Skip to content

P
Paddle

PaddlePaddle / Paddle
接近 2 年 前同步成功

通知 2323
Star 20933
Fork 5424
P
Paddle
提交 ee760d1c 编写于 作者: C chenweihang
浏览文件
操作

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

上级 0cef33a4 f7fd711e
展开全部 隐藏空白更改
内联 并排
Showing with 63812 addition and 52622 deletion

要显示的变更太多。

To preserve performance only 1000 of 1000+ files are displayed.
文本差异 电子邮件补丁
CMakeLists.txt
浏览文件 @ ee760d1c
......@@ -61,8 +61,11 @@ option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen" OFF)
option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" OFF)
option(WITH_FAST_BUNDLE_TEST "Bundle tests that can be run in a single process together to reduce launch overhead" OFF)
option(WITH_CONTRIB "Compile the third-party contributation" OFF)
option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF)
option(WITH_ANAKIN "Compile with Anakin library" OFF)
option(WITH_GRPC "Use grpc as the default rpc framework" ${WITH_DISTRIBUTE})
option(WITH_BRPC_RDMA "Use brpc rdma as the rpc protocal" OFF)
option(WITH_SYSTEM_BLAS "Use system blas library" OFF)
# CMAKE_BUILD_TYPE
if(NOT CMAKE_BUILD_TYPE)
......@@ -131,6 +134,10 @@ if (NOT DEFINED WITH_MKLDNN)
set(WITH_MKLDNN OFF)
endif()
endif()
if (REPLACE_ENFORCE_GLOG)
add_definitions("-DREPLACE_ENFORCE_GLOG")
endif()
########################################################################################
include(external/mklml) # download mklml package
......@@ -153,12 +160,24 @@ include(external/cares)
if(WITH_DISTRIBUTE)
if(WITH_GRPC)
include(external/grpc)
message(STATUS "Use grpc framework.")
else()
message(STATUS "Use brpc framework.")
include(external/leveldb)
include(external/brpc)
endif()
endif()
if(WITH_BRPC_RDMA)
message(STATUS "Use brpc with rdma.")
if(WITH_GRPC)
message(FATAL_ERROR "Can't use grpc with brpc rdma.")
endif()
if(NOT WITH_DISTRIBUTE)
message(FATAL_ERROR "Can't use brpc rdma in no distribute env.")
endif()
endif()
include(external/snappy) # download snappy
include(external/snappystream)
include(external/threadpool)
......@@ -178,7 +197,7 @@ include(inference_lib) # add paddle fluid inference libraries
include_directories("${PADDLE_SOURCE_DIR}")
include_directories("${PADDLE_SOURCE_DIR}/paddle/cuda/include")
include_directories("${PADDLE_SOURCE_DIR}/paddle/legacy/cuda/include")
include_directories("${CMAKE_CURRENT_BINARY_DIR}/proto")
include_directories("${CMAKE_CURRENT_BINARY_DIR}/go/pserver/client/c")
......@@ -222,7 +241,7 @@ add_subdirectory(proto)
if(NOT MOBILE_INFERENCE AND NOT WITH_FLUID_ONLY)
# "add_subdirectory(go)" should be placed after the following loine,
# because it depends on paddle/optimizer.
add_subdirectory(paddle/optimizer)
add_subdirectory(paddle/legacy/optimizer)
endif()
# "add_subdirectory(paddle)" and "add_subdirectory(python)" should be
......
CONTRIBUTING.md
浏览文件 @ ee760d1c
......@@ -159,4 +159,4 @@ This will enable VLOG messages generated by `buddy_allocator.{h,cc}` and in the
- verbose level 1: [framework](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework)
- verbose level 3: [operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)
- verbose level 5: [memory](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory), [platform](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/platform)
- verbose level 7: [math](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/math)
- verbose level 7: [math](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/math)
README.md
浏览文件 @ ee760d1c
......@@ -4,7 +4,6 @@
[![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/index_cn.html)
[![Coverage Status](https://coveralls.io/repos/github/PaddlePaddle/Paddle/badge.svg?branch=develop)](https://coveralls.io/github/PaddlePaddle/Paddle?branch=develop)
[![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
[![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
......
benchmark/fluid/args.py
浏览文件 @ ee760d1c
......@@ -125,6 +125,10 @@ def parse_args():
parser.add_argument(
'--use_inference_transpiler',
action='store_true',
help='If set, uses inference transpiler to optimize the program.')
help='If set, use inference transpiler to optimize the program.')
parser.add_argument(
'--no_random',
action='store_true',
help='If set, keep the random seed and do not shuffle the data.')
args = parser.parse_args()
return args
benchmark/fluid/fluid_benchmark.py 100755 → 100644
浏览文件 @ ee760d1c
......@@ -132,10 +132,6 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
exe.run(startup_prog)
# Use inference_transpiler to speedup
if args.use_inference_transpiler:
t = fluid.InferenceTranspiler()
t.transpile(infer_prog, place)
if not args.use_reader_op:
feed_var_list = [
var for var in train_prog.global_block().vars.itervalues()
......@@ -186,6 +182,10 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
print("Pass: %d, Loss: %f" % (pass_id, np.mean(train_losses))),
# evaluation
if not args.no_test and batch_acc and not args.use_reader_op:
if args.use_inference_transpiler:
t = fluid.InferenceTranspiler()
t.transpile(infer_prog, place)
pass_test_acc = test(exe, infer_prog, test_reader, feeder,
batch_acc)
print(", Test Accuracy: %f" % pass_test_acc)
......@@ -316,6 +316,8 @@ def main():
args = parse_args()
print_arguments(args)
print_paddle_envs()
if args.no_random:
fluid.default_startup_program().random_seed = 1
# the unique trainer id, starting from 0, needed by trainer
# only
......
benchmark/fluid/models/resnet.py
浏览文件 @ ee760d1c
......@@ -197,12 +197,12 @@ def get_model(args):
optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
batched_train_reader = paddle.batch(
paddle.reader.shuffle(
train_reader if args.no_random else paddle.reader.shuffle(
train_reader, buf_size=5120),
batch_size=args.batch_size * args.gpus,
drop_last=True)
batched_test_reader = paddle.batch(
train_reader, batch_size=args.batch_size, drop_last=True)
test_reader, batch_size=args.batch_size, drop_last=True)
return avg_cost, inference_program, optimizer, batched_train_reader,\
batched_test_reader, batch_acc
cmake/cblas.cmake
浏览文件 @ ee760d1c
......@@ -83,18 +83,20 @@ else()
set(REFERENCE_CBLAS_LIB_SEARCH_PATHS ${REFERENCE_CBLAS_ROOT}/lib)
endif()
find_path(REFERENCE_CBLAS_INCLUDE_DIR NAMES cblas.h PATHS
if(WITH_SYSTEM_BLAS)
find_path(REFERENCE_CBLAS_INCLUDE_DIR NAMES cblas.h PATHS
${REFERENCE_CBLAS_INCLUDE_SEARCH_PATHS})
find_library(REFERENCE_CBLAS_LIBRARY NAMES cblas PATHS
find_library(REFERENCE_CBLAS_LIBRARY NAMES cblas PATHS
${REFERENCE_CBLAS_LIB_SEARCH_PATHS})
if(REFERENCE_CBLAS_INCLUDE_DIR AND REFERENCE_CBLAS_LIBRARY)
set(CBLAS_FOUND ON)
set(CBLAS_PROVIDER REFERENCE)
set(CBLAS_INC_DIR ${REFERENCE_CBLAS_INCLUDE_DIR})
set(CBLAS_LIBRARIES ${REFERENCE_CBLAS_LIBRARY})
add_definitions(-DPADDLE_USE_REFERENCE_CBLAS)
message(STATUS "Found reference-cblas (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
if(REFERENCE_CBLAS_INCLUDE_DIR AND REFERENCE_CBLAS_LIBRARY)
set(CBLAS_FOUND ON)
set(CBLAS_PROVIDER REFERENCE)
set(CBLAS_INC_DIR ${REFERENCE_CBLAS_INCLUDE_DIR})
set(CBLAS_LIBRARIES ${REFERENCE_CBLAS_LIBRARY})
add_definitions(-DPADDLE_USE_REFERENCE_CBLAS)
message(STATUS "Found reference-cblas (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
endif()
endif()
if(IOS_USE_VECLIB_FOR_BLAS AND VECLIB_FOUND)
......
cmake/configure.cmake
浏览文件 @ ee760d1c
......@@ -174,3 +174,7 @@ endif(WITH_GOLANG)
if(WITH_GRPC)
add_definitions(-DPADDLE_WITH_GRPC)
endif(WITH_GRPC)
if(WITH_BRPC_RDMA)
add_definitions(-DPADDLE_WITH_BRPC_RDMA)
endif(WITH_BRPC_RDMA)
cmake/external/brpc.cmake
浏览文件 @ ee760d1c
......@@ -14,6 +14,15 @@
INCLUDE(ExternalProject)
find_library(SSL_LIBRARY NAMES ssl)
ADD_LIBRARY(ssl SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET ssl PROPERTY IMPORTED_LOCATION ${SSL_LIBRARY})
find_library(CRYPTO_LIBRARY NAMES crypto)
ADD_LIBRARY(crypto SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET crypto PROPERTY IMPORTED_LOCATION ${CRYPTO_LIBRARY})
SET(BRPC_SOURCES_DIR ${THIRD_PARTY_PATH}/brpc)
SET(BRPC_INSTALL_DIR ${THIRD_PARTY_PATH}/install/brpc)
SET(BRPC_INCLUDE_DIR "${BRPC_INSTALL_DIR}/include" CACHE PATH "brpc include directory." FORCE)
......@@ -22,14 +31,14 @@ SET(BRPC_LIBRARIES "${BRPC_INSTALL_DIR}/lib/libbrpc.a" CACHE FILEPATH "brpc libr
INCLUDE_DIRECTORIES(${BRPC_INCLUDE_DIR})
# Reference https://stackoverflow.com/questions/45414507/pass-a-list-of-prefix-paths-to-externalproject-add-in-cmake-args
set(prefix_path "${THIRD_PARTY_PATH}/install/gflags|${THIRD_PARTY_PATH}/install/leveldb|${THIRD_PARTY_PATH}/install/snappy|${THIRD_PARTY_PATH}/install/gtest|${THIRD_PARTY_PATH}/install/protobuf")
set(prefix_path "${THIRD_PARTY_PATH}/install/gflags|${THIRD_PARTY_PATH}/install/leveldb|${THIRD_PARTY_PATH}/install/snappy|${THIRD_PARTY_PATH}/install/gtest|${THIRD_PARTY_PATH}/install/protobuf|${THIRD_PARTY_PATH}/install/zlib")
# If minimal .a is need, you can set WITH_DEBUG_SYMBOLS=OFF
ExternalProject_Add(
extern_brpc
${EXTERNAL_PROJECT_LOG_ARGS}
GIT_REPOSITORY "https://github.com/brpc/brpc"
GIT_TAG "6d153dd7ff00f960ae6895c9c5fff0ce9f07aff2"
GIT_REPOSITORY "https://github.com/gongweibao/brpc"
GIT_TAG "7dc04defad1fd4173aae170c3fcbde131b65155a"
PREFIX ${BRPC_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
......@@ -42,6 +51,8 @@ ExternalProject_Add(
-DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
-DCMAKE_PREFIX_PATH=${prefix_path}
-DBRPC_WITH_GLOG=ON
-DIOBUF_WITH_HUGE_BLOCK=ON
-DBRPC_WITH_RDMA=${WITH_BRPC_RDMA}
${EXTERNAL_OPTIONAL_ARGS}
LIST_SEPARATOR |
CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${BRPC_INSTALL_DIR}
......@@ -49,7 +60,7 @@ ExternalProject_Add(
-DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
-DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
)
ADD_DEPENDENCIES(extern_brpc protobuf leveldb gflags glog gtest snappy)
ADD_DEPENDENCIES(extern_brpc protobuf ssl crypto leveldb gflags glog gtest snappy)
ADD_LIBRARY(brpc STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET brpc PROPERTY IMPORTED_LOCATION ${BRPC_LIBRARIES})
ADD_DEPENDENCIES(brpc extern_brpc)
......
cmake/generic.cmake
浏览文件 @ ee760d1c
......@@ -96,6 +96,20 @@ if(NOT APPLE AND NOT ANDROID)
set(CMAKE_CXX_LINK_EXECUTABLE "${CMAKE_CXX_LINK_EXECUTABLE} -pthread -ldl -lrt")
endif(NOT APPLE AND NOT ANDROID)
set_property(GLOBAL PROPERTY FLUID_MODULES "")
# find all fluid modules is used for paddle fluid static library
# for building inference libs
function(find_fluid_modules TARGET_NAME)
get_filename_component(__target_path ${TARGET_NAME} ABSOLUTE)
string(REGEX REPLACE "^${PADDLE_SOURCE_DIR}/" "" __target_path ${__target_path})
string(FIND "${__target_path}" "fluid" pos)
if(pos GREATER 1)
get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
set(fluid_modules ${fluid_modules} ${TARGET_NAME})
set_property(GLOBAL PROPERTY FLUID_MODULES "${fluid_modules}")
endif()
endfunction(find_fluid_modules)
function(merge_static_libs TARGET_NAME)
set(libs ${ARGN})
list(REMOVE_DUPLICATES libs)
......@@ -250,6 +264,7 @@ function(cc_test TARGET_NAME)
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
if (${cc_test_SERIAL})
set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
endif()
endif()
endfunction(cc_test)
......@@ -314,6 +329,7 @@ function(nv_test TARGET_NAME)
add_test(${TARGET_NAME} ${TARGET_NAME})
if (nv_test_SERIAL)
set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
endif()
endif()
endfunction(nv_test)
......@@ -561,7 +577,7 @@ function(py_test TARGET_NAME)
set(multiValueArgs SRCS DEPS ARGS ENVS)
cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
add_test(NAME ${TARGET_NAME}
COMMAND env PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
COMMAND env FLAGS_init_allocated_mem=true PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
endif()
......
cmake/inference_lib.cmake
浏览文件 @ ee760d1c
......@@ -12,19 +12,6 @@
# See the License for the specific language governing permissions and
# limitations under the License.
set_property(GLOBAL PROPERTY FLUID_MODULES "")
# find all fluid modules is used for paddle fluid static library
function(find_fluid_modules TARGET_NAME)
get_filename_component(__target_path ${TARGET_NAME} ABSOLUTE)
string(REGEX REPLACE "^${PADDLE_SOURCE_DIR}/" "" __target_path ${__target_path})
string(FIND "${__target_path}" "fluid" pos)
if(pos GREATER 1)
get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
set(fluid_modules ${fluid_modules} ${TARGET_NAME})
set_property(GLOBAL PROPERTY FLUID_MODULES "${fluid_modules}")
endif()
endfunction(find_fluid_modules)
# make package for paddle fluid shared and static library
function(copy TARGET)
set(options "")
......@@ -154,7 +141,7 @@ set(inference_deps paddle_fluid_shared paddle_fluid)
if(WITH_CONTRIB)
message(STATUS "installing contrib")
set(contrib_dst_dir "${FLUID_INSTALL_DIR}/contrib/inference")
if (WITH_ANAKIN)
if (WITH_ANAKIN AND WITH_GPU)
copy(contrib_anakin_inference_lib DEPS paddle_inference_api inference_anakin_api
SRCS
${PADDLE_BINARY_DIR}/paddle/contrib/inference/libinference_anakin_api* # compiled anakin api
......@@ -163,9 +150,9 @@ if(WITH_CONTRIB)
list(APPEND inference_deps contrib_anakin_inference_lib)
endif()
copy(contrib_inference_lib DEPS paddle_inference_api
copy(contrib_inference_lib DEPS paddle_inference_api paddle_inference_api_shared
SRCS ${PADDLE_SOURCE_DIR}/paddle/contrib/inference/paddle_inference_api.h
${PADDLE_BINARY_DIR}/paddle/contrib/inference/libpaddle_inference_api.*
${PADDLE_BINARY_DIR}/paddle/contrib/inference/libpaddle_inference_api*
DSTS ${contrib_dst_dir} ${contrib_dst_dir})
list(APPEND inference_deps contrib_inference_lib)
endif()
......
doc/fluid/api/layers.rst
浏览文件 @ ee760d1c
......@@ -1468,6 +1468,14 @@ argmax
.. autofunction:: paddle.fluid.layers.argmax
:noindex:
.. _api_fluid_layers_argsort:
argsort
-------
.. autofunction:: paddle.fluid.layers.argsort
:noindex:
.. _api_fluid_layers_ones:
ones
......
doc/fluid/design/multi_devices/kernel_selection.md
浏览文件 @ ee760d1c
......@@ -74,10 +74,10 @@ void OperatorWithKernel::Run(
auto kernel_type_for_var = this->GetKernelTypeForVar(...);
if (kernel_type_for_var.place_ != expected_kernel_key.place_) {
auto* trans_var = new_scope.Var(var_name);
auto* out = DataTransform(expected_kernel_key,
auto* out = TransformData(expected_kernel_key,
kernel_type_for_var,
*tensor_in);
CopyVariableWithTensor(...);
SetTensorToVariable(...);
}
}
......
doc/v2/design/interface/00.why_plain_c.md
浏览文件 @ ee760d1c
......@@ -65,7 +65,7 @@ paddle_error paddle_matrix_get_shape(paddle_matrix matrix,
而在CPP里面实现这个C的接口，文件 `paddle_matrix.cpp`
```cpp
#include "paddle/math/matrix.h"
#include "paddle/legacy/math/matrix.h"
extern "C"
paddle_error paddle_matrix_shape(paddle_matrix matrix,
uint64_t *width,
......
doc/v2/dev/new_layer_cn.rst
浏览文件 @ ee760d1c
......@@ -58,7 +58,7 @@ PaddlePaddle的base layer类可以自动计算上面的导数。
实现C++类
===================
一个网络层的C++类需要实现初始化，前向和后向。全连接层的实现位于:code:`paddle/gserver/layers/FullyConnectedLayer.h`及:code:`paddle/gserver/layers/FullyConnectedLayer.cpp`。这里我们展示一份简化过的代码。
一个网络层的C++类需要实现初始化，前向和后向。全连接层的实现位于:code:`paddle/legacy/gserver/layers/FullyConnectedLayer.h`及:code:`paddle/legacy/gserver/layers/FullyConnectedLayer.cpp`。这里我们展示一份简化过的代码。
这个类需要继承 :code:`paddle::Layer` 这个基类，并且需要重写基类中的以下几个虚函数：
......@@ -153,7 +153,7 @@ PaddlePaddle的base layer类可以自动计算上面的导数。
- 每个层在其 :code:`forward` 函数的开头必须调用 :code:`Layer::forward(passType);` 。
- 之后使用 :code:`reserveOutput(batchSize, size);` 为输出分配内存。由于我们支持训练数据有不同的批次大小，所以这一步是必要的。 :code:`reserveOutput` 会相应地改变输出的尺寸。为了保证效率，如果需要扩大矩阵，我们会重新分配内存；如果需要缩减矩阵，我们会继续使用现有的内存块。
- 之后使用矩阵运算函数来计算 :math:`\sum_i W_i x + b`。:code:`getInput(i).value` 返回第i个输入矩阵。每个输入都是一个 :math:`batchSize \times dim` 的矩阵，每行表示一个批次中的单个输入。对于我们支持的全部矩阵操作，请参考 :code:`paddle/math/Matrix.h`和:code:`paddle/math/BaseMatrix.h` 。
- 之后使用矩阵运算函数来计算 :math:`\sum_i W_i x + b`。:code:`getInput(i).value` 返回第i个输入矩阵。每个输入都是一个 :math:`batchSize \times dim` 的矩阵，每行表示一个批次中的单个输入。对于我们支持的全部矩阵操作，请参考 :code:`paddle/legacy/math/Matrix.h`和:code:`paddle/legacy/math/BaseMatrix.h` 。
- 最终，使用 :code:`forwardActivation();` 进行激活操作。这会自动进行网络配置中声明的激活操作。
......@@ -262,7 +262,7 @@ PaddlePaddle的base layer类可以自动计算上面的导数。
REGISTER_LAYER(fc, FullyConnectedLayer);
}
若 :code:`cpp` 被放在 :code:`paddle/gserver/layers` 目录下，其会自动被加入编译列表。
若 :code:`cpp` 被放在 :code:`paddle/legacy/gserver/layers` 目录下，其会自动被加入编译列表。
写梯度检查单元测试
......@@ -270,7 +270,7 @@ PaddlePaddle的base layer类可以自动计算上面的导数。
写梯度检查单元测试是一个验证新实现的层是否正确的相对简单的办法。梯度检查单元测试通过有限差分法来验证一个层的梯度。首先对输入做一个小的扰动 :math:`\Delta x` ，然后观察到输出的变化为 :math:`\Delta y` ，那么，梯度就可以通过这个方程计算得到 :math:`\frac{\Delta y}{\Delta x }` 。之后，再用这个梯度去和 :code:`backward` 函数得到的梯度去对比，以保证梯度计算的正确性。需要注意的是梯度检查仅仅验证了梯度的计算，并不保证 :code:`forward` 和 :code:`backward` 函数的实现是正确的。你需要一些更复杂的单元测试来保证你实现的网络层是正确的。
所有网络层的梯度检查单测都位于 :code:`paddle/gserver/tests/test_LayerGrad.cpp` 。我们建议你在写新网络层时把测试代码放入新的文件中。下面列出了全连接层的梯度检查单元测试。它包含以下几步：
所有网络层的梯度检查单测都位于 :code:`paddle/legacy/gserver/tests/test_LayerGrad.cpp` 。我们建议你在写新网络层时把测试代码放入新的文件中。下面列出了全连接层的梯度检查单元测试。它包含以下几步：
+ 生成网络层配置。网络层配置包含以下几项：
- 偏置参数的大小。（例子中是4096）
......@@ -322,7 +322,7 @@ PaddlePaddle的base layer类可以自动计算上面的导数。
}
}
如果你要为了测试而增加新的文件，例如 :code:`paddle/gserver/tests/testFCGrad.cpp` ，你需要把该文件加入 :code:`paddle/gserver/tests/CMakeLists.txt` 中。下面给出了一个例子。当你执行命令 :code:`make tests` 时，所有的单测都会被执行一次。注意，有些层可能需要高精度来保证梯度检查单测正确执行。你需要在配置cmake时将 :code:`WITH_DOUBLE` 设置为 `ON` 。
如果你要为了测试而增加新的文件，例如 :code:`paddle/legacy/gserver/tests/testFCGrad.cpp` ，你需要把该文件加入 :code:`paddle/legacy/gserver/tests/CMakeLists.txt` 中。下面给出了一个例子。当你执行命令 :code:`make tests` 时，所有的单测都会被执行一次。注意，有些层可能需要高精度来保证梯度检查单测正确执行。你需要在配置cmake时将 :code:`WITH_DOUBLE` 设置为 `ON` 。
.. code-block:: bash
......
doc/v2/dev/new_layer_en.rst
浏览文件 @ ee760d1c
......@@ -58,7 +58,7 @@ Finally we can use chain rule to calculate :math:`\frac{\partial z}{\partial x}`
Implement C++ Class
===================
The C++ class of the layer implements the initialization, forward, and backward part of the layer. The fully connected layer is at :code:`paddle/gserver/layers/FullyConnectedLayer.h` and :code:`paddle/gserver/layers/FullyConnectedLayer.cpp`. We list simplified version of the code below.
The C++ class of the layer implements the initialization, forward, and backward part of the layer. The fully connected layer is at :code:`paddle/legacy/gserver/layers/FullyConnectedLayer.h` and :code:`paddle/legacy/gserver/layers/FullyConnectedLayer.cpp`. We list simplified version of the code below.
It needs to derive the base class :code:`paddle::Layer`, and it needs to override the following functions:
......@@ -154,7 +154,7 @@ The implementation of the forward part has the following steps.
- Every layer must call :code:`Layer::forward(passType);` at the beginning of its :code:`forward` function.
- Then it allocates memory for the output using :code:`reserveOutput(batchSize, size);`. This step is necessary because we support the batches to have different batch sizes. :code:`reserveOutput` will change the size of the output accordingly. For the sake of efficiency, we will allocate new memory if we want to expand the matrix, but we will reuse the existing memory block if we want to shrink the matrix.
- Then it computes :math:`\sum_i W_i x + b` using Matrix operations. :code:`getInput(i).value` retrieve the matrix of the i-th input. Each input is a :math:`batchSize \times dim` matrix, where each row represents an single input in a batch. For a complete lists of supported matrix operations, please refer to :code:`paddle/math/Matrix.h` and :code:`paddle/math/BaseMatrix.h`.
- Then it computes :math:`\sum_i W_i x + b` using Matrix operations. :code:`getInput(i).value` retrieve the matrix of the i-th input. Each input is a :math:`batchSize \times dim` matrix, where each row represents an single input in a batch. For a complete lists of supported matrix operations, please refer to :code:`paddle/legacy/math/Matrix.h` and :code:`paddle/legacy/math/BaseMatrix.h`.
- Finally it applies the activation function using :code:`forwardActivation();`. It will automatically applies the corresponding activation function specifies in the network configuration.
......@@ -263,7 +263,7 @@ Finally, you can use :code:`REGISTER_LAYER(fc, FullyConnectedLayer);` to registe
REGISTER_LAYER(fc, FullyConnectedLayer);
}
If the :code:`cpp` file is put into :code:`paddle/gserver/layers`, it will be automatically added to the compilation list.
If the :code:`cpp` file is put into :code:`paddle/legacy/gserver/layers`, it will be automatically added to the compilation list.
Write Gradient Check Unit Test
......@@ -271,7 +271,7 @@ Write Gradient Check Unit Test
An easy way to verify the correctness of new layer's implementation is to write a gradient check unit test. Gradient check unit test utilizes finite difference method to verify the gradient of a layer. It modifies the input with a small perturbation :math:`\Delta x` and observes the changes of output :math:`\Delta y`, the gradient can be computed as :math:`\frac{\Delta y}{\Delta x }`. This gradient can be compared with the gradient computed by the :code:`backward` function of the layer to ensure the correctness of the gradient computation. Notice that the gradient check only tests the correctness of the gradient computation, it does not necessarily guarantee the correctness of the implementation of the :code:`forward` and :code:`backward` function. You need to write more sophisticated unit tests to make sure your layer is implemented correctly.
All the gradient check unit tests are located in :code:`paddle/gserver/tests/test_LayerGrad.cpp`. You are recommended to put your test into a new test file if you are planning to write a new layer. The gradient test of the gradient check unit test of the fully connected layer is listed below. It has the following steps.
All the gradient check unit tests are located in :code:`paddle/legacy/gserver/tests/test_LayerGrad.cpp`. You are recommended to put your test into a new test file if you are planning to write a new layer. The gradient test of the gradient check unit test of the fully connected layer is listed below. It has the following steps.
+ Create layer configuration. A layer configuration can include the following attributes:
- size of the bias parameter. (4096 in our example)
......@@ -323,7 +323,7 @@ All the gradient check unit tests are located in :code:`paddle/gserver/tests/tes
}
}
If you are creating a new file for the test, such as :code:`paddle/gserver/tests/testFCGrad.cpp`, you need to add the file to :code:`paddle/gserver/tests/CMakeLists.txt`. An example is given below. All the unit tests will run when you execute the command :code:`make tests`. Notice that some layers might need high accuracy for the gradient check unit tests to work well. You need to configure :code:`WITH_DOUBLE` to `ON` when configuring cmake.
If you are creating a new file for the test, such as :code:`paddle/legacy/gserver/tests/testFCGrad.cpp`, you need to add the file to :code:`paddle/legacy/gserver/tests/CMakeLists.txt`. An example is given below. All the unit tests will run when you execute the command :code:`make tests`. Notice that some layers might need high accuracy for the gradient check unit tests to work well. You need to configure :code:`WITH_DOUBLE` to `ON` when configuring cmake.
.. code-block:: bash
......
doc/v2/faq/parameter/index_cn.rst
浏览文件 @ ee760d1c
......@@ -196,6 +196,6 @@ PaddlePaddle保存的模型参数文件内容由16字节头信息和网络参数
obj="process",
args={"src_dict_path": src_dict_path})
完整源码可参考 `sequence_recurrent <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/gserver/tests/sequence_recurrent.py>`_ 示例。
完整源码可参考 `sequence_recurrent <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/gserver/tests/sequence_recurrent.py>`_ 示例。
doc/v2/howto/optimization/gpu_profiling_cn.rst
浏览文件 @ ee760d1c
......@@ -50,12 +50,12 @@ GPU则还需要高并行性，才能发挥其全部能力。这正是它们速
**nvprof** 是Nvidia性能分析工具， **nvvp** 则是带GUI的Nvidia可视化性能分析工具。
在这个教程中，我们主要会介绍nvprof和nvvp。
:code:`test_GpuProfiler` from :code:`paddle/math/tests` directory will be used to evaluate
:code:`test_GpuProfiler` from :code:`paddle/legacy/math/tests` directory will be used to evaluate
above profilers.
:code:`paddle/math/test` 目录中的 :code:`test_GpuProfiler` 就是用于展示上述分析工具的用法。
:code:`paddle/legacy/math/test` 目录中的 :code:`test_GpuProfiler` 就是用于展示上述分析工具的用法。
.. literalinclude:: ../../../../paddle/math/tests/test_GpuProfiler.cpp
.. literalinclude:: ../../../../paddle/legacy/math/tests/test_GpuProfiler.cpp
:language: c++
:lines: 137-151
:linenos:
......@@ -83,7 +83,7 @@ program crashes when CPU version of PaddlePaddle invokes them.
1. 加入 :code:`REGISTER_TIMER_INFO` 和 :code:`printAllStatus` 函数（如高亮部分）。
.. literalinclude:: ../../../../paddle/math/tests/test_GpuProfiler.cpp
.. literalinclude:: ../../../../paddle/legacy/math/tests/test_GpuProfiler.cpp
:language: c++
:lines: 137-151
:emphasize-lines: 8-12,14
......@@ -101,8 +101,8 @@ program crashes when CPU version of PaddlePaddle invokes them.
.. code-block:: bash
:emphasize-lines: 1,12-15
> ./paddle/math/tests/test_GpuProfiler
I1117 11:13:42.313065 2522362816 Util.cpp:155] commandline: ./paddle/math/tests/test_GpuProfiler
> ./paddle/legacy/math/tests/test_GpuProfiler
I1117 11:13:42.313065 2522362816 Util.cpp:155] commandline: ./paddle/legacy/math/tests/test_GpuProfiler
I1117 11:13:42.845065 2522362816 Util.cpp:130] Calling runInitFunctions
I1117 11:13:42.845208 2522362816 Util.cpp:143] Call runInitFunctions done.
[==========] Running 1 test from 1 test case.
......@@ -130,7 +130,7 @@ nvprof 工具
1. 将 :code:`REGISTER_GPU_PROFILER` 函数加到代码中（参考强调部分）。
.. literalinclude:: ../../../../paddle/math/tests/test_GpuProfiler.cpp
.. literalinclude:: ../../../../paddle/legacy/math/tests/test_GpuProfiler.cpp
:language: c++
:lines: 137-151
:emphasize-lines: 6-7
......@@ -147,13 +147,13 @@ nvprof 工具
.. code-block:: bash
nvprof ./paddle/math/tests/test_GpuProfiler
nvprof ./paddle/legacy/math/tests/test_GpuProfiler
然后，您就能获得如下的分析结果：
.. code-block:: bash
==78544== Profiling application: ./paddle/math/tests/test_GpuProfiler
==78544== Profiling application: ./paddle/legacy/math/tests/test_GpuProfiler
==78544== Profiling result:
Time(%) Time Calls Avg Min Max Name
27.60% 9.6305ms 5 1.9261ms 3.4560us 6.4035ms [CUDA memcpy HtoD]
......
doc/v2/howto/optimization/gpu_profiling_en.rst
浏览文件 @ ee760d1c
......@@ -51,10 +51,10 @@ For general GPU profiling, a bunch of tools are provided from both NVIDIA and th
**nvprof** is Nvidia profiler and **nvvp** is (GUI based) Nvidia visual profiler.
In this tutorial, we will focus on nvprof and nvvp.
:code:`test_GpuProfiler` from :code:`paddle/math/tests` directory will be used to evaluate
:code:`test_GpuProfiler` from :code:`paddle/legacy/math/tests` directory will be used to evaluate
above profilers.
.. literalinclude:: ../../../../paddle/math/tests/test_GpuProfiler.cpp
.. literalinclude:: ../../../../paddle/legacy/math/tests/test_GpuProfiler.cpp
:language: c++
:lines: 137-151
:linenos:
......@@ -80,7 +80,7 @@ As a simple example, consider the following:
1. Add :code:`REGISTER_TIMER_INFO` and :code:`printAllStatus` functions (see the emphasize-lines).
.. literalinclude:: ../../../../paddle/math/tests/test_GpuProfiler.cpp
.. literalinclude:: ../../../../paddle/legacy/math/tests/test_GpuProfiler.cpp
:language: c++
:lines: 137-151
:emphasize-lines: 8-12,14
......@@ -98,8 +98,8 @@ As a simple example, consider the following:
.. code-block:: bash
:emphasize-lines: 1,12-15
> ./paddle/math/tests/test_GpuProfiler
I1117 11:13:42.313065 2522362816 Util.cpp:155] commandline: ./paddle/math/tests/test_GpuProfiler
> ./paddle/legacy/math/tests/test_GpuProfiler
I1117 11:13:42.313065 2522362816 Util.cpp:155] commandline: ./paddle/legacy/math/tests/test_GpuProfiler
I1117 11:13:42.845065 2522362816 Util.cpp:130] Calling runInitFunctions
I1117 11:13:42.845208 2522362816 Util.cpp:143] Call runInitFunctions done.
[==========] Running 1 test from 1 test case.
......@@ -127,7 +127,7 @@ To use this command line profiler **nvprof**, you can simply issue the following
1. Add :code:`REGISTER_GPU_PROFILER` function (see the emphasize-lines).
.. literalinclude:: ../../../../paddle/math/tests/test_GpuProfiler.cpp
.. literalinclude:: ../../../../paddle/legacy/math/tests/test_GpuProfiler.cpp
:language: c++
:lines: 137-151
:emphasize-lines: 6-7
......@@ -144,13 +144,13 @@ To use this command line profiler **nvprof**, you can simply issue the following
.. code-block:: bash
nvprof ./paddle/math/tests/test_GpuProfiler
nvprof ./paddle/legacy/math/tests/test_GpuProfiler
Then, you can get the following profiling result:
.. code-block:: bash
==78544== Profiling application: ./paddle/math/tests/test_GpuProfiler
==78544== Profiling application: ./paddle/legacy/math/tests/test_GpuProfiler
==78544== Profiling result:
Time(%) Time Calls Avg Min Max Name
27.60% 9.6305ms 5 1.9261ms 3.4560us 6.4035ms [CUDA memcpy HtoD]
......
doc/v2/howto/rnn/hrnn_rnn_api_compare_cn.rst
浏览文件 @ ee760d1c
......@@ -4,7 +4,7 @@
单双层RNN API对比介绍
#####################
本文以PaddlePaddle的双层RNN单元测试为示例，用多对效果完全相同的、分别使用单双层RNN作为网络配置的模型，来讲解如何使用双层RNN。本文中所有的例子，都只是介绍双层RNN的API接口，并不是使用双层RNN解决实际的问题。如果想要了解双层RNN在具体问题中的使用，请参考\ :ref:`algo_hrnn_demo`\ 。本文中示例所使用的单元测试文件是\ `test_RecurrentGradientMachine.cpp <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/gserver/tests/test_RecurrentGradientMachine.cpp>`_\ 。
本文以PaddlePaddle的双层RNN单元测试为示例，用多对效果完全相同的、分别使用单双层RNN作为网络配置的模型，来讲解如何使用双层RNN。本文中所有的例子，都只是介绍双层RNN的API接口，并不是使用双层RNN解决实际的问题。如果想要了解双层RNN在具体问题中的使用，请参考\ :ref:`algo_hrnn_demo`\ 。本文中示例所使用的单元测试文件是\ `test_RecurrentGradientMachine.cpp <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/gserver/tests/test_RecurrentGradientMachine.cpp>`_\ 。
示例1：双层RNN，子序列间无Memory
================================
......@@ -13,8 +13,8 @@
在本示例中，单层RNN和双层RNN的网络配置，都是将每一句分好词后的句子，使用LSTM作为encoder，压缩成一个向量。区别是RNN使用两层序列模型，将多句话看成一个整体同时使用encoder压缩。二者语意上完全一致。这组语义相同的示例配置如下：
* 单层RNN\: `sequence_layer_group.conf <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/gserver/tests/sequence_layer_group.conf>`_
* 双层RNN\: `sequence_nest_layer_group.conf <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/gserver/tests/sequence_nest_layer_group.conf>`_
* 单层RNN\: `sequence_layer_group.conf <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/gserver/tests/sequence_layer_group.conf>`_
* 双层RNN\: `sequence_nest_layer_group.conf <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/gserver/tests/sequence_nest_layer_group.conf>`_
读取双层序列数据
......@@ -24,18 +24,18 @@
- 本例中的原始数据一共有10个样本。每个样本由两部分组成，一个label（此处都为2）和一个已经分词后的句子。这个数据也被单层RNN网络直接使用。
.. literalinclude:: ../../../../paddle/gserver/tests/Sequence/tour_train_wdseg
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/Sequence/tour_train_wdseg
:language: text
- 双层序列数据一共有4个样本。 每个样本间用空行分开，整体数据和原始数据完全一样。但于双层序列的LSTM来说，第一个样本同时encode两条数据成两个向量。这四条数据同时处理的句子数量为\ :code:`[2, 3, 2, 3]`\ 。
.. literalinclude:: ../../../../paddle/gserver/tests/Sequence/tour_train_wdseg.nest
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/Sequence/tour_train_wdseg.nest
:language: text
其次，对于两种不同的输入数据类型，不同DataProvider对比如下(`sequenceGen.py <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/gserver/tests/sequenceGen.py>`_)\：
其次，对于两种不同的输入数据类型，不同DataProvider对比如下(`sequenceGen.py <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/gserver/tests/sequenceGen.py>`_)\：
.. literalinclude:: ../../../../paddle/gserver/tests/sequenceGen.py
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequenceGen.py
:language: python
:lines: 21-39
:linenos:
......@@ -47,7 +47,7 @@
- words是原始数据中的每一句话，所对应的词表index数组。它是integer_value_sequence类型的，即整数数组。words即为这个数据中的单层时间序列。
- label是原始数据中对于每一句话的分类标签，它是integer_value类型的。
.. literalinclude:: ../../../../paddle/gserver/tests/sequenceGen.py
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequenceGen.py
:language: python
:lines: 42-71
:linenos:
......@@ -64,7 +64,7 @@
首先，我们看一下单层RNN的配置。代码中9-15行(高亮部分)即为单层RNN序列的使用代码。这里使用了PaddlePaddle预定义好的RNN处理函数。在这个函数中，RNN对于每一个时间步通过了一个LSTM网络。
.. literalinclude:: ../../../../paddle/gserver/tests/sequence_layer_group.conf
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_layer_group.conf
:language: python
:lines: 38-63
:linenos:
......@@ -85,7 +85,7 @@
* 至此，\ :code:`lstm_last`\ 便和单层RNN配置中的\ :code:`lstm_last`\ 具有相同的结果了。
.. literalinclude:: ../../../../paddle/gserver/tests/sequence_nest_layer_group.conf
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_nest_layer_group.conf
:language: python
:lines: 38-64
:linenos:
......@@ -107,7 +107,7 @@
- 单层RNN：过了一个很简单的recurrent_group。每一个时间步，当前的输入y和上一个时间步的输出rnn_state做了一个全链接。
.. literalinclude:: ../../../../paddle/gserver/tests/sequence_rnn.conf
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_rnn.conf
:language: python
:lines: 36-48
......@@ -116,7 +116,7 @@
- 内层inner_step的recurrent_group和单层序列的几乎一样。除了boot_layer=outer_mem，表示将外层的outer_mem作为内层memory的初始状态。外层outer_step中，outer_mem是一个子句的最后一个向量，即整个双层group是将前一个子句的最后一个向量，作为下一个子句memory的初始状态。
- 从输入数据上看，单双层序列的句子是一样的，只是双层序列将其又做了子序列划分。因此双层序列的配置中，必须将前一个子句的最后一个元素，作为boot_layer传给下一个子句的memory，才能保证和单层序列的配置中“每个时间步都用了上一个时间步的输出结果”一致。
.. literalinclude:: ../../../../paddle/gserver/tests/sequence_nest_rnn.conf
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_nest_rnn.conf
:language: python
:lines: 39-66
......@@ -134,7 +134,7 @@
**输入不等长** 是指recurrent_group的多个输入序列，在每个时间步的子序列长度可以不相等。但序列输出时，需要指定与某一个输入的序列信息是一致的。使用\ :red:`targetInlink`\ 可以指定哪一个输入和输出序列信息一致，默认指定第一个输入。
示例3的配置分别为\ `单层不等长RNN <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/gserver/tests/sequence_rnn_multi_unequalength_inputs.py>`_\ 和\ `双层不等长RNN <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py>`_\ 。
示例3的配置分别为\ `单层不等长RNN <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/gserver/tests/sequence_rnn_multi_unequalength_inputs.py>`_\ 和\ `双层不等长RNN <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py>`_\ 。
示例3对于单层RNN和双层RNN数据完全相同。
......@@ -152,14 +152,14 @@
* 单层RNN\:
.. literalinclude:: ../../../../paddle/gserver/tests/sequence_rnn_multi_unequalength_inputs.py
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_rnn_multi_unequalength_inputs.py
:language: python
:lines: 42-59
:linenos:
* 双层RNN\ \:
.. literalinclude:: ../../../../paddle/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py
:language: python
:lines: 41-80
:linenos:
......
doc/v2/howto/rnn/hrnn_rnn_api_compare_en.rst
浏览文件 @ ee760d1c
......@@ -4,7 +4,7 @@
API comparision between RNN and hierarchical RNN
#####################
This article takes PaddlePaddle's hierarchical RNN unit test as an example. We will use several examples to illestrate the usage of single-layer and hierarchical RNNs. Each example has two model configurations, one for single-layer, and the other for hierarchical RNN. Although the implementations are different, both the two model configurations' effects are the same. All of the examples in this article only describe the API interface of the hierarchical RNN, while we do not use this hierarchical RNN to solve practical problems. If you want to understand the use of hierarchical RNN in specific issues, please refer to \ :ref:`algo_hrnn_demo`\ 。The unit test file used in this article's example is \ `test_RecurrentGradientMachine.cpp <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/gserver/tests/test_RecurrentGradientMachine.cpp>`_\ 。
This article takes PaddlePaddle's hierarchical RNN unit test as an example. We will use several examples to illestrate the usage of single-layer and hierarchical RNNs. Each example has two model configurations, one for single-layer, and the other for hierarchical RNN. Although the implementations are different, both the two model configurations' effects are the same. All of the examples in this article only describe the API interface of the hierarchical RNN, while we do not use this hierarchical RNN to solve practical problems. If you want to understand the use of hierarchical RNN in specific issues, please refer to \ :ref:`algo_hrnn_demo`\ 。The unit test file used in this article's example is \ `test_RecurrentGradientMachine.cpp <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/gserver/tests/test_RecurrentGradientMachine.cpp>`_\ 。
Example 1：Hierarchical RNN without Memory between subsequences
================================
......@@ -13,8 +13,8 @@ The classical case in the hierarchical RNN is to perform sequence operations on
In this example, the network configuration of single-layer RNNs and hierarchical RNNs are all to use LSTM as en encoder to compress a word-segmented sentence into a vector. The difference is that, RNN uses a hierarchical RNN model, treating multiple sentences as a whole to use encoder to compress simultaneously. They are completely consistent in their semantic meanings. This pair of semantically identical example configurations is as follows：
* RNN\: `sequence_layer_group.conf <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/gserver/tests/sequence_layer_group.conf>`_
* Hierarchical RNN\: `sequence_nest_layer_group.conf <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/gserver/tests/sequence_nest_layer_group.conf>`_
* RNN\: `sequence_layer_group.conf <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/gserver/tests/sequence_layer_group.conf>`_
* Hierarchical RNN\: `sequence_nest_layer_group.conf <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/gserver/tests/sequence_nest_layer_group.conf>`_
Reading hierarchical sequence data
......@@ -24,18 +24,18 @@ Firstly, the original data in this example is as follows \:
- The original data in this example has 10 samples. Each of the sample includes two components: a lable(all 2 here), and a word-segmented sentence. This data is used by single RNN as well.
.. literalinclude:: ../../../../paddle/gserver/tests/Sequence/tour_train_wdseg
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/Sequence/tour_train_wdseg
:language: text
- The data for hierarchical RNN has 4 samples. Every sample is seperated by a blank line, while the content of the data is the same as the original data. But as for hierarchical LSTM, the first sample will encode two sentences into two vectors simultaneously. The sentence count dealed simultaneously by this 4 samples are \ :code:`[2, 3, 2, 3]`\ .
.. literalinclude:: ../../../../paddle/gserver/tests/Sequence/tour_train_wdseg.nest
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/Sequence/tour_train_wdseg.nest
:language: text
Secondly, as for these two types of different input data formats, the contrast of different DataProviders are as follows (`sequenceGen.py <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/gserver/tests/sequenceGen.py>`_)\：
Secondly, as for these two types of different input data formats, the contrast of different DataProviders are as follows (`sequenceGen.py <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/gserver/tests/sequenceGen.py>`_)\：
.. literalinclude:: ../../../../paddle/gserver/tests/sequenceGen.py
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequenceGen.py
:language: python
:lines: 21-39
:linenos:
......@@ -47,7 +47,7 @@ Secondly, as for these two types of different input data formats, the contrast o
- "words" is a list of word table indices corresponding to each word in the sentence in the original data. Its data type is integer_value_sequence, that is integer list. So, "words" is a singler-layer time series in the data.
- "label" is the categorical label of each sentence, whose data type is integer_value.
.. literalinclude:: ../../../../paddle/gserver/tests/sequenceGen.py
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequenceGen.py
:language: python
:lines: 42-71
:linenos:
......@@ -64,7 +64,7 @@ Model configuration
Firstly, let's look at the configuration of single-layer RNN. The hightlighted part of line 9 to line 15 is the usage of single-layer RNN. Here we use the pre-defined RNN process function in PaddlePaddle. In this function, for each time step, RNN passes through an LSTM network.
.. literalinclude:: ../../../../paddle/gserver/tests/sequence_layer_group.conf
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_layer_group.conf
:language: python
:lines: 38-63
:linenos:
......@@ -85,7 +85,7 @@ Secondly, let's look at the model configuration of hierarchical RNN which has th
* Till now, \ :code:`lstm_last`\ has the same result as \ :code:`lstm_last`\ in single-layer RNN configuration.
.. literalinclude:: ../../../../paddle/gserver/tests/sequence_nest_layer_group.conf
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_nest_layer_group.conf
:language: python
:lines: 38-64
:linenos:
......@@ -107,7 +107,7 @@ We select the different parts between single-layer RNN and hierarchical RNN conf
- single-layer RNN：passes through a simple recurrent_group. For each time step, the current input y and the last time step's output rnn_state pass through a fully-connected layer.
.. literalinclude:: ../../../../paddle/gserver/tests/sequence_rnn.conf
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_rnn.conf
:language: python
:lines: 36-48
......@@ -116,7 +116,7 @@ We select the different parts between single-layer RNN and hierarchical RNN conf
- The recurrent_group of inner layer's inner_step is nearly the same as single-layer sequence, except for the case of boot_layer=outer_mem, which means using the outer layer's outer_mem as the initial state for the inner layer's memory. In the outer layer's out_step, outer_mem is the last vector of a subsequence, that is, the whole hierarchical group uses the last vector of the previous subsequence as the initial state for the next subsequence's memory.
- From the aspect of the input data, sentences from single-layer and hierarchical RNN are the same. The only difference is that, hierarchical RNN disassembes the sequence into subsequences. So in the hierarchical RNN configuration, we must use the last element of the previous subsequence as a boot_layer for the memory of the next subsequence, so that it makes no difference with "every time step uses the output of last time step" in the sigle-layer RNN configuration.
.. literalinclude:: ../../../../paddle/gserver/tests/sequence_nest_rnn.conf
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_nest_rnn.conf
:language: python
:lines: 39-66
......@@ -134,7 +134,7 @@ Example 3：hierarchical RNN with unequal length inputs
**unequal length inputs** means in the multiple input sequences of recurrent_group, the lengths of subsequences can be unequal. But the output of the sequence, needs to be consistent with one of the input sequences. Using \ :red:`targetInlink`\ can help you specify which of the input sequences and the output sequence can be consistent, by default is the first input.
The configurations of Example 3 are \ `sequence_rnn_multi_unequalength_inputs <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/gserver/tests/sequence_rnn_multi_unequalength_inputs.py>`_ \ and \ `sequence_nest_rnn_multi_unequalength_inputs <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py>`_\ .
The configurations of Example 3 are \ `sequence_rnn_multi_unequalength_inputs <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/gserver/tests/sequence_rnn_multi_unequalength_inputs.py>`_ \ and \ `sequence_nest_rnn_multi_unequalength_inputs <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py>`_\ .
The data for the configurations of Example 3's single-layer RNN and hierarchical RNN are exactly the same.
......@@ -152,14 +152,14 @@ Similar to Example 2's configuration, Example 3's configuration uses single-laye
* single-layer RNN\:
.. literalinclude:: ../../../../paddle/gserver/tests/sequence_rnn_multi_unequalength_inputs.py
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_rnn_multi_unequalength_inputs.py
:language: python
:lines: 42-59
:linenos:
* hierarchical RNN\ \:
.. literalinclude:: ../../../../paddle/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py
.. literalinclude:: ../../../../paddle/legacy/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py
:language: python
:lines: 41-80
:linenos:
......
go/pserver/optimizer.go
浏览文件 @ ee760d1c
......@@ -16,7 +16,7 @@ package pserver
// #cgo CFLAGS: -I ../../
// #cgo LDFLAGS: ${SRCDIR}/client/c/libpaddle_go_optimizer.a -lstdc++ -lm
// #include "paddle/optimizer/optimizer.h"
// #include "paddle/legacy/optimizer/optimizer.h"
// #include <stdlib.h>
// #include <string.h>
import "C"
......
paddle/CMakeLists.txt
浏览文件 @ ee760d1c
if(NOT WITH_FLUID_ONLY)
add_subdirectory(cuda)
add_subdirectory(function)
add_subdirectory(legacy/cuda)
add_subdirectory(legacy/function)
add_subdirectory(utils)
add_subdirectory(math)
add_subdirectory(gserver)
add_subdirectory(parameter)
add_subdirectory(legacy/math)
add_subdirectory(legacy/gserver)
add_subdirectory(legacy/parameter)
if(MOBILE_INFERENCE)
add_subdirectory(capi)
add_subdirectory(legacy/capi)
else()
add_subdirectory(pserver)
add_subdirectory(legacy/pserver)
add_subdirectory(trainer)
add_subdirectory(scripts)
if(WITH_C_API)
add_subdirectory(capi)
add_subdirectory(legacy/capi)
endif()
if(WITH_SWIG_PY)
add_subdirectory(api)
add_subdirectory(legacy/api)
endif()
endif()
endif()
......
paddle/api/Arguments.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
/* 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 "PaddleAPI.h"
#include "PaddleAPIPrivate.h"
#include "paddle/parameter/Argument.h"
size_t Arguments::getSlotNum() const { return m->outputs.size(); }
Arguments* Arguments::createArguments(size_t slotNum) {
auto args = new Arguments();
args->m->outputs.resize(slotNum);
return args;
}
void Arguments::resize(size_t slotNum) { m->outputs.resize(slotNum); }
Arguments::Arguments() : m(new ArgumentsPrivate()) {}
Arguments::~Arguments() { delete m; }
Arguments* Arguments::createByPaddleArgumentVector(void* ptr) {
auto p = (std::vector<paddle::Argument>*)(ptr);
auto args = new Arguments();
args->m->outputs = *p;
return args;
}
Arguments* Arguments::createByPaddleArgument(const void* ptr) {
auto p = (paddle::Argument*)(ptr);
auto args = new Arguments();
args->m->outputs.push_back(*p);
return args;
}
Matrix* Arguments::getSlotValue(size_t idx) const throw(RangeError) {
auto& a = m->getArg(idx);
return Matrix::createByPaddleMatrixPtr(&a.value);
}
Matrix* Arguments::getSlotGrad(size_t idx) const throw(RangeError) {
auto& a = m->getArg(idx);
return Matrix::createByPaddleMatrixPtr(&a.grad);
}
IVector* Arguments::getSlotIds(size_t idx) const throw(RangeError) {
auto& a = m->getArg(idx);
return IVector::createByPaddleVectorPtr(&a.ids);
}
Matrix* Arguments::getSlotIn(size_t idx) const throw(RangeError) {
auto& a = m->getArg(idx);
return Matrix::createByPaddleMatrixPtr(&a.in);
}
void Arguments::setSlotValue(size_t idx, Matrix* mat) throw(RangeError) {
auto& a = m->getArg(idx);
a.value = m->cast<paddle::Matrix>(mat->getSharedPtr());
}
void Arguments::setSlotGrad(size_t idx, Matrix* mat) throw(RangeError) {
auto& a = m->getArg(idx);
a.grad = m->cast<paddle::Matrix>(mat->getSharedPtr());
}
void Arguments::setSlotIn(size_t idx, Matrix* mat) throw(RangeError) {
auto& a = m->getArg(idx);
a.in = m->cast<paddle::Matrix>(mat->getSharedPtr());
}
void Arguments::setSlotIds(size_t idx, IVector* vec) throw(RangeError) {
auto& a = m->getArg(idx);
auto& v = m->cast<paddle::IVector>(vec->getSharedPtr());
a.ids = v;
}
template <typename T1>
static inline void doCopyFromSafely(std::shared_ptr<T1>& dest,
std::shared_ptr<T1>& src) {
if (src) {
if (dest) {
dest->copyFrom(*src);
} else {
dest = src;
}
}
}
IVector* Arguments::getSlotSequenceStartPositions(size_t idx) const
throw(RangeError) {
auto& a = m->getArg(idx);
if (a.sequenceStartPositions) {
return IVector::createByPaddleVectorPtr(
&a.sequenceStartPositions->getMutableVector(false));
} else {
return nullptr;
}
}
IVector* Arguments::getSlotSubSequenceStartPositions(size_t idx) const
throw(RangeError) {
auto& a = m->getArg(idx);
if (a.subSequenceStartPositions) {
return IVector::createByPaddleVectorPtr(
&a.subSequenceStartPositions->getMutableVector(false));
} else {
return nullptr;
}
}
void Arguments::setSlotSequenceStartPositions(size_t idx,
IVector* vec) throw(RangeError) {
auto& a = m->getArg(idx);
auto& v = m->cast<paddle::IVector>(vec->getSharedPtr());
a.sequenceStartPositions = std::make_shared<paddle::ICpuGpuVector>(v);
}
void Arguments::setSlotSubSequenceStartPositions(
size_t idx, IVector* vec) throw(RangeError) {
auto& a = m->getArg(idx);
auto& v = m->cast<paddle::IVector>(vec->getSharedPtr());
a.subSequenceStartPositions = std::make_shared<paddle::ICpuGpuVector>(v);
}
IVector* Arguments::getSlotSequenceDim(size_t idx) const throw(RangeError) {
auto& a = m->getArg(idx);
return IVector::createByPaddleVectorPtr(&a.cpuSequenceDims);
}
void Arguments::setSlotSequenceDim(size_t idx, IVector* vec) throw(RangeError) {
auto& a = m->getArg(idx);
a.cpuSequenceDims = m->cast<paddle::IVector>(vec->getSharedPtr());
}
float Arguments::sum() const { return paddle::Argument::sum(m->outputs); }
int64_t Arguments::getBatchSize(size_t idx) const throw(RangeError) {
auto& a = m->getArg(idx);
return a.getBatchSize();
}
void Arguments::setSlotFrameHeight(size_t idx, size_t h) throw(RangeError) {
auto& a = m->getArg(idx);
a.setFrameHeight(h);
}
void Arguments::setSlotFrameWidth(size_t idx, size_t w) throw(RangeError) {
auto& a = m->getArg(idx);
a.setFrameWidth(w);
}
size_t Arguments::getSlotFrameHeight(size_t idx) const throw(RangeError) {
auto& a = m->getArg(idx);
return a.getFrameHeight();
}
size_t Arguments::getSlotFrameWidth(size_t idx) const throw(RangeError) {
auto& a = m->getArg(idx);
return a.getFrameWidth();
}
void* Arguments::getInternalArgumentsPtr() const { return &m->outputs; }
paddle/api/GradientMachine.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
/* 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 "PaddleAPI.h"
#include "PaddleAPIPrivate.h"
#include "Internal.h"
#include "paddle/gserver/gradientmachines/NeuralNetwork.h"
std::vector<int> GradientMachine::defaultParamTypes = {
PARAMETER_VALUE, PARAMETER_GRADIENT, PARAMETER_MOMENTUM};
GradientMachine::GradientMachine() : m(new GradientMachinePrivate()) {}
GradientMachine::~GradientMachine() { delete m; }
GradientMachine* GradientMachine::createFromPaddleModelPtr(
const void* confPtr,
GradientMatchineCreateMode mode,
const std::vector<int>& types) {
auto& conf = *(const paddle::ModelConfig*)(confPtr);
std::vector<ParameterType> realTypes;
staticCastVector(&realTypes, types);
auto machineRawPtr = paddle::GradientMachine::create(conf, mode, realTypes);
auto machinePtr = std::shared_ptr<paddle::GradientMachine>(machineRawPtr);
if (machinePtr != nullptr) {
auto machine = new GradientMachine();
machine->m->machine = machinePtr;
return machine;
} else {
return nullptr;
}
}
GradientMachine* GradientMachine::createByConfigProtoStr(
const std::string& protoStr,
GradientMatchineCreateMode mode,
const std::vector<int>& types) {
paddle::ModelConfig conf;
conf.ParseFromString(protoStr);
if (conf.IsInitialized()) {
return GradientMachine::createFromPaddleModelPtr(&conf, mode, types);
} else {
return nullptr;
}
}
GradientMachine* GradientMachine::createByModelConfig(
ModelConfig* conf,
GradientMatchineCreateMode mode,
const std::vector<int>& types) {
auto confPtr = &conf->m->conf->getModelConfig();
return GradientMachine::createFromPaddleModelPtr(confPtr, mode, types);
}
void GradientMachine::start() { m->machine->start(); }
void GradientMachine::finish() { m->machine->finish(); }
void GradientMachine::onPassEnd() { m->machine->onPassEnd(); }
void GradientMachine::prefetch(const Arguments& inArgs) {
auto& in =
m->cast<std::vector<paddle::Argument>>(inArgs.getInternalArgumentsPtr());
m->machine->prefetch(in);
}
void GradientMachine::forward(const Arguments& inArgs,
Arguments* outArgs,
PassType passType) {
auto& in =
m->cast<std::vector<paddle::Argument>>(inArgs.getInternalArgumentsPtr());
auto& out = m->cast<std::vector<paddle::Argument>>(
outArgs->getInternalArgumentsPtr());
paddle::PassType pt = (paddle::PassType)(passType);
m->machine->forward(in, &out, pt);
}
UpdateCallback::~UpdateCallback() {}
void UpdateCallback::apply(Parameter* p) {
// UNUSED(p);
}
class UpdateCallbackWrapper {
public:
explicit UpdateCallbackWrapper(const UpdateCallback& callback)
: callback(const_cast<UpdateCallback&>(callback)) {}
void operator()(paddle::Parameter* param) {
auto p = Parameter::createFromRawPtr(&param);
// @TODO Use Stack variable instead.
callback.apply(p);
delete p;
}
private:
UpdateCallback& callback;
};
void GradientMachine::backward(const UpdateCallback& callback) {
m->machine->backward(UpdateCallbackWrapper(callback));
}
void GradientMachine::forwardBackward(const Arguments& inArgs,
Arguments* outArgs,
PassType passType,
const UpdateCallback& callback) {
auto& in =
m->cast<std::vector<paddle::Argument>>(inArgs.getInternalArgumentsPtr());
auto& out = m->cast<std::vector<paddle::Argument>>(
outArgs->getInternalArgumentsPtr());
paddle::PassType pt = (paddle::PassType)(passType);
m->machine->forwardBackward(in, &out, pt, UpdateCallbackWrapper(callback));
}
void GradientMachine::loadParameters(const std::string& path) {
m->machine->loadParameters(path);
}
size_t GradientMachine::getParameterSize() const {
return m->machine->getParameters().size();
}
Parameter* GradientMachine::getParameter(size_t i) throw(RangeError) {
auto params = m->machine->getParameters();
if (i < params.size()) {
return Parameter::createFromSharedPtr(&m->machine->getParameters()[i]);
} else {
throw RangeError();
}
}
size_t GradientMachine::getNonStaticParameterSize() const {
return m->machine->getNonStaticParameters().size();
}
Parameter* GradientMachine::getNonStaticParameter(size_t i) throw(RangeError) {
auto params = m->machine->getNonStaticParameters();
if (i < params.size()) {
return Parameter::createFromSharedPtr(
&m->machine->getNonStaticParameters()[i]);
} else {
throw RangeError();
}
}
void GradientMachine::randParameters() { m->machine->randParameters(); }
Arguments* GradientMachine::getLayerOutput(const std::string& layerName) const
throw(UnsupportError) {
auto nn = m->machine;
if (nn) {
auto arg = nn->getLayerOutput(layerName);
return Arguments::createByPaddleArgument(&arg);
} else {
throw UnsupportError();
}
}
SequenceGenerator* GradientMachine::asSequenceGenerator(
const std::vector<std::string>& dict,
size_t begin_id,
size_t end_id,
size_t max_length,
size_t beam_size) {
SequenceGenerator* r =
SequenceGenerator::createByGradientMachineSharedPtr(&m->machine);
r->setDict(dict);
r->setBos(begin_id);
r->setEos(end_id);
r->setMaxLength(max_length);
r->setBeamSize(beam_size);
return r;
}
Evaluator* GradientMachine::makeEvaluator() {
auto ev = new Evaluator();
ev->m->rawPtr = m->machine->makeEvaluator();
return ev;
}
void GradientMachine::eval(Evaluator* evaluator) {
m->machine->eval(evaluator->m->rawPtr);
}
paddle/api/Matrix.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/math/Matrix.h"
#include <cstring>
#include <iostream>
#include "PaddleAPI.h"
#include "paddle/math/CpuSparseMatrix.h"
#include "paddle/math/SparseMatrix.h"
struct MatrixPrivate {
std::shared_ptr<paddle::Matrix> mat;
};
Matrix::Matrix() : m(new MatrixPrivate()) {}
Matrix* Matrix::createByPaddleMatrixPtr(void* sharedPtr) {
auto* mat = reinterpret_cast<paddle::MatrixPtr*>(sharedPtr);
if ((*mat) != nullptr) {
auto m = new Matrix();
m->m->mat = *mat;
return m;
} else {
return nullptr;
}
}
Matrix* Matrix::createZero(size_t height, size_t width, bool useGpu) {
auto m = new Matrix();
m->m->mat = paddle::Matrix::create(height, width, useGpu);
m->m->mat->zero();
return m;
}
Matrix* Matrix::createDense(const std::vector<float>& data,
size_t height,
size_t width,
bool useGpu) {
auto m = new Matrix();
m->m->mat = paddle::Matrix::create(height, width, useGpu);
m->m->mat->copyFrom(data.data(), data.size());
return m;
}
Matrix* Matrix::createDenseFromNumpy(float* data,
int dim1,
int dim2,
bool copy,
bool useGpu) throw(UnsupportError) {
if (useGpu) {
/// Gpu mode only supports copy=True
if (!copy) {
throw UnsupportError("Gpu mode only supports copy=True");
}
return Matrix::createGpuDenseFromNumpy(data, dim1, dim2);
} else {
return Matrix::createCpuDenseFromNumpy(data, dim1, dim2, copy);
}
}
Matrix* Matrix::createCpuDenseFromNumpy(float* data,
int dim1,
int dim2,
bool copy) {
auto m = new Matrix();
if (copy) {
m->m->mat = paddle::Matrix::create(dim1, dim2);
m->m->mat->copyFrom(data, dim1 * dim2);
} else {
m->m->mat = paddle::Matrix::create(data, dim1, dim2, false);
}
return m;
}
Matrix* Matrix::createGpuDenseFromNumpy(float* data, int dim1, int dim2) {
auto m = new Matrix();
m->m->mat = paddle::Matrix::create(dim1, dim2, false, true);
m->m->mat->copyFrom(data, dim1 * dim2);
return m;
}
Matrix* Matrix::createSparse(size_t height,
size_t width,
size_t nnz,
bool isNonVal,
bool isTrans,
bool useGpu) {
auto m = new Matrix();
m->m->mat = paddle::Matrix::createSparseMatrix(
height,
width,
nnz,
isNonVal ? paddle::NO_VALUE : paddle::FLOAT_VALUE,
isTrans,
useGpu);
return m;
}
Matrix::~Matrix() { delete m; }
size_t Matrix::getHeight() const { return m->mat->getHeight(); }
size_t Matrix::getWidth() const { return m->mat->getWidth(); }
float Matrix::get(size_t x, size_t y) const throw(RangeError) {
if (x > this->getWidth() || y > this->getHeight()) {
RangeError e;
throw e;
}
return m->mat->getElement(x, y);
}
void Matrix::set(size_t x, size_t y, float val) throw(RangeError,
UnsupportError) {
if (x > this->getWidth() || y > this->getHeight()) {
RangeError e;
throw e;
}
auto rawMat = m->mat.get();
if (auto cDenseMat = dynamic_cast<paddle::CpuMatrix*>(rawMat)) {
*(cDenseMat->getData() + x + y * cDenseMat->getWidth()) = val;
} else {
UnsupportError e;
throw e;
}
}
bool Matrix::isSparse() const {
auto raw_mat = m->mat.get();
return dynamic_cast<paddle::CpuSparseMatrix*>(raw_mat) != nullptr ||
dynamic_cast<paddle::GpuSparseMatrix*>(raw_mat) != nullptr;
}
SparseValueType Matrix::getSparseValueType() const throw(UnsupportError) {
auto cpuSparseMat =
std::dynamic_pointer_cast<paddle::CpuSparseMatrix>(m->mat);
if (cpuSparseMat != nullptr) {
return (SparseValueType)cpuSparseMat->getValueType();
} else {
auto gpuSparseMat =
std::dynamic_pointer_cast<paddle::GpuSparseMatrix>(m->mat);
if (gpuSparseMat != nullptr) {
return (SparseValueType)gpuSparseMat->getValueType();
} else {
UnsupportError e;
throw e;
}
}
}
SparseFormatType Matrix::getSparseFormat() const throw(UnsupportError) {
auto cpuSparseMat =
std::dynamic_pointer_cast<paddle::CpuSparseMatrix>(m->mat);
if (cpuSparseMat != nullptr) {
return (SparseFormatType)cpuSparseMat->getFormat();
} else {
auto gpuSparseMat =
std::dynamic_pointer_cast<paddle::GpuSparseMatrix>(m->mat);
if (gpuSparseMat != nullptr) {
return SPARSE_CSR;
} else {
UnsupportError e;
throw e;
}
}
}
IntArray Matrix::getSparseRowCols(size_t i) const
throw(UnsupportError, RangeError) {
auto cpuSparseMat =
std::dynamic_pointer_cast<paddle::CpuSparseMatrix>(m->mat);
if (cpuSparseMat != nullptr &&
cpuSparseMat->getFormat() == paddle::SPARSE_CSR) {
if (i < cpuSparseMat->getHeight()) {
// cpuSparseMat->print(std::cout);
size_t len = cpuSparseMat->getColNum(i);
return IntArray(cpuSparseMat->getRowCols(i), len);
} else {
RangeError e;
throw e;
}
} else {
UnsupportError e;
throw e;
}
}
IntWithFloatArray Matrix::getSparseRowColsVal(size_t i) const
throw(UnsupportError, RangeError) {
auto cpuSparseMat =
std::dynamic_pointer_cast<paddle::CpuSparseMatrix>(m->mat);
if (cpuSparseMat != nullptr &&
cpuSparseMat->getValueType() == paddle::FLOAT_VALUE) {
if (i < cpuSparseMat->getHeight()) {
return IntWithFloatArray(cpuSparseMat->getRowValues(i),
cpuSparseMat->getRowCols(i),
cpuSparseMat->getColNum(i));
} else {
RangeError e;
throw e;
}
} else {
UnsupportError e;
throw e;
}
}
FloatArray Matrix::getData() const {
auto rawMat = m->mat.get();
if (dynamic_cast<paddle::GpuMemoryHandle*>(rawMat->getMemoryHandle().get())) {
// is gpu. then copy data
float* data = rawMat->getData();
size_t len = rawMat->getElementCnt();
float* cpuData = new float[len];
hl_memcpy_device2host(cpuData, data, len * sizeof(float));
FloatArray ret_val(cpuData, len);
ret_val.needFree = true;
return ret_val;
} else {
FloatArray ret_val(rawMat->getData(), rawMat->getElementCnt());
return ret_val;
}
}
void Matrix::sparseCopyFrom(
const std::vector<int>& rows,
const std::vector<int>& cols,
const std::vector<float>& vals) throw(UnsupportError) {
auto cpuSparseMat =
std::dynamic_pointer_cast<paddle::CpuSparseMatrix>(m->mat);
if (cpuSparseMat != nullptr) {
// LOG(INFO) <<"RowSize = "<<rows.size()
// <<" ColSize = "<<cols.size()
// <<" ValSize = "<<vals.size();
cpuSparseMat->copyFrom(const_cast<std::vector<int>&>(rows),
const_cast<std::vector<int>&>(cols),
const_cast<std::vector<float>&>(vals));
} else {
UnsupportError e;
throw e;
}
}
void* Matrix::getSharedPtr() const { return &m->mat; }
void Matrix::toNumpyMatInplace(float** view_data,
int* dim1,
int* dim2) throw(UnsupportError) {
auto cpuMat = std::dynamic_pointer_cast<paddle::CpuMatrix>(m->mat);
if (cpuMat) {
*dim1 = cpuMat->getHeight();
*dim2 = cpuMat->getWidth();
*view_data = cpuMat->getData();
} else {
throw UnsupportError();
}
}
void Matrix::copyToNumpyMat(float** view_m_data,
int* dim1,
int* dim2) throw(UnsupportError) {
static_assert(sizeof(paddle::real) == sizeof(float),
"Currently PaddleAPI only support for single "
"precision version of paddle.");
if (this->isSparse()) {
throw UnsupportError();
} else {
*dim1 = m->mat->getHeight();
*dim2 = m->mat->getWidth();
*view_m_data = new float[(*dim1) * (*dim2)];
if (auto cpuMat = dynamic_cast<paddle::CpuMatrix*>(m->mat.get())) {
auto src = cpuMat->getData();
auto dest = *view_m_data;
std::memcpy(dest, src, sizeof(paddle::real) * (*dim1) * (*dim2));
} else if (auto gpuMat = dynamic_cast<paddle::GpuMatrix*>(m->mat.get())) {
auto src = gpuMat->getData();
auto dest = *view_m_data;
hl_memcpy_device2host(
dest, src, sizeof(paddle::real) * (*dim1) * (*dim2));
} else {
LOG(WARNING) << "Unexpected Situation";
throw UnsupportError();
}
}
}
void Matrix::copyFromNumpyMat(float* data,
int dim1,
int dim2) throw(UnsupportError, RangeError) {
if (isSparse()) {
throw UnsupportError();
} else {
if (this->getHeight() == (size_t)dim1 && this->getWidth() == (size_t)dim2) {
if (m->mat->getData() != data) {
m->mat->copyFrom(data, dim1 * dim2);
}
} else {
throw RangeError();
}
}
}
bool Matrix::isGpu() const {
auto rawPtr = m->mat.get();
return dynamic_cast<paddle::GpuMatrix*>(rawPtr) != nullptr ||
dynamic_cast<paddle::GpuSparseMatrix*>(rawPtr) != nullptr;
}
paddle/api/Paddle.i 已删除 100644 → 0
浏览文件 @ 0cef33a4
%module(directors="1") swig_paddle
%include "std_string.i"
%{
#define SWIG_FILE_WITH_INIT
#include "api/PaddleAPI.h"
%}
%include "exception.i"
%typemap(throws) UnsupportError %{
SWIG_exception(SWIG_RuntimeError, $1.what());
SWIG_fail;
%}
%include "std_vector.i"
%include "std_pair.i"
#ifdef SWIGPYTHON
%include "numpy.i"
#endif
%init %{
#ifdef SWIGPYTHON
import_array();
#endif
%}
namespace std {
%template(vector_int) vector<int>;
%template(vector_uint) vector<unsigned int>;
%template(vector_float) vector<float>;
%template(vector_string) vector<string>;
%template(vector_vec_star) vector<Vector*>;
}
#ifdef SWIGPYTHON
%typemap(in) (int argc, char** argv) {
int i = 0;
if (!PyList_Check($input)) {
PyErr_SetString(PyExc_ValueError, "Expecting a list");
return NULL;
}
$1 = PyList_Size($input);
$2 = (char **) malloc(($1+1)*sizeof(char *));
for (i = 0; i < $1; i++) {
PyObject *s = PyList_GetItem($input,i);
if (!PyString_Check(s)) {
free($2);
PyErr_SetString(PyExc_ValueError, "List items must be strings");
return NULL;
}
$2[i] = PyString_AsString(s);
}
$2[i] = 0;
}
%typemap(freearg) (int argc, char** argv) {
if ($2) free($2);
}
%typemap(out) FloatArray {
$result = PyList_New($1.length);
for (size_t i=0; i<$1.length; ++i) {
PyList_SetItem($result, i, PyFloat_FromDouble($1.buf[i]));
}
if($1.needFree) {
delete [] $1.buf;
}
}
%typemap(out) IntArray {
$result = PyList_New($1.length);
for (size_t i=0; i<$1.length; ++i) {
PyList_SetItem($result, i, PyInt_FromLong($1.buf[i]));
}
if ($1.needFree) {
delete [] $1.buf;
}
}
%typemap(out) IntWithFloatArray {
$result = PyList_New($1.length);
for (size_t i=0; i<$1.length; ++i) {
PyList_SetItem($result, i, PyTuple_Pack(2,
PyInt_FromLong($1.idxBuf[i]),
PyFloat_FromDouble($1.valBuf[i])
));
}
if ($1.needFree) {
delete [] $1.idxBuf;
delete [] $1.valBuf;
}
}
%rename(__getitem__) IVector::get;
%rename(__setitem__) IVector::set;
%rename(__len__) IVector::getSize;
%rename(__getitem__) Vector::get;
%rename(__setitem__) Vector::set;
%rename(__len__) Vector::getSize;
%rename(__len__) Parameter::getSize;
%rename(__call__) ParameterTraverseCallback::apply;
%rename(__repr__) Evaluator::toString;
%apply (float* INPLACE_ARRAY2, int DIM1, int DIM2) {
(float* data, int dim1, int dim2)
}
%apply (float** ARGOUTVIEW_ARRAY2, int* DIM1, int* DIM2) {
(float** view_data, int* dim1, int* dim2)
}
%apply (float** ARGOUTVIEWM_ARRAY2, int* DIM1, int* DIM2) {
(float** view_m_data, int* dim1, int* dim2)
}
%apply (int** ARGOUTVIEWM_ARRAY1, int* DIM1) {
(int** view_m_data, int* dim1)
}
%apply (int* INPLACE_ARRAY1, int DIM1) {
(int* data, int dim)
}
%apply (int** ARGOUTVIEW_ARRAY1, int* DIM1) {
(int** view_data, int* dim1)
}
%apply (float* INPLACE_ARRAY1, int DIM1) {
(float* data, int dim)
}
%apply (float** ARGOUTVIEW_ARRAY1, int* DIM1) {
(float** view_data, int* dim1)
}
%apply (float** ARGOUTVIEWM_ARRAY1, int* DIM1) {
(float** view_m_data, int* dim1)
}
#endif
// The below functions internally create object by "new", so it should use
// use SWIG to handle gc. There are hints for SWIG to handle GC.
%newobject Matrix::createZero;
%newobject Matrix::createSparse;
%newobject Matrix::createDense;
%newobject Matrix::createDenseFromNumpy;
%newobject Matrix::createCpuDenseFromNumpy;
%newobject Matrix::createGpuDenseFromNumpy;
%newobject Vector::createZero;
%newobject Vector::create;
%newobject Vector::createVectorFromNumpy;
%newobject Vector::createCpuVectorFromNumpy;
%newobject Vector::createGpuVectorFromNumpy;
%newobject IVector::createZero;
%newobject IVector::create;
%newobject IVector::createVectorFromNumpy;
%newobject IVector::createCpuVectorFromNumpy;
%newobject IVector::createGpuVectorFromNumpy;
%newobject Trainer::createByCommandLine;
%newobject Trainer::getForwardOutput;
%newobject Trainer::getLayerOutput;
%newobject Arguments::getSlotValue;
%newobject Arguments::getSlotIds;
%newobject Arguments::getSlotIn;
%newobject Arguments::getSlotSequenceStartPositions;
%newobject Arguments::getSlotSequenceDim;
%newobject Arguments::createArguments;
%newobject GradientMachine::createByConfigProtoStr;
%newobject GradientMachine::createByModelConfig;
%newobject GradientMachine::asSequenceGenerator;
%newobject GradientMachine::getParameter;
%newobject GradientMachine::getLayerOutput;
%newobject GradientMachine::makeEvaluator;
%newobject TrainerConfig::createFromTrainerConfigFile;
%newobject TrainerConfig::getModelConfig;
%newobject TrainerConfig::getOptimizationConfig;
%newobject Parameter::getBuf;
%newobject Parameter::getConfig;
%newobject ParameterOptimizer::create;
%newobject ParameterOptimizer::needSpecialTraversal;
%newobject ParameterUpdater::createLocalUpdater;
%newobject ParameterUpdater::createRemoteUpdater;
%newobject ParameterUpdater::createNewRemoteUpdater;
%feature("director") UpdateCallback;
%feature("autodoc", 1); // To generate method stub, for code hint in ide
// Ignore many private class, and method cannot be handled by swig.
%ignore MatrixPrivate;
%ignore TrainerPrivate;
%ignore IVector::operator[];
%ignore ArgumentsPrivate;
%ignore GradientMachinePrivate;
%ignore TrainerConfigPrivate;
%ignore ModelConfigPrivate;
%ignore ParameterPrivate;
%ignore SequenceGeneratorPrivate;
%ignore VectorPrivate;
%ignore ParameterConfigPrivate;
%ignore OptimizationConfigPrivate;
%ignore ParameterTraverseCallbackPrivate;
%include "utils/GlobalConstants.h"
%include "api/PaddleAPI.h"
paddle/api/PaddleAPI.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/api/PaddleAPIPrivate.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
/* 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 <memory>
#include "PaddleAPI.h"
#include "paddle/gserver/evaluators/Evaluator.h"
#include "paddle/gserver/gradientmachines/GradientMachine.h"
#include "paddle/parameter/ParameterUpdaterBase.h"
#include "paddle/trainer/TrainerConfigHelper.h"
struct GradientMachinePrivate {
std::shared_ptr<paddle::GradientMachine> machine;
template <typename T>
inline T& cast(void* ptr) {
return *(T*)(ptr);
}
};
struct OptimizationConfigPrivate {
std::shared_ptr<paddle::TrainerConfigHelper> trainer_config;
paddle::OptimizationConfig config;
const paddle::OptimizationConfig& getConfig() {
if (trainer_config != nullptr) {
return trainer_config->getOptConfig();
} else {
return config;
}
}
};
struct TrainerConfigPrivate {
std::shared_ptr<paddle::TrainerConfigHelper> conf;
TrainerConfigPrivate() {}
};
struct ModelConfigPrivate {
std::shared_ptr<paddle::TrainerConfigHelper> conf;
};
struct ArgumentsPrivate {
std::vector<paddle::Argument> outputs;
inline paddle::Argument& getArg(size_t idx) throw(RangeError) {
if (idx < outputs.size()) {
return outputs[idx];
} else {
RangeError e;
throw e;
}
}
template <typename T>
std::shared_ptr<T>& cast(void* rawPtr) const {
return *(std::shared_ptr<T>*)(rawPtr);
}
};
struct ParameterUpdaterPrivate {
std::unique_ptr<paddle::ParameterUpdater> updater;
};
struct ParameterPrivate {
std::shared_ptr<paddle::Parameter> sharedPtr;
paddle::Parameter* rawPtr; // rawPtr only used in ParameterUpdater,
// in other situation sharedPtr should
// contains value.
ParameterPrivate() : sharedPtr(nullptr), rawPtr(nullptr) {}
paddle::Parameter* getPtr() {
if (sharedPtr) {
return sharedPtr.get();
} else {
return rawPtr;
}
}
};
struct EvaluatorPrivate {
paddle::Evaluator* rawPtr;
EvaluatorPrivate() : rawPtr(nullptr) {}
~EvaluatorPrivate() { delete rawPtr; }
};
paddle/api/Parameter.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/parameter/Parameter.h"
#include "PaddleAPI.h"
#include "PaddleAPIPrivate.h"
Parameter::Parameter() : m(new ParameterPrivate()) {}
Parameter::~Parameter() { delete m; }
Parameter* Parameter::createFromRawPtr(void* ptr) {
auto p = new Parameter();
p->m->rawPtr = *static_cast<paddle::Parameter**>(ptr);
return p;
}
Parameter* Parameter::createFromSharedPtr(void* ptr) {
auto& p = *(paddle::ParameterPtr*)(ptr);
if (p == nullptr) {
return nullptr;
} else {
auto retParam = new Parameter();
retParam->m->sharedPtr = p;
return retParam;
}
}
std::string Parameter::getName() const { return m->getPtr()->getName(); }
Vector* Parameter::getBuf(ParameterType type) {
auto buf = m->getPtr()->getBuf(type);
return Vector::createByPaddleVectorPtr(&buf);
}
ParameterConfig* Parameter::getConfig() {
if (m->sharedPtr) {
return ParameterConfig::createParameterConfigFromParameterSharedPtr(
&m->sharedPtr);
} else {
return ParameterConfig::createParameterConfigFromParameterPtr(m->rawPtr);
}
}
size_t Parameter::getID() const { return m->getPtr()->getID(); }
void Parameter::setValueUpdated() { m->getPtr()->setValueUpdated(); }
bool Parameter::save(const std::string& filename) const {
return m->getPtr()->save(filename);
}
bool Parameter::load(const std::string& filename) const {
return m->getPtr()->load(filename);
}
size_t Parameter::getSize() const { return m->getPtr()->getSize(); }
paddle/api/ParameterOptimizer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/parameter/ParameterOptimizer.h"
#include <algorithm>
#include "Internal.h"
#include "PaddleAPI.h"
#include "PaddleAPIPrivate.h"
struct ParameterOptimizerPrivate {
std::unique_ptr<paddle::ParameterOptimizer> optimizer;
};
struct ParameterTraverseCallbackPrivate {
paddle::ParameterOptimizer::TraverseCallback callback;
ParameterTraverseCallbackPrivate() {}
ParameterTraverseCallbackPrivate(
const paddle::ParameterOptimizer::TraverseCallback& callback)
: callback(callback) {}
void apply(const std::vector<Vector*>& vecs,
const ParameterConfig& conf,
size_t sparseId) {
std::vector<paddle::VectorPtr> real_vecs;
real_vecs.resize(vecs.size());
std::transform(vecs.begin(), vecs.end(), real_vecs.begin(), [](Vector* v) {
if (v) {
return *(paddle::VectorPtr*)(v->getSharedPtr());
} else {
return paddle::VectorPtr();
}
});
paddle::ParameterConfig& real_conf =
*(paddle::ParameterConfig*)(const_cast<ParameterConfig&>(conf)
.getRawPtr());
callback(real_vecs.data(), real_conf, sparseId);
}
};
ParameterOptimizer::ParameterOptimizer() : m(new ParameterOptimizerPrivate()) {}
ParameterOptimizer::~ParameterOptimizer() { delete m; }
ParameterOptimizer* ParameterOptimizer::create(OptimizationConfig* config) {
CHECK(config != nullptr);
auto retOptimizer = new ParameterOptimizer();
retOptimizer->m->optimizer.reset(
paddle::ParameterOptimizer::create(config->m->getConfig(), false));
return retOptimizer;
}
void ParameterOptimizer::init(size_t numRows, const ParameterConfig* config) {
auto& conf = *(paddle::ParameterConfig*)(const_cast<ParameterConfig*>(config)
->getRawPtr());
m->optimizer->init(numRows, &conf);
}
void ParameterOptimizer::startPass() { m->optimizer->startPass(); }
void ParameterOptimizer::finishPass() { m->optimizer->finishPass(); }
void ParameterOptimizer::startBatch(size_t numSamplesProcessed) {
constexpr size_t high_1 = 1UL << (sizeof(size_t) * 8 - 1);
CHECK_EQ(numSamplesProcessed & high_1, 0UL); // Safely cast.
m->optimizer->startBatch((int64_t)numSamplesProcessed);
}
void ParameterOptimizer::finishBatch() { m->optimizer->finishBatch(); }
void ParameterOptimizer::update(const std::vector<Vector*>& vecs,
const ParameterConfig& conf,
size_t sparseId) {
ParameterTraverseCallbackPrivate invoker(
[&](const paddle::VectorPtr _vecs[],
const paddle::ParameterConfig& config,
size_t sid = -1UL) { m->optimizer->update(_vecs, config, sid); });
invoker.apply(vecs, conf, sparseId);
}
std::vector<int> ParameterOptimizer::getParameterTypes() const {
std::vector<int> returnValue;
staticCastVector(&returnValue, m->optimizer->getParameterTypes());
return returnValue;
}
ParameterTraverseCallback::ParameterTraverseCallback()
: m(new ParameterTraverseCallbackPrivate()) {}
ParameterTraverseCallback::~ParameterTraverseCallback() { delete m; }
void ParameterTraverseCallback::apply(const std::vector<Vector*>& vecs,
const ParameterConfig& conf,
size_t sparseId) {
m->apply(vecs, conf, sparseId);
}
ParameterTraverseCallback* ParameterOptimizer::needSpecialTraversal(
const ParameterConfig& config) const {
auto& param_config =
*(paddle::ParameterConfig*)const_cast<ParameterConfig&>(config)
.getRawPtr();
auto callback = m->optimizer->needSpecialTraversal(param_config);
if (callback) {
auto retCallback = new ParameterTraverseCallback();
retCallback->m->callback = callback;
return retCallback;
} else {
return nullptr;
}
}
paddle/api/SequenceGenerator.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
/* 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 <algorithm>
#include <iterator>
#include <sstream>
#include <vector>
#include "PaddleAPI.h"
#include "paddle/gserver/gradientmachines/GradientMachine.h"
#include "paddle/parameter/Argument.h"
#include "paddle/utils/Flags.h"
// used to represent partial sequence
struct Path {
std::vector<int> ids;
float logProb;
paddle::MachineState machineState;
Path() { logProb = 0; }
Path(std::vector<int>& ids, float logProb, paddle::MachineState& machineState)
: ids(ids), logProb(logProb), machineState(machineState) {}
bool operator<(const Path& other) const { return (logProb > other.logProb); }
};
// Return top k (k == beam_size) optimal paths using beam search. The last
// element of inArgs is the Argument of feedback. gradMachine has MaxIdLayer
// as output and outArgs thus stores top k labels and their probabilities per
// position
static void findNBest(paddle::GradientMachine* gradMachine,
std::vector<paddle::Argument>& inArgs,
std::vector<Path>& finalPaths,
size_t bos_id,
size_t eos_id,
size_t max_length) {
std::vector<Path> paths;
Path emptyPath;
paths.push_back(emptyPath);
finalPaths.clear();
gradMachine->resetState();
paddle::Argument feedback = inArgs.back();
feedback.ids->setElement(0, (int)(bos_id));
float minFinalPathLogProb = 0;
size_t beam = 0;
int id;
std::vector<paddle::Argument> outArgs;
while (true) { // iterate over each generated word
std::vector<Path> newPaths;
paddle::MachineState machineState;
for (size_t j = 0; j < paths.size(); j++) {
Path& path = paths[j];
if (path.machineState.size() > 0) {
gradMachine->setState(path.machineState);
feedback.ids->setElement(0, path.ids.back());
}
gradMachine->forward(inArgs, &outArgs, paddle::PASS_TEST);
gradMachine->getState(machineState);
beam = outArgs[0].ids->getSize();
for (size_t k = 0; k < beam; k++) {
id = outArgs[0].ids->getElement(k);
float prob = outArgs[0].in->getElement(0, k);
std::vector<int> nids(path.ids);
nids.push_back(id);
float newLogProb = path.logProb + log(prob);
Path newPath(nids, newLogProb, machineState);
if (id == (int)eos_id || nids.size() >= max_length) {
finalPaths.push_back(newPath);
if (minFinalPathLogProb > newPath.logProb) {
minFinalPathLogProb = newPath.logProb;
}
} else {
newPaths.push_back(newPath);
}
}
}
if (newPaths.size() == 0) {
break;
}
std::nth_element(newPaths.begin(),
newPaths.begin() + std::min(beam, newPaths.size()),
newPaths.end());
if (newPaths.size() > beam) {
newPaths.resize(beam);
}
// pathA < pathB means pathA.logProb > pathB.logProb
float maxPathLogProb =
std::min_element(newPaths.begin(), newPaths.end())->logProb;
if (finalPaths.size() >= beam && minFinalPathLogProb >= maxPathLogProb) {
break;
}
paths = newPaths;
} // end while
std::partial_sort(finalPaths.begin(),
finalPaths.begin() + std::min(beam, finalPaths.size()),
finalPaths.end());
if (finalPaths.size() > beam) {
finalPaths.resize(beam);
}
}
struct SequenceGeneratorPrivate {
std::shared_ptr<paddle::GradientMachine> machine;
std::shared_ptr<std::vector<std::string>> dict;
size_t beginPos;
size_t endPos;
size_t maxLength;
paddle::Argument feedback;
template <typename T>
inline T& cast(void* ptr) {
return *(T*)(ptr);
}
inline void findNBest(std::vector<paddle::Argument>& inArgs,
std::vector<Path>& path) {
::findNBest(machine.get(), inArgs, path, beginPos, endPos, maxLength);
}
SequenceGeneratorPrivate()
: dict(std::make_shared<std::vector<std::string>>()),
beginPos(0UL),
endPos(0UL),
maxLength(0UL),
feedback(__create_feedback__()) {}
private:
static paddle::Argument __create_feedback__() {
paddle::Argument feedback;
feedback.ids = paddle::IVector::create(/* size= */ 1, FLAGS_use_gpu);
feedback.sequenceStartPositions =
paddle::ICpuGpuVector::create(/* size= */ 2, /* useGpu= */ false);
feedback.sequenceStartPositions->getMutableData(false)[0] = 0;
feedback.sequenceStartPositions->getMutableData(false)[1] = 1;
return feedback;
}
};
SequenceGenerator::SequenceGenerator() : m(new SequenceGeneratorPrivate()) {}
SequenceGenerator::~SequenceGenerator() { delete m; }
class PathSequenceResults : public ISequenceResults {
// ISequenceResults interface
public:
PathSequenceResults(const std::shared_ptr<std::vector<Path>>& path,
const std::shared_ptr<std::vector<std::string>>& dict)
: path_(path), dict_(dict) {}
size_t getSize() const { return path_->size(); }
std::string getSentence(size_t id, bool split) const throw(RangeError) {
if (id < getSize()) {
Path& p = (*path_)[id];
std::ostringstream sout;
std::transform(p.ids.begin(),
p.ids.end(),
std::ostream_iterator<std::string>(sout, split ? " " : ""),
[&](int id) { return (*dict_)[id]; });
return sout.str();
} else {
RangeError e;
throw e;
}
}
std::vector<int> getSequence(size_t id) const throw(RangeError) {
if (id < getSize()) {
Path& p = (*path_)[id];
return p.ids;
} else {
RangeError e;
throw e;
}
}
float getScore(size_t id) const throw(RangeError) {
if (id < getSize()) {
Path& p = (*path_)[id];
return p.logProb;
} else {
RangeError e;
throw e;
}
}
private:
std::shared_ptr<std::vector<Path>> path_;
std::shared_ptr<std::vector<std::string>> dict_;
};
ISequenceResults* SequenceGenerator::generateSequence(
const Arguments& inArgs) const {
auto& in_args =
m->cast<std::vector<paddle::Argument>>(inArgs.getInternalArgumentsPtr());
for (auto& arg : in_args) {
arg.sequenceStartPositions = m->feedback.sequenceStartPositions;
}
in_args.push_back(m->feedback);
auto path = std::make_shared<std::vector<Path>>();
m->findNBest(in_args, *path);
return new PathSequenceResults(path, m->dict);
}
SequenceGenerator* SequenceGenerator::createByGradientMachineSharedPtr(
void* ptr) {
SequenceGenerator* r = new SequenceGenerator();
r->m->machine = r->m->cast<std::shared_ptr<paddle::GradientMachine>>(ptr);
return r;
}
void SequenceGenerator::setDict(const std::vector<std::string>& dict) {
*m->dict = dict;
}
void SequenceGenerator::setBos(size_t bos) { m->beginPos = bos; }
void SequenceGenerator::setEos(size_t eos) { m->endPos = eos; }
void SequenceGenerator::setMaxLength(size_t maxLength) {
m->maxLength = maxLength;
}
void SequenceGenerator::setBeamSize(size_t beamSize) {
if (beamSize != -1UL) {
FLAGS_beam_size = beamSize;
}
}
ISequenceResults::~ISequenceResults() {}
paddle/api/Trainer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
/* 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 "PaddleAPI.h"
#include "PaddleAPIPrivate.h"
#include <stdlib.h>
#include <atomic>
#include <memory>
#include "paddle/gserver/gradientmachines/NeuralNetwork.h"
#include "paddle/trainer/ParamUtil.h"
#include "paddle/trainer/Trainer.h"
#include "paddle/trainer/TrainerInternal.h"
#include "paddle/utils/Flags.h"
using paddle::real;
DECLARE_string(config);
DECLARE_string(init_model_path);
DECLARE_int32(start_pass);
struct TrainerPrivate : public paddle::Trainer {
bool _trainOneBatch(size_t batchSize);
bool forwardOneBatch(size_t batchSize);
void forwardOneDataBatch(const std::vector<paddle::Argument>& inArgs);
void setBatchSize(size_t batchSize);
std::vector<paddle::Argument>& getForwardOutput();
void startTestPeriod();
void finishTestPeriod();
void testOneDataBatch(const paddle::DataBatch& dataBatch);
TrainerPrivate() : paddle::Trainer() {}
};
Trainer::Trainer() : m(new TrainerPrivate()) {
auto conf = paddle::TrainerConfigHelper::createFromFlags();
if (conf != nullptr) {
m->init(conf);
}
}
Trainer::~Trainer() { delete m; }
Trainer* Trainer::createByCommandLine() throw(IOError) {
auto retv = new Trainer();
if (retv->m->getConfig().IsInitialized()) {
return retv;
} else {
throw IOError();
}
}
Trainer::Trainer(TrainerConfig* config, GradientMachine* gm)
: m(new TrainerPrivate()) {
m->init(config->m->conf, /* testing= */ false, gm ? gm->m->machine : nullptr);
}
Trainer* Trainer::create(TrainerConfig* config,
GradientMachine* gm) throw(IOError) {
auto retv = new Trainer(config, gm);
if (retv->m->getConfig().IsInitialized()) {
return retv;
} else {
retv->m->getConfig().CheckInitialized();
throw IOError();
}
}
void Trainer::startTrain() { m->startTrain(); }
void Trainer::finishTrain() { m->finishTrain(); }
void Trainer::startTrainPass() { m->startTrainPass(); }
void Trainer::finishTrainPass() { m->finishTrainPass(); }
void Trainer::trainOneDataBatch(size_t batchSize, const Arguments& inArgs) {
paddle::DataBatch dataBatch;
dataBatch.getStreams() = inArgs.m->outputs;
dataBatch.setSize(batchSize);
m->trainOneDataBatch(dataBatch);
}
bool Trainer::trainOneBatch(size_t batchSize) {
return m->_trainOneBatch(batchSize);
}
bool TrainerPrivate::_trainOneBatch(size_t batchSize) {
paddle::DataBatch dataBatch;
CHECK(dataProvider_) << "data_provider is not specified";
int num = dataProvider_->getNextBatch(batchSize, &dataBatch);
if (num == 0) {
return false;
}
trainOneDataBatch(dataBatch);
return false;
}
void TrainerPrivate::startTestPeriod() {
if (!tester_) {
createTester();
}
tester_->startTestPeriod();
}
void Trainer::startTestPeriod() { m->startTestPeriod(); }
void TrainerPrivate::testOneDataBatch(const paddle::DataBatch& dataBatch) {
tester_->testOneDataBatch(dataBatch, &forwardOutput_);
}
void Trainer::testOneDataBatch(size_t batchSize, const Arguments& args) {
paddle::DataBatch dataBatch;
dataBatch.getStreams() = args.m->outputs;
dataBatch.setSize(batchSize);
m->testOneDataBatch(dataBatch);
}
void TrainerPrivate::finishTestPeriod() { tester_->finishTestPeriod(); }
void Trainer::finishTestPeriod() { m->finishTestPeriod(); }
Arguments* Trainer::getLayerOutput(const std::string& layerName) const {
auto nn = this->m->getGradientMachine();
CHECK(nn) << "trainerInternal_.getGradientMachine() is not NeuralNetwork";
auto arg = nn->getLayerOutput(layerName);
return Arguments::createByPaddleArgument(&arg);
}
void Trainer::forwardOneBatch(size_t batchSize) {
m->forwardOneBatch(batchSize);
}
bool TrainerPrivate::forwardOneBatch(size_t batchSize) {
CHECK(dataProvider_) << "data_provider is not specified";
paddle::DataBatch dataBatch;
int num = dataProvider_->getNextBatch(batchSize, &dataBatch);
if (num == 0) {
return false;
}
forwardOneDataBatch(dataBatch.getStreams());
return true;
}
void TrainerPrivate::forwardOneDataBatch(
const std::vector<paddle::Argument>& inArgs) {
std::vector<paddle::Argument>& outArgs = forwardOutput_;
if (config_->getOptConfig().use_sparse_remote_updater()) {
trainerInternal_.getGradientMachine()->prefetch(inArgs);
trainerInternal_.getParameterUpdater()->getParametersRemote();
}
trainerInternal_.getGradientMachine()->forward(
inArgs, &outArgs, paddle::PASS_TEST);
}
Arguments* Trainer::getForwardOutput() {
return Arguments::createByPaddleArgumentVector(&m->getForwardOutput());
}
std::vector<paddle::Argument>& TrainerPrivate::getForwardOutput() {
return forwardOutput_;
}
paddle/api/Util.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
/* 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 "PaddleAPI.h"
#include "paddle/parameter/Parameter.h"
#include "paddle/utils/Common.h"
#include "paddle/utils/Flags.h"
#include "paddle/utils/PythonUtil.h"
#include "paddle/utils/Util.h"
#include <algorithm>
#include <iostream>
#include <iterator>
void initPaddle(int argc, char** argv) {
paddle::initMain(argc, argv);
paddle::initPython(argc, argv);
feenableexcept(FE_INVALID | FE_DIVBYZERO | FE_OVERFLOW);
}
FloatArray::FloatArray(const float* b, const size_t l)
: buf(b), length(l), needFree(false) {}
IntArray::IntArray(const int* b, const size_t l, bool f)
: buf(b), length(l), needFree(f) {}
IntWithFloatArray::IntWithFloatArray(const float* v,
const int* i,
size_t l,
bool f)
: valBuf(v), idxBuf(i), length(l), needFree(f) {}
bool isUsingGpu() { return FLAGS_use_gpu; }
void setUseGpu(bool useGpu) { FLAGS_use_gpu = useGpu; }
bool isGpuVersion() {
#ifndef PADDLE_WITH_CUDA
return false;
#else
return true;
#endif
}
int getTrainerCount() { return FLAGS_trainer_count; }
static_assert(NUM_PARAMETER_TYPES == paddle::NUM_PARAMETER_TYPES,
"The Parameter Type should be same in core/api and core/common");
paddle/api/Vector.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
/* 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 "PaddleAPI.h"
#include "paddle/math/Vector.h"
#include <cstring>
struct IVectorPrivate {
paddle::IVectorPtr vec;
};
IVector::IVector() : m(new IVectorPrivate()) {}
IVector* IVector::createZero(size_t sz, bool useGpu) {
auto v = new IVector();
v->m->vec = paddle::IVector::create(sz, useGpu);
v->m->vec->zeroMem();
return v;
}
IVector* IVector::create(const std::vector<int>& data, bool useGpu) {
auto v = new IVector();
v->m->vec = paddle::IVector::create(data.size(), useGpu);
v->m->vec->copyFrom(data.data(), data.size());
return v;
}
IVector* IVector::createVectorFromNumpy(int* data,
int dim,
bool copy,
bool useGpu) throw(UnsupportError) {
if (useGpu) {
/// if use gpu only copy=true is supported
if (!copy) {
throw UnsupportError("Gpu mode only supports copy=True");
}
return IVector::createGpuVectorFromNumpy(data, dim);
} else {
return IVector::createCpuVectorFromNumpy(data, dim, copy);
}
}
IVector* IVector::createCpuVectorFromNumpy(int* data, int dim, bool copy) {
auto v = new IVector();
if (copy) {
v->m->vec = paddle::IVector::create(dim, false);
v->m->vec->copyFrom(data, dim);
} else {
v->m->vec = paddle::IVector::create(data, dim, false);
}
return v;
}
IVector* IVector::createGpuVectorFromNumpy(int* data, int dim) {
auto v = new IVector();
v->m->vec = paddle::IVector::create(dim, true);
v->m->vec->copyFrom(data, dim);
return v;
}
bool IVector::isGpu() const {
return dynamic_cast<paddle::GpuIVector*>(m->vec.get()) != nullptr;
}
IntArray IVector::getData() const {
if (this->isGpu()) {
int* src = m->vec->getData();
size_t len = m->vec->getSize();
int* dest = new int[len];
hl_memcpy_device2host(dest, src, len * sizeof(int));
return IntArray(dest, len, true);
} else {
return IntArray(m->vec->getData(), m->vec->getSize());
}
}
int& IVector::operator[](const size_t idx) throw(RangeError, UnsupportError) {
if (this->isGpu()) {
UnsupportError e;
throw e;
} else {
if (idx >= m->vec->getSize()) {
RangeError e;
throw e;
}
}
return m->vec->getData()[idx];
}
const int& IVector::operator[](const size_t idx) const
throw(RangeError, UnsupportError) {
return (*const_cast<IVector*>(this))[idx];
}
IVector* IVector::createByPaddleVectorPtr(void* ptr) {
auto* p = (paddle::IVectorPtr*)ptr;
if ((*p) != nullptr) {
IVector* vec = new IVector();
vec->m->vec = *p;
return vec;
} else {
return nullptr;
}
}
IVector::~IVector() { delete m; }
void* IVector::getSharedPtr() const { return &m->vec; }
size_t IVector::getSize() const { return m->vec->getSize(); }
void IVector::toNumpyArrayInplace(int** data, int* dim1) throw(UnsupportError) {
auto v = std::dynamic_pointer_cast<paddle::CpuIVector>(m->vec);
if (v) {
*data = v->getData();
*dim1 = v->getSize();
} else {
throw UnsupportError();
}
}
void IVector::copyToNumpyArray(int** view_m_data, int* dim1) {
*dim1 = m->vec->getSize();
*view_m_data = new int[*dim1];
if (auto cpuVec = dynamic_cast<paddle::CpuIVector*>(m->vec.get())) {
std::memcpy(*view_m_data, cpuVec->getData(), sizeof(int) * (*dim1));
} else if (auto gpuVec = dynamic_cast<paddle::GpuIVector*>(m->vec.get())) {
hl_memcpy_device2host(
*view_m_data, gpuVec->getData(), sizeof(int) * (*dim1));
} else {
LOG(INFO) << "Unexpected situation";
}
}
void IVector::copyFromNumpyArray(int* data, int dim) {
m->vec->resize(dim);
m->vec->copyFrom(data, dim);
}
struct VectorPrivate {
paddle::VectorPtr vec;
void safeAccessData(const size_t idx,
const std::function<void(float&)>& func) const
throw(RangeError, UnsupportError) {
auto cpuVec = std::dynamic_pointer_cast<const paddle::CpuVector>(vec);
if (cpuVec != nullptr) {
if (idx < vec->getSize()) {
func(vec->getData()[idx]);
} else {
throw RangeError();
}
} else {
throw UnsupportError();
}
}
};
Vector::Vector() : m(new VectorPrivate()) {}
Vector::~Vector() { delete m; }
Vector* Vector::createZero(size_t sz, bool useGpu) {
auto retVec = new Vector();
retVec->m->vec = paddle::Vector::create(sz, useGpu);
retVec->m->vec->zero();
return retVec;
}
Vector* Vector::create(const std::vector<float>& data, bool useGpu) {
auto retVec = new Vector();
retVec->m->vec = paddle::Vector::create(data.size(), useGpu);
retVec->m->vec->copyFrom(data.data(), data.size());
return retVec;
}
Vector* Vector::createByPaddleVectorPtr(void* ptr) {
auto& v = *(paddle::VectorPtr*)(ptr);
if (v == nullptr) {
return nullptr;
} else {
auto retVec = new Vector();
retVec->m->vec = v;
return retVec;
}
}
Vector* Vector::createVectorFromNumpy(float* data,
int dim,
bool copy,
bool useGpu) throw(UnsupportError) {
if (useGpu) {
/// if use gpu only copy=True is supported
if (!copy) {
throw UnsupportError("Gpu mode only supports copy=True");
}
return Vector::createGpuVectorFromNumpy(data, dim);
} else {
return Vector::createCpuVectorFromNumpy(data, dim, copy);
}
}
Vector* Vector::createCpuVectorFromNumpy(float* data, int dim, bool copy) {
CHECK_GT(dim, 0);
auto retVec = new Vector();
if (copy) {
retVec->m->vec = paddle::Vector::create((size_t)dim, false);
retVec->m->vec->copyFrom(data, dim);
} else {
retVec->m->vec = paddle::Vector::create(data, (size_t)dim, false);
}
return retVec;
}
Vector* Vector::createGpuVectorFromNumpy(float* data, int dim) {
CHECK_GT(dim, 0);
auto retVec = new Vector();
retVec->m->vec = paddle::Vector::create((size_t)dim, true);
retVec->m->vec->copyFrom(data, (size_t)dim);
return retVec;
}
void Vector::toNumpyArrayInplace(float** view_data,
int* dim1) throw(UnsupportError) {
auto v = std::dynamic_pointer_cast<paddle::CpuVector>(m->vec);
if (v != nullptr) {
*view_data = v->getData();
*dim1 = (int)v->getSize();
} else {
throw UnsupportError();
}
}
void Vector::copyToNumpyArray(float** view_m_data, int* dim1) {
*dim1 = m->vec->getSize();
*view_m_data = new float[*dim1];
if (auto cpuVec = dynamic_cast<paddle::CpuVector*>(m->vec.get())) {
std::memcpy(*view_m_data, cpuVec->getData(), sizeof(float) * (*dim1));
} else if (auto gpuVec = dynamic_cast<paddle::GpuVector*>(m->vec.get())) {
hl_memcpy_device2host(
*view_m_data, gpuVec->getData(), sizeof(float) * (*dim1));
} else {
LOG(INFO) << "Unexpected situation";
}
}
void Vector::copyFromNumpyArray(float* data, int dim) {
m->vec->resize(dim);
m->vec->copyFrom(data, dim);
}
FloatArray Vector::getData() const {
if (this->isGpu()) {
float* src = m->vec->getData();
size_t len = m->vec->getSize();
float* dest = new float[len];
hl_memcpy_device2host(dest, src, len * sizeof(float));
FloatArray ret_val(dest, len);
ret_val.needFree = true;
return ret_val;
} else {
FloatArray ret_val(m->vec->getData(), m->vec->getSize());
return ret_val;
}
}
void Vector::copyFrom(Vector* src) throw(RangeError) {
if (src->m->vec->getSize() != m->vec->getSize()) {
throw RangeError();
}
m->vec->copyFrom(*src->m->vec);
}
bool Vector::isGpu() const {
return std::dynamic_pointer_cast<paddle::GpuVector>(m->vec) != nullptr;
}
float Vector::get(const size_t idx) const throw(RangeError, UnsupportError) {
float r;
m->safeAccessData(idx, [&](float& o) { r = o; });
return r;
}
void Vector::set(const size_t idx, float val) throw(RangeError,
UnsupportError) {
m->safeAccessData(idx, [&](float& o) { o = val; });
}
size_t Vector::getSize() const { return m->vec->getSize(); }
void* Vector::getSharedPtr() { return &m->vec; }
paddle/capi/capi_private.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
/* 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 "capi.h"
#include "paddle/gserver/gradientmachines/GradientMachine.h"
#include "paddle/math/Matrix.h"
#include "paddle/math/Vector.h"
#include "paddle/parameter/Argument.h"
#pragma once
namespace paddle {
namespace capi {
enum CType { kIVECTOR = 0, kMATRIX, kARGUMENTS, kGRADIENT_MACHINE };
#define STRUCT_HEADER CType type;
struct CHeader {
STRUCT_HEADER
};
struct CIVector {
STRUCT_HEADER
IVectorPtr vec;
CIVector() : type(kIVECTOR) {}
};
struct CMatrix {
STRUCT_HEADER
MatrixPtr mat;
CMatrix() : type(kMATRIX) {}
};
struct CArguments {
STRUCT_HEADER
std::vector<paddle::Argument> args;
CArguments() : type(kARGUMENTS) {}
template <typename T>
paddle_error accessSeqPos(uint64_t ID, uint32_t nestedLevel, T callback) {
if (ID >= args.size()) return kPD_OUT_OF_RANGE;
switch (nestedLevel) {
case 0:
callback(args[ID].sequenceStartPositions);
break;
case 1:
callback(args[ID].subSequenceStartPositions);
break;
default:
return kPD_OUT_OF_RANGE;
}
return kPD_NO_ERROR;
}
};
struct CGradientMachine {
STRUCT_HEADER
paddle::GradientMachinePtr machine;
CGradientMachine() : type(kGRADIENT_MACHINE) {}
};
template <typename T>
inline T* cast(void* ptr) {
return reinterpret_cast<T*>(ptr);
}
} // namespace capi
} // namespace paddle
paddle/capi/gradient_machine.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
/* 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 "gradient_machine.h"
#include "capi_private.h"
#include "paddle/gserver/gradientmachines/NeuralNetwork.h"
#define cast(v) paddle::capi::cast<paddle::capi::CGradientMachine>(v)
enum GradientMatchineCreateMode {
CREATE_MODE_NORMAL = 0,
CREATE_MODE_TESTING = 4
};
namespace paddle {
class MyNeuralNetwork : public NeuralNetwork {
public:
MyNeuralNetwork(const std::string& name, NeuralNetwork* network)
: NeuralNetwork(name, network) {}
};
NeuralNetwork* newCustomNerualNetwork(const std::string& name,
NeuralNetwork* network) {
return new MyNeuralNetwork(name, network);
}
} // namespace paddle
extern "C" {
paddle_error paddle_gradient_machine_create_for_inference(
paddle_gradient_machine* machine, void* modelConfigProtobuf, int size) {
if (modelConfigProtobuf == nullptr) return kPD_NULLPTR;
paddle::ModelConfig config;
if (!config.ParseFromArray(modelConfigProtobuf, size) ||
!config.IsInitialized()) {
return kPD_PROTOBUF_ERROR;
}
auto ptr = new paddle::capi::CGradientMachine();
ptr->machine.reset(paddle::GradientMachine::create(
config, CREATE_MODE_TESTING, {paddle::PARAMETER_VALUE}));
*machine = ptr;
return kPD_NO_ERROR;
}
paddle_error paddle_gradient_machine_create_for_inference_with_parameters(
paddle_gradient_machine* machine, void* mergedModel, uint64_t size) {
if (mergedModel == nullptr) return kPD_NULLPTR;
std::istringstream is(std::string(static_cast<char*>(mergedModel), size));
int64_t modelConfigSize = 0;
is.read((char*)(&modelConfigSize), sizeof(modelConfigSize));
std::string modelConfigProtobuf;
modelConfigProtobuf.resize(modelConfigSize);
is.read(&modelConfigProtobuf[0], modelConfigSize);
paddle::TrainerConfig config;
paddle::ModelConfig modelConfig;
if (!config.ParseFromString(modelConfigProtobuf) || !config.IsInitialized()) {
if (!modelConfig.ParseFromString(modelConfigProtobuf) ||
!modelConfig.IsInitialized()) {
return kPD_PROTOBUF_ERROR;
}
} else {
modelConfig = config.model_config();
}
auto ptr = new paddle::capi::CGradientMachine();
ptr->machine.reset(paddle::GradientMachine::create(
modelConfig, CREATE_MODE_TESTING, {paddle::PARAMETER_VALUE}));
std::vector<paddle::ParameterPtr>& parameters = ptr->machine->getParameters();
for (auto& para : parameters) {
para->load(is);
}
*machine = ptr;
return kPD_NO_ERROR;
}
paddle_error paddle_gradient_machine_destroy(paddle_gradient_machine machine) {
delete cast(machine);
return kPD_NO_ERROR;
}
paddle_error paddle_gradient_machine_load_parameter_from_disk(
paddle_gradient_machine machine, const char* path) {
auto m = cast(machine);
if (m == nullptr || path == nullptr || m->machine == nullptr)
return kPD_NULLPTR;
m->machine->loadParameters(path);
return kPD_NO_ERROR;
}
paddle_error paddle_gradient_machine_forward(paddle_gradient_machine machine,
paddle_arguments inArgs,
paddle_arguments outArgs,
bool isTrain) {
auto m = cast(machine);
auto in = paddle::capi::cast<paddle::capi::CArguments>(inArgs);
auto out = paddle::capi::cast<paddle::capi::CArguments>(outArgs);
if (m == nullptr || in == nullptr || out == nullptr || m->machine == nullptr)
return kPD_NULLPTR;
m->machine->forward(
in->args, &out->args, isTrain ? paddle::PASS_TRAIN : paddle::PASS_TEST);
return kPD_NO_ERROR;
}
paddle_error paddle_gradient_machine_create_shared_param(
paddle_gradient_machine origin,
void* modelConfigProtobuf,
int size,
paddle_gradient_machine* slave) {
auto o = cast(origin);
if (origin == nullptr || slave == nullptr || o->machine == nullptr) {
return kPD_NULLPTR;
}
paddle::ModelConfig config;
if (!config.ParseFromArray(modelConfigProtobuf, size) ||
!config.IsInitialized()) {
return kPD_PROTOBUF_ERROR;
}
std::unique_ptr<paddle::capi::CGradientMachine> ptr(
new paddle::capi::CGradientMachine());
auto nn = paddle::NeuralNetwork::create(config);
nn->init(config,
[&o](int paramId, paddle::Parameter* param) {
auto p = o->machine->getParameters()[paramId];
param->enableSharedType(paddle::PARAMETER_VALUE,
p->getBuf(paddle::PARAMETER_VALUE));
},
{paddle::PARAMETER_VALUE},
false);
ptr->machine.reset(nn);
*slave = ptr.release();
return kPD_NO_ERROR;
}
}
paddle_error paddle_gradient_machine_randomize_param(
paddle_gradient_machine machine) {
auto m = cast(machine);
if (m == nullptr || m->machine == nullptr) return kPD_NULLPTR;
m->machine->randParameters();
return kPD_NO_ERROR;
}
paddle_error paddle_gradient_machine_get_layer_output(
paddle_gradient_machine machine,
const char* layerName,
paddle_arguments args) {
auto m = cast(machine);
auto out = paddle::capi::cast<paddle::capi::CArguments>(args);
if (m == nullptr || layerName == nullptr || out == nullptr ||
m->machine == nullptr) {
return kPD_NULLPTR;
}
auto layerOutput = m->machine->getLayerOutput(layerName);
out->args.push_back(layerOutput);
return kPD_NO_ERROR;
}
paddle_error paddle_gradient_machine_release_layer_output(
paddle_gradient_machine machine) {
auto m = cast(machine);
if (m == nullptr || m->machine == nullptr) {
return kPD_NULLPTR;
}
m->machine->releaseOutput();
return kPD_NO_ERROR;
}
paddle/capi/tests/test_GradientMachine.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
/* 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 <gtest/gtest.h>
#include <paddle/gserver/gradientmachines/GradientMachine.h>
#include <paddle/trainer/TrainerConfigHelper.h>
#include <stdlib.h>
#include <string.h>
#include <type_traits>
#include "capi.h"
#include "paddle/utils/ThreadLocal.h"
static std::vector<paddle_real> randomBuffer(size_t bufSize) {
auto& eng = paddle::ThreadLocalRandomEngine::get();
std::uniform_real_distribution<paddle_real> dist(-1.0, 1.0);
std::vector<paddle_real> retv;
retv.reserve(bufSize);
for (size_t i = 0; i < bufSize; ++i) {
retv.push_back(dist(eng));
}
return retv;
}
TEST(GradientMachine, testPredict) {
//! TODO(yuyang18): Test GPU Code.
paddle::TrainerConfigHelper config("./test_predict_network.py");
std::string buffer;
ASSERT_TRUE(config.getModelConfig().SerializeToString(&buffer));
paddle_gradient_machine machine;
ASSERT_EQ(kPD_NO_ERROR,
paddle_gradient_machine_create_for_inference(
&machine, &buffer[0], (int)buffer.size()));
std::unique_ptr<paddle::GradientMachine> gm(
paddle::GradientMachine::create(config.getModelConfig()));
ASSERT_NE(nullptr, gm);
gm->randParameters();
gm->saveParameters("./");
ASSERT_EQ(kPD_NO_ERROR,
paddle_gradient_machine_load_parameter_from_disk(machine, "./"));
paddle_gradient_machine machineSlave;
ASSERT_EQ(kPD_NO_ERROR,
paddle_gradient_machine_create_shared_param(
machine, &buffer[0], (int)buffer.size(), &machineSlave));
std::swap(machineSlave, machine);
paddle_arguments outArgs = paddle_arguments_create_none();
paddle_arguments inArgs = paddle_arguments_create_none();
ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_resize(inArgs, 1));
paddle_matrix mat = paddle_matrix_create(1, 100, false);
static_assert(std::is_same<paddle_real, paddle::real>::value, "");
auto data = randomBuffer(100);
paddle_real* rowPtr;
ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_row(mat, 0, &rowPtr));
memcpy(rowPtr, data.data(), data.size() * sizeof(paddle_real));
ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_set_value(inArgs, 0, mat));
ASSERT_EQ(kPD_NO_ERROR,
paddle_gradient_machine_forward(machine, inArgs, outArgs, false));
uint64_t sz;
ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_get_size(outArgs, &sz));
ASSERT_EQ(1UL, sz);
ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_get_value(outArgs, 0, mat));
std::vector<paddle::Argument> paddleInArgs;
std::vector<paddle::Argument> paddleOutArgs;
paddleInArgs.resize(1);
paddleInArgs[0].value =
paddle::Matrix::create(data.data(), 1, 100, false, false);
gm->forward(paddleInArgs, &paddleOutArgs, paddle::PASS_TEST);
auto matPaddle = paddleOutArgs[0].value;
uint64_t height, width;
ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_shape(mat, &height, &width));
ASSERT_EQ(matPaddle->getHeight(), height);
ASSERT_EQ(matPaddle->getWidth(), width);
ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_get_row(mat, 0, &rowPtr));
for (size_t i = 0; i < width; ++i) {
ASSERT_NEAR(matPaddle->getData()[i], rowPtr[i], 1e-5);
}
ASSERT_EQ(kPD_NO_ERROR, paddle_matrix_destroy(mat));
ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_destroy(inArgs));
ASSERT_EQ(kPD_NO_ERROR, paddle_arguments_destroy(outArgs));
std::swap(machineSlave, machine);
ASSERT_EQ(kPD_NO_ERROR, paddle_gradient_machine_destroy(machineSlave));
ASSERT_EQ(kPD_NO_ERROR, paddle_gradient_machine_destroy(machine));
}
int main(int argc, char** argv) {
testing::InitGoogleTest(&argc, argv);
std::vector<char*> argvs;
argvs.push_back(strdup("--use_gpu=false"));
paddle_init((int)argvs.size(), argvs.data());
for (auto each : argvs) {
free(each);
}
return RUN_ALL_TESTS();
}
paddle/contrib/inference/CMakeLists.txt
浏览文件 @ ee760d1c
......@@ -46,6 +46,10 @@ cc_library(paddle_inference_api
SRCS paddle_inference_api.cc paddle_inference_api_impl.cc
DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
cc_library(paddle_inference_api_shared SHARED
SRCS paddle_inference_api.cc paddle_inference_api_impl.cc
DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
cc_test(test_paddle_inference_api
SRCS test_paddle_inference_api.cc
DEPS paddle_inference_api)
......
paddle/cuda/include/hl_base.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
/* 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 <cstddef>
#ifdef PADDLE_TYPE_DOUBLE
#define HL_FLOAT_MAX 3.40282347e+38F
#define HL_FLOAT_MIN 1.17549435e-38F
using real = double;
#else
#define HL_FLOAT_MAX 1.7976931348623157e+308
#define HL_FLOAT_MIN 2.2250738585072014e-308
using real = float;
#endif
/**
* The maximum input value for exp, used to avoid overflow problem.
* currently only used for tanh function.
*/
#define EXP_MAX_INPUT 40.0
/**
* @brief DIVUP(x, y) is similar to ceil(x / y).
* @note For CUDA, DIVUP will be used to specify
* the size of blockDim.
*/
#ifndef DIVUP
#define DIVUP(x, y) (((x) + (y)-1) / (y))
#endif
/**
* HPPL is an internal high performance parallel computing library
* for high-level neural network routines, which can support many
* heterogeneous compute architectures, such as GPU, FPGA, etc.
*/
/**
* @brief HPPL CUDA Stream.
*
* @note Each thread can use HPPL_STREAM_* after calling hl_init.
* HPPL_STREAM_DEFAULT is HPPL default stream.
*/
typedef enum {
HPPL_STREAM_DEFAULT = 0, /* Thread Default Stream*/
HPPL_STREAM_1 = 1,
HPPL_STREAM_2 = 2,
HPPL_STREAM_3 = 3,
HPPL_STREAM_4 = 4,
HPPL_THREAD_STREAM_1 = 5,
HPPL_THREAD_STREAM_2 = 6,
HPPL_THREAD_STREAM_3 = 7,
HPPL_THREAD_STREAM_4 = 8,
HPPL_STREAM_END
} hl_stream_t;
/**
* @brief HPPL activation mode.
*/
typedef enum {
HL_ACTIVATION_SIGMOID = 0,
HL_ACTIVATION_RELU = 1,
HL_ACTIVATION_TANH = 2,
HL_ACTIVATION_LINEAR = 3,
HL_ACTIVATION_END
} hl_activation_mode_t;
/**
* @brief Transpose type.
*/
typedef enum {
HPPL_OP_N = 0, /* transpose */
HPPL_OP_T = 1, /* non transpose */
HPPL_OP_END
} hl_trans_op_t;
/**
* @brief Lstm value.
*
* @param gateValue input value.
* @param prevStateValue previous state value.
* @param stateValue state value.
* @param stateActiveValue state active value.
* @param outputValue output value.
*/
typedef struct {
real *gateValue;
real *prevStateValue;
real *stateValue;
real *stateActiveValue;
real *outputValue;
real *checkIg;
real *checkFg;
real *checkOg;
} hl_lstm_value;
/**
* @brief Lstm gradient.
*
* @param gateGrad input gradient.
* @param prevStateGrad previous state gradient.
* @param stateGrad state gradient.
* @param stateActiveGrad state active gradient.
* @param outputGrad output gradient.
*/
typedef struct {
real *gateGrad;
real *prevStateGrad;
real *stateGrad;
real *stateActiveGrad;
real *outputGrad;
real *checkIgGrad;
real *checkFgGrad;
real *checkOgGrad;
} hl_lstm_grad;
/**
* @brief Gru value.
*
* @param gateWeight gate weight (updateGate + resetGate).
* @param stateWeight frame state weight.
* @param gateValue gate value results.
* @param resetOutputValue resetOutput value.
* @param outputValue output value.
* @param prevOutValue previous output value.
*
*/
typedef struct {
real *gateWeight;
real *stateWeight;
real *gateValue;
real *resetOutputValue;
real *outputValue;
real *prevOutValue;
} hl_gru_value;
/**
* @brief Gru gradient.
*
* @param gateWeightGrad gate weight gradient.
* @param stateWeightGrad frame state weight gradient.
* @param gateGrad gate gradient results.
* @param resetOutputGrad resetOutput gradient.
* @param outputGrad output gradient.
* @param prevOutGrad previous output gradient.
*/
typedef struct {
real *gateWeightGrad;
real *stateWeightGrad;
real *gateGrad;
real *resetOutputGrad;
real *outputGrad;
real *prevOutGrad;
} hl_gru_grad;
/**
* @brief Sparse matrix value type.
*/
typedef enum {
HL_NO_VALUE = 0, /* matrix values only 0 or 1 */
HL_FLOAT_VALUE = 1,
HL_VALUE_END
} hl_matrix_value_t;
/**
* @brief HPPL matrix format.
*/
typedef enum {
HL_SPARSE_CSR = 0,
HL_SPARSE_CSC = 1,
HL_SPARSE_END
} hl_matrix_format_t;
typedef struct _hl_matrix_s *hl_matrix_s;
/**
* @brief HPPL sparse matrix.
*
* @param matrix sparse matrix.
* @param format matrix format.
* @param type the type of matrix values.
* @param rows matrix rows.
* @param cols matrix columns.
* @param nnz nonzero values of sparse matrix.
*/
typedef struct {
hl_matrix_s matrix;
hl_matrix_format_t format;
hl_matrix_value_t type;
int rows;
int cols;
size_t nnz;
} _hl_sparse_matrix_s, *hl_sparse_matrix_s;
#ifdef __NVCC__
#include <cuda_runtime.h>
#include "paddle/cuda/include/hl_cuda.h"
#include "paddle/utils/Logging.h"
extern __thread bool g_sync_flag;
extern __thread cudaStream_t default_stream;
#define STREAM_DEFAULT default_stream
/**
* @brief Check cuda kernel execution.
* @param msg error string
*/
#define CHECK_SYNC(msg) \
if (true == g_sync_flag) { \
hl_stream_synchronize(HPPL_STREAM_DEFAULT); \
cudaError_t err = (cudaError_t)hl_get_device_last_error(); \
CHECK_EQ(cudaSuccess, err) \
<< "[" << msg << "] " \
<< "CUDA error: " << hl_get_device_error_string((size_t)err); \
}
// __shfl has been deprecated as of CUDA 9.0.
#if CUDA_VERSION < 9000
template <typename T>
__forceinline__ __device__ T __shfl_down_sync(unsigned, T val, int delta) {
return __shfl_down(val, delta);
}
template <typename T>
__forceinline__ __device__ T
__shfl_sync(unsigned, T val, int src_line, int width) {
return __shfl(val, src_line, width);
}
#define CREATE_SHFL_MASK(mask, predicate) mask = 0u;
#else
#define FULL_WARP_MASK 0xFFFFFFFF
#define CREATE_SHFL_MASK(mask, predicate) \
mask = __ballot_sync(FULL_WARP_MASK, (predicate))
#endif
#endif // __NVCC__
paddle/cuda/src/hl_top_k.cu 已删除 100644 → 0
浏览文件 @ 0cef33a4
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/cuda/include/hl_base.h"
#include "paddle/cuda/include/hl_sparse.ph"
#include "paddle/cuda/include/hl_top_k.h"
#include "paddle/utils/Logging.h"
// using namespace hppl;
struct Pair {
__device__ __forceinline__ Pair() {}
__device__ __forceinline__ Pair(real value, int id) : v_(value), id_(id) {}
__device__ __forceinline__ void set(real value, int id) {
v_ = value;
id_ = id;
}
__device__ __forceinline__ void operator=(const Pair& in) {
v_ = in.v_;
id_ = in.id_;
}
__device__ __forceinline__ bool operator<(const real value) const {
return (v_ < value);
}
__device__ __forceinline__ bool operator<(const Pair& in) const {
return (v_ < in.v_) || ((v_ == in.v_) && (id_ > in.id_));
}
__device__ __forceinline__ bool operator>(const Pair& in) const {
return (v_ > in.v_) || ((v_ == in.v_) && (id_ < in.id_));
}
real v_;
int id_;
};
__device__ __forceinline__ void addTo(Pair topK[],
const Pair& p,
int beamSize) {
for (int k = beamSize - 2; k >= 0; k--) {
if (topK[k] < p) {
topK[k + 1] = topK[k];
} else {
topK[k + 1] = p;
return;
}
}
topK[0] = p;
}
template <int beamSize>
__device__ __forceinline__ void addTo(Pair topK[], const Pair& p) {
for (int k = beamSize - 2; k >= 0; k--) {
if (topK[k] < p) {
topK[k + 1] = topK[k];
} else {
topK[k + 1] = p;
return;
}
}
topK[0] = p;
}
template <int blockSize>
__device__ __forceinline__ void getTopK(
Pair topK[], real* src, int idx, int dim, int beamSize) {
while (idx < dim) {
if (topK[beamSize - 1] < src[idx]) {
Pair tmp(src[idx], idx);
addTo(topK, tmp, beamSize);
}
idx += blockSize;
}
}
template <int blockSize>
__device__ __forceinline__ void getTopK(
Pair topK[], real* src, int idx, int dim, const Pair& max, int beamSize) {
while (idx < dim) {
if (topK[beamSize - 1] < src[idx]) {
Pair tmp(src[idx], idx);
if (tmp < max) {
addTo(topK, tmp, beamSize);
}
}
idx += blockSize;
}
}
template <int blockSize>
__device__ __forceinline__ void getTopK(
Pair topK[], real* val, int* col, int idx, int dim, int beamSize) {
while (idx < dim) {
if (topK[beamSize - 1] < val[idx]) {
Pair tmp(val[idx], col[idx]);
addTo(topK, tmp, beamSize);
}
idx += blockSize;
}
}
template <int blockSize>
__device__ __forceinline__ void getTopK(Pair topK[],
real* val,
int* col,
int idx,
int dim,
const Pair& max,
int beamSize) {
while (idx < dim) {
if (topK[beamSize - 1] < val[idx]) {
Pair tmp(val[idx], col[idx]);
if (tmp < max) {
addTo(topK, tmp, beamSize);
}
}
idx += blockSize;
}
}
template <int maxLength, int blockSize>
__device__ __forceinline__ void threadGetTopK(Pair topK[],
int& beam,
int beamSize,
real* src,
bool& firstStep,
bool& isEmpty,
Pair& max,
int dim,
const int tid) {
if (beam > 0) {
int length = beam < beamSize ? beam : beamSize;
if (firstStep) {
firstStep = false;
getTopK<blockSize>(topK, src, tid, dim, length);
} else {
for (int k = 0; k < maxLength; k++) {
if (k < maxLength - beam) {
topK[k] = topK[k + beam];
} else {
topK[k].set(-HL_FLOAT_MAX, -1);
}
}
if (!isEmpty) {
getTopK<blockSize>(topK + maxLength - beam, src, tid, dim, max, length);
}
}
max = topK[maxLength - 1];
if (max.id_ == -1) isEmpty = true;
beam = 0;
}
}
template <int maxLength, int blockSize>
__device__ __forceinline__ void threadGetTopK(Pair topK[],
int& beam,
int beamSize,
real* val,
int* col,
bool& firstStep,
bool& isEmpty,
Pair& max,
int dim,
const int tid) {
if (beam > 0) {
int length = beam < beamSize ? beam : beamSize;
if (firstStep) {
firstStep = false;
getTopK<blockSize>(topK, val, col, tid, dim, length);
} else {
for (int k = 0; k < maxLength; k++) {
if (k < maxLength - beam) {
topK[k] = topK[k + beam];
} else {
topK[k].set(-HL_FLOAT_MAX, -1);
}
}
if (!isEmpty) {
getTopK<blockSize>(
topK + maxLength - beam, val, col, tid, dim, max, length);
}
}
max = topK[maxLength - 1];
if (max.id_ == -1) isEmpty = true;
beam = 0;
}
}
template <int maxLength, int blockSize>
__device__ __forceinline__ void blockReduce(Pair* shTopK,
int* maxId,
Pair topK[],
real** topVal,
int** topIds,
int& beam,
int& beamSize,
const int tid,
const int warp) {
while (true) {
__syncthreads();
if (tid < blockSize / 2) {
if (shTopK[tid] < shTopK[tid + blockSize / 2]) {
maxId[tid] = tid + blockSize / 2;
} else {
maxId[tid] = tid;
}
}
__syncthreads();
for (int stride = blockSize / 4; stride > 0; stride = stride / 2) {
if (tid < stride) {
if (shTopK[maxId[tid]] < shTopK[maxId[tid + stride]]) {
maxId[tid] = maxId[tid + stride];
}
}
__syncthreads();
}
__syncthreads();
if (tid == 0) {
**topVal = shTopK[maxId[0]].v_;
**topIds = shTopK[maxId[0]].id_;
(*topVal)++;
(*topIds)++;
}
if (tid == maxId[0]) beam++;
if (--beamSize == 0) break;
__syncthreads();
// NOTE(zcd): temporary solution
unsigned mask = 0u;
CREATE_SHFL_MASK(mask, true);
if (tid == maxId[0]) {
if (beam < maxLength) {
shTopK[tid] = topK[beam];
}
}
if (maxId[0] / 32 == warp) {
if (__shfl_sync(mask, beam, (maxId[0]) % 32, 32) == maxLength) break;
}
}
}
/**
* Each block compute one sample.
* In a block:
* 1. every thread get top maxLength value;
* 2. merge to shTopK, block reduce and get max value;
* 3. go to the second setp, until one thread's topK value is null;
* 4. go to the first setp, until get the topK value.
*/
template <int maxLength, int blockSize>
__global__ void KeMatrixTopK(real* topVal,
int ldv,
int* topIds,
real* src,
int lds,
int dim,
int beamSize) {
__shared__ Pair shTopK[blockSize];
__shared__ int maxId[blockSize / 2];
const int tid = threadIdx.x;
const int warp = threadIdx.x / 32;
src += blockIdx.x * lds;
topVal += blockIdx.x * ldv;
topIds += blockIdx.x * beamSize;
Pair topK[maxLength]; // NOLINT
int beam = maxLength;
Pair max;
bool isEmpty = false;
bool firstStep = true;
for (int k = 0; k < maxLength; k++) {
topK[k].set(-HL_FLOAT_MAX, -1);
}
while (beamSize) {
threadGetTopK<maxLength, blockSize>(
topK, beam, beamSize, src, firstStep, isEmpty, max, dim, tid);
shTopK[tid] = topK[0];
blockReduce<maxLength, blockSize>(
shTopK, maxId, topK, &topVal, &topIds, beam, beamSize, tid, warp);
}
}
template <int maxLength, int blockSize>
__global__ void KeSMatrixTopK(real* topVal,
int ldv,
int* topIds,
real* val,
int* row,
int* col,
int beamSize) {
__shared__ Pair shTopK[blockSize];
__shared__ int maxId[blockSize / 2];
const int tid = threadIdx.x;
const int warp = threadIdx.x / 32;
topVal += blockIdx.x * ldv;
topIds += blockIdx.x * beamSize;
Pair topK[maxLength]; // NOLINT
int beam = maxLength;
Pair max;
bool isEmpty = false;
bool firstStep = true;
int start = row[blockIdx.x];
int end = row[blockIdx.x + 1];
int dim = end - start;
val += start;
col += start;
if (beamSize > dim) {
// if the number of values to sort are less than the output size,
// use -1 to indicate the end of valid sorted values.
if (tid == 0) {
topIds[dim] = -1;
}
beamSize = dim;
}
for (int k = 0; k < maxLength; k++) {
topK[k].set(-HL_FLOAT_MAX, -1);
}
while (beamSize) {
threadGetTopK<maxLength, blockSize>(
topK, beam, beamSize, val, col, firstStep, isEmpty, max, dim, tid);
shTopK[tid] = topK[0];
blockReduce<maxLength, blockSize>(
shTopK, maxId, topK, &topVal, &topIds, beam, beamSize, tid, warp);
}
}
void hl_matrix_top_k(real* topVal,
int ldv,
int* topIds,
real* src,
int lds,
int dim,
int beamSize,
int numSamples) {
CHECK_NOTNULL(topVal);
CHECK_NOTNULL(topIds);
CHECK_NOTNULL(src);
if (beamSize > dim) beamSize = dim;
dim3 threads(256, 1);
dim3 grid(numSamples, 1);
KeMatrixTopK<5, 256><<<grid, threads, 0, STREAM_DEFAULT>>>(
topVal, ldv, topIds, src, lds, dim, beamSize);
CHECK_SYNC("hl_matrix_top_k failed");
}
void hl_sparse_matrix_top_k(real* topVal,
int ldv,
int* topIds,
hl_sparse_matrix_s src,
int beamSize,
int numSamples) {
CHECK_NOTNULL(topVal);
CHECK_NOTNULL(topIds);
CHECK_NOTNULL(src);
CHECK_EQ(src->format, HL_SPARSE_CSR) << "sparse matrix format error!";
hl_csr_matrix csr = (hl_csr_matrix)src->matrix;
if (csr->csr_val == NULL || csr->csr_row == NULL || csr->csr_col == NULL) {
LOG(FATAL) << "parameter src is null!";
}
dim3 threads(256, 1);
dim3 grid(numSamples, 1);
KeSMatrixTopK<5, 256><<<grid, threads, 0, STREAM_DEFAULT>>>(
topVal, ldv, topIds, csr->csr_val, csr->csr_row, csr->csr_col, beamSize);
CHECK_SYNC("hl_sparse_matrix_top_k failed");
}
/**
* Each block compute one sample.
* In a block:
* 1. every thread get top maxLength value;
* 2. merge to shTopK, block reduce and get max value;
* 3. go to the second setp, until one thread's topK value is null;
* 4. go to the first setp, until get the topK value.
*/
template <int maxLength, int blockSize>
__global__ void KeMatrixTopKClassificationError(real* topVal,
int ldv,
int* topIds,
real* src,
int lds,
int dim,
int beamSize,
int* label,
real* recResult) {
__shared__ Pair shTopK[blockSize];
__shared__ int maxId[blockSize / 2];
const int tid = threadIdx.x;
const int warp = threadIdx.x / 32;
src += blockIdx.x * lds;
topVal += blockIdx.x * ldv;
topIds += blockIdx.x * beamSize;
Pair topK[maxLength]; // NOLINT
int beam = maxLength;
Pair max;
bool isEmpty = false;
bool firstStep = true;
int topkSize = beamSize;
for (int k = 0; k < maxLength; k++) {
topK[k].set(-HL_FLOAT_MAX, -1);
}
while (beamSize) {
threadGetTopK<maxLength, blockSize>(
topK, beam, beamSize, src, firstStep, isEmpty, max, dim, tid);
shTopK[tid] = topK[0];
blockReduce<maxLength, blockSize>(
shTopK, maxId, topK, &topVal, &topIds, beam, beamSize, tid, warp);
}
__syncthreads();
if (tid == 0) {
for (int i = 0; i < topkSize; i++) {
if (*--topIds == label[blockIdx.x]) {
recResult[blockIdx.x] = 0;
break;
}
recResult[blockIdx.x] = 1.0f;
}
}
}
void hl_matrix_classification_error(real* topVal,
int ldv,
int* topIds,
real* src,
int lds,
int dim,
int topkSize,
int numSamples,
int* label,
real* recResult) {
CHECK_NOTNULL(topVal);
CHECK_NOTNULL(topIds);
CHECK_NOTNULL(src);
if (topkSize > dim) topkSize = dim;
dim3 threads(256, 1);
dim3 grid(numSamples, 1);
KeMatrixTopKClassificationError<5, 256><<<grid, threads, 0, STREAM_DEFAULT>>>(
topVal, ldv, topIds, src, lds, dim, topkSize, label, recResult);
CHECK_SYNC("hl_matrix_top_k classification error failed");
}
paddle/fluid/framework/data_layout_transform.cc
浏览文件 @ ee760d1c
......@@ -147,9 +147,9 @@ void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
"Input tensor type is not supported: ", in.type().name());
memory::data_type out_type = in_type;
auto in_format = MKLDNNFormatForSize(in_tz.size(), in.format());
auto in_format = platform::MKLDNNFormatForSize(in_tz.size(), in.format());
auto out_format =
MKLDNNFormatForSize(in_tz.size(), ToMKLDNNFormat(out_layout));
platform::MKLDNNFormatForSize(in_tz.size(), ToMKLDNNFormat(out_layout));
void* in_data = GetDataFromTensor(in, in_type);
......
paddle/fluid/framework/data_layout_transform.h
浏览文件 @ ee760d1c
......@@ -62,12 +62,6 @@ inline MKLDNNDataType ToMKLDNNDataType(const std::type_index type) {
return MKLDNNDataType::data_undef;
}
inline MKLDNNFormat MKLDNNFormatForSize(size_t dims_size,
MKLDNNFormat default_format) {
return (dims_size == 1
? mkldnn::memory::format::x
: dims_size == 2 ? mkldnn::memory::format::nc : default_format);
}
#endif
void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
......
paddle/fluid/framework/data_transform.cc
浏览文件 @ ee760d1c
......@@ -18,17 +18,21 @@ limitations under the License. */
#include "paddle/fluid/framework/data_layout_transform.h"
#include "paddle/fluid/framework/data_type_transform.h"
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif
namespace paddle {
namespace framework {
static void PassTensorData(Tensor* from, Tensor* to) {
static void PassTensorData(Tensor *from, Tensor *to) {
to->ShareDataWith(*from);
*from = Tensor();
}
void DataTransform(const OpKernelType& expected_kernel_type,
const OpKernelType& kernel_type_for_var,
const Tensor& input_tensor, Tensor* output_tensor) {
void TransformData(const OpKernelType &expected_kernel_type,
const OpKernelType &kernel_type_for_var,
const Tensor &input_tensor, Tensor *output_tensor) {
bool transformed = false;
Tensor in;
in.ShareDataWith(input_tensor);
......@@ -48,8 +52,8 @@ void DataTransform(const OpKernelType& expected_kernel_type,
// Case1 - transform from Non-MKLDNN OPKernel to MKLDNN OPKernel
// Just set layout/format. No real transform occur
auto out_format =
MKLDNNFormatForSize(in.dims().size(), ToMKLDNNFormat(lin));
auto out_format = platform::MKLDNNFormatForSize(in.dims().size(),
ToMKLDNNFormat(lin));
out.ShareDataWith(input_tensor);
out.set_layout(DataLayout::kMKLDNN);
......@@ -89,17 +93,17 @@ void DataTransform(const OpKernelType& expected_kernel_type,
output_tensor->ShareDataWith(in);
}
void CopyVariableWithTensor(const Variable& in_var, const Tensor& tensor,
Variable* out_var) {
void SetTensorToVariable(const Variable &in_var, const Tensor &tensor,
Variable *out_var) {
if (in_var.IsType<LoDTensor>()) {
auto& in_lod_tensor = in_var.Get<LoDTensor>();
auto* tran_lod_tensor = out_var->GetMutable<LoDTensor>();
auto &in_lod_tensor = in_var.Get<LoDTensor>();
auto *tran_lod_tensor = out_var->GetMutable<LoDTensor>();
tran_lod_tensor->set_lod(in_lod_tensor.lod());
tran_lod_tensor->set_layout(in_lod_tensor.layout());
tran_lod_tensor->ShareDataWith(tensor);
} else if (in_var.IsType<SelectedRows>()) {
auto& in_selected_rows = in_var.Get<SelectedRows>();
auto* trans_selected_rows = out_var->GetMutable<SelectedRows>();
auto &in_selected_rows = in_var.Get<SelectedRows>();
auto *trans_selected_rows = out_var->GetMutable<SelectedRows>();
trans_selected_rows->set_height(in_selected_rows.height());
trans_selected_rows->set_rows(in_selected_rows.rows());
trans_selected_rows->mutable_value()->ShareDataWith(tensor);
......
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @ ee760d1c
......@@ -25,11 +25,12 @@ else()
cc_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
endif()
cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_base scope lod_tensor)
cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope)
cc_library(multi_devices_graph_builder SRCS multi_devices_graph_builder.cc DEPS ssa_graph_builder computation_op_handle
scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle)
scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle)
cc_library(ssa_graph_builder_factory SRCS ssa_graph_builder_factory.cc DEPS multi_devices_graph_builder ssa_graph_printer ssa_graph_checker)
......
paddle/fluid/framework/details/data_balance_op_handle.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/fluid/framework/executor.cc
浏览文件 @ ee760d1c
此差异已折叠。
paddle/fluid/framework/executor.h
浏览文件 @ ee760d1c
此差异已折叠。
paddle/fluid/framework/operator.cc
浏览文件 @ ee760d1c
此差异已折叠。
paddle/fluid/framework/operator.h
浏览文件 @ ee760d1c
此差异已折叠。
paddle/fluid/framework/var_type.h
浏览文件 @ ee760d1c
此差异已折叠。
paddle/fluid/inference/analysis/README.md 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/fluid/memory/malloc.cc
浏览文件 @ ee760d1c
此差异已折叠。
paddle/fluid/operators/argsort_op.cc 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/fluid/operators/argsort_op.cu 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/fluid/operators/argsort_op.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/fluid/operators/detection/anchor_generator_op.cc 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/fluid/operators/detection/anchor_generator_op.cu 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/fluid/operators/detection/anchor_generator_op.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/fluid/operators/print_op.cc
浏览文件 @ ee760d1c
此差异已折叠。
paddle/fluid/operators/read_op.cc
浏览文件 @ ee760d1c
此差异已折叠。
paddle/fluid/operators/reshape_op.cu 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/fluid/operators/reshape_op.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/fluid/operators/while_op.cc
浏览文件 @ ee760d1c
此差异已折叠。
paddle/fluid/platform/enforce.h
浏览文件 @ ee760d1c
此差异已折叠。
paddle/fluid/pybind/pybind.cc
浏览文件 @ ee760d1c
此差异已折叠。
paddle/function/BufferArg.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/BufferArg.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/BufferArgTest.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/ContextProjectionOp.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/ContextProjectionOpTest.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/CosSimOp.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/CosSimOpTest.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/CropOp.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/CrossMapNormalOp.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/DepthwiseConvOpGpu.cu 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/EigenGemm.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/Function.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/FunctionTest.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/FunctionTest.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/GemmConvOp.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/GemmFunctor.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/Im2ColTest.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/MulOp.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/MulOp.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/MulOpGpu.cu 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/MulOpTest.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/PadOp.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/RowConvOp.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/RowConvOpGpu.cu 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/ScaleSubRegionOp.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/SwitchOp.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/TensorType.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/neon/NeonDepthwiseConv.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/nnpack/NNPACKConvOp.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/function/nnpack/NNPACKConvOpTest.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/activations/MKLDNNActivation.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/dataproviders/DataProvider.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/evaluators/ChunkEvaluator.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/evaluators/Evaluator.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/evaluators/Evaluator.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/AddtoLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/AgentLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/AverageLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/BilinearInterpLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/BlockExpandLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/Conv3DLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/ConvBaseLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/ConvBaseLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/ConvBaseOperator.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/ConvBaseOperator.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/ConvBaseProjection.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/ConvOperator.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/ConvOperator.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/ConvProjection.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/ConvShiftLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/ConvTransOperator.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/ConvTransOperator.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/ConvTransProjection.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/CosSimLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/CosSimVecMatLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/CostLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/CudnnBatchNormLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/CudnnConvBaseLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/CudnnPoolLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/DataNormLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/DeConv3DLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/DetectionUtil.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/DotProdLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/ExpandConvLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/ExpandLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/FullyConnectedLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/FullyConnectedLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/GatedRecurrentLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/GruCompute.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/InterpolationLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/L2DistanceLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/Layer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/Layer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/LinearChainCRF.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/LinearChainCTC.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/LstmLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/LstmLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/MDLstmLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/MKLDNNConvLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/MKLDNNLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/MKLDNNPoolLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/MKLPackedWeight.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/MaxLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/MaxOutLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/MaxPoolWithMaskLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/MultiplexLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/NCELayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/NormLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/NormProjectionLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/Operator.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/OuterProdLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/ParameterReluLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/Pool3DLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/PoolLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/PoolProjection.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/PoolProjectionLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/PowerLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/PriorBox.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/Projection.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/RecurrentLayerGroup.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/ResizeLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/RotateLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/ScalingLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/SequenceConcatLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/SequencePoolLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/SequenceSliceLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/SequenceToBatch.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/SlopeInterceptLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/SubSequenceLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/SumToOneNormLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/TensorLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/TransLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/UpsampleLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/layers/ValidationLayer.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/CMakeLists.txt 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/LayerGradUtil.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/MKLDNNTester.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/MKLDNNTester.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/Sequence/train.list 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/Sequence/train.list.nest 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/sequence_lstm.conf 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/sequence_nest_rnn.conf 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/sequence_recurrent.py 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/sequence_rnn.conf 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/test_ActivationGrad.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/test_BatchNorm.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/test_CRFLayerGrad.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/test_CompareSparse.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/test_CompareTwoNets.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/test_ConvTrans.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/test_ConvUnify.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/test_KmaxSeqScore.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/test_LayerGrad.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/test_LinearChainCRF.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/test_MKLDNN.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/test_NetworkCompare.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/test_PyDataProvider.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/test_PyDataProvider2.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/test_RecurrentLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/test_Upsample.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/gserver/tests/test_WarpCTCLayer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/legacy/api/Arguments.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/api/GradientMachine.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/api/Matrix.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/api/Paddle.i 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/api/PaddleAPI.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/api/PaddleAPIPrivate.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/api/Parameter.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/api/ParameterOptimizer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/api/SequenceGenerator.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/api/Trainer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/api/Util.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/api/Vector.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/capi/capi_private.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/capi/gradient_machine.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/capi/tests/test_GradientMachine.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/cuda/include/hl_base.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/cuda/src/hl_top_k.cu 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/BufferArg.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/BufferArg.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/BufferArgTest.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/ContextProjectionOp.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/ContextProjectionOpTest.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/CosSimOp.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/CosSimOpTest.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/CropOp.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/CrossMapNormalOp.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/DepthwiseConvOpGpu.cu 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/EigenGemm.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/Function.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/FunctionTest.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/FunctionTest.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/GemmConvOp.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/GemmFunctor.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/Im2ColTest.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/MulOp.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/MulOp.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/MulOpGpu.cu 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/MulOpTest.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/PadOp.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/RowConvOp.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/RowConvOpGpu.cu 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/ScaleSubRegionOp.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/SwitchOp.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/TensorType.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/neon/NeonDepthwiseConv.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/nnpack/NNPACKConvOp.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/function/nnpack/NNPACKConvOpTest.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/activations/MKLDNNActivation.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/dataproviders/DataProvider.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/evaluators/CTCErrorEvaluator.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/evaluators/ChunkEvaluator.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/evaluators/Evaluator.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/evaluators/Evaluator.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/gradientmachines/NeuralNetwork.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/AddtoLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/AgentLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/AverageLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/BilinearInterpLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/BlockExpandLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/Conv3DLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ConvBaseLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ConvBaseLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ConvBaseOperator.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ConvBaseOperator.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ConvBaseProjection.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ConvOperator.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ConvOperator.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ConvProjection.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ConvShiftLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ConvTransOperator.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ConvTransOperator.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ConvTransProjection.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ConvexCombinationLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/CosSimLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/CosSimVecMatLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/CostLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/CrossChannelNormLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/CudnnBatchNormLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/CudnnConvBaseLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/CudnnPoolLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/DataNormLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/DeConv3DLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/DetectionUtil.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/DotProdLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ExpandConvLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ExpandLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/FeatureMapExpandLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/FullyConnectedLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/FullyConnectedLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/GatedRecurrentLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/GruCompute.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/InterpolationLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/L2DistanceLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/Layer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/Layer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/LinearChainCRF.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/LinearChainCTC.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/LstmLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/LstmLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/MDLstmLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/MKLDNNConvLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/MKLDNNLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/MKLDNNPoolLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/MKLPackedWeight.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/MaxLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/MaxOutLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/MaxPoolWithMaskLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/MultiplexLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/NCELayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/NormLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/NormProjectionLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/Operator.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/OuterProdLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ParameterReluLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/Pool3DLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/PoolLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/PoolProjection.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/PoolProjectionLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/PowerLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/PriorBox.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/Projection.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/RecurrentLayerGroup.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ResizeLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/RotateLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ScalingLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/SequenceConcatLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/SequencePoolLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/SequenceReshapeLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/SequenceSliceLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/SequenceToBatch.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/SlopeInterceptLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/SpatialPyramidPoolLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/SubNestedSequenceLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/SubSequenceLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/SumToOneNormLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/TensorLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/TransLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/UpsampleLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/layers/ValidationLayer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/CMakeLists.txt 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/LayerGradUtil.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/MKLDNNTester.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/MKLDNNTester.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/Sequence/train.list 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/Sequence/train.list.nest 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/sequence_layer_group.conf 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/sequence_lstm.conf 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/sequence_nest_rnn.conf 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/sequence_recurrent.py 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/sequence_recurrent_group.py 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/sequence_rnn.conf 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_ActivationGrad.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_BatchNorm.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_CRFLayerGrad.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_CompareSparse.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_CompareTwoNets.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_ConvTrans.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_ConvUnify.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_KmaxSeqScore.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_LayerGrad.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_LinearChainCRF.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_MKLDNN.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_MultinomialSampler.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_NetworkCompare.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_PyDataProvider.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_PyDataProvider2.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_RecurrentLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_SelectiveFCLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_SeqSliceLayerGrad.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_Upsample.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/gserver/tests/test_WarpCTCLayer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/CMakeLists.txt 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/CpuSparseMatrix.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/MKLDNNMatrix.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/MathFunctions.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/Matrix.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/Matrix.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/tests/OriginalOptimizerApi.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/tests/TensorCheck.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/tests/TestUtils.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/tests/test_Allocator.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/tests/test_BaseMatrix.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/tests/test_CpuGpuVector.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/tests/test_ExecViaCpu.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/tests/test_FPException.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/tests/test_GpuProfiler.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/tests/test_RowBuffer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/tests/test_SIMDFunctions.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/tests/test_Tensor.cu 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/tests/test_TrainingAlgorithm.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/tests/test_lazyAssign.cu 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/tests/test_matrixCompare.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/tests/test_matrixUtil.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/math/tests/test_sparseMatrixCompare.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/parameter/Argument.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/parameter/Argument.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/parameter/FirstOrderOptimizer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/parameter/Parameter.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/parameter/Parameter.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/parameter/ParameterUpdateFunctions.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/parameter/ParameterUpdaterHook.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/parameter/ThreadLocalBuffer.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/parameter/Weight.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/parameter/tests/test_argument.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/parameter/tests/test_common.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/pserver/BaseClient.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/pserver/ParameterClient2.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/pserver/ParameterClient2.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/pserver/ParameterServer2.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/pserver/ParameterServer2.h 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/pserver/test/SocketTest.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/pserver/test/test_ParameterServer2.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/pserver/test/test_ProtoServer.cpp 0 → 100644
浏览文件 @ ee760d1c
此差异已折叠。
paddle/legacy/pserver/test/test_ProtoServer.sh 0 → 100755
浏览文件 @ ee760d1c
此差异已折叠。
paddle/math/CMakeLists.txt 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/math/CpuSparseMatrix.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/math/MKLDNNMatrix.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/math/MathFunctions.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/math/Matrix.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/math/Matrix.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/math/tests/OriginalOptimizerApi.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/math/tests/TensorCheck.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/math/tests/TestUtils.h 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/math/tests/test_Allocator.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/math/tests/test_BaseMatrix.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/math/tests/test_CpuGpuVector.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/math/tests/test_ExecViaCpu.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/math/tests/test_FPException.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/math/tests/test_GpuProfiler.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/math/tests/test_RowBuffer.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/math/tests/test_SIMDFunctions.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/math/tests/test_Tensor.cu 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
paddle/math/tests/test_TrainingAlgorithm.cpp 已删除 100644 → 0
浏览文件 @ 0cef33a4
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或