Skip to content

P
Paddle

机器未来 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 c27e71e2 编写于 作者: L Liu Yiqun
浏览文件
操作

Merge branch 'develop' into cmake_protobuf

上级 e6785463 d10f6cfb
隐藏空白更改
内联 并排
Showing with 432 addition and 247 deletion
RELEASE.md
浏览文件 @ c27e71e2
# Release v0.10.0
We are glad to release version 0.10.0. In this version, we are happy to release the new
[Python API](http://research.baidu.com/paddlepaddles-new-api-simplifies-deep-learning-programs/).
- Our old Python API is kind of out of date. It's hard to learn and hard to
use. To write a PaddlePaddle program using the old API, we'd have to write
at least two Python files: one `data provider` and another one that defines
the network topology. Users start a PaddlePaddle job by running the
`paddle_trainer` C++ program, which calls Python interpreter to run the
network topology configuration script and then start the training loop,
which iteratively calls the data provider function to load minibatches.
This prevents us from writing a Python program in a modern way, e.g., in the
Jupyter Notebook.
- The new API, which we often refer to as the *v2 API*, allows us to write
much shorter Python programs to define the network and the data in a single
.py file. Also, this program can run in Jupyter Notebook, since the entry
point is in Python program and PaddlePaddle runs as a shared library loaded
and invoked by this Python program.
Basing on the new API, we delivered an online interative
book, [Deep Learning 101](http://book.paddlepaddle.org/index.en.html)
and [its Chinese version](http://book.paddlepaddle.org/).
We also worked on updating our online documentation to describe the new API.
But this is an ongoing work. We will release more documentation improvements
in the next version.
We also worked on bring the new API to distributed model training (via MPI and
Kubernetes). This work is ongoing. We will release more about it in the next
version.
## New Features
* We release [new Python API](http://research.baidu.com/paddlepaddles-new-api-simplifies-deep-learning-programs/).
* Deep Learning 101 book in [English](http://book.paddlepaddle.org/index.en.html) and [Chinese](http://book.paddlepaddle.org/).
* Support rectangle input for CNN.
* Support stride pooling for seqlastin and seqfirstin.
* Expose `seq_concat_layer/seq_reshape_layer` in `trainer_config_helpers`.
* Add dataset package: CIFAR, MNIST, IMDB, WMT14, CONLL05, movielens, imikolov.
* Add Priorbox layer for Single Shot Multibox Detection.
* Add smooth L1 cost.
* Add data reader creator and data reader decorator for v2 API.
* Add the CPU implementation of cmrnorm projection.
## Improvements
* Support Python virtualenv for `paddle_trainer`.
* Add pre-commit hooks, used for automatically format our code.
* Upgrade protobuf to version 3.x.
* Add an option to check data type in Python data provider.
* Speedup the backward of average layer on GPU.
* Documentation refinement.
* Check dead links in documents using Travis-CI.
* Add a example for explaining `sparse_vector`.
* Add ReLU in layer_math.py
* Simplify data processing flow for Quick Start.
* Support CUDNN Deconv.
* Add data feeder in v2 API.
* Support predicting the samples from sys.stdin for sentiment demo.
* Provide multi-proccess interface for image preprocessing.
* Add benchmark document for v1 API.
* Add ReLU in `layer_math.py`.
* Add packages for automatically downloading public datasets.
* Rename `Argument::sumCost` to `Argument::sum` since class `Argument` is nothing with cost.
* Expose Argument::sum to Python
* Add a new `TensorExpression` implementation for matrix-related expression evaluations.
* Add lazy assignment for optimizing the calculation of a batch of multiple expressions.
* Add abstract calss `Function` and its implementation:
* `PadFunc` and `PadGradFunc`.
* `ContextProjectionForwardFunc` and `ContextProjectionBackwardFunc`.
* `CosSimBackward` and `CosSimBackwardFunc`.
* `CrossMapNormalFunc` and `CrossMapNormalGradFunc`.
* `MulFunc`.
* Add class `AutoCompare` and `FunctionCompare`, which make it easier to write unit tests for comparing gpu and cpu version of a function.
* Generate `libpaddle_test_main.a` and remove the main function inside the test file.
* Support dense numpy vector in PyDataProvider2.
* Clean code base, remove some copy-n-pasted code snippets:
* Extract `RowBuffer` class for `SparseRowMatrix`.
* Clean the interface of `GradientMachine`.
* Use `override` keyword in layer.
* Simplify `Evaluator::create`, use `ClassRegister` to create `Evaluator`s.
* Check MD5 checksum when downloading demo's dataset.
* Add `paddle::Error` which intentially replace `LOG(FATAL)` in Paddle.
## Bug Fixes
* Check layer input types for `recurrent_group`.
* Don't run `clang-format` with .cu source files.
* Fix bugs with `LogActivation`.
* Fix the bug that runs `test_layerHelpers` multiple times.
* Fix the bug that the seq2seq demo exceeds protobuf message size limit.
* Fix the bug in dataprovider converter in GPU mode.
* Fix a bug in `GatedRecurrentLayer`.
* Fix bug for `BatchNorm` when testing more than one models.
* Fix broken unit test of paramRelu.
* Fix some compile-time warnings about `CpuSparseMatrix`.
* Fix `MultiGradientMachine` error when `trainer_count > batch_size`.
* Fix bugs that prevents from asynchronous data loading in `PyDataProvider2`.
# Release v0.9.0
......
cmake/cblas.cmake
浏览文件 @ c27e71e2
......@@ -44,7 +44,6 @@ if(MKL_INC_DIR AND MKL_CORE_LIB AND MKL_SEQUENTIAL_LIB AND MKL_INTEL_LP64)
message(STATUS "Found MKL (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
set(CBLAS_FOUND ON)
if(${MKL_LAPACK_INC_DIR})
add_definitions(-DPADDLE_USE_LAPACK)
message(STATUS "Found lapack in MKL (include: ${MKL_LAPACK_INC_DIR})")
endif()
return() # return file.
......@@ -80,7 +79,6 @@ if(ATLAS_INC_DIR AND ATLAS_CBLAS_LIB AND ATLAS_LIB AND NOT CBLAS_FOUND)
message(STATUS "Found ATLAS (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
set(CBLAS_FOUND ON)
if(ATLAS_CLAPACK_INC_DIR)
add_definitions(-DPADDLE_USE_LAPACK)
set(CBLAS_INC_DIR ${CBLAS_INC_DIR} ${ATLAS_CLAPACK_INC_DIR})
message(STATUS "Found lapack in ATLAS (include: ${ATLAS_CLAPACK_INC_DIR})")
endif()
......@@ -115,7 +113,6 @@ if(OPENBLAS_INC_DIR AND OPENBLAS_LIB)
message(STATUS "Found OpenBLAS (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
set(CBLAS_FOUND ON)
if(OPENBLAS_LAPACKE_INC_DIR)
add_definitions(-DPADDLE_USE_LAPACK)
message(STATUS "Found lapack in OpenBLAS (include: ${OPENBLAS_LAPACKE_INC_DIR})")
endif()
return()
......
cmake/external/openblas.cmake
浏览文件 @ c27e71e2
......@@ -24,45 +24,17 @@ IF(NOT ${CBLAS_FOUND})
SET(CBLAS_LIBRARIES "${CBLAS_INSTALL_DIR}/lib/${LIBRARY_PREFIX}openblas${STATIC_LIBRARY_SUFFIX}"
CACHE FILEPATH "openblas library." FORCE)
# check fortran compiler and library
SET(COMMON_ARGS CC=${CMAKE_C_COMPILER} NO_LAPACK=1 NO_SHARED=1)
IF(ANDROID)
SET(OPENBLAS_COMMIT "b5c96fcfcdc82945502a2303116a64d89985daf5")
SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER} TARGET=ARMV7 ARM_SOFTFP_ABI=1 NOFORTRAN=1 USE_THREAD=0 libs)
SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER} TARGET=ARMV7 ARM_SOFTFP_ABI=1 USE_THREAD=0 libs)
ELSEIF(RPI)
SET(OPENBLAS_COMMIT "v0.2.19")
SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER} TARGET=ARMV7 NOFORTRAN=1 USE_THREAD=0 libs)
SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER} TARGET=ARMV7 USE_THREAD=0 libs)
ELSE()
IF(CMAKE_COMPILER_IS_GNUCC)
ENABLE_LANGUAGE(Fortran)
if (NOT CMAKE_Fortran_COMPILER_VERSION)
# cmake < 3.4 cannot get CMAKE_Fortran_COMPILER_VERSION directly.
execute_process(COMMAND ${CMAKE_Fortran_COMPILER} -dumpversion
OUTPUT_VARIABLE CMAKE_Fortran_COMPILER_VERSION)
endif()
string(REGEX MATCHALL "[0-9]+" Fortran_VERSION ${CMAKE_Fortran_COMPILER_VERSION})
list(GET Fortran_VERSION 0 Fortran_MAJOR)
list(GET Fortran_VERSION 1 Fortran_MINOR)
find_library(GFORTRAN_LIBRARY NAMES gfortran PATHS
/lib
/usr/lib
/usr/lib/gcc/x86_64-linux-gnu/${Fortran_MAJOR}.${Fortran_MINOR}/
/usr/lib/gcc/x86_64-linux-gnu/${Fortran_MAJOR}/)
if (NOT GFORTRAN_LIBRARY)
message(FATAL_ERROR "Cannot found gfortran library which it is used by openblas")
endif()
find_package(Threads REQUIRED)
LIST(APPEND CBLAS_LIBRARIES ${GFORTRAN_LIBRARY} ${CMAKE_THREAD_LIBS_INIT})
ENDIF(CMAKE_COMPILER_IS_GNUCC)
IF(NOT CMAKE_Fortran_COMPILER)
MESSAGE(FATAL_ERROR "To build lapack in libopenblas, "
"you need to set gfortran compiler: cmake .. -DCMAKE_Fortran_COMPILER=...")
ENDIF(NOT CMAKE_Fortran_COMPILER)
ADD_DEFINITIONS(-DPADDLE_USE_LAPACK)
SET(OPENBLAS_COMMIT "v0.2.19")
SET(OPENBLAS_ARGS FC=${CMAKE_Fortran_COMPILER} DYNAMIC_ARCH=1 libs netlib)
SET(OPENBLAS_ARGS DYNAMIC_ARCH=1 libs)
ENDIF()
ExternalProject_Add(
......@@ -73,7 +45,7 @@ IF(NOT ${CBLAS_FOUND})
PREFIX ${CBLAS_SOURCES_DIR}
INSTALL_DIR ${CBLAS_INSTALL_DIR}
BUILD_IN_SOURCE 1
BUILD_COMMAND ${CMAKE_MAKE_PROGRAM} CC=${CMAKE_C_COMPILER} NO_SHARED=1 ${OPTIONAL_ARGS}
BUILD_COMMAND ${CMAKE_MAKE_PROGRAM} ${COMMON_ARGS} ${OPTIONAL_ARGS}
INSTALL_COMMAND ${CMAKE_MAKE_PROGRAM} install NO_SHARED=1 PREFIX=<INSTALL_DIR>
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
......
cmake/package.cmake
浏览文件 @ c27e71e2
set(CPACK_PACKAGE_NAME paddle)
set(CPACK_PACKAGE_DESCRIPTION_SUMMARY "")
set(CPACK_PACKAGE_VERSION_MAJOR ${PADDLE_MAJOR_VERSION})
set(CPACK_PACKAGE_VERSION_MINOR ${PADDLE_MINOR_VERSION})
set(CPACK_PACKAGE_VERSION_PATCH ${PADDLE_PATCH_VERSION})
......@@ -10,8 +9,9 @@ set(CPACK_DEBIAN_PACKAGE_ARCHITECTURE amd64)
set(CPACK_DEBIAN_PACKAGE_MAINTAINER PaddlePaddle Dev <paddle-dev@baidu.com>)
set(CPACK_PACKAGE_DESCRIPTION_SUMMARY "Paddle")
set(CPACK_PACKAGE_DESCRIPTION "")
set(CPACK_DEBIAN_PACKAGE_DEPENDS "libatlas3-base, libgflags2, libgoogle-glog0, libprotobuf8, libpython2.7, libstdc++6, python-numpy, python-pip, python-pip-whl, python-protobuf")
set(CPACK_DEBIAN_PACKAGE_DEPENDS "libpython2.7-dev, libstdc++6, python-pip, curl, libgfortran3, python-pip-whl")
set(CPACK_DEBIAN_PACKAGE_SECTION Devel)
set(CPACK_DEBIAN_PACKAGE_VERSION ${PADDLE_VERSION})
set(CPACK_DEBIAN_PACKAGE_CONTROL_EXTRA "${PROJ_ROOT}/paddle/scripts/deb/postinst")
#set(CPACK_GENERATOR "DEB")
# Start cpack
......
demo/sentiment/trainer_config.py
浏览文件 @ c27e71e2
......@@ -29,7 +29,7 @@ settings(
batch_size=128,
learning_rate=2e-3,
learning_method=AdamOptimizer(),
average_window=0.5,
model_average=ModelAverage(0.5),
regularization=L2Regularization(8e-4),
gradient_clipping_threshold=25)
......
demo/seqToseq/seqToseq_net.py
浏览文件 @ c27e71e2
......@@ -69,7 +69,8 @@ def gru_encoder_decoder(data_conf,
encoder_size=512,
decoder_size=512,
beam_size=3,
max_length=250):
max_length=250,
error_clipping=50):
"""
A wrapper for an attention version of GRU Encoder-Decoder network
is_generating: whether this config is used for generating
......@@ -90,9 +91,19 @@ def gru_encoder_decoder(data_conf,
input=src_word_id,
size=word_vector_dim,
param_attr=ParamAttr(name='_source_language_embedding'))
src_forward = simple_gru(input=src_embedding, size=encoder_size)
src_forward = simple_gru(
input=src_embedding,
size=encoder_size,
naive=True,
gru_layer_attr=ExtraLayerAttribute(
error_clipping_threshold=error_clipping))
src_backward = simple_gru(
input=src_embedding, size=encoder_size, reverse=True)
input=src_embedding,
size=encoder_size,
reverse=True,
naive=True,
gru_layer_attr=ExtraLayerAttribute(
error_clipping_threshold=error_clipping))
encoded_vector = concat_layer(input=[src_forward, src_backward])
with mixed_layer(size=decoder_size) as encoded_proj:
......@@ -117,11 +128,13 @@ def gru_encoder_decoder(data_conf,
decoder_inputs += full_matrix_projection(input=context)
decoder_inputs += full_matrix_projection(input=current_word)
gru_step = gru_step_layer(
gru_step = gru_step_naive_layer(
name='gru_decoder',
input=decoder_inputs,
output_mem=decoder_mem,
size=decoder_size)
size=decoder_size,
layer_attr=ExtraLayerAttribute(
error_clipping_threshold=error_clipping))
with mixed_layer(
size=target_dict_dim, bias_attr=True,
......
doc/getstarted/index_cn.rst
浏览文件 @ c27e71e2
......@@ -2,7 +2,8 @@
============
.. toctree::
:maxdepth: 2
:maxdepth: 1
build_and_install/index_cn.rst
basic_usage/index_cn.rst
- `深度学习入门课程 <http://book.paddlepaddle.org/>`_
doc/getstarted/index_en.rst
浏览文件 @ c27e71e2
......@@ -2,7 +2,8 @@ GET STARTED
============
.. toctree::
:maxdepth: 2
:maxdepth: 1
build_and_install/index_en.rst
basic_usage/index_en.rst
- `Deep Learning 101 <http://book.paddlepaddle.org/index.en.html>`_
doc/howto/deep_model/rnn/hierarchical_layer_cn.rst
浏览文件 @ c27e71e2
......@@ -19,18 +19,18 @@
在 PaddlePaddle中，下面这些Layer能够接受双层序列作为输入，完成相应的计算。
pooling_layer
==============
pooling
========
pooling_layer 的使用示例如下，详细见 :ref:`api_trainer_config_helpers_layers_pooling_layer` 配置API。
pooling 的使用示例如下，详细见 :ref:`api_v2.layer_pooling` 配置API。
.. code-block:: bash
seq_pool = pooling_layer(input=layer,
pooling_type=AvgPooling(),
agg_level=AggregateLevel.EACH_SEQUENCE)
seq_pool = pooling(input=layer,
pooling_type=pooling.Max(),
agg_level=AggregateLevel.EACH_SEQUENCE)
- `pooling_type` 目前支持两种，分别是：MaxPooling()和AvgPooling()。
- `pooling_type` 目前支持两种，分别是：pooling.Max()和pooling.Avg()。
- `agg_level=AggregateLevel.EACH_TIMESTEP` 时（默认值）：
......@@ -47,7 +47,7 @@ pooling_layer 的使用示例如下，详细见 :ref:`api_trainer_config_helpers
last_seq 和 first_seq
=====================
last_seq 的使用示例如下（ :ref:`api_trainer_config_helpers_layers_first_seq` 类似），详细见 :ref:`api_trainer_config_helpers_layers_last_seq` 配置API。
last_seq 的使用示例如下（ :ref:`api_v2.layer_first_seq` 类似），详细见 :ref:`api_v2.layer_last_seq` 配置API。
.. code-block:: bash
......@@ -65,16 +65,16 @@ last_seq 的使用示例如下（ :ref:`api_trainer_config_helpers_layers_first_
- 输入：必须是一个双层序列
- 输出：一个单层序列，其中每个元素是双层序列中每个subseq最后一个（或第一个）元素。
expand_layer
============
expand
======
expand_layer 的使用示例如下，详细见 :ref:`api_trainer_config_helpers_layers_expand_layer` 配置API。
expand 的使用示例如下，详细见 :ref:`api_v2.layer_expand` 配置API。
.. code-block:: bash
expand = expand_layer(input=layer1,
expand_as=layer2,
expand_level=ExpandLevel.FROM_TIMESTEP)
ex = expand(input=layer1,
expand_as=layer2,
expand_level=ExpandLevel.FROM_TIMESTEP)
- `expand_level=ExpandLevel.FROM_TIMESTEP` 时（默认值）：
......
doc/howto/deep_model/rnn/index_cn.rst
浏览文件 @ c27e71e2
......@@ -4,7 +4,6 @@ RNN相关模型
.. toctree::
:maxdepth: 1
rnn_config_cn.rst
recurrent_group_cn.md
hierarchical_layer_cn.rst
hrnn_rnn_api_compare_cn.rst
doc/howto/deep_model/rnn/index_en.rst
浏览文件 @ c27e71e2
RNN Models
==========
.. toctree::
:maxdepth: 1
rnn_config_en.rst
doc/index_cn.rst
浏览文件 @ c27e71e2
......@@ -5,7 +5,6 @@ PaddlePaddle 文档
:maxdepth: 1
getstarted/index_cn.rst
tutorials/index_cn.md
howto/index_cn.rst
api/index_cn.rst
faq/index_cn.rst
doc/index_en.rst
浏览文件 @ c27e71e2
......@@ -5,8 +5,6 @@ PaddlePaddle Documentation
:maxdepth: 1
getstarted/index_en.rst
tutorials/index_en.md
howto/index_en.rst
api/index_en.rst
about/index_en.rst
\ No newline at end of file
doc_theme/templates/layout.html
浏览文件 @ c27e71e2
......@@ -114,10 +114,7 @@
</ul>
</div>
<ul class="site-page-links">
<li><a>Home</a></li>
<li><a>Get Started</a></li>
<li class="active"><a>Documentation</a></li>
<li><a>About Us</a></li>
<li><a href="/">Home</a></li>
</ul>
</div>
<div class="doc-module">
......@@ -137,7 +134,7 @@
{{ toctree }}
{% endblock %}
</nav>
{% if toc %}
{% if False %}
<nav class="local-toc">{{ toc }}</nav>
{% endif %}
<section class="doc-content-wrap">
......@@ -168,7 +165,8 @@
VERSION:'{{ release|e }}',
COLLAPSE_INDEX:false,
FILE_SUFFIX:'{{ '' if no_search_suffix else file_suffix }}',
HAS_SOURCE: {{ has_source|lower }}
HAS_SOURCE: {{ has_source|lower }},
SOURCELINK_SUFFIX: ".txt",
};
</script>
{%- for scriptfile in script_files %}
......
paddle/cuda/CMakeLists.txt
浏览文件 @ c27e71e2
......@@ -21,16 +21,13 @@ set(CUDA_CXX_WITH_GPU_SOURCES
if(WITH_GPU)
set(CUDA_CXX_SOURCES
src/hl_dso_loader.cc
src/hl_warpctc_wrap.cc
${CUDA_CXX_WITH_GPU_SOURCES})
set_source_files_properties(${CUDA_CXX_SOURCES}
PROPERTIES COMPILE_FLAGS "-D__NVCC__")
else()
set(CUDA_CXX_SOURCES
src/hl_dso_loader.cc
src/hl_warpctc_wrap.cc)
set(CUDA_CXX_SOURCES src/hl_warpctc_wrap.cc)
endif()
set(CUDA_CU_SOURCES
......@@ -47,7 +44,6 @@ set(CUDA_CU_SOURCES
set(CUDA_HEADERS
include/hl_time.h
include/hl_dso_loader.h
include/hl_warpctc_wrap.h
include/hl_sequence.h
include/hl_cuda_cublas.h
......
paddle/cuda/include/hl_activation_functions.h
浏览文件 @ c27e71e2
......@@ -40,18 +40,18 @@ public:
namespace gpu {
static __device__ Active<real>::forward forward[] = HPPL_ACTIVE_FUNCTION;
static __device__ Active<real>::backward backward[] = HPPL_ACTIVE_FUNCTION;
}
} // namespace gpu
#else
namespace cpu {
static Active<real>::forward forward[] = HPPL_ACTIVE_FUNCTION;
static Active<real>::backward backward[] = HPPL_ACTIVE_FUNCTION;
}
} // namespace cpu
#ifdef __AVX__
namespace avx {
static Active<__m256>::forward forward[] = HPPL_ACTIVE_FUNCTION;
static Active<__m256>::backward backward[] = HPPL_ACTIVE_FUNCTION;
}
} // namespace avx
#endif
#endif
......
paddle/cuda/include/hl_cnn.h
浏览文件 @ c27e71e2
......@@ -273,23 +273,23 @@ extern void hl_bilinear_forward(const real* inData,
const real ratioW);
/**
* @brief Bilinear interpolation backward.
*
* @param[out] inGrad input gradient.
* @param[in] inImgH input image height.
* @param[in] inImgW input image width.
* @param[in] inputH input batchSize.
* @param[in] inputW input image data dim.
* @param[in] outGrad output gradient.
* @param[in] outImgH output image height.
* @param[in] outImgW output image width.
* @param[in] outputH output batchSize.
* @param[in] outputW output image data dim.
* @param[in] numChannels number of channels.
* @param[in] ratioH inImgH / outImgH.
* @param[in] ratioW inImgW / outImgW.
*
*/
* @brief Bilinear interpolation backward.
*
* @param[out] inGrad input gradient.
* @param[in] inImgH input image height.
* @param[in] inImgW input image width.
* @param[in] inputH input batchSize.
* @param[in] inputW input image data dim.
* @param[in] outGrad output gradient.
* @param[in] outImgH output image height.
* @param[in] outImgW output image width.
* @param[in] outputH output batchSize.
* @param[in] outputW output image data dim.
* @param[in] numChannels number of channels.
* @param[in] ratioH inImgH / outImgH.
* @param[in] ratioW inImgW / outImgW.
*
*/
extern void hl_bilinear_backward(real* inGrad,
const size_t inImgH,
const size_t inImgW,
......
paddle/cuda/src/hl_cuda_cublas.cc
浏览文件 @ c27e71e2
......@@ -14,10 +14,9 @@ limitations under the License. */
#include "hl_cuda_cublas.h"
#include <sys/time.h>
#include <mutex>
#include "hl_cuda.h"
#include "hl_dso_loader.h"
#include "hl_thread.ph"
#include "paddle/utils/DynamicLoader.h"
#include "paddle/utils/Logging.h"
namespace dynload {
......
paddle/cuda/src/hl_cuda_cudnn.cc
浏览文件 @ c27e71e2
......@@ -15,10 +15,9 @@ limitations under the License. */
#include "hl_cuda_cudnn.h"
#include <cudnn.h>
#include <gflags/gflags.h>
#include <mutex>
#include "hl_cuda_cudnn.ph"
#include "hl_dso_loader.h"
#include "hl_thread.ph"
#include "paddle/utils/DynamicLoader.h"
#include "paddle/utils/Logging.h"
DEFINE_int32(cudnn_conv_workspace_limit_in_mb,
......
paddle/cuda/src/hl_cuda_device.cc
浏览文件 @ c27e71e2
......@@ -21,11 +21,10 @@ limitations under the License. */
#include <sys/syscall.h>
#include <sys/time.h>
#include <unistd.h>
#include <mutex>
#include "hl_cuda.ph"
#include "hl_thread.ph"
#include "hl_dso_loader.h"
#include "paddle/utils/Logging.h"
#include "paddle/utils/DynamicLoader.h"
// clang-format on
namespace dynload {
......
paddle/cuda/src/hl_warpctc_wrap.cc
浏览文件 @ c27e71e2
......@@ -14,7 +14,7 @@ limitations under the License. */
#include "hl_warpctc_wrap.h"
#include <mutex>
#include "hl_dso_loader.h"
#include "paddle/utils/DynamicLoader.h"
#include "paddle/utils/Logging.h"
namespace dynload {
......
paddle/function/CMakeLists.txt
浏览文件 @ c27e71e2
......@@ -12,7 +12,7 @@ endif()
add_library(paddle_function STATIC ${cpp_files} ${cu_objs})
add_dependencies(paddle_function ${external_project_dependencies})
add_dependencies(paddle_function gen_proto_cpp)
if(WITH_GPU)
if(WITH_TESTING)
......
paddle/function/MulOpTest.cpp
浏览文件 @ c27e71e2
......@@ -74,9 +74,9 @@ TEST(MulOp, DDDMatrixMul) {
}
/**
* C += A * B, B, C dense, A sparse
* dense = sparse * dense
*/
* C += A * B, B, C dense, A sparse
* dense = sparse * dense
*/
void testFuncDSparseDMatrix(
size_t dimM, size_t dimN, size_t dimK, size_t nnz, SparseFormat FORMAT) {
real scaleT = 1.0;
......@@ -119,9 +119,9 @@ TEST(MuLOp, DSparseDMul) {
}
/**
* C += A * B, A, C dense, B sparse
* dense = dense * sparse
*/
* C += A * B, A, C dense, B sparse
* dense = dense * sparse
*/
void testFuncDDSparseMatrix(
size_t dimM, size_t dimN, size_t dimK, size_t nnz, SparseFormat FORMAT) {
real scaleT = 1.0;
......@@ -165,9 +165,9 @@ TEST(MulOp, DDSparseMul) {
}
/**
* C += A * B, A sparse, B, C dense
* sparse = dense * dense
*/
* C += A * B, A sparse, B, C dense
* sparse = dense * dense
*/
void testFuncSparseDDMatrix(
size_t dimM, size_t dimN, size_t dimK, size_t nnz, SparseFormat FORMAT) {
real scaleT = 1.0;
......
paddle/gserver/gradientmachines/GradientMachine.cpp
浏览文件 @ c27e71e2
......@@ -21,7 +21,6 @@ limitations under the License. */
#include "MultiGradientMachine.h"
#include "MultiNetwork.h"
#include "NeuralNetwork.h"
#include "NeuralNetwork.h"
#include "ParallelNeuralNetwork.h"
#include "hl_gpu.h"
......
paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
浏览文件 @ c27e71e2
......@@ -637,7 +637,7 @@ void RecurrentGradientMachine::removeBeamSearchStatisticsCallbacks() {
/* create scattered id infomation for all realLayer of inFrameLines one time.
* If hasSubseq, will also create scattered sequenceStartPositions infomation
* for all realLayer of inFrameLines one time.
*/
*/
void RecurrentGradientMachine::createInFrameInfo(int inlinkId,
const Argument& input,
......
paddle/gserver/gradientmachines/RecurrentGradientMachine.h
浏览文件 @ c27e71e2
......@@ -107,18 +107,18 @@ public:
DropCallback;
/**
* @brief NormOrDropNodeCallback
*
* Normalize a path's probabilities or just drop it by modifying path.logProb
*
* The first parameter is sequence index in a batch
*
* The second parameter is path.ids
*
* The third parameter is probabilites for each node in this path.
*
* The fourth parameter is the probability of the whole path.
*/
* @brief NormOrDropNodeCallback
*
* Normalize a path's probabilities or just drop it by modifying path.logProb
*
* The first parameter is sequence index in a batch
*
* The second parameter is path.ids
*
* The third parameter is probabilites for each node in this path.
*
* The fourth parameter is the probability of the whole path.
*/
typedef std::function<void(
int seqId, const std::vector<int>&, std::vector<real>&, real*)>
NormOrDropNodeCallback;
......@@ -348,9 +348,9 @@ protected:
int targetInfoInlinkId_;
/* create scattered id infomation for all realLayer of inFrameLines one time.
* If hasSubseq, will also create scattered sequenceStartPositions infomation
* for all realLayer of inFrameLines one time.
*/
* If hasSubseq, will also create scattered sequenceStartPositions infomation
* for all realLayer of inFrameLines one time.
*/
void createInFrameInfo(int inlinks_id,
const Argument& input,
PassType passType);
......
paddle/gserver/layers/Layer.h
浏览文件 @ c27e71e2
......@@ -106,9 +106,9 @@ protected:
public:
/**
* Wait until all input value ready.
* Called before Layer::forward() function.
*/
* Wait until all input value ready.
* Called before Layer::forward() function.
*/
virtual void waitInputValue();
/**
......@@ -118,9 +118,9 @@ public:
virtual void copyOutputToOtherDevice();
/**
* Wait until all output grad ready and merge them to output_.grad.
* Called before Layer::backward() function.
*/
* Wait until all output grad ready and merge them to output_.grad.
* Called before Layer::backward() function.
*/
virtual void waitAndMergeOutputGrad();
/**
......
paddle/gserver/layers/RotateLayer.h
浏览文件 @ c27e71e2
......@@ -29,7 +29,7 @@ namespace paddle {
*
* The config file api is rotate_layer
*
*/
*/
class RotateLayer : public Layer {
public:
......
paddle/gserver/layers/SequencePoolLayer.cpp
浏览文件 @ c27e71e2
......@@ -60,7 +60,7 @@ void SequencePoolLayer::forward(PassType passType) {
* thus, in this case, output_ has no sequenceStartPositions.
* If type_ = kSeq, seq has sub-seq degrades to a seq, thus, only in this
* case, we should compute the new sequenceStartPositions.
*/
*/
if (type_) {
CHECK(input.subSequenceStartPositions)
<< "when trans_type = seq, input must hasSubseq";
......
paddle/gserver/tests/sequence_layer_group.conf
浏览文件 @ c27e71e2
......@@ -48,8 +48,7 @@ lstm = lstmemory_group(
size=hidden_dim,
act=TanhActivation(),
gate_act=SigmoidActivation(),
state_act=TanhActivation(),
lstm_layer_attr=ExtraLayerAttribute(error_clipping_threshold=50))
state_act=TanhActivation())
lstm_last = last_seq(input=lstm)
......
paddle/gserver/tests/sequence_nest_layer_group.conf
浏览文件 @ c27e71e2
......@@ -51,8 +51,7 @@ def lstm_group(lstm_group_input):
size=hidden_dim,
act=TanhActivation(),
gate_act=SigmoidActivation(),
state_act=TanhActivation(),
lstm_layer_attr=ExtraLayerAttribute(error_clipping_threshold=50))
state_act=TanhActivation())
return lstm_output
......
paddle/math/MathFunctions.cpp
浏览文件 @ c27e71e2
......@@ -15,6 +15,54 @@ limitations under the License. */
#include "MathFunctions.h"
#include "hl_matrix_apply.cuh"
#include "hl_matrix_ops.cuh"
#include "paddle/utils/DynamicLoader.h"
namespace dynload {
std::once_flag lapack_dso_flag;
void* lapack_dso_handle = nullptr;
/**
* The following macro definition can generate structs
* (for each function) to dynamic load lapack routine
* via operator overloading.
*
* note: default dynamic linked libs
*/
#define DYNAMIC_LOAD_LAPACK_WRAP(__name) \
struct DynLoad__##__name { \
template <typename... Args> \
auto operator()(Args... args) -> decltype(__name(args...)) { \
using lapack_func = decltype(__name(args...)) (*)(Args...); \
std::call_once(lapack_dso_flag, GetLapackDsoHandle, &lapack_dso_handle); \
void* p_##__name = dlsym(lapack_dso_handle, #__name); \
return reinterpret_cast<lapack_func>(p_##__name)(args...); \
} \
} __name; // struct DynLoad__##__name
// clang-format off
#ifdef PADDLE_USE_ATLAS
#define PADDLE_SGETRF clapack_sgetrf
#define PADDLE_DGETRF clapack_dgetrf
#define PADDLE_SGETRI clapack_sgetri
#define PADDLE_DGETRI clapack_dgetri
#else
#define PADDLE_SGETRF LAPACKE_sgetrf
#define PADDLE_DGETRF LAPACKE_dgetrf
#define PADDLE_SGETRI LAPACKE_sgetri
#define PADDLE_DGETRI LAPACKE_dgetri
#endif
#define LAPACK_ROUTINE_EACH(__macro) \
__macro(PADDLE_SGETRF) \
__macro(PADDLE_DGETRF) \
__macro(PADDLE_SGETRI) \
__macro(PADDLE_DGETRI)
// clang-format on
LAPACK_ROUTINE_EACH(DYNAMIC_LOAD_LAPACK_WRAP)
} // namespace dynload
namespace paddle {
......@@ -85,16 +133,7 @@ int getrf<float>(const CBLAS_ORDER order,
float* A,
const int lda,
int* ipiv) {
#ifdef PADDLE_USE_LAPACK
#ifdef PADDLE_USE_ATLAS
return clapack_sgetrf(order, M, N, A, lda, ipiv);
#else
return LAPACKE_sgetrf(order, M, N, A, lda, ipiv);
#endif
#else
LOG(FATAL) << "Not implemented";
#endif
return 0;
return dynload::PADDLE_SGETRF(order, M, N, A, lda, ipiv);
}
template <>
......@@ -104,16 +143,7 @@ int getrf<double>(const CBLAS_ORDER order,
double* A,
const int lda,
int* ipiv) {
#ifdef PADDLE_USE_LAPACK
#ifdef PADDLE_USE_ATLAS
return clapack_dgetrf(order, M, N, A, lda, ipiv);
#else
return LAPACKE_dgetrf(order, M, N, A, lda, ipiv);
#endif
#else
LOG(FATAL) << "Not implemented";
#endif
return 0;
return dynload::PADDLE_DGETRF(order, M, N, A, lda, ipiv);
}
template <>
......@@ -122,16 +152,7 @@ int getri<float>(const CBLAS_ORDER order,
float* A,
const int lda,
const int* ipiv) {
#ifdef PADDLE_USE_LAPACK
#ifdef PADDLE_USE_ATLAS
return clapack_sgetri(order, N, A, lda, ipiv);
#else
return LAPACKE_sgetri(order, N, A, lda, ipiv);
#endif
#else
LOG(FATAL) << "Not implemented";
#endif
return 0;
return dynload::PADDLE_SGETRI(order, N, A, lda, ipiv);
}
template <>
......@@ -140,15 +161,7 @@ int getri<double>(const CBLAS_ORDER order,
double* A,
const int lda,
const int* ipiv) {
#ifdef PADDLE_USE_LAPACK
#ifdef PADDLE_USE_ATLAS
return clapack_dgetri(order, N, A, lda, ipiv);
#else
return LAPACKE_dgetri(order, N, A, lda, ipiv);
#endif
#else
LOG(FATAL) << "Not implemented";
#endif
return dynload::PADDLE_DGETRI(order, N, A, lda, ipiv);
return 0;
}
......
paddle/math/MathFunctions.h
浏览文件 @ c27e71e2
......@@ -17,14 +17,11 @@ limitations under the License. */
#ifdef PADDLE_USE_MKL
#include <mkl.h>
#ifdef PADDLE_USE_LAPACK
#include <mkl_lapacke.h>
#endif
#else
extern "C" {
#include <cblas.h>
}
#ifdef PADDLE_USE_LAPACK
#ifdef PADDLE_USE_ATLAS
extern "C" {
#include <clapack.h>
......@@ -33,7 +30,6 @@ extern "C" {
#include <lapacke.h>
#endif
#endif
#endif
#include <cmath>
......
paddle/math/tests/TestUtils.h
浏览文件 @ c27e71e2
......@@ -37,7 +37,7 @@ limitations under the License. */
*
* AutoCompare test;
* test.cmpWithoutArg<I...>(function, height, width)
*/
*/
#include <gtest/gtest.h>
#include "TensorCheck.h"
......
paddle/math/tests/test_matrixCompare.cpp
浏览文件 @ c27e71e2
......@@ -21,6 +21,7 @@ limitations under the License. */
#include "paddle/math/Matrix.h"
#include "paddle/math/SparseMatrix.h"
#include "paddle/testing/TestUtil.h"
#include "paddle/utils/DynamicLoader.h"
#include "paddle/utils/Stat.h"
#include "paddle/utils/Util.h"
......@@ -235,10 +236,15 @@ TEST(Matrix, unary) {
testMatrixTranspose(height, width);
testMatrixRotate(height, width);
}
// inverse
#ifdef PADDLE_USE_LAPACK
testMatrixInverse(height);
#endif
// inverse matrix
void** dso_handler = nullptr;
GetLapackDsoHandle(dso_handler);
if (nullptr == *dso_handler) {
LOG(WARNING) << "Failed to find liblapack.so, please specify its path "
"using LD_LIBRARY_PATH.";
} else {
testMatrixInverse(height);
}
}
}
......
paddle/parameter/FirstOrderOptimizer.h
浏览文件 @ c27e71e2
......@@ -126,7 +126,7 @@ protected:
/*
* AdaDelta Optimization.
* http://www.matthewzeiler.com/pubs/googleTR2012/googleTR2012.pdf
*/
*/
class AdaDeltaParameterOptimizer : public ParameterOptimizer {
public:
explicit AdaDeltaParameterOptimizer(const OptimizationConfig& optConfig)
......
paddle/scripts/deb/postinst 0 → 100644
浏览文件 @ c27e71e2
#!/bin/bash
set -e
echo "Post install paddle debian package."
echo "Install some python package used for paddle. You can run "
echo " pip install /usr/opt/paddle/share/wheels/*.whl to install them."
find /usr/ -name '*paddle*.whl' | xargs pip install
paddle/scripts/docker/build.sh
浏览文件 @ c27e71e2
......@@ -5,13 +5,8 @@ set -e
# Set BASE_IMAGE according to env variables
if [ ${WITH_GPU} == "ON" ]; then
BASE_IMAGE="nvidia/cuda:8.0-cudnn5-runtime-ubuntu14.04"
# additional packages to install when building gpu images
GPU_DOCKER_PKG="python-pip python-dev"
else
BASE_IMAGE="python:2.7.13-slim"
# FIXME: python base image uses different python version than WITH_GPU
# need to change PYTHONHOME to /usr/local when using python base image
CPU_DOCKER_PYTHON_HOME_ENV="ENV PYTHONHOME /usr/local"
BASE_IMAGE="ubuntu:14.04"
fi
DOCKERFILE_GPU_ENV=""
......@@ -66,10 +61,7 @@ if [ ${WITH_DOC} == "ON" ]; then
rm -rf /paddle/build_doc
fi
# generate deb package for current build
# FIXME(typhoonzero): should we remove paddle/scripts/deb ?
# FIXME: CPACK_DEBIAN_PACKAGE_DEPENDS removes all dev dependencies, must
# install them in docker
cpack -D CPACK_GENERATOR='DEB' -D CPACK_DEBIAN_PACKAGE_DEPENDS="" ..
cpack -D CPACK_GENERATOR='DEB' ..
if [[ ${WOBOQ:-OFF} == 'ON' ]]; then
apt-get install -y clang-3.8 llvm-3.8 libclang-3.8-dev
......@@ -97,32 +89,30 @@ fi
paddle version
if [[ -n ${APT_MIRROR} ]]; then
MIRROR_UPDATE="sed -i '${APT_MIRROR}' /etc/apt/sources.list && \\"
else
MIRROR_UPDATE="\\"
fi
cat > /paddle/build/Dockerfile <<EOF
FROM ${BASE_IMAGE}
MAINTAINER PaddlePaddle Authors <paddle-dev@baidu.com>
ENV HOME /root
ENV LANG en_US.UTF-8
# Use Fix locales to en_US.UTF-8
RUN ${MIRROR_UPDATE}
apt-get update && \
apt-get install -y libgfortran3 libpython2.7 ${GPU_DOCKER_PKG} && \
apt-get clean -y && \
pip install --upgrade pip && \
pip install -U 'protobuf==3.1.0' requests numpy
EOF
if [[ -n ${APT_MIRROR} ]]; then
cat >> /paddle/build/Dockerfile <<EOF
RUN sed -i '${APT_MIRROR}' /etc/apt/sources.list
EOF
fi
cat >> /paddle/build/Dockerfile <<EOF
# Use different deb file when building different type of images
ADD *.deb /usr/local/opt/paddle/deb/
ADD *.deb /
# run paddle version to install python packages first
RUN dpkg -i /usr/local/opt/paddle/deb/*.deb && \
rm -f /usr/local/opt/paddle/deb/*.deb && \
find /usr/ -name '*paddle-*.whl' | xargs pip install && \
RUN apt-get update &&\
apt-get install -y python-pip && pip install -U pip && \
dpkg -i /*.deb ; apt-get install -f -y && \
apt-get clean -y && \
rm -f /*.deb && \
paddle version
${CPU_DOCKER_PYTHON_HOME_ENV}
${DOCKERFILE_CUDNN_DSO}
${DOCKERFILE_GPU_ENV}
# default command shows the paddle version and exit
......
paddle/scripts/travis/docs.sh
浏览文件 @ c27e71e2
......@@ -60,6 +60,7 @@ function deploy_docs() {
deploy_docs "master" "."
deploy_docs "develop" "./develop/"
deploy_docs "release/0.10.0" "./release/0.10.0/"
# Check is there anything changed.
set +e
......
paddle/setup.py.in
浏览文件 @ c27e71e2
......@@ -23,7 +23,7 @@ setup(name="py_paddle",
install_requires = [
'nltk>=3.2.2',
'numpy>=1.8.0', # The numpy is required.
'protobuf>=${PROTOBUF_VERSION}' # The paddle protobuf version
'protobuf==${PROTOBUF_VERSION}' # The paddle protobuf version
],
url='http://www.paddlepaddle.org/',
license='Apache 2.0',
......
paddle/trainer/tests/picojson.h
浏览文件 @ c27e71e2
......@@ -1059,14 +1059,14 @@ inline bool operator==(const value& x, const value& y) {
}
inline bool operator!=(const value& x, const value& y) { return !(x == y); }
}
} // namespace picojson
namespace std {
template <>
inline void swap(picojson::value& x, picojson::value& y) {
x.swap(y);
}
}
} // namespace std
inline std::istream& operator>>(std::istream& is, picojson::value& x) {
picojson::set_last_error(std::string());
......
paddle/cuda/src/hl_dso_loader.cc → paddle/utils/DynamicLoader.cpp
浏览文件 @ c27e71e2
......@@ -12,9 +12,9 @@ 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 "hl_dso_loader.h"
#include "DynamicLoader.h"
#include <gflags/gflags.h>
#include "paddle/utils/Logging.h"
#include "Logging.h"
DEFINE_string(cudnn_dir,
"",
......@@ -30,6 +30,8 @@ DEFINE_string(cuda_dir,
DEFINE_string(warpctc_dir, "", "Specify path for loading libwarpctc.so.");
DEFINE_string(lapack_dir, "", "Specify path for loading liblapack.so.");
static inline std::string join(const std::string& part1,
const std::string& part2) {
// directory separator
......@@ -160,3 +162,11 @@ void GetWarpCTCDsoHandle(void** dso_handle) {
GetDsoHandleFromSearchPath(FLAGS_warpctc_dir, "libwarpctc.so", dso_handle);
#endif
}
void GetLapackDsoHandle(void** dso_handle) {
#if defined(__APPLE__) || defined(__OSX__)
GetDsoHandleFromSearchPath(FLAGS_lapack_dir, "liblapack.dylib", dso_handle);
#else
GetDsoHandleFromSearchPath(FLAGS_lapack_dir, "liblapack.so", dso_handle);
#endif
}
paddle/cuda/include/hl_dso_loader.h → paddle/utils/DynamicLoader.h
浏览文件 @ c27e71e2
......@@ -12,13 +12,13 @@ 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. */
#ifndef HL_DSO_LOADER_H_
#define HL_DSO_LOADER_H_
#ifndef DYNAMIC_LOAD_H_
#define DYNAMIC_LOAD_H_
#include <dlfcn.h>
#include <memory>
#include <mutex>
#include <string>
#include "hl_base.h"
/**
* @brief load the DSO of CUBLAS
......@@ -52,4 +52,12 @@ void GetCurandDsoHandle(void** dso_handle);
*/
void GetWarpCTCDsoHandle(void** dso_handle);
#endif // HL_DSO_LOADER_H_
/**
* @brief load the DSO of lapack
*
* @param **dso_handle dso handler
*
*/
void GetLapackDsoHandle(void** dso_handle);
#endif // DYNAMIC_LOAD_H_
python/paddle/trainer_config_helpers/attrs.py
浏览文件 @ c27e71e2
......@@ -208,12 +208,15 @@ class ExtraLayerAttribute(object):
drop_rate=None,
device=None):
self.attr = dict()
if isinstance(error_clipping_threshold, float):
assert error_clipping_threshold > 0
self.attr["error_clipping_threshold"] = error_clipping_threshold
if isinstance(drop_rate, float):
assert drop_rate > 0
if error_clipping_threshold is not None:
error_clipping_threshold = float(error_clipping_threshold)
if error_clipping_threshold < 0:
raise ValueError("Error clipping must > 0")
self.attr['error_clipping_threshold'] = error_clipping_threshold
if drop_rate is not None:
drop_rate = float(drop_rate)
if drop_rate < 0:
raise ValueError("Dropout rate must > 0")
self.attr["drop_rate"] = drop_rate
if isinstance(device, int):
......
python/paddle/trainer_config_helpers/layers.py
浏览文件 @ c27e71e2
......@@ -84,6 +84,7 @@ __all__ = [
'GeneratedInput',
'SubsequenceInput',
'gru_step_layer',
'gru_step_naive_layer',
'recurrent_layer',
'BaseGeneratedInput',
'conv_operator',
......@@ -2286,7 +2287,7 @@ def img_pool_layer(input,
type_name = pool_type.name + '-projection' \
if (
isinstance(pool_type, AvgPooling) or isinstance(pool_type, MaxPooling)) \
isinstance(pool_type, AvgPooling) or isinstance(pool_type, MaxPooling)) \
else pool_type.name
pool_size_y = pool_size if pool_size_y is None else pool_size_y
......@@ -3086,6 +3087,78 @@ def gru_step_layer(input,
activation=act)
@wrap_bias_attr_default()
@wrap_param_attr_default()
@wrap_act_default(param_names=['gate_act'], act=SigmoidActivation())
@wrap_act_default(act=TanhActivation())
@wrap_name_default('gru_step_naive')
@layer_support(ERROR_CLIPPING, DROPOUT)
def gru_step_naive_layer(input,
output_mem,
size=None,
name=None,
act=None,
gate_act=None,
bias_attr=None,
param_attr=None,
layer_attr=None):
"""
GRU Step Layer, but using MixedLayer to generate. It support ERROR_CLIPPING
and DROPOUT.
:param input:
:param output_mem:
:param size:
:param name:
:param act:
:param gate_act:
:param bias_attr:
:param param_attr:
:param layer_attr:
:return:
"""
if input.size % 3 != 0:
raise ValueError("GruStep input size must be divided by 3")
if size is None:
size = input.size / 3
def __gate__(gate_name, offset):
with mixed_layer(
name=name + "_" + gate_name,
size=size,
layer_attr=layer_attr,
bias_attr=bias_attr,
act=gate_act) as gate:
gate += identity_projection(input=input, offset=offset)
gate += full_matrix_projection(
input=output_mem, param_attr=param_attr)
return gate
update_gate = __gate__("update", 0)
reset_gate = __gate__("reset", size)
with mixed_layer(
name=name + "_reset_output", bias_attr=False) as reset_output:
reset_output += dotmul_operator(a=output_mem, b=reset_gate)
with mixed_layer(
name=name + "_output_candidate",
size=size,
layer_attr=layer_attr,
bias_attr=bias_attr,
act=act) as output_candidate:
output_candidate += identity_projection(input=input, offset=2 * size)
output_candidate += full_matrix_projection(
input=reset_output, param_attr=param_attr)
with mixed_layer(name=name) as output:
output += identity_projection(output_mem)
output += dotmul_operator(a=output_mem, b=update_gate, scale=-1.0)
output += dotmul_operator(a=output_candidate, b=update_gate)
return output
@wrap_name_default()
@layer_support()
def get_output_layer(input, arg_name, name=None, layer_attr=None):
......
python/paddle/trainer_config_helpers/networks.py
浏览文件 @ c27e71e2
......@@ -825,7 +825,8 @@ def gru_unit(input,
gru_param_attr=None,
act=None,
gate_act=None,
gru_layer_attr=None):
gru_layer_attr=None,
naive=False):
"""
Define calculations that a gated recurrent unit performs in a single time
step. This function itself is not a recurrent layer, so that it can not be
......@@ -857,7 +858,12 @@ def gru_unit(input,
out_mem = memory(name=name, size=size)
gru_out = gru_step_layer(
if naive:
__step__ = gru_step_naive_layer
else:
__step__ = gru_step_layer
gru_out = __step__(
name=name,
input=input,
output_mem=out_mem,
......@@ -879,7 +885,8 @@ def gru_group(input,
gru_param_attr=None,
act=None,
gate_act=None,
gru_layer_attr=None):
gru_layer_attr=None,
naive=False):
"""
gru_group is a recurrent layer group version of Gated Recurrent Unit. It
does exactly the same calculation as the grumemory layer does. A promising
......@@ -928,7 +935,8 @@ def gru_group(input,
gru_param_attr=gru_param_attr,
act=act,
gate_act=gate_act,
gru_layer_attr=gru_layer_attr)
gru_layer_attr=gru_layer_attr,
naive=naive)
return recurrent_group(
name='%s_recurrent_group' % name,
......@@ -949,7 +957,8 @@ def simple_gru(input,
gru_param_attr=None,
act=None,
gate_act=None,
gru_layer_attr=None):
gru_layer_attr=None,
naive=False):
"""
You maybe see gru_step_layer, grumemory in layers.py, gru_unit, gru_group,
simple_gru in network.py. The reason why there are so many interfaces is
......@@ -1018,7 +1027,8 @@ def simple_gru(input,
gru_param_attr=gru_param_attr,
act=act,
gate_act=gate_act,
gru_layer_attr=gru_layer_attr)
gru_layer_attr=gru_layer_attr,
naive=naive)
@wrap_name_default('simple_gru2')
......
python/paddle/trainer_config_helpers/tests/configs/protostr/projections.protostr
浏览文件 @ c27e71e2
......@@ -320,6 +320,7 @@ layers {
}
}
drop_rate: 0.5
error_clipping_threshold: 40.0
}
parameters {
name: "___embedding_0__.w0"
......
python/paddle/v2/layer.py
浏览文件 @ c27e71e2
......@@ -356,6 +356,9 @@ def mixed(size=0,
return MixedLayerV2(size, input, name, act, bias_attr, layer_attr)
mixed.__doc__ = conf_helps.mixed_layer.__doc__
class RecurrentLayerInput(Layer):
def __init__(self, recurrent_name, index, parent_layers):
parents_len = len(parent_layers)
......@@ -404,6 +407,8 @@ data.__name__ = 'data'
AggregateLevel = conf_helps.layers.AggregateLevel
ExpandLevel = conf_helps.layers.ExpandLevel
memory = MemoryV2
memory.__name__ = 'memory'
memory.__doc__ = conf_helps.memory.__doc__
def __layer_name_mapping__(inname):
......@@ -512,6 +517,9 @@ def recurrent_group(step, input, name=None):
return retv
recurrent_group.__doc__ = conf_helps.recurrent_group.__doc__
@wrap_name_default()
def beam_search(step,
input,
......@@ -579,6 +587,8 @@ def beam_search(step,
return tmp
beam_search.__doc__ = conf_helps.beam_search.__doc__
__projection_names__ = filter(lambda x: x.endswith('_projection'),
dir(conf_helps))
......
python/setup.py.in
浏览文件 @ c27e71e2
......@@ -15,6 +15,9 @@ setup(name='paddle',
description='Parallel Distributed Deep Learning',
install_requires=[
"requests",
"numpy",
"protobuf==${PROTOBUF_VERSION}",
"matplotlib",
],
packages=packages,
package_dir={
......
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