_recv_impl.h"
#include "paddle/operators/detail/simple_block_queue.h"
namespace paddle {
namespace operators {
void RunServer(Server **rpc_server,
std::shared_ptr<detail::SendRecvServerImpl> service,
const std::string &server_address) {
ServerBuilder builder;
builder.AddListeningPort(server_address, grpc::InsecureServerCredentials());
builder.RegisterService(service.get());
std::unique_ptr<Server> server(builder.BuildAndStart());
*rpc_server = server.get();
LOG(INFO) << "Server listening on " << server_address << std::endl;
server->Wait();
}
class RecvOp : public framework::OperatorBase {
public:
RecvOp(const std::string &type, const framework::VariableNameMap &inputs,
const framework::VariableNameMap &outputs,
const framework::AttributeMap &attrs)
: OperatorBase(type, inputs, outputs, attrs) {
if (!rpc_service_) {
rpc_service_.reset(new detail::SendRecvServerImpl());
std::string endpoint = Attr<std::string>("endpoint");
server_thread_.reset(
new std::thread(RunServer, &rpc_server_, rpc_service_, endpoint));
}
}
virtual ~RecvOp() {
rpc_server_->Shutdown();
server_thread_->join();
}
void Run(const framework::Scope &scope,
const platform::DeviceContext &dev_ctx) const override {
// blocking get one var from client.
const framework::LoDTensor &t = rpc_service_->Get();
framework::Scope &recv_scope = scope.NewScope();
// set graph input var
auto *var = recv_scope.Var(Input("RX"));
auto *tensor = var->GetMutable<framework::LoDTensor>();
// FIXME(typhoonzero): do not copy
framework::CopyFrom(t, dev_ctx.GetPlace(), dev_ctx, tensor);
auto *block = Attr<framework::BlockDescBind *>("OptimizeBlock");
auto *program = block->Program();
framework::Executor executor(dev_ctx);
// Run sub graph to get optimized tensor
executor.Run(*program, &recv_scope, block->ID(),
false /*create_local_scope*/);
auto *out_var = recv_scope.FindVar("Out");
// push back
rpc_service_->Push(out_var->Get<framework::LoDTensor>());
}
protected:
// grpc server instance to track status and gracefully shutdown.
// borrow an pointer from server thread.
Server *rpc_server_{nullptr};
// grpc send/recv service implement to register.
std::shared_ptr<detail::SendRecvServerImpl> rpc_service_;
std::shared_ptr<std::thread> server_thread_;
};
class RecvOpMaker : public framework::OpProtoAndCheckerMaker {
public:
RecvOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("RX", "(Tensor) Input tensor to be saved");
AddComment(R"DOC(
Recv operator
This operator will recv tensor from send_op
)DOC");
AddAttr<std::string>("endpoint",
"(string, default 127.0.0.1:6164)"
"IP address to listen on.")
.SetDefault("127.0.0.1:6164")
.AddCustomChecker([](const std::string &ip) { return !ip.empty(); });
AddAttr<framework::BlockDescBind *>("OptimizeBlock", "type BlockDescBind*",
"optimize network run in server");
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(recv, ops::RecvOp, ops::RecvOpMaker);
paddle/operators/roi_pool_op.cc 100755 → 100644
浏览文件 @ a1bf4c28
......@@ -43,8 +43,8 @@ class ROIPoolOp : public framework::OperatorWithKernel {
"ROIs should be a 2-D tensor of shape (num_rois, 5)"
"given as [[batch_id, x1, y1, x2, y2], …].");
PADDLE_ENFORCE(rois_dims[1] == kROISize,
"ROIs should be a 2-D tensor of shape (num_rois, 5)"
"given as [[batch_id, x1, y1, x2, y2], …].");
"ROIs should be a 2-D tensor of shape (num_rois, 5)"
"given as [[batch_id, x1, y1, x2, y2], …].");
int pooled_height = ctx->Attrs().Get<int>("pooled_height");
int pooled_width = ctx->Attrs().Get<int>("pooled_width");
......@@ -65,7 +65,7 @@ class ROIPoolOp : public framework::OperatorWithKernel {
ctx->SetOutputDim("Out", out_dims);
ctx->SetOutputDim("Argmax", out_dims);
}
}
protected:
framework::OpKernelType GetKernelType(
......@@ -100,7 +100,7 @@ class ROIPoolGradOp : public framework::OperatorWithKernel {
class ROIPoolOpMaker : public framework::OpProtoAndCheckerMaker {
public:
ROIPoolOpMaker(framework::OpProto* proto,
framework::OpAttrChecker* op_checker)
framework::OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X",
"(Tensor), "
......@@ -125,21 +125,22 @@ class ROIPoolOpMaker : public framework::OpProtoAndCheckerMaker {
"(Tensor), "
"Argmaxes corresponding to indices in X used "
"for gradient computation. Only output "
"if arg “is_test” is false.").AsIntermediate();
"if arg “is_test” is false.")
.AsIntermediate();
AddAttr<float>("spatial_scale",
"(float, default 1.0), "
"Multiplicative spatial scale factor "
"to translate ROI coords from their input scale "
"to the scale used when pooling.")
.SetDefault(1.0);
.SetDefault(1.0);
AddAttr<int>("pooled_height",
"(int, default 1), "
"The pooled output height.")
.SetDefault(1);
.SetDefault(1);
AddAttr<int>("pooled_width",
"(int, default 1), "
"The pooled output width.")
.SetDefault(1);
.SetDefault(1);
AddComment(R"DOC(
ROIPool operator
......@@ -153,11 +154,10 @@ https://stackoverflow.com/questions/43430056/what-is-roi-layer-in-fast-rcnn
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(roi_pool, ops::ROIPoolOp, ops::ROIPoolOpMaker,
roi_pool_grad, ops::ROIPoolGradOp);
REGISTER_OP(roi_pool, ops::ROIPoolOp, ops::ROIPoolOpMaker, roi_pool_grad,
ops::ROIPoolGradOp);
REGISTER_OP_CPU_KERNEL(
roi_pool,
ops::CPUROIPoolOpKernel<paddle::platform::CPUPlace, float>,
roi_pool, ops::CPUROIPoolOpKernel<paddle::platform::CPUPlace, float>,
ops::CPUROIPoolOpKernel<paddle::platform::CPUPlace, double>);
REGISTER_OP_CPU_KERNEL(
roi_pool_grad,
......
paddle/operators/roi_pool_op.cu 100755 → 100644
浏览文件 @ a1bf4c28
......@@ -29,101 +29,95 @@ static inline int NumBlocks(const int N) {
kNumMaxinumNumBlocks);
}
template <typename T>
__global__ void GPUROIPoolForward(
const int nthreads, const T* input_data, const int64_t* input_rois,
const float spatial_scale, const int channels, const int height,
const int width, const int pooled_height, const int pooled_width,
T* output_data, int64_t* argmax_data) {
int index = blockIdx.x * blockDim.x + threadIdx.x;
int offset = blockDim.x * gridDim.x;
for (size_t i = index; i < nthreads; i += offset) {
int pw = index % pooled_width;
int ph = (index / pooled_width) % pooled_height;
int c = (index / pooled_width / pooled_height) % channels;
int n = index / pooled_width / pooled_height / channels;
const int64_t* offset_input_rois = input_rois + n * kROISize;
int roi_batch_ind = offset_input_rois[0];
int roi_start_w = round(offset_input_rois[1] * spatial_scale);
int roi_start_h = round(offset_input_rois[2] * spatial_scale);
int roi_end_w = round(offset_input_rois[3] * spatial_scale);
int roi_end_h = round(offset_input_rois[4] * spatial_scale);
int roi_width = max(roi_end_w - roi_start_w + 1, 1);
int roi_height = max(roi_end_h - roi_start_h + 1, 1);
T bin_size_h = static_cast<T>(roi_height) / static_cast<T>(pooled_height);
T bin_size_w = static_cast<T>(roi_width) / static_cast<T>(pooled_width);
int hstart = static_cast<int>(floor(static_cast<T>(ph) * bin_size_h));
int wstart = static_cast<int>(floor(static_cast<T>(pw) * bin_size_w));
int hend = static_cast<int>(ceil(static_cast<T>(ph + 1) * bin_size_h));
int wend = static_cast<int>(ceil(static_cast<T>(pw + 1) * bin_size_w));
hstart = min(max(hstart + roi_start_h, 0), height);
hend = min(max(hend + roi_start_h, 0), height);
wstart = min(max(wstart + roi_start_w, 0), width);
wend = min(max(wend + roi_start_w, 0), width);
bool is_empty = (hend <= hstart) || (wend <= wstart);
T maxval = is_empty ? 0 : -std::numeric_limits<T>::max();
int maxidx = -1;
const T* offset_input_data =
input_data + (roi_batch_ind * channels + c) * height * width;
for (int h = hstart; h < hend; ++h) {
for (int w = wstart; w < wend; ++w) {
int input_data_index = h * width + w;
if (offset_input_data[input_data_index] > maxval) {
maxval = offset_input_data[input_data_index];
maxidx = input_data_index;
}
template <typename T>
__global__ void GPUROIPoolForward(const int nthreads, const T* input_data,
const int64_t* input_rois,
const float spatial_scale, const int channels,
const int height, const int width,
const int pooled_height,
const int pooled_width, T* output_data,
int64_t* argmax_data) {
int index = blockIdx.x * blockDim.x + threadIdx.x;
int offset = blockDim.x * gridDim.x;
for (size_t i = index; i < nthreads; i += offset) {
int pw = index % pooled_width;
int ph = (index / pooled_width) % pooled_height;
int c = (index / pooled_width / pooled_height) % channels;
int n = index / pooled_width / pooled_height / channels;
const int64_t* offset_input_rois = input_rois + n * kROISize;
int roi_batch_ind = offset_input_rois[0];
int roi_start_w = round(offset_input_rois[1] * spatial_scale);
int roi_start_h = round(offset_input_rois[2] * spatial_scale);
int roi_end_w = round(offset_input_rois[3] * spatial_scale);
int roi_end_h = round(offset_input_rois[4] * spatial_scale);
int roi_width = max(roi_end_w - roi_start_w + 1, 1);
int roi_height = max(roi_end_h - roi_start_h + 1, 1);
T bin_size_h = static_cast<T>(roi_height) / static_cast<T>(pooled_height);
T bin_size_w = static_cast<T>(roi_width) / static_cast<T>(pooled_width);
int hstart = static_cast<int>(floor(static_cast<T>(ph) * bin_size_h));
int wstart = static_cast<int>(floor(static_cast<T>(pw) * bin_size_w));
int hend = static_cast<int>(ceil(static_cast<T>(ph + 1) * bin_size_h));
int wend = static_cast<int>(ceil(static_cast<T>(pw + 1) * bin_size_w));
hstart = min(max(hstart + roi_start_h, 0), height);
hend = min(max(hend + roi_start_h, 0), height);
wstart = min(max(wstart + roi_start_w, 0), width);
wend = min(max(wend + roi_start_w, 0), width);
bool is_empty = (hend <= hstart) || (wend <= wstart);
T maxval = is_empty ? 0 : -std::numeric_limits<T>::max();
int maxidx = -1;
const T* offset_input_data =
input_data + (roi_batch_ind * channels + c) * height * width;
for (int h = hstart; h < hend; ++h) {
for (int w = wstart; w < wend; ++w) {
int input_data_index = h * width + w;
if (offset_input_data[input_data_index] > maxval) {
maxval = offset_input_data[input_data_index];
maxidx = input_data_index;
}
}
output_data[index] = maxval;
if (argmax_data) {
argmax_data[index] = maxidx;
}
}
output_data[index] = maxval;
if (argmax_data) {
argmax_data[index] = maxidx;
}
}
}
template <typename T>
__global__ void GPUROIPoolBackward(
const int nthreads,
const int64_t* input_rois,
const T* output_grad,
const int64_t* argmax_data,
const int num_rois,
const float spatial_scale,
const int channels,
const int height,
const int width,
const int pooled_height,
const int pooled_width,
T* input_grad) {
int index = blockIdx.x * blockDim.x + threadIdx.x;
int offset = blockDim.x * gridDim.x;
for (int i = index; i < nthreads; i += offset) {
int pw = index % pooled_width;
int ph = (index / pooled_width) % pooled_height;
int c = (index / pooled_width / pooled_height) % channels;
int n = index / pooled_width / pooled_height / channels;
const int64_t* offset_input_rois = input_rois + n * kROISize;
int roi_batch_ind = offset_input_rois[0];
int input_offset = (roi_batch_ind * channels + c) * height * width;
int output_offset = (n * channels + c) * pooled_height * pooled_width;
const T* offset_output_grad = output_grad + output_offset;
T* offset_input_grad = input_grad + input_offset;
const int64_t* offset_argmax_data = argmax_data + output_offset;
int argmax = offset_argmax_data[ph * pooled_width + pw];
if (argmax != -1) {
platform::CudaAtomicAdd(offset_input_grad + argmax,
const int nthreads, const int64_t* input_rois, const T* output_grad,
const int64_t* argmax_data, const int num_rois, const float spatial_scale,
const int channels, const int height, const int width,
const int pooled_height, const int pooled_width, T* input_grad) {
int index = blockIdx.x * blockDim.x + threadIdx.x;
int offset = blockDim.x * gridDim.x;
for (int i = index; i < nthreads; i += offset) {
int pw = index % pooled_width;
int ph = (index / pooled_width) % pooled_height;
int c = (index / pooled_width / pooled_height) % channels;
int n = index / pooled_width / pooled_height / channels;
const int64_t* offset_input_rois = input_rois + n * kROISize;
int roi_batch_ind = offset_input_rois[0];
int input_offset = (roi_batch_ind * channels + c) * height * width;
int output_offset = (n * channels + c) * pooled_height * pooled_width;
const T* offset_output_grad = output_grad + output_offset;
T* offset_input_grad = input_grad + input_offset;
const int64_t* offset_argmax_data = argmax_data + output_offset;
int argmax = offset_argmax_data[ph * pooled_width + pw];
if (argmax != -1) {
platform::CudaAtomicAdd(
offset_input_grad + argmax,
static_cast<T>(offset_output_grad[ph * pooled_width + pw]));
}
}
}
}
template <typename Place, typename T>
class GPUROIPoolOpKernel : public framework::OpKernel<T> {
......@@ -145,25 +139,18 @@ class GPUROIPoolOpKernel : public framework::OpKernel<T> {
int width = in_dims[3];
size_t rois_num = rois->dims()[0];
if (rois_num== 0) return;
if (rois_num == 0) return;
int output_size = out->numel();
int blocks = NumBlocks(output_size);
int threads = kNumCUDAThreads;
GPUROIPoolForward<T>
<<<blocks, threads, 0, ctx.cuda_device_context().stream()>>>(
output_size,
in->data<T>(),
rois->data<int64_t>(),
spatial_scale,
channels,
height,
width,
pooled_height,
pooled_width,
out->mutable_data<T>(ctx.GetPlace()),
argmax->mutable_data<int64_t>(ctx.GetPlace()));
GPUROIPoolForward<
T><<<blocks, threads, 0, ctx.cuda_device_context().stream()>>>(
output_size, in->data<T>(), rois->data<int64_t>(), spatial_scale,
channels, height, width, pooled_height, pooled_width,
out->mutable_data<T>(ctx.GetPlace()),
argmax->mutable_data<int64_t>(ctx.GetPlace()));
}
};
......@@ -175,10 +162,8 @@ class GPUROIPoolGradOpKernel : public framework::OpKernel<T> {
auto* rois = ctx.Input<Tensor>("ROIs");
auto* argmax = ctx.Input<Tensor>("Argmax");
auto* out_grad =
ctx.Input<Tensor>(framework::GradVarName("Out"));
auto* x_grad =
ctx.Output<Tensor>(framework::GradVarName("X"));
auto* out_grad = ctx.Input<Tensor>(framework::GradVarName("Out"));
auto* x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
auto pooled_height = ctx.Attr<int>("pooled_height");
auto pooled_width = ctx.Attr<int>("pooled_width");
......@@ -199,21 +184,13 @@ class GPUROIPoolGradOpKernel : public framework::OpKernel<T> {
int threads = kNumCUDAThreads;
if (output_grad_size > 0) {
GPUROIPoolBackward<T>
<<<blocks, threads, 0, ctx.cuda_device_context().stream()>>>(
output_grad_size,
rois->data<int64_t>(),
out_grad->data<T>(),
argmax->data<int64_t>(),
rois_num,
spatial_scale,
channels,
height,
width,
pooled_height,
pooled_width,
x_grad->mutable_data<T>(ctx.GetPlace()));
}
GPUROIPoolBackward<
T><<<blocks, threads, 0, ctx.cuda_device_context().stream()>>>(
output_grad_size, rois->data<int64_t>(), out_grad->data<T>(),
argmax->data<int64_t>(), rois_num, spatial_scale, channels, height,
width, pooled_height, pooled_width,
x_grad->mutable_data<T>(ctx.GetPlace()));
}
}
}
};
......@@ -223,8 +200,7 @@ class GPUROIPoolGradOpKernel : public framework::OpKernel<T> {
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(
roi_pool,
ops::GPUROIPoolOpKernel<paddle::platform::GPUPlace, float>,
roi_pool, ops::GPUROIPoolOpKernel<paddle::platform::GPUPlace, float>,
ops::GPUROIPoolOpKernel<paddle::platform::GPUPlace, double>);
REGISTER_OP_GPU_KERNEL(
roi_pool_grad,
......
paddle/operators/roi_pool_op.h 100755 → 100644
浏览文件 @ a1bf4c28
......@@ -133,54 +133,47 @@ class CPUROIPoolGradOpKernel : public framework::OpKernel<T> {
auto* in = ctx.Input<framework::Tensor>("X");
auto* rois = ctx.Input<framework::Tensor>("ROIs");
auto* argmax = ctx.Input<framework::Tensor>("Argmax");
auto* out_grad =
ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
auto* x_grad =
ctx.Output<framework::Tensor>(framework::GradVarName("X"));
auto* in_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
auto pooled_height = ctx.Attr<int>("pooled_height");
auto pooled_width = ctx.Attr<int>("pooled_width");
if (x_grad) {
int channels = in->dims()[1];
auto in_stride = framework::stride(in->dims());
auto roi_stride = framework::stride(rois->dims());
if (in_grad) {
const int64_t* rois_data = rois->data<int64_t>();
int rois_num = rois->dims()[0];
T* x_grad_data = x_grad->mutable_data<T>(ctx.GetPlace());
const T* out_grad_data = out_grad->data<T>();
const int64_t* argmax_data = argmax->data<int64_t>();
T* in_grad_data = in_grad->mutable_data<T>(ctx.GetPlace());
math::SetConstant<Place, T> set_zero;
set_zero(ctx.device_context(), x_grad, static_cast<T>(0));
set_zero(ctx.device_context(), in_grad, static_cast<T>(0));
size_t roi_offset = roi_stride[0];
size_t batch_offset = in_stride[0];
size_t channel_offset = in_stride[1];
auto in_stride = framework::stride(in->dims());
auto argmax_stride = framework::stride(argmax->dims());
auto roi_stride = framework::stride(rois->dims());
auto out_stride = framework::stride(out_grad->dims());
const T* out_grad_data = out_grad->data<T>();
size_t pool_channel_offset = pooled_height * pooled_width;
const int64_t* argmax_data = argmax->data<int64_t>();
int rois_num = rois->dims()[0];
int channels = in->dims()[1];
for (size_t n = 0; n < rois_num; ++n) {
size_t roi_batch_idx = rois_data[0];
T* batch_grad_data = x_grad_data + batch_offset * roi_batch_idx;
for (int n = 0; n < rois_num; ++n) {
int roi_batch_idx = rois_data[0];
T* batch_grad_data = in_grad_data + roi_batch_idx * in_stride[0];
for (int c = 0; c < channels; ++c) {
for (int ph = 0; ph < pooled_height; ++ph) {
for (int pw = 0; pw < pooled_width; ++pw) {
size_t pool_index = ph * pooled_width + pw;
int pool_index = ph * pooled_width + pw;
if (argmax_data[pool_index] >= 0) {
size_t index = static_cast<size_t>(argmax_data[pool_index]);
auto index = argmax_data[pool_index];
batch_grad_data[index] += out_grad_data[pool_index];
}
}
}
batch_grad_data += channel_offset;
out_grad_data += pool_channel_offset;
argmax_data += pool_channel_offset;
batch_grad_data += in_stride[1];
out_grad_data += out_stride[1];
argmax_data += argmax_stride[1];
}
rois_data += roi_offset;
rois_data += roi_stride[0];
}
}
}
......
paddle/operators/save_op.cc
浏览文件 @ a1bf4c28
......@@ -88,73 +88,7 @@ class SaveOp : public framework::OperatorBase {
"SaveOp only support LoDTensor, %s has wrong type", iname);
auto &tensor = var->Get<framework::LoDTensor>();
{ // the 1st field, uint32_t version
constexpr uint32_t version = 0;
fout.write(reinterpret_cast<const char *>(&version), sizeof(version));
}
{ // the 2nd field, tensor description
// int32_t size
// void* protobuf message
framework::TensorDesc desc;
desc.set_data_type(framework::ToDataType(tensor.type()));
auto dims = framework::vectorize(tensor.dims());
auto *pb_dims = desc.mutable_dims();
pb_dims->Resize(static_cast<int>(dims.size()), 0);
std::copy(dims.begin(), dims.end(), pb_dims->begin());
int32_t size = desc.ByteSize();
fout.write(reinterpret_cast<const char *>(&size), sizeof(size));
auto out = desc.SerializeAsString();
fout.write(out.data(), size);
}
{ // the 3rd field, tensor data
uint64_t size = tensor.memory_size();
auto *data_ptr = tensor.data<void>();
PADDLE_ENFORCE(size < std::numeric_limits<std::streamsize>::max(),
"Index overflow when writing tensor");
if (platform::is_gpu_place(tensor.place())) {
#ifdef PADDLE_WITH_CUDA
constexpr size_t kBufSize = 1024 * 1024 * 64; // 64MB
std::unique_ptr<char[]> buf(new char[kBufSize]);
auto &gpu_dev_ctx =
static_cast<const platform::CUDADeviceContext &>(dev_ctx);
platform::CPUPlace cpu;
uintptr_t data = reinterpret_cast<uintptr_t>(data_ptr);
while (size != 0) {
size_t size_to_write = std::min(kBufSize, static_cast<size_t>(size));
memory::Copy(cpu, buf.get(),
boost::get<platform::GPUPlace>(tensor.place()),
reinterpret_cast<const void *>(data), size_to_write,
gpu_dev_ctx.stream());
gpu_dev_ctx.Wait();
fout.write(buf.get(), size_to_write);
data += size_to_write;
size -= size_to_write;
}
#else
PADDLE_THROW("Unexpected branch");
#endif
} else {
fout.write(static_cast<const char *>(data_ptr),
static_cast<std::streamsize>(size));
}
}
{ // the 4th field, lod information
// uint64_t lod_level
// uint64_t lod_level_1 size in byte.
// int* lod_level_1 data
// ...
auto lod = tensor.lod();
uint64_t size = lod.size();
fout.write(reinterpret_cast<const char *>(&size), sizeof(size));
for (auto &each : lod) {
size = each.size() * sizeof(framework::LoD::value_type::value_type);
fout.write(reinterpret_cast<const char *>(&size), sizeof(size));
fout.write(reinterpret_cast<const char *>(each.data()),
static_cast<std::streamsize>(size));
}
}
framework::SerializeToStream(fout, tensor, dev_ctx);
}
};
......
paddle/operators/scale_op.cc
浏览文件 @ a1bf4c28
......@@ -77,4 +77,6 @@ REGISTER_OPERATOR(scale, ops::ScaleOp, ops::ScaleOpMaker<float>,
ops::ScaleGradMaker);
REGISTER_OP_CPU_KERNEL(scale,
ops::ScaleKernel<paddle::platform::CPUPlace, float>,
ops::ScaleKernel<paddle::platform::CPUPlace, double>);
ops::ScaleKernel<paddle::platform::CPUPlace, double>,
ops::ScaleKernel<paddle::platform::CPUPlace, int>,
ops::ScaleKernel<paddle::platform::CPUPlace, int64_t>);
paddle/operators/scale_op.cu
浏览文件 @ a1bf4c28
......@@ -16,4 +16,6 @@
REGISTER_OP_GPU_KERNEL(
scale, paddle::operators::ScaleKernel<paddle::platform::GPUPlace, float>,
paddle::operators::ScaleKernel<paddle::platform::GPUPlace, double>);
paddle::operators::ScaleKernel<paddle::platform::GPUPlace, double>,
paddle::operators::ScaleKernel<paddle::platform::GPUPlace, int>,
paddle::operators::ScaleKernel<paddle::platform::GPUPlace, int64_t>);
paddle/operators/send_op.cc 0 → 100644
浏览文件 @ a1bf4c28
/* 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 <ostream>
#include "paddle/framework/data_type.h"
#include "paddle/framework/framework.pb.h"
#include "paddle/framework/lod_tensor.h"
#include "paddle/framework/op_registry.h"
#include "paddle/operators/detail/send_recv_impl.h"
#include "paddle/operators/detail/simple_block_queue.h"
namespace paddle {
namespace operators {
// TODO(typhoonzero): this is a simple implementation which only send
// one tensor
class SendOp : public framework::OperatorBase {
public:
SendOp(const std::string &type, const framework::VariableNameMap &inputs,
const framework::VariableNameMap &outputs,
const framework::AttributeMap &attrs)
: OperatorBase(type, inputs, outputs, attrs) {
// init client when the operator is created at runtime.
if (!client_) {
std::string endpoint = Attr<std::string>("endpoint");
client_.reset(new detail::RPCClient(
grpc::CreateChannel(endpoint, grpc::InsecureChannelCredentials())));
// TODO(typhoonzero): how to call InitVariables
}
}
void Run(const framework::Scope &scope,
const platform::DeviceContext &dev_ctx) const override {
auto iname = Input("X");
auto oname = Output("Out");
// TODO(typhoonzero): currently it's non-blocking,
// should block until server responds.
bool ret = client_->SendVariable(scope, iname, oname);
if (!ret) {
LOG(ERROR) << "send variable error";
}
}
protected:
std::shared_ptr<detail::RPCClient> client_{nullptr};
};
class SendOpMaker : public framework::OpProtoAndCheckerMaker {
public:
SendOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "(Tensor) Input tensor to be saved");
AddOutput("Out", "(Tensor) Output fetched from server");
AddComment(R"DOC(
Recv operator
This operator will recv tensor from send_op
)DOC");
AddAttr<std::string>("endpoint",
"(string, default 127.0.0.1:6164)"
"IP address to listen on.")
.SetDefault("127.0.0.1:6164")
.AddCustomChecker([](const std::string &ip) { return !ip.empty(); });
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(send, ops::SendOp, ops::SendOpMaker);
paddle/operators/send_recv_op_test.cc 0 → 100644
浏览文件 @ a1bf4c28
/* 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. */
// TODO(typhoonzero): add python bindings for this test as
// a RemoteOptimizer.
#include <unistd.h>
#include <thread>
#include "gtest/gtest.h"
#include "paddle/framework/op_registry.h"
#include "paddle/framework/operator.h"
#include "paddle/framework/program_desc.h"
USE_NO_KERNEL_OP(send);
USE_NO_KERNEL_OP(recv);
USE_OP(sum);
// global for simplicity.
std::unique_ptr<paddle::framework::OperatorBase> recv_op;
void InitTensorsInScope(paddle::framework::Scope &scope,
paddle::platform::CPUPlace &place) {
paddle::platform::CPUDeviceContext ctx(place);
auto var = scope.Var("X");
auto tensor = var->GetMutable<paddle::framework::LoDTensor>();
tensor->Resize({10, 10});
float *expect = tensor->mutable_data<float>(place);
for (int64_t i = 0; i < tensor->numel(); ++i) {
expect[i] = static_cast<float>(i);
}
auto out_var = scope.Var("Out");
auto out_tensor = out_var->GetMutable<paddle::framework::LoDTensor>();
out_tensor->Resize({10, 10});
tensor->mutable_data<float>(place); // allocate
}
void AddOp(const std::string &type,
const paddle::framework::VariableNameMap &inputs,
const paddle::framework::VariableNameMap &outputs,
paddle::framework::AttributeMap attrs,
paddle::framework::BlockDescBind *block) {
// insert output
for (auto kv : outputs) {
for (auto v : kv.second) {
auto var = block->Var(v);
var->SetDataType(paddle::framework::DataType::FP32);
}
}
// insert op
auto op = block->AppendOp();
op->SetType(type);
for (auto &kv : inputs) {
op->SetInput(kv.first, kv.second);
}
for (auto &kv : outputs) {
op->SetOutput(kv.first, kv.second);
}
op->SetAttrMap(attrs);
}
void StartServerNet() {
paddle::framework::Scope scope;
paddle::platform::CPUPlace place;
InitTensorsInScope(scope, place);
// sub program run in recv_op, for simple test we use sum
paddle::framework::ProgramDescBind program;
paddle::framework::BlockDescBind *block = program.MutableBlock(0);
// X for server side tensors, RX for received tensers, must be of same shape.
AddOp("sum", {{"X", {"X", "RX"}}}, {{"Out", {"Out"}}}, {}, block);
paddle::framework::AttributeMap attrs;
attrs.insert({"endpoint", std::string("127.0.0.1:6174")});
attrs.insert({"OptimizeBlock", block});
recv_op = paddle::framework::OpRegistry::CreateOp("recv", {{"RX", {"RX"}}},
{{"Out", {"Out"}}}, attrs);
paddle::platform::CPUDeviceContext ctx(place);
recv_op->Run(scope, ctx);
}
TEST(SendRecvOp, CPU) {
std::thread server_thread(StartServerNet);
sleep(5); // wait server to start
// local net
paddle::framework::Scope scope;
paddle::platform::CPUPlace place;
InitTensorsInScope(scope, place);
paddle::framework::AttributeMap attrs;
attrs.insert({"endpoint", std::string("127.0.0.1:6174")});
auto send_op = paddle::framework::OpRegistry::CreateOp(
"send", {{"X", {"X"}}}, {{"Out", {"Out"}}}, attrs);
paddle::platform::CPUDeviceContext ctx(place);
send_op->Run(scope, ctx);
auto in_var = scope.Var("X");
auto tensor = in_var->GetMutable<paddle::framework::LoDTensor>();
float *expected = tensor->data<float>();
auto out_var = scope.Var("Out");
auto target = out_var->GetMutable<paddle::framework::LoDTensor>();
// send fail cause output is none.
EXPECT_NE(target->memory_size(), size_t(0));
float *actual = target->data<float>();
for (int64_t i = 0; i < target->numel(); ++i) {
EXPECT_EQ(expected[i] * 2, actual[i]);
}
recv_op.reset(); // dtor can shutdown and join server thread.
server_thread.join();
}
paddle/operators/sequence_slice_op.cc 100755 → 100644
浏览文件 @ a1bf4c28
......@@ -45,7 +45,7 @@ class SequenceSliceOp : public framework::OperatorWithKernel {
// Initialize the output's dims to maximum,
// and re-set to real dims by the value of Offset and Length at kernel
ctx->SetOutputDim("Out", input_dims);
}
}
protected:
framework::OpKernelType GetKernelType(
......@@ -93,8 +93,7 @@ class SequenceSliceOpMaker : public framework::OpProtoAndCheckerMaker {
"(Tensor), "
"a vector<int> to describe the length of every input sequence for "
"sub sequence item.");
AddOutput("Out",
"(LoDTensor), the output of SequenceSliceOp.");
AddOutput("Out", "(LoDTensor), the output of SequenceSliceOp.");
AddComment(R"DOC(
Sequence slice operator
......
paddle/operators/sgd_op.cc
浏览文件 @ a1bf4c28
......@@ -55,7 +55,7 @@ SGD operator
This operator implements one step of the stochastic gradient descent algorithm.
$$param_out = param - learning_rate * grad$$
$$param\_out = param - learning\_rate * grad$$
)DOC");
}
......
paddle/operators/shrink_rnn_memory_op.cc
浏览文件 @ a1bf4c28
......@@ -57,11 +57,21 @@ class ShrinkRNNMemoryOpProtoMaker : public framework::OpProtoAndCheckerMaker {
ShrinkRNNMemoryOpProtoMaker(framework::OpProto *proto,
framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "");
AddInput("RankTable", "");
AddInput("I", "");
AddOutput("Out", "");
AddComment("");
AddInput("X", "(LoDTensor) The RNN step memory to be shrinked.");
AddInput("RankTable", "(LoDRankTable) The lod_rank_table of dynamic RNN.");
AddInput("I",
"(LoDTensor) The step index. The RNN step memory 'X' will be "
"shrinked to match the size of the input of the index'th step.");
AddOutput("Out", "(LoDTensor) The shrinked RNN step memory.");
AddComment(
R"DOC(
In dynamic RNN, we are able to handle sequences of different lengths.
Because of the multiple lengths, the size of each step input can be
different, which may lead to a mismatching between the input of
the current step and the memory generated by the previous one. This
operator shrinks memory according to the size of the next step input,
to make sure that they can match each other.
)DOC");
}
};
......
paddle/scripts/docker/build.sh
浏览文件 @ a1bf4c28
......@@ -16,11 +16,13 @@ function cmake_gen() {
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}
PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27m/bin/python
-DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27m/include/python2.7
-DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs2/lib/libpython2.7.so"
elif [ "$1" == "cp27-cp27mu" ]; then
export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs4/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs2/lib:}
export PATH=/opt/python/cp27-cp27mu/bin/:${PATH}
PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27mu/bin/python
-DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27mu/include/python2.7
-DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs4/lib/libpython2.7.so"
......
paddle/scripts/travis/build_doc.sh
浏览文件 @ a1bf4c28
......@@ -11,8 +11,9 @@ make -j `nproc` gen_proto_py
make -j `nproc` paddle_docs paddle_docs_cn
# check websites for broken links
linkchecker doc/en/html/index.html
linkchecker doc/cn/html/index.html
# It will be failed now!
#linkchecker doc/en/html/index.html
#linkchecker doc/cn/html/index.html
# Parse Github URL
REPO=`git config remote.origin.url`
......
proto/ModelConfig.proto
浏览文件 @ a1bf4c28
......@@ -544,6 +544,9 @@ message LayerConfig {
// for batch normalization layer
// The small constant added to the variance to improve numeric stability.
optional double epsilon = 60 [ default = 0.00001 ];
// for factorization machine layer
optional uint32 factor_size = 61;
}
message EvaluatorConfig {
......
python/paddle/trainer/config_parser.py
浏览文件 @ a1bf4c28
......@@ -2400,6 +2400,14 @@ class CropLayer(LayerBase):
image_conf.img_size_y = input_layer.height
image_conf.channels = input_layer.size / (input_layer.width *
input_layer.height)
# only support for 4-dims inputs and NCHW order
if (len(self.config.inputs) == 2):
self.set_layer_height_width(
self.get_input_layer(1).height, self.get_input_layer(1).width)
self.set_layer_size(self.get_input_layer(1).size)
else:
self.set_layer_height_width(shape[-2], shape[-1])
self.set_layer_size(reduce(lambda x, y: x * y, shape[1:]))
@config_layer('batch_norm')
......@@ -3849,6 +3857,26 @@ class SwitchOrderLayer(LayerBase):
name, 'switch_order', 0, inputs=inputs, **xargs)
self.config.reshape_conf.height_axis.extend(reshape['height'])
self.config.reshape_conf.width_axis.extend(reshape['width'])
input_layer = self.get_input_layer(0)
if reshape is None:
self.set_layer_size(input_layer.size)
else:
in_h = input_layer.height
in_w = input_layer.width
out_dims = None
if input_layer.has_depth():
in_d = input_layer.depth
in_c = input_layer.size / in_h / in_w / in_d
# batch_size, depth, height, width, channel
out_dims = [0, in_d, in_h, in_w, in_c]
else:
in_c = input_layer.size / in_h / in_w
# batch_size, height, width, channel
out_dims = [0, in_h, in_w, in_c]
# Because (reshape['width'][0] > 0) always be true.
# So out_dims[0] won't be used.
size = reduce(lambda x, y: x * y, out_dims[reshape['width'][0]:])
self.set_layer_size(size)
@config_layer('scale_sub_region')
......@@ -3870,6 +3898,21 @@ class ScaleSubRegionLayer(LayerBase):
image_conf.channels)
@config_layer('factorization_machine')
class FactorizationMachineLayer(LayerBase):
def __init__(self, name, inputs, factor_size, **xargs):
super(FactorizationMachineLayer, self).__init__(
name, 'factorization_machine', size=1, inputs=inputs, **xargs)
config_assert(
len(self.inputs) == 1,
'factorization machine layer must have one and only one input.')
self.config.factor_size = factor_size
input_layer = self.get_input_layer(0)
psize = input_layer.size * factor_size
dims = [input_layer.size, factor_size]
self.create_input_parameter(0, psize, dims)
# Deprecated, use a new layer specific class instead
@config_func
def Layer(name, type, **xargs):
......
python/paddle/trainer_config_helpers/tests/configs/file_list.sh
浏览文件 @ a1bf4c28
......@@ -11,6 +11,7 @@ test_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_l
test_kmax_seq_socre_layer test_sub_nested_seq_select_layer test_scale_shift_layer
test_seq_slice_layer test_cross_entropy_over_beam test_roi_pool_layer test_pooling3D_layer
test_conv3d_layer test_deconv3d_layer test_BatchNorm3D test_resize_layer
test_scale_sub_region_layer test_dot_prod_layer test_l2_distance_layer)
test_scale_sub_region_layer test_dot_prod_layer test_l2_distance_layer
test_factorization_machine)
export whole_configs=(test_split_datasource)
python/paddle/trainer_config_helpers/tests/configs/protostr/test_factorization_machine.protostr 0 → 100644
浏览文件 @ a1bf4c28
type: "nn"
layers {
name: "data"
type: "data"
size: 1024
active_type: ""
}
layers {
name: "__factorization_machine_0__"
type: "factorization_machine"
size: 1
active_type: ""
inputs {
input_layer_name: "data"
input_parameter_name: "___factorization_machine_0__.w0"
}
factor_size: 10
}
parameters {
name: "___factorization_machine_0__.w0"
size: 10240
initial_mean: 0.0
initial_std: 0.03125
dims: 1024
dims: 10
initial_strategy: 0
initial_smart: true
}
input_layer_names: "data"
output_layer_names: "__factorization_machine_0__"
sub_models {
name: "root"
layer_names: "data"
layer_names: "__factorization_machine_0__"
input_layer_names: "data"
output_layer_names: "__factorization_machine_0__"
is_recurrent_layer_group: false
}
python/paddle/trainer_config_helpers/tests/configs/test_factorization_machine.py 0 → 100644
浏览文件 @ a1bf4c28
from paddle.trainer_config_helpers import *
data = data_layer(name='data', size=1024)
fm = factorization_machine(input=data, factor_size=10)
outputs(fm)
python/paddle/v2/__init__.py
浏览文件 @ a1bf4c28
......@@ -83,11 +83,10 @@ def set_omp_mkl_env_vars(trainer_count):
'''Get the number of physical cores'''
if platform.system() == "Linux":
num_sockets = int(
os.popen("lscpu |grep \"Socket\" |awk -F':' '{print $2}'|xargs")
os.popen("grep 'physical id' /proc/cpuinfo | sort -u | wc -l")
.read())
num_cores_per_socket = int(
os.popen(
"lscpu |grep \"per socket\" |awk -F':' '{print $2}'|xargs")
os.popen("grep 'core id' /proc/cpuinfo | sort -u | wc -l")
.read())
return num_sockets * num_cores_per_socket
else:
......
python/paddle/v2/dataset/uci_housing.py
浏览文件 @ a1bf4c28
......@@ -38,6 +38,7 @@ UCI_TEST_DATA = None
URL_MODEL = 'https://github.com/PaddlePaddle/book/raw/develop/01.fit_a_line/fit_a_line.tar'
MD5_MODEL = '52fc3da8ef3937822fcdd87ee05c0c9b'
def feature_range(maximums, minimums):
import matplotlib
matplotlib.use('Agg')
......@@ -114,7 +115,8 @@ def test():
def model():
tar_file = paddle.v2.dataset.common.download(URL_MODEL, 'fit_a_line.tar', MD5_MODEL)
tar_file = paddle.v2.dataset.common.download(URL_MODEL, 'fit_a_line.tar',
MD5_MODEL)
with open(tar_file, 'r') as f:
parameters = Parameters.from_tar(f)
return parameters
......
python/paddle/v2/fluid/evaluator.py
浏览文件 @ a1bf4c28
......@@ -26,9 +26,9 @@ class Evaluator(object):
name(str): The name of evaluator. such as, "accuracy". Used for generate
temporary variable name.
main_program(Program, optional): The evaluator should be added to this
main_program. Default g_main_program
main_program. Default default_main_program()
startup_program(Program, optional):The parameter should be added to this
startup_program. Default g_startup_program
startup_program. Default default_startup_program()
Attributes:
states(list): The list of state variables. states will be reset to zero
......
python/paddle/v2/fluid/executor.py
浏览文件 @ a1bf4c28
import numpy as np
from . import core
from framework import Program, g_main_program
from framework import Program, default_main_program
__all__ = ['Executor', 'g_scope']
......@@ -103,7 +103,7 @@ class Executor(object):
fetch_list = []
if program is None:
program = g_main_program
program = default_main_program()
if not isinstance(program, Program):
raise TypeError()
......
python/paddle/v2/fluid/framework.py
浏览文件 @ a1bf4c28
......@@ -6,7 +6,7 @@ import proto.framework_pb2 as framework_pb2
__all__ = [
'Block', 'Variable', 'Program', 'Operator', 'default_startup_program',
'default_main_program', 'g_startup_program', 'g_main_program'
'default_main_program'
]
......@@ -395,7 +395,11 @@ class Block(object):
return v
def all_parameters(self):
return {v for k, v in self.vars.iteritems() if isinstance(v, Parameter)}
return list(self.iter_parameters())
def iter_parameters(self):
return (item[1] for item in self.vars.iteritems()
if isinstance(item[1], Parameter))
def create_var(self, *args, **kwargs):
var = Variable(self, *args, **kwargs)
......@@ -469,6 +473,37 @@ class Block(object):
for index in range(len(self.ops)):
assert self.ops[index].desc == ops_in_cpp[index]
def copy_param_info_from(self, other):
"""
Copy the information of parameters from other block
Args:
other(Block): other block
Returns:
None
"""
if not isinstance(other, Block):
raise TypeError("copy_param_info_from should be invoked with Block")
for p in other.iter_parameters():
assert isinstance(p, Parameter)
v = self.vars.get(p.name, None)
if v is None:
raise ValueError("copy_param_info_from should be invoked with "
"same topology")
assert isinstance(v, Variable)
new_p = Parameter(
block=self,
shape=v.shape,
dtype=v.dtype,
type=v.type,
lod_level=v.lod_level,
stop_gradient=p.stop_gradient,
trainable=p.trainable,
optimize_attr=p.optimize_attr,
regularizer=p.regularizer,
name=v.name)
self.vars[new_p.name] = new_p
class Program(object):
def __init__(self):
......@@ -489,6 +524,7 @@ class Program(object):
p.desc = core.ProgramDesc(self.desc)
p.blocks = [Block(p, i) for i in xrange(self.desc.num_blocks())]
p.sync_with_cpp()
p.copy_param_info_from(self)
return p
def prune(self, targets):
......@@ -572,6 +608,24 @@ class Program(object):
for block in self.blocks:
block.sync_with_cpp()
def copy_param_info_from(self, other):
"""
Copy the information of parameters from other program.
Args:
other(Program): Other program
Returns:
None
"""
if not isinstance(other, Program):
raise TypeError("copy_param_info_from should be invoked with "
"Program")
if len(self.blocks) != len(other.blocks):
raise ValueError("copy_param_info_from should be invoked with two "
"program, with represent the same topology")
self.global_block().copy_param_info_from(other.global_block())
def list_vars(self):
for each_block in self.blocks:
for each_var in each_block.vars.itervalues():
......@@ -600,13 +654,13 @@ class Parameter(Variable):
# program is a global instance.
g_main_program = Program()
g_startup_program = Program()
_main_program_ = Program()
_startup_program_ = Program()
def default_startup_program():
return g_startup_program
return _startup_program_
def default_main_program():
return g_main_program
return _main_program_
python/paddle/v2/fluid/io.py
浏览文件 @ a1bf4c28
import os
import cPickle as pickle
from paddle.v2.fluid.framework import Program, Parameter, g_main_program, \
Variable
from paddle.v2.fluid.framework import Program, Parameter, default_main_program, Variable
__all__ = [
'save_vars', 'save_params', 'save_persistables', 'load_vars', 'load_params',
......@@ -46,7 +45,7 @@ def save_vars(executor, dirname, main_program=None, vars=None, predicate=None):
"""
if vars is None:
if main_program is None:
main_program = g_main_program
main_program = default_main_program()
if not isinstance(main_program, Program):
raise TypeError("program should be as Program type or None")
......@@ -98,7 +97,7 @@ def load_vars(executor, dirname, main_program=None, vars=None, predicate=None):
:param executor: executor that save variable
:param dirname: directory path
:param main_program: program. If vars is None, then filter all variables in this
program which fit `predicate`. Default g_program.
program which fit `predicate`. Default default_main_program().
:param predicate: The Predicate describes a callable that returns a variable
as a bool. If it returns true, the variables will be loaded.
:param vars: variables need to be loaded. If specify vars, program &
......@@ -107,7 +106,7 @@ def load_vars(executor, dirname, main_program=None, vars=None, predicate=None):
"""
if vars is None:
if main_program is None:
main_program = g_main_program
main_program = default_main_program()
if not isinstance(main_program, Program):
raise TypeError("program's type should be Program")
......@@ -154,7 +153,7 @@ def load_persistables(executor, dirname, main_program=None):
def get_inference_program(target_vars, main_program=None):
if main_program is None:
main_program = g_main_program
main_program = default_main_program()
if not isinstance(target_vars, list):
target_vars = [target_vars]
......@@ -177,12 +176,12 @@ def save_inference_model(dirname,
:param target_vars: Variables from which we can get inference results.
:param executor: executor that save inference model
:param main_program: original program, which will be pruned to build the inference model.
Default g_main_program.
Default default_main_program().
:return: None
"""
if main_program is None:
main_program = g_main_program
main_program = default_main_program()
if not isinstance(target_vars, list):
target_vars = [target_vars]
......@@ -272,10 +271,10 @@ def get_parameter_value_by_name(name, executor, program=None):
:param executor: executor for retrieving the value
:param name: the name of the parameter
:param program: the program where the variable is found
Default g_main_program.
Default default_main_program().
:return: the LoDTensor for the variable
"""
if program is None:
program = g_main_program
program = default_main_program()
var = program.global_block().var(name)
return get_parameter_value(var, executor)
python/paddle/v2/fluid/layer_helper.py
浏览文件 @ a1bf4c28
import copy
import itertools
from framework import Variable, g_main_program, \
g_startup_program, unique_name, dtype_is_floating
from framework import Variable, default_main_program, default_startup_program, unique_name, dtype_is_floating
from paddle.v2.fluid.initializer import Constant, Xavier
......@@ -22,7 +21,7 @@ class LayerHelper(object):
def main_program(self):
prog = self.kwargs.get('main_program', None)
if prog is None:
return g_main_program
return default_main_program()
else:
return prog
......@@ -30,7 +29,7 @@ class LayerHelper(object):
def startup_program(self):
prog = self.kwargs.get('startup_program', None)
if prog is None:
return g_startup_program
return default_startup_program()
else:
return prog
......
python/paddle/v2/fluid/layers.py
浏览文件 @ a1bf4c28
from . import core
import core
import proto.framework_pb2 as framework_pb2
from framework import OpProtoHolder, Variable, Program, Operator
from initializer import Constant, Normal, Xavier
from initializer import Constant, Normal, Xavier, Initializer
from paddle.v2.fluid.layer_helper import LayerHelper, unique_name
import re
import cStringIO
......@@ -1587,6 +1587,97 @@ def array_length(array, main_program=None):
return tmp
def conv2d_transpose(input,
num_filters,
output_size=None,
filter_size=None,
padding=None,
stride=None,
param_attr=None,
param_initializer=None,
main_program=None,
startup_program=None):
"""
The transpose of conv2d layer.
This layer is also known as deconvolution layer.
Args:
input(Variable): The input image with [N, C, H, W] format.
num_filters(int): The number of filter. It is as same as the output
image channel.
output_size(int|tuple|None): The output image size. If output size is a
tuple, it must contain two integers, (image_H, image_W). This
parameter only works when filter_size is None.
filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
it must contain two integers, (filter_size_H, filter_size_W).
Otherwise, the filter will be a square. None if use output size to
calculate filter_size
padding(int|tuple): The padding size. If padding is a tuple, it must
contain two integers, (padding_H, padding_W). Otherwise, the
padding_H = padding_W = padding.
stride(int|tuple): The stride size. If stride is a tuple, it must
contain two integers, (stride_H, stride_W). Otherwise, the
stride_H = stride_W = stride.
param_attr: Parameter Attribute.
param_initializer(Initializer): Parameter Initializer. Default is Xavier
main_program(Program): the main program
startup_program(Program): the startup program
Returns:
Variable: Output image.
"""
helper = LayerHelper("conv2d_transpose", **locals())
if not isinstance(input, Variable):
raise TypeError("Input of conv2d_transpose must be Variable")
input_channel = input.shape[1]
op_attr = dict()
if isinstance(padding, int):
op_attr['paddings'] = [padding, padding]
elif padding is not None:
op_attr['paddings'] = padding
if isinstance(stride, int):
op_attr['strides'] = stride
elif stride is not None:
op_attr['strides'] = stride
if filter_size is None:
if output_size is None:
raise ValueError("output_size must be set when filter_size is None")
if isinstance(output_size, int):
output_size = [output_size, output_size]
padding = op_attr.get('paddings', [0, 0])
stride = op_attr.get('strides', [1, 1])
h_in = input.shape[2]
w_in = input.shape[3]
filter_size_h = output_size[0] - (h_in - 1) * stride[0] + 2 * padding[0]
filter_size_w = output_size[1] - (w_in - 1) * stride[1] + 2 * padding[1]
filter_size = [filter_size_h, filter_size_w]
elif isinstance(filter_size, int):
filter_size = [filter_size, filter_size]
filter_shape = [input_channel, num_filters] + filter_size
img_filter = helper.create_parameter(
dtype=input.dtype,
shape=filter_shape,
attr=helper.param_attr,
initializer=param_initializer)
out = helper.create_tmp_variable(dtype=input.dtype)
helper.append_op(
type='conv2d_transpose',
inputs={'Input': [input],
'Filter': [img_filter]},
outputs={'Output': out},
attrs=op_attr)
return out
class ConditionalBlockGuard(BlockGuard):
def __init__(self, block):
if not isinstance(block, ConditionalBlock):
......
python/paddle/v2/fluid/tests/book/CMakeLists.txt
浏览文件 @ a1bf4c28
file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
list(REMOVE_ITEM TEST_OPS test_image_classification_train)
py_test(test_image_classification_train_resnet SRCS test_image_classification_train.py ARGS resnet)
py_test(test_image_classification_train_vgg SRCS test_image_classification_train.py ARGS vgg)
# default test
foreach(src ${TEST_OPS})
py_test(${src} SRCS ${src}.py)
endforeach()
python/paddle/v2/fluid/tests/book/test_image_classification_train.py
浏览文件 @ a1bf4c28
from __future__ import print_function
import numpy as np
import paddle.v2 as paddle
import paddle.v2.fluid as fluid
import sys
def resnet_cifar10(input, depth=32):
......@@ -67,8 +69,7 @@ def vgg16_bn_drop(input):
drop = fluid.layers.dropout(x=conv5, dropout_prob=0.5)
fc1 = fluid.layers.fc(input=drop, size=512, act=None)
reshape1 = fluid.layers.reshape(x=fc1, shape=list(fc1.shape + (1, 1)))
bn = fluid.layers.batch_norm(input=reshape1, act='relu')
bn = fluid.layers.batch_norm(input=fc1, act='relu')
drop2 = fluid.layers.dropout(x=bn, dropout_prob=0.5)
fc2 = fluid.layers.fc(input=drop2, size=512, act=None)
return fc2
......@@ -80,11 +81,18 @@ data_shape = [3, 32, 32]
images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
# Add neural network config
# option 1. resnet
# net = resnet_cifar10(images, 32)
# option 2. vgg
net = vgg16_bn_drop(images)
net_type = "vgg"
if len(sys.argv) >= 2:
net_type = sys.argv[1]
if net_type == "vgg":
print("train vgg net")
net = vgg16_bn_drop(images)
elif net_type == "resnet":
print("train resnet")
net = resnet_cifar10(images, 32)
else:
raise ValueError("%s network is not supported" % net_type)
predict = fluid.layers.fc(input=net, size=classdim, act='softmax')
cost = fluid.layers.cross_entropy(input=predict, label=label)
......
python/paddle/v2/fluid/tests/book/test_recognize_digits_mlp.py
浏览文件 @ a1bf4c28
......@@ -35,6 +35,13 @@ opts = optimizer.minimize(avg_cost)
accuracy = fluid.evaluator.Accuracy(input=predict, label=label)
inference_program = fluid.default_main_program().clone()
test_accuracy = fluid.evaluator.Accuracy(
input=predict, label=label, main_program=inference_program)
test_target = [avg_cost] + test_accuracy.metrics + test_accuracy.states
inference_program = fluid.io.get_inference_program(
test_target, main_program=inference_program)
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.mnist.train(), buf_size=8192),
......@@ -69,11 +76,6 @@ for pass_id in range(PASS_NUM):
acc = np.array(outs[1])
pass_acc = accuracy.eval(exe)
test_accuracy = fluid.evaluator.Accuracy(input=predict, label=label)
test_target = [avg_cost] + test_accuracy.metrics + test_accuracy.states
inference_program = fluid.io.get_inference_program(test_target)
test_accuracy.reset(exe)
for data in test_reader():
x_data = np.array(map(lambda x: x[0], data)).astype("float32")
......
python/paddle/v2/fluid/tests/test_array_read_write_op.py
浏览文件 @ a1bf4c28
......@@ -3,7 +3,7 @@ import paddle.v2.fluid.core as core
import paddle.v2.fluid.layers as layers
from paddle.v2.fluid.executor import Executor
from paddle.v2.fluid.backward import append_backward_ops
from paddle.v2.fluid.framework import g_main_program
from paddle.v2.fluid.framework import default_main_program
import numpy
......@@ -66,7 +66,7 @@ class TestArrayReadWrite(unittest.TestCase):
append_backward_ops(total_sum_scaled)
g_vars = map(g_main_program.global_block().var,
g_vars = map(default_main_program().global_block().var,
[each_x.name + "@GRAD" for each_x in x])
g_out = [
item.sum()
......
python/paddle/v2/fluid/tests/test_batch_norm_op.py
浏览文件 @ a1bf4c28
......@@ -21,6 +21,13 @@ def get_backward_op(scope, op, no_grad_set):
def _reference_training(x, scale, offset, epsilon, data_format):
x_shape = x.shape
if len(x_shape) == 2:
if data_format == "NCHW":
x = np.reshape(x, (x.shape[0], x.shape[1], 1, 1))
else:
x = np.reshape(x, (x.shape[0], 1, 1, x.shape[1]))
if data_format == "NCHW":
n, c, h, w = x.shape
x_square = x * x
......@@ -39,6 +46,8 @@ def _reference_training(x, scale, offset, epsilon, data_format):
offset_tile = np.reshape(offset, (1, c, 1, 1))
offset_tile = np.reshape(offset_tile, (1, c, 1, 1))
y = normalized * scale_tile + offset_tile
if len(x_shape) == 2:
y = np.reshape(y, (y.shape[0], y.shape[1]))
return y, mean, var
elif data_format == "NHWC":
x_square = x * x
......@@ -48,7 +57,10 @@ def _reference_training(x, scale, offset, epsilon, data_format):
mean = x_sum / element_count
var = x_square_sum / element_count - mean * mean
normalized = (x - mean) / np.sqrt(var + epsilon)
return (normalized * scale + offset), mean, var
y = normalized * scale + offset
if len(x_shape) == 2:
y = np.reshape(y, x_shape)
return y, mean, var
else:
raise ValueError("Unknown data order.")
......@@ -65,6 +77,18 @@ def _reference_grad(x, grad_y, scale, mean, var, epsilon, data_format):
# (x - mean) * sum(grad_y * (x - mean)) / (var + epsilon))
# transfer from (N, C, H, W) to (N, H, W, C) to simplify computation
x_shape = x.shape
if len(x_shape) == 2:
if data_format == "NCHW":
x = np.reshape(x, (x.shape[0], x.shape[1], 1, 1))
grad_y = np.reshape(grad_y,
(grad_y.shape[0], grad_y.shape[1], 1, 1))
else:
x = np.reshape(x, (x.shape[0], 1, 1, x.shape[1]))
grad_y = np.reshape(grad_y,
(grad_y.shape[0], 1, 1, grad_y.shape[1]))
if data_format == "NCHW":
x = np.transpose(x, (0, 2, 3, 1))
grad_y = np.transpose(grad_y, (0, 2, 3, 1))
......@@ -83,6 +107,9 @@ def _reference_grad(x, grad_y, scale, mean, var, epsilon, data_format):
grad_x = np.transpose(grad_x, (0, 3, 1, 2))
x = np.transpose(x, (0, 3, 1, 2))
grad_y = np.transpose(grad_y, (0, 3, 1, 2))
if len(x_shape) == 2:
grad_x = np.reshape(grad_x, x_shape)
return grad_x, grad_scale, grad_offset
......@@ -127,7 +154,7 @@ class TestBatchNormOp(OpTest):
momentum = 0.9
# N, H, W, C: 2, 3, 4, 2
n, h, w, c = 2, 3, 4, 2
n, h, w, c = 2, 3, 4, 5
x_shape = [n, h, w, c]
scale_shape = [c]
......@@ -184,20 +211,23 @@ class TestBatchNormOp(OpTest):
print 'python: NHWC, NCHW, backward checking passed'
def test_forward_backward(self):
def test_with_place(place, tensor_format):
def test_with_place(place, tensor_format, shape):
# attr
epsilon = 0.00001
momentum = 0.9
# N, H, W, C: 12, 3, 4, 2
n, h, w, c = 2, 3, 4, 2
if data_format == "NHWC":
x_shape = [n, h, w, c]
elif data_format == "NCHW":
x_shape = [n, c, h, w]
if len(shape) == 2:
x_shape = shape
c = shape[1]
else:
raise ValueError("Unknown data type.")
# n, h, w, c = 2, 3, 4, 2
n, h, w, c = shape[0], shape[1], shape[2], shape[3]
if data_format == "NHWC":
x_shape = [n, h, w, c]
elif data_format == "NCHW":
x_shape = [n, c, h, w]
else:
raise ValueError("Unknown data type.")
scale_shape = [c]
x_val = np.random.random_sample(x_shape).astype(np.float32)
......@@ -219,7 +249,10 @@ class TestBatchNormOp(OpTest):
# for gradient test
# y_grad = np.ones(x_shape).astype(np.float32)
y_grad = np.zeros(x_shape).astype(np.float32)
y_grad[0, 0, 0, 0] = 1.
if len(y_grad.shape) == 2:
y_grad[0, 0] = 1.
else:
y_grad[0, 0, 0, 0] = 1.
# y_grad = np.random.random_sample(x_shape).astype(np.float32)
x_grad_ref, scale_grad_ref, bias_grad_ref = _reference_grad(
x_val, y_grad, scale_val, saved_mean, var_ref, epsilon,
......@@ -313,7 +346,8 @@ class TestBatchNormOp(OpTest):
places.append(core.GPUPlace(0))
for place in places:
for data_format in ["NCHW", "NHWC"]:
test_with_place(place, data_format)
test_with_place(place, data_format, [2, 3, 4, 5])
test_with_place(place, data_format, [2, 3])
if __name__ == '__main__':
......
python/paddle/v2/fluid/tests/test_conditional_block.py
浏览文件 @ a1bf4c28
import unittest
import paddle.v2.fluid.layers as layers
import paddle.v2.fluid.core as core
from paddle.v2.fluid.framework import g_startup_program, g_main_program
from paddle.v2.fluid.framework import default_startup_program, default_main_program
from paddle.v2.fluid.executor import Executor
from paddle.v2.fluid.backward import append_backward_ops
import numpy
......@@ -19,7 +19,7 @@ class ConditionalBlock(unittest.TestCase):
cpu = core.CPUPlace()
exe = Executor(cpu)
exe.run(g_startup_program)
exe.run(default_startup_program())
x = numpy.random.random(size=(10, 1)).astype('float32')
......@@ -29,7 +29,9 @@ class ConditionalBlock(unittest.TestCase):
append_backward_ops(loss=loss)
outs = exe.run(
feed={'X': x},
fetch_list=[g_main_program.block(0).var(data.name + "@GRAD")])[0]
fetch_list=[
default_main_program().block(0).var(data.name + "@GRAD")
])[0]
print outs
......
python/paddle/v2/fluid/tests/test_executor_and_mul.py
浏览文件 @ a1bf4c28
import unittest
from paddle.v2.fluid.layers import mul, data, sequence_pool
import numpy
import paddle.v2.fluid.core as core
from paddle.v2.fluid.executor import Executor
from paddle.v2.fluid.framework import g_main_program
import numpy
from paddle.v2.fluid.layers import mul, data
class TestExecutor(unittest.TestCase):
......@@ -19,10 +20,7 @@ class TestExecutor(unittest.TestCase):
a_np = numpy.random.random((100, 784)).astype('float32')
b_np = numpy.random.random((784, 100)).astype('float32')
exe = Executor(place)
outs = exe.run(g_main_program,
feed={'a': a_np,
'b': b_np},
fetch_list=[out])
outs = exe.run(feed={'a': a_np, 'b': b_np}, fetch_list=[out])
out = outs[0]
self.assertEqual((100, 100), out.shape)
self.assertTrue(numpy.allclose(out, numpy.dot(a_np, b_np)))
......
python/paddle/v2/fluid/tests/test_image_classification_layer.py
浏览文件 @ a1bf4c28
import unittest
import paddle.v2.fluid.layers as layers
import paddle.v2.fluid as fluid
import paddle.v2.fluid.nets as nets
from paddle.v2.fluid.framework import Program
......@@ -29,27 +29,35 @@ class TestLayer(unittest.TestCase):
def test_batch_norm_layer(self):
main_program = Program()
startup_program = Program()
images = layers.data(
images = fluid.layers.data(
name='pixel',
shape=[3, 48, 48],
dtype='float32',
main_program=main_program)
layers.batch_norm(
hidden1 = fluid.layers.batch_norm(
input=images,
main_program=main_program,
startup_program=startup_program)
hidden2 = fluid.layers.fc(input=hidden1,
size=128,
act='relu',
main_program=main_program)
hidden3 = fluid.layers.batch_norm(
input=hidden2,
main_program=main_program,
startup_program=startup_program)
# print str(main_program)
print str(main_program)
def test_dropout_layer(self):
main_program = Program()
startup_program = Program()
images = layers.data(
images = fluid.layers.data(
name='pixel',
shape=[3, 48, 48],
dtype='float32',
main_program=main_program)
layers.dropout(
fluid.layers.dropout(
x=images,
dropout_prob=0.5,
main_program=main_program,
......@@ -61,7 +69,7 @@ class TestLayer(unittest.TestCase):
main_program = Program()
startup_program = Program()
images = layers.data(
images = fluid.layers.data(
name='pixel',
shape=[3, 48, 48],
dtype='float32',
......@@ -77,19 +85,19 @@ class TestLayer(unittest.TestCase):
def test_elementwise_add_with_act(self):
main_program = Program()
startup_program = Program()
image1 = layers.data(
image1 = fluid.layers.data(
name='pixel1',
shape=[3, 48, 48],
dtype='float32',
main_program=main_program,
startup_program=startup_program)
image2 = layers.data(
image2 = fluid.layers.data(
name='pixel2',
shape=[3, 48, 48],
dtype='float32',
main_program=main_program,
startup_program=startup_program)
out = layers.elementwise_add(
out = fluid.layers.elementwise_add(
x=image1,
y=image2,
act='relu',
......
python/paddle/v2/fluid/tests/test_layers.py
浏览文件 @ a1bf4c28
......@@ -65,6 +65,15 @@ class TestBook(unittest.TestCase):
print str(program)
def test_conv2d_transpose(self):
program = Program()
kwargs = {'main_program': program}
img = layers.data(
name='pixel', shape=[3, 2, 2], dtype='float32', **kwargs)
layers.conv2d_transpose(
input=img, num_filters=10, output_size=28, **kwargs)
print str(program)
def test_recognize_digits_conv(self):
program = Program()
......
python/paddle/v2/fluid/tests/test_lod_rank_table.py
浏览文件 @ a1bf4c28
from paddle.v2.fluid.layers import lod_rank_table, data
from paddle.v2.fluid.executor import Executor
from paddle.v2.fluid.framework import g_main_program
import paddle.v2.fluid.core as core
import numpy
import unittest
......@@ -18,7 +17,7 @@ class TestLoDRankTable(unittest.TestCase):
tensor = core.LoDTensor()
tensor.set(numpy.random.random(size=(17, 100)), cpu)
tensor.set_lod([[0, 1, 3], [0, 5, 6, 7], [0, 3, 4, 9, 10, 13, 16, 17]])
exe.run(g_main_program, scope=scope, feed={'x': tensor})
exe.run(scope=scope, feed={'x': tensor})
var = scope.find_var(rank_table.name)
table = var.get_lod_rank_table()
self.assertEqual([(0, 5), (1, 1), (2, 1)], table.items())
......
python/paddle/v2/fluid/tests/test_log_loss_op.py 0 → 100644
浏览文件 @ a1bf4c28
import unittest
import numpy as np
from op_test import OpTest
class TestLogLossOp(OpTest):
def setUp(self):
self.op_type = 'log_loss'
samples_num = 32
predicted = np.random.uniform(0.1, 1.0,
(samples_num, 1)).astype("float32")
labels = np.random.randint(0, 2, (samples_num, 1)).astype("float32")
epsilon = 1e-4
self.inputs = {
'Predicted': predicted,
'Labels': labels,
}
self.attrs = {'epsilon': epsilon}
loss = -labels * np.log(predicted + epsilon) - (
1 - labels) * np.log(1 - predicted + epsilon)
self.outputs = {'Loss': loss}
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(['Predicted'], 'Loss', max_relative_error=0.03)
if __name__ == '__main__':
unittest.main()
python/paddle/v2/fluid/tests/test_maxout_op.py
浏览文件 @ a1bf4c28
......@@ -30,9 +30,7 @@ class TestMaxOutOp(OpTest):
def init_test_case(self):
self.MaxOut_forward_naive = maxout_forward_naive
self.shape = [100, 6, 2, 2]
self.groups=2
self.groups = 2
if __name__ == '__main__':
......
python/paddle/v2/fluid/tests/test_operator_desc.py
浏览文件 @ a1bf4c28
import unittest
from paddle.v2.fluid.framework import Variable, Program, g_main_program
import paddle.v2.fluid.core as core
from paddle.v2.fluid.framework import Program, default_startup_program
main_program = default_startup_program()
class TestOperator(unittest.TestCase):
def test_error_type(self):
block = g_main_program.create_block()
block = main_program.create_block()
try:
block.append_op()
self.assertFail()
......
python/paddle/v2/fluid/tests/test_parameter.py
浏览文件 @ a1bf4c28
import unittest
from paddle.v2.fluid.framework import g_main_program
from paddle.v2.fluid.framework import default_main_program
import paddle.v2.fluid.core as core
from paddle.v2.fluid.executor import Executor
import paddle.v2.fluid.io as io
from paddle.v2.fluid.initializer import ConstantInitializer
import numpy as np
main_program = default_main_program()
class TestParameter(unittest.TestCase):
def test_param(self):
shape = [784, 100]
val = 1.0625
b = g_main_program.global_block()
b = main_program.global_block()
param = b.create_parameter(
name='fc.w',
shape=shape,
......@@ -23,9 +25,9 @@ class TestParameter(unittest.TestCase):
self.assertEqual(core.DataType.FP32, param.dtype)
self.assertEqual(0, param.block.idx)
exe = Executor(core.CPUPlace())
p = exe.run(g_main_program, fetch_list=[param])[0]
p = exe.run(main_program, fetch_list=[param])[0]
self.assertTrue(np.allclose(p, np.ones(shape) * val))
p = io.get_parameter_value_by_name('fc.w', exe, g_main_program)
p = io.get_parameter_value_by_name('fc.w', exe, main_program)
self.assertTrue(np.allclose(np.array(p), np.ones(shape) * val))
......
python/paddle/v2/fluid/tests/test_profiler.py
浏览文件 @ a1bf4c28
......@@ -18,7 +18,7 @@ class TestProfiler(unittest.TestCase):
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
with profiler.cuda_profiler('cuda_profiler.txt', 'kvp') as nvprof:
with profiler.cuda_profiler("cuda_profiler.txt", 'csv') as nvprof:
for i in range(epoc):
input = np.random.random(dshape).astype('float32')
exe.run(fluid.default_main_program(), feed={'data': input})
......
python/paddle/v2/fluid/tests/test_program.py
浏览文件 @ a1bf4c28
from __future__ import print_function
import unittest
from paddle.v2.fluid.framework import Program
from paddle.v2.fluid.framework import g_main_program
from paddle.v2.fluid.framework import Program, default_main_program
import paddle.v2.fluid.layers as layers
main_program = default_main_program()
class TestProgram(unittest.TestCase):
def test_program(self):
b = g_main_program.current_block()
b = main_program.current_block()
self.assertEqual(-1, b.parent_idx)
self.assertEqual(0, b.idx)
b = g_main_program.create_block()
b = main_program.create_block()
self.assertEqual(1, b.idx)
self.assertEqual(0, b.parent_idx)
b = g_main_program.create_block()
b = main_program.create_block()
self.assertEqual(2, b.idx)
self.assertEqual(1, b.parent_idx)
g_main_program.rollback()
main_program.rollback()
b = g_main_program.current_block()
b = main_program.current_block()
self.assertEqual(1, b.idx)
self.assertEqual(0, b.parent_idx)
b = g_main_program.create_block()
b = main_program.create_block()
self.assertEqual(3, b.idx)
self.assertEqual(1, b.parent_idx)
g_main_program.rollback()
b = g_main_program.current_block()
main_program.rollback()
b = main_program.current_block()
self.assertEqual(1, b.idx)
self.assertEqual(0, b.parent_idx)
......@@ -48,8 +51,8 @@ class TestProgram(unittest.TestCase):
# FIXME(yuyang18): We manual compare the output string, since the order
# of variable could be changed.
print prog
print prog.clone()
print(prog)
print(prog.clone())
def test_parse_program_from_string(self):
prog = Program()
......@@ -67,8 +70,8 @@ class TestProgram(unittest.TestCase):
binary_str = prog.desc.serialize_to_string()
prog_restored = Program.parse_from_string(binary_str)
print prog
print prog_restored
print(prog)
print(prog_restored)
def test_append_backward(self):
prog = Program()
......@@ -123,6 +126,20 @@ class TestProgram(unittest.TestCase):
actual_ops.append(op.type)
self.assertEqual(actual_ops, expect_ops)
def test_program_clone_with_parameter(self):
main_program = Program()
startup_program = Program()
kwargs = {
'main_program': main_program,
'startup_program': startup_program
}
d = layers.data(name='x', shape=[784], dtype='float32', **kwargs)
hidden = layers.fc(input=d, size=100, **kwargs)
layers.fc(input=hidden, size=100, **kwargs)
new_program = main_program.clone()
self.assertNotEqual(0, len(new_program.blocks[0].all_parameters()))
if __name__ == '__main__':
unittest.main()
python/paddle/v2/fluid/tests/test_roi_pool_op.py
浏览文件 @ a1bf4c28
......@@ -4,24 +4,22 @@ import math
import sys
from op_test import OpTest
class TestROIPoolOp(OpTest):
def set_data(self):
self.init_test_case()
self.make_rois()
self.calc_roi_pool()
self.inputs = {
'X': self.x,
'ROIs': self.rois}
self.inputs = {'X': self.x, 'ROIs': self.rois}
self.attrs = {
'spatial_scale': self.spatial_scale,
'pooled_height': self.pooled_height,
'pooled_width': self.pooled_width}
'pooled_width': self.pooled_width
}
self.outputs = {
'Out': self.outs,
'Argmax': self.argmaxes}
self.outputs = {'Out': self.outs, 'Argmax': self.argmaxes}
def init_test_case(self):
self.batch_size = 5
......@@ -30,10 +28,9 @@ class TestROIPoolOp(OpTest):
self.width = 4
# n, c, h, w
self.x_dim = (self.batch_size, self.channels,
self.height, self.width)
self.x_dim = (self.batch_size, self.channels, self.height, self.width)
self.spatial_scale = 1.0/4.0
self.spatial_scale = 1.0 / 4.0
self.pooled_height = 2
self.pooled_width = 2
self.rois_num = 2
......@@ -41,13 +38,11 @@ class TestROIPoolOp(OpTest):
self.x = np.random.random(self.x_dim).astype('float32')
def calc_roi_pool(self):
out_data = np.zeros(
(self.rois_num, self.channels,
self.pooled_height, self.pooled_width))
argmax_data = np.zeros(
(self.rois_num, self.channels,
self.pooled_height, self.pooled_width))
out_data = np.zeros((self.rois_num, self.channels, self.pooled_height,
self.pooled_width))
argmax_data = np.zeros((self.rois_num, self.channels,
self.pooled_height, self.pooled_width))
for i in range(self.rois_num):
roi = self.rois[i]
roi_batch_id = roi[0]
......@@ -56,8 +51,8 @@ class TestROIPoolOp(OpTest):
roi_end_w = int(round(roi[3] * self.spatial_scale))
roi_end_h = int(round(roi[4] * self.spatial_scale))
roi_height = int(max(roi_end_h - roi_start_h + 1, 1));
roi_width = int(max(roi_end_w - roi_start_w + 1, 1));
roi_height = int(max(roi_end_h - roi_start_h + 1, 1))
roi_width = int(max(roi_end_w - roi_start_w + 1, 1))
x_i = self.x[roi_batch_id]
......@@ -84,7 +79,7 @@ class TestROIPoolOp(OpTest):
out_data[i, c, ph, pw] = -sys.float_info.max
argmax_data[i, c, ph, pw] = -1
for h in range(hstart, hend):
for w in range(wstart, wend):
if x_i[c, h, w] > out_data[i, c, ph, pw]:
......@@ -104,11 +99,11 @@ class TestROIPoolOp(OpTest):
y1 = np.random.random_integers(
0, self.height / self.spatial_scale - self.pooled_height)
x2 = np.random.random_integers(
x1 + self.pooled_width, self.width / self.spatial_scale)
y2 = np.random.random_integers(
y1 + self.pooled_height, self.height / self.spatial_scale)
x2 = np.random.random_integers(x1 + self.pooled_width,
self.width / self.spatial_scale)
y2 = np.random.random_integers(y1 + self.pooled_height,
self.height / self.spatial_scale)
roi = [batch_ids[i], x1, y1, x2, y2]
rois.append(roi)
self.rois = np.array(rois).astype("int64")
......@@ -123,5 +118,6 @@ class TestROIPoolOp(OpTest):
def test_check_grad(self):
self.check_grad(['X'], 'Out')
if __name__ == '__main__':
unittest.main()
python/paddle/v2/fluid/tests/test_shrink_rnn_memory.py
浏览文件 @ a1bf4c28
......@@ -3,9 +3,11 @@ import paddle.v2.fluid.core as core
from paddle.v2.fluid.executor import Executor
import paddle.v2.fluid.layers as layers
from paddle.v2.fluid.backward import append_backward_ops
from paddle.v2.fluid.framework import g_main_program
from paddle.v2.fluid.framework import default_main_program
import numpy
main_program = default_main_program()
class TestShrinkRNNMemory(unittest.TestCase):
def test_shrink_rnn_memory(self):
......@@ -36,7 +38,7 @@ class TestShrinkRNNMemory(unittest.TestCase):
append_backward_ops(loss=mem3_mean)
x_grad = exe.run(
feed={'x': tensor},
fetch_list=[g_main_program.global_block().var('x@GRAD')])[0]
fetch_list=[main_program.global_block().var('x@GRAD')])[0]
self.assertAlmostEqual(1.0, x_grad.sum(), delta=0.1)
......
python/paddle/v2/fluid/tests/test_variable.py
浏览文件 @ a1bf4c28
import unittest
from paddle.v2.fluid.framework import g_main_program, Program, convert_np_dtype_to_dtype_
from paddle.v2.fluid.framework import default_main_program, Program, convert_np_dtype_to_dtype_
import paddle.v2.fluid.core as core
import numpy as np
......@@ -18,7 +18,7 @@ class TestVariable(unittest.TestCase):
self.assertRaises(ValueError, lambda: convert("int8"))
def test_var(self):
b = g_main_program.current_block()
b = default_main_program().current_block()
w = b.create_var(
dtype="float64", shape=[784, 100], lod_level=0, name="fc.w")
self.assertNotEqual(str(w), "")
......
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