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 f85e16f1 编写于 作者: S sneaxiy
浏览文件
操作

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

上级 c53c5d45 26ae6111
隐藏空白更改
内联 并排
Showing with 998 addition and 671 deletion
.travis.yml
浏览文件 @ f85e16f1
......@@ -31,7 +31,7 @@ script:
if [[ "$JOB" != "doc" ]]; then exit 0; fi;
# For document only
if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi;
if [[ "$TRAVIS_BRANCH" != "develop" && ! "$TRAVIS_BRANCH" =~ ^v[[:digit:]]+\.[[:digit:]]+(\.[[:digit:]]+)?(-\S*)?$ ]]; then exit 0; fi;
if [[ "$TRAVIS_BRANCH" != "develop" && ! "$TRAVIS_BRANCH" =~ ^v|release/[[:digit:]]+\.[[:digit:]]+(\.[[:digit:]]+)?(-\S*)?$ ]]; then exit 0; fi;
export DEPLOY_DOCS_SH=https://raw.githubusercontent.com/PaddlePaddle/PaddlePaddle.org/master/scripts/deploy/deploy_docs.sh
export DOCS_DIR=`pwd`
cd ..
......
Dockerfile
浏览文件 @ f85e16f1
......@@ -23,7 +23,7 @@ ENV HOME /root
COPY ./paddle/scripts/docker/root/ /root/
RUN apt-get update && \
apt-get install -y --allow-downgrades \
apt-get install -y --allow-downgrades patchelf \
git python-pip python-dev python-opencv openssh-server bison \
libnccl2=2.1.2-1+cuda8.0 libnccl-dev=2.1.2-1+cuda8.0 \
wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
......
cmake/external/anakin.cmake
浏览文件 @ f85e16f1
......@@ -7,7 +7,17 @@ set(ANAKIN_INSTALL_DIR "${THIRD_PARTY_PATH}/install/anakin" CACHE PATH
set(ANAKIN_INCLUDE "${ANAKIN_INSTALL_DIR}" CACHE STRING "root of Anakin header files")
set(ANAKIN_LIBRARY "${ANAKIN_INSTALL_DIR}" CACHE STRING "path of Anakin library")
set(ANAKIN_COMPILE_EXTRA_FLAGS -Wno-error=unused-variable -Wno-error=format-extra-args -Wno-error=comment -Wno-error=format -Wno-error=switch -Wno-error=return-type -Wno-error=non-virtual-dtor -Wno-reorder -Wno-error=cpp)
set(ANAKIN_COMPILE_EXTRA_FLAGS
-Wno-error=unused-variable -Wno-unused-variable
-Wno-error=format-extra-args -Wno-format-extra-args
-Wno-error=comment -Wno-comment
-Wno-error=format -Wno-format
-Wno-error=switch -Wno-switch
-Wno-error=return-type -Wno-return-type
-Wno-error=non-virtual-dtor -Wno-non-virtual-dtor
-Wno-sign-compare
-Wno-reorder
-Wno-error=cpp)
set(ANAKIN_LIBRARY_URL "https://github.com/pangge/Anakin/releases/download/3.0/anakin_release_simple.tar.gz")
......
cmake/generic.cmake
浏览文件 @ f85e16f1
......@@ -257,8 +257,8 @@ function(cc_test TARGET_NAME)
set(multiValueArgs SRCS DEPS ARGS)
cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
add_executable(${TARGET_NAME} ${cc_test_SRCS})
target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main memory gtest gflags glog)
add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main memory gtest gflags glog)
target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
add_test(NAME ${TARGET_NAME}
COMMAND ${TARGET_NAME} ${cc_test_ARGS}
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
......@@ -324,8 +324,8 @@ function(nv_test TARGET_NAME)
set(multiValueArgs SRCS DEPS)
cmake_parse_arguments(nv_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
cuda_add_executable(${TARGET_NAME} ${nv_test_SRCS})
target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main memory gtest gflags glog)
add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main memory gtest gflags glog)
target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
add_test(${TARGET_NAME} ${TARGET_NAME})
if (nv_test_SERIAL)
set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1)
......
cmake/version.cmake
浏览文件 @ f85e16f1
# Get the latest git tag.
set(PADDLE_VERSION $ENV{PADDLE_VERSION})
set(tmp_version "HEAD")
set(TAG_VERSION_REGEX "[0-9]+\\.[0-9]+\\.[0-9]+(\\.(a|b|rc)\\.[0-9]+)?")
set(COMMIT_VERSION_REGEX "[0-9a-f]+[0-9a-f]+[0-9a-f]+[0-9a-f]+[0-9a-f]+")
while ("${PADDLE_VERSION}" STREQUAL "")
execute_process(
COMMAND ${GIT_EXECUTABLE} describe --tags --abbrev=0 ${tmp_version}
COMMAND ${GIT_EXECUTABLE} describe --tags --abbrev=0 --always ${tmp_version}
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
OUTPUT_VARIABLE GIT_TAG_NAME
RESULT_VARIABLE GIT_RESULT
ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
if (NOT ${GIT_RESULT})
# Check the tag is a correct version
if (${GIT_TAG_NAME} MATCHES "v[0-9]+\\.[0-9]+\\.[0-9]+(\\.(a|b|rc)\\.[0-9]+)?")
if (${GIT_TAG_NAME} MATCHES "${COMMIT_VERSION_REGEX}")
# if no tag was found, set PADDLE_VERSION to latest
set(PADDLE_VERSION "latest")
elseif (${GIT_TAG_NAME} MATCHES "v${TAG_VERSION_REGEX}")
string(REPLACE "v" "" PADDLE_VERSION ${GIT_TAG_NAME})
else() # otherwise, get the previous git tag name.
set(tmp_version "${GIT_TAG_NAME}~1")
......
doc/fluid/api/transpiler.rst
浏览文件 @ f85e16f1
......@@ -14,6 +14,15 @@ DistributeTranspiler
:members:
:noindex:
.. _api_fluid_transpiler_InferenceTranspiler:
InferenceTranspiler
-------------------
.. autoclass:: paddle.fluid.transpiler.InferenceTranspiler
:members:
:noindex:
.. _api_fluid_transpiler_memory_optimize:
memory_optimize
......
doc/fluid/design/dist_train/dist_train_nccl2.md 0 → 100644
浏览文件 @ f85e16f1
# Distributed Training with NCCL2
We design a pattern that can enable training with `ParallelExecutor` and
using [NCCL2](https://developer.nvidia.com/nccl) as it's collective
communication library.
In `ParallelExecutor` we can use `AllReduce` or `Reduce` and `Broadcast`
to do multi GPU training. And if we initialize NCCL2 communicators as
ranks in a distributed environment, we can simply run the `ParallelExecutor`
as a distributed program! The only thing that may be different than in
the single node version is that we need to broadcast the NCCL unique ID
to all the nodes, and initialize communicators using that ID, so NCCL2
will know each other as ranks.
To achieve this feature, we introduce a new operator: `gen_nccl_id` op,
so we are ***not*** "bind to" running NCCL2 with MPI, we can run it in
what ever platform you like.
It have two running modes:
1. Generate and broadcast mode, which should be used on trainer 0;
1. Listen and fetch mode, which should be used on trainers other than 0.
In both two modes, this op can save the NCCL ID into current scope as a
persistable variable, Then we can insert this op at the end of
"startup program" of fluid, so that all workers can get the same ID to
initialize NCCL communicator objects.
<img src="src/ncc2_design.png">
The above figure indicates the general process when training with NCCL2
distributed. Each trainer have the number of communicators equal to the
number of GPUs, but the ranks should match the global ranks number: here
we have total 8 GPUs, so `nranks==8`, for each trainer, the ranks should
be from 0 ~ 3 on trainer 0 and 4 ~ 7 on trainer 1.
doc/fluid/design/dist_train/distributed_lookup_table_design.md
浏览文件 @ f85e16f1
# Design Doc: Distributed Lookup Table Operator
A lookup table operator in PaddlePaddle where the table could be out
A distribute lookup table operator in PaddlePaddle where the table could be out
of the memory of a computer.
## Background
......@@ -24,14 +24,14 @@ memory, so we'd need a distributed storage service, which supports the
lookup of rows.
The following figure illustrates the multiplication of x with two
non-zero elements, or say, two symbols, and a lookup table W:
non-zero elements, or say two symbols, and a lookup table W:
![lookup table](./src/lookup_table.png)
### The Backward Algorithm
The backward algorithm computes W'(x) using W(x). W'(x) has the same
scale of size as W(x) and is much smaller than W.
the scale of size as W(x) and is much smaller than W.
To optimize W given W', we can do simple SGD update:
......@@ -44,85 +44,46 @@ $$W = f(W, W')$$
The following figure illustrates the backward pass of the lookup
operator: ![lookup table training](./src/lookup_table_training.png)
## Distributed Storage Service
The forward algorithm requires a distributed storage service for W.
The backward algorithm prefers that the storage system can apply the
optimization algorithm on W. The following two sections describe two
solutions -- the former doesn't require that the storage service can
do optimization, the latter does.
### Storage Service Doesn't Optimize
In this design, we use highly-optimized distributed storage, e.g.,
memcached, as the storage service, and we run the optimization
algorithm on parameter servers of PaddlePaddle. The following figure
illustrates the training process.
<!--
Note: please update the following URL when update this digraph.
<img src='https://g.gravizo.com/svg?
digraph G {
rankdir="LR";
subgraph cluster1 {
P1 [label="pserver 1"];
P2 [label="pserver 2"];
T1 [label="trainer 1"];
T2 [label="trainer 2"];
T3 [label="trainer 3"];
}
KV [label="memcached"];
T1 -> P1;
T1 -> P2;
T2 -> P1;
T2 -> P2;
T3 -> P1;
T3 -> P2;
P1 -> KV [color=gray, weight=0.1];
KV -> P1 [color=gray, weight=0.1];
P2 -> KV [color=gray, weight=0.1];
KV -> P2 [color=gray, weight=0.1];
KV -> T1 [color=gray, weight=0.1];
KV -> T2 [color=gray, weight=0.1];
KV -> T3 [color=gray, weight=0.1];
}
)
'/>
-->
<img src='https://g.gravizo.com/svg?%20digraph%20G%20{%20rankdir=%22LR%22;%20subgraph%20cluster1%20{%20P1%20[label=%22pserver%201%22];%20P2%20[label=%22pserver%202%22];%20T1%20[label=%22trainer%201%22];%20T2%20[label=%22trainer%202%22];%20T3%20[label=%22trainer%203%22];%20}%20KV%20[label=%22memcached%22];%20T1%20-%3E%20P1;%20T1%20-%3E%20P2;%20T2%20-%3E%20P1;%20T2%20-%3E%20P2;%20T3%20-%3E%20P1;%20T3%20-%3E%20P2;%20P1%20-%3E%20KV%20[color=gray,%20weight=0.1];%20KV%20-%3E%20P1%20[color=gray,%20weight=0.1];%20P2%20-%3E%20KV%20[color=gray,%20weight=0.1];%20KV%20-%3E%20P2%20[color=gray,%20weight=0.1];%20KV%20-%3E%20T1%20[color=gray,%20weight=0.1];%20KV%20-%3E%20T2%20[color=gray,%20weight=0.1];%20KV%20-%3E%20T3%20[color=gray,%20weight=0.1];%20}'/>
Each trainer runs the forward and backward passes using their local
data:
1. In the forward pass, when a trainer runs the forward algorithm of a
lookup operator, it retrieves W(x) from the storage service.
1. The trainer computes W'(x) in the backward pass using W(x).
During the global update process:
1. Each trainer uploads its W'(x) to parameter servers.
1. The parameter server runs the optimization algorithm, e.g., the
Adam optimization algorithm, which requires that
1. The parameter server retrieves W(x) from memcached, and
1. The parameter server pushes $\Delta W(x)=f(W(x), lambda \sum_j
W'(x))$ to memcached, where $f$ denotes the optimization
algorithm.
### Storage Service Does Optimize
This design is very similar to the above one, except that the
optimization algorithm $f$ runs on the storage service.
- Pro: parameter servers do not retrieve W(x) from the storage
service, thus saves half network communication.
- Con: the storage service needs to be able to run the optimization
algorithm.
## Conclusion
Let us do the "storage service does not optimize" solution first, as a
baseline at least, because it is easier to use a well-optimized
distributed storage service like memcached. We can do the "storage
service does optimize" solution later or at the same time, which, if
implemented carefully, should have better performance than the former.
## Distributed Lookup Table
### Problem 1: The lookup table may be very large.
In the condition like the search engine and recommendation system, the number of feature Id may be very large, say 100,000,000,000, then for a float value lookup table of size 8, the total size of the table is:
```
100,000,000,000 * 8 * 4(Bytes) = 2980.23 GB
```
### Solution: Distributed storage
1. Paddle use [SelectedRows](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/selected_rows.md) as the storage format for the lookup table, the lookup table parameter will be split to multi-machine according to the hash of the feature ID, and data will also be split and send to the same machine to prefetch the parameter.
1. For common parameters, the trainer will get the whole parameter for training, but for the big lookup table, the trainer can not store the whole parameter. Because the input data feature is very sparse, every time we only need a few parameters for training, so we use `prefetch_op` to only prefetch the parameter needed to trainer.
### Problem 2. The Id in the lookup table is not sure before training.
The feature Id is calculated by the hash function because the feature data source is so large, we can not get all the Id before training. So we can not initialize the table before training.
### Solution: Id auto growth
At the beginning of training, paddle only malloc the memory for the lookup table at parameter server side, the Id and it's value will not be initialized. During training, when a parameter server received an Id, if it is already in the lookup table, it will return the existing parameter, if the Id does not exist, paddle will add it into the lookup table and initialize the value for it.
### Problem 3: parameter load and save
For common parameters, paddle use trainer to save and load them. But for distributed lookup table, trainer cannot do this because it's large size.
### Solution: Parameter server side save and load
Paddle support parameter server side save and load for distribute lookup table. Each machine of parameter servers will only save and load part of the whole table.
## Architecture
The whole architecture of the distribute lookup table is as below:
### Training steps:
1. Read a batch of data, the data is feature ids.
1. The input ids will be split by `split_ids_op` with the same hash function of the lookup table.
1. The `prefetch_op` use the split result to prefetch parameters back from the lookup table.
1. Run forward-backward to get the gradient of the lookup table.
1. `split_ids_op` split the gradient and then use `send_op` to the parameter server.
1. parameter server update the table with the received gradient.
![distribute lookup table](./src/distributed_lookup_table.jpeg)
doc/v2/howto/capi/workflow_of_capi_cn.md
浏览文件 @ f85e16f1
......@@ -28,9 +28,9 @@
### 准备预测模型
准备预测模型部分，我们以手写数字识别任务为例进行介绍。手写数字识别任务定义了一个含有[两个隐层的简单全连接网络](https://github.com/PaddlePaddle/book/blob/develop/02.recognize_digits/README.cn.md#softmax回归softmax-regression)，网络接受一幅图片作为输入，将图片分类到 0 ~ 9 类别标签之一。完整代码可以查看[此目录](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense) 中的相关脚本。
准备预测模型部分，我们以手写数字识别任务为例进行介绍。手写数字识别任务定义了一个含有[两个隐层的简单全连接网络](https://github.com/PaddlePaddle/book/blob/develop/02.recognize_digits/README.cn.md#softmax回归softmax-regression)，网络接受一幅图片作为输入，将图片分类到 0 ~ 9 类别标签之一。完整代码可以查看[此目录](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/capi/examples/model_inference/dense) 中的相关脚本。
调用C-API开发预测程序需要一个训练好的模型，运行[MNIST手写数字识别目录](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense)下的[mnist_v2.py](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/examples/model_inference/dense/mnist_v2.py)脚本，在终端执行`python mnist_v2.py`，会使用 PaddlePaddle 内置的 [MNIST 数据集](http://yann.lecun.com/exdb/mnist/)进行训练。训练好的模型默认保存在当前运行目录下的`models`目录中。
调用C-API开发预测程序需要一个训练好的模型，运行[MNIST手写数字识别目录](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/capi/examples/model_inference/dense)下的[mnist_v2.py](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/capi/examples/model_inference/dense/mnist_v2.py)脚本，在终端执行`python mnist_v2.py`，会使用 PaddlePaddle 内置的 [MNIST 数据集](http://yann.lecun.com/exdb/mnist/)进行训练。训练好的模型默认保存在当前运行目录下的`models`目录中。
下面，我们将训练结束后存储下来的模型转换成预测模型。
......@@ -48,7 +48,7 @@
dump_v2_config(predict, "trainer_config.bin", True)
```
对[手写数字识别](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense)这个示例，[`mnist_v2.py`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense/mnist_v2.py)脚本集成了序列化神经网络结构的过程，可以直接运行 `python mnist_v2.py --task dump_config` 对神经网络结构进行序列化，结果会写入当前运行目录下的`trainer_config.bin`文件中。
对[手写数字识别](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/capi/examples/model_inference/dense)这个示例，[`mnist_v2.py`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/capi/examples/model_inference/dense/mnist_v2.py)脚本集成了序列化神经网络结构的过程，可以直接运行 `python mnist_v2.py --task dump_config` 对神经网络结构进行序列化，结果会写入当前运行目录下的`trainer_config.bin`文件中。
使用这种方式，需要**在运行时将神经网络的多个可学习参数放在同一个目录中**，C-API可以通过分别指定序列化后的网络结构文件和参数目录来加载训练好的模型。
......@@ -68,7 +68,7 @@
merge_v2_model(net, param_file, output_file)
```
对[手写数字识别](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense)这个示例，可直接运行 `python` [merge_v2_model.py](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference/dense/merge_v2_model.py)。序列化结果会写入当前运行目录下的`output.paddle.model`文件中。使用这种方式，运行时C-API可以通过指定`output.paddle.model`文件的路径来加载预测模型。
对[手写数字识别](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/capi/examples/model_inference/dense)这个示例，可直接运行 `python` [merge_v2_model.py](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/capi/examples/model_inference/dense/merge_v2_model.py)。序列化结果会写入当前运行目录下的`output.paddle.model`文件中。使用这种方式，运行时C-API可以通过指定`output.paddle.model`文件的路径来加载预测模型。
#### 注意事项
1. 为使用C-API，在调用`dump_v2_config`序列化神经网络结构时，参数`binary`必须指定为`True`。
......@@ -77,10 +77,10 @@
### 编写预测代码
预测代码更多详细示例代码请参考[C-API使用示例](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/capi/examples/model_inference) 目录下的代码示例。这一节对图1中预测代码编写的5个步骤进行介绍和说明。
预测代码更多详细示例代码请参考[C-API使用示例](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/capi/examples/model_inference) 目录下的代码示例。这一节对图1中预测代码编写的5个步骤进行介绍和说明。
#### step 1. 初始化PaddlePaddle运行环境
第一步需调用[`paddle_init`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/main.h#L27) 初始化PaddlePaddle运行环境，该接口接受两个参数：参数的个数和参数列表。
第一步需调用[`paddle_init`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/capi/main.h#L27) 初始化PaddlePaddle运行环境，该接口接受两个参数：参数的个数和参数列表。
#### step2. 加载模型
......@@ -88,8 +88,8 @@
概念上，在 PaddlePaddle 内部，一个GradientMachine类的对象管理着一组计算层（PaddlePaddle Layers）来完成前向和反向计算，并处理与之相关的所有细节。在调用C-API预测时，只需进行前向计算而无需调用反向计算。这篇文档之后部分会使用`gradient machine`来特指调用PaddlePaddle C-API创建的GradientMachine类的对象。每一个 `gradient machine` 都会管理维护一份训练好的模型，下面是C-API提供的，两种常用的模型加载方式：
1. 调用[`paddle_gradient_machine_load_parameter_from_disk`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/gradient_machine.h#L61)接口，从磁盘加载预测模型。这时`gradient machine`会独立拥有一份训练好的模型；
1. 调用[`paddle_gradient_machine_create_shared_param`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/gradient_machine.h#L88)接口，与其它`gradient machine`的共享已经加载的预测模型。这种情况多出现在使用多线程预测时，通过多个线程共享同一个模型来减少内存开销。可参考[此示例](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/examples/model_inference/multi_thread/main.c)。
1. 调用[`paddle_gradient_machine_load_parameter_from_disk`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/capi/gradient_machine.h#L61)接口，从磁盘加载预测模型。这时`gradient machine`会独立拥有一份训练好的模型；
1. 调用[`paddle_gradient_machine_create_shared_param`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/capi/gradient_machine.h#L88)接口，与其它`gradient machine`的共享已经加载的预测模型。这种情况多出现在使用多线程预测时，通过多个线程共享同一个模型来减少内存开销。可参考[此示例](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/capi/examples/model_inference/multi_thread/main.c)。
- 注意事项
......@@ -117,7 +117,7 @@ C-API支持的所有输入数据类型和他们的组织方式，请参考“输
#### step 4. 前向计算
完成上述准备之后，通过调用 [`paddle_gradient_machine_forward`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/capi/gradient_machine.h#L73) 接口完成神经网络的前向计算。
完成上述准备之后，通过调用 [`paddle_gradient_machine_forward`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/legacy/capi/gradient_machine.h#L73) 接口完成神经网络的前向计算。
#### step 5. 清理
......
paddle/contrib/inference/CMakeLists.txt
浏览文件 @ f85e16f1
......@@ -46,9 +46,14 @@ cc_library(paddle_inference_api
SRCS paddle_inference_api.cc paddle_inference_api_impl.cc
DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
# Here the shared library doesn't depend on other fluid libraries, or double free will occur.
cc_library(paddle_inference_api_shared SHARED
SRCS paddle_inference_api.cc paddle_inference_api_impl.cc
DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
SRCS paddle_inference_api.cc paddle_inference_api_impl.cc)
set_target_properties(paddle_inference_api_shared PROPERTIES OUTPUT_NAME paddle_inference_api)
if(NOT APPLE)
set(LINK_FLAGS "-fPIC -fvisibility=hidden")
set_target_properties(paddle_inference_api_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
endif()
cc_test(test_paddle_inference_api
SRCS test_paddle_inference_api.cc
......
paddle/contrib/inference/paddle_inference_api.cc
浏览文件 @ f85e16f1
......@@ -23,7 +23,6 @@ int PaddleDtypeSize(PaddleDType dtype) {
case PaddleDType::INT64:
return sizeof(int64_t);
default:
//
assert(false);
return -1;
}
......
paddle/contrib/inference/paddle_inference_api_impl.h
浏览文件 @ f85e16f1
......@@ -22,9 +22,9 @@
#include "paddle/contrib/inference/paddle_inference_api.h"
#include "paddle/fluid/framework/ddim.h"
#include "paddle/fluid/framework/init.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/inference/io.h"
#include "paddle/fluid/platform/init.h"
#include "paddle/fluid/platform/profiler.h"
namespace paddle {
......
paddle/contrib/inference/test_paddle_inference_api_impl.cc
浏览文件 @ f85e16f1
......@@ -249,7 +249,7 @@ void MainThreadsImageClassification(bool use_gpu) {
const size_t len = local_outputs[0].data.length();
float* data = static_cast<float*>(local_outputs[0].data.data());
float* ref_data = refs[tid].data<float>();
EXPECT_EQ(refs[tid].numel(), len / sizeof(float));
EXPECT_EQ((size_t)refs[tid].numel(), len / sizeof(float));
for (int i = 0; i < refs[tid].numel(); ++i) {
EXPECT_NEAR(ref_data[i], data[i], 1e-3);
}
......
paddle/fluid/framework/CMakeLists.txt
浏览文件 @ f85e16f1
......@@ -21,12 +21,13 @@ endif()
cc_test(eigen_test SRCS eigen_test.cc DEPS tensor)
nv_test(mixed_vector_test SRCS mixed_vector_test.cu DEPS place memory device_context init)
nv_test(mixed_vector_test SRCS mixed_vector_test.cu DEPS place memory device_context tensor)
cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto recordio)
cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor memory)
nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor init)
nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor)
cc_library(reader SRCS reader.cc DEPS lod_tensor ddim)
cc_test(reader_test SRCS reader_test.cc DEPS reader)
cc_test(variable_test SRCS variable_test.cc)
......@@ -38,7 +39,7 @@ cc_test(scope_test SRCS scope_test.cc DEPS scope)
cc_library(data_device_transform SRCS data_device_transform.cc DEPS tensor)
nv_test(data_device_transform_test SRCS data_device_transform_test.cu
DEPS operator op_registry init math_function)
DEPS operator op_registry device_context math_function)
if(WITH_GPU)
nv_library(data_type_transform SRCS data_type_transform.cu DEPS tensor)
......@@ -63,7 +64,7 @@ cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context)
cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
shape_inference data_transform lod_tensor profiler)
cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry init)
cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry device_context)
cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog)
cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc)
......@@ -101,14 +102,14 @@ cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry
cc_library(selected_rows SRCS selected_rows.cc DEPS tensor)
cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
cc_library(init SRCS init.cc DEPS gflags device_context place stringpiece operator)
cc_test(init_test SRCS init_test.cc DEPS init)
cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto)
cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc)
# cc_test(channel_test SRCS channel_test.cc)
cc_test(tuple_test SRCS tuple_test.cc )
cc_test(concurrency_test SRCS concurrency_test.cc DEPS go_op channel_close_op channel_create_op
channel_send_op channel_recv_op sum_op select_op elementwise_add_op compare_op
conditional_block_op while_op assign_op print_op executor proto_desc)
# disable test temporarily.
# TODO https://github.com/PaddlePaddle/Paddle/issues/11971
# cc_test(concurrency_test SRCS concurrency_test.cc DEPS go_op channel_close_op channel_create_op
# channel_send_op channel_recv_op sum_op select_op elementwise_add_op compare_op
# conditional_block_op while_op assign_op print_op executor proto_desc)
paddle/fluid/framework/data_device_transform_test.cu
浏览文件 @ f85e16f1
......@@ -14,13 +14,13 @@ limitations under the License. */
#include "gtest/gtest.h"
#include "paddle/fluid/framework/init.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_info.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/elementwise_op_function.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/init.h"
namespace paddle {
namespace framework {
......
paddle/fluid/framework/details/build_strategy.h
浏览文件 @ f85e16f1
......@@ -34,7 +34,7 @@ struct BuildStrategy {
std::string debug_graphviz_path_{""};
bool enable_data_balance_{true};
bool enable_data_balance_{false};
};
} // namespace details
......
paddle/fluid/framework/details/data_balance_op_handle.cc
浏览文件 @ f85e16f1
......@@ -86,9 +86,9 @@ std::vector<std::array<int, 3>> DataBalanceOpHandle::GetBalancePlan(
}
void DataBalanceOpHandle::RunImpl() {
if (places_.size() == 1) {
return;
}
PADDLE_ENFORCE_GT(places_.size(), 1,
"Data balance can only be enabled when the number of "
"places to run larger than 1.");
auto in_var_handles = DynamicCast<VarHandle>(inputs_);
auto out_var_handles = DynamicCast<VarHandle>(outputs_);
PADDLE_ENFORCE(in_var_handles.size() % places_.size() == 0);
......
paddle/fluid/framework/details/multi_devices_graph_builder.cc
浏览文件 @ f85e16f1
......@@ -59,6 +59,11 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
grad_names_.insert(GradVarName(p));
}
balance_vars_.resize(places_.size(), 0);
if (strategy_.enable_data_balance_ && places_.size() == 1) {
LOG(WARNING) << "It is no need to enable data balance when there is only "
"one place. enable_data_balance is set to False.";
strategy_.enable_data_balance_ = false;
}
}
void MultiDevSSAGraphBuilder::CreateOpHandleIOs(SSAGraph *result,
......
paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc
浏览文件 @ f85e16f1
......@@ -13,6 +13,7 @@
// limitations under the License.
#include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h"
#include <stdexcept>
#include <string>
#include <vector>
#include "paddle/fluid/framework/executor.h"
......@@ -54,7 +55,13 @@ FeedFetchList ScopeBufferedSSAGraphExecutor::Run(
}
}
auto fetch_data = underlying_executor_->Run(fetch_tensors);
std::vector<framework::LoDTensor> fetch_data;
std::exception_ptr eptr;
try {
fetch_data = underlying_executor_->Run(fetch_tensors);
} catch (...) {
eptr = std::current_exception();
}
drop_scope_counter_ += 1;
if (!fetch_tensors.empty() ||
drop_scope_counter_ == strategy_.num_iteration_per_drop_scope_) {
......@@ -69,7 +76,12 @@ FeedFetchList ScopeBufferedSSAGraphExecutor::Run(
scope->DeleteScope(local_scope);
}
}
return fetch_data;
if (eptr) {
std::rethrow_exception(eptr);
} else {
return fetch_data;
}
}
} // namespace details
} // namespace framework
......
paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
浏览文件 @ f85e16f1
......@@ -78,6 +78,9 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
set.clear();
};
run_op_futures_.clear();
exception_.reset();
// Step 3. Execution
while (!pending_vars.empty()) {
// 1. Run All Ready ops
......@@ -98,6 +101,9 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
if (timeout) {
std::lock_guard<std::mutex> l(exception_mu_);
if (exception_) {
for (auto &run_op_future : run_op_futures_) {
run_op_future.wait();
}
std::exception *exp = exception_.get();
if (dynamic_cast<platform::EOFException *>(exp)) {
auto e = *static_cast<platform::EOFException *>(exp);
......@@ -222,7 +228,7 @@ void ThreadedSSAGraphExecutor::RunOp(
}
};
if (pool_) {
pool_->enqueue(op_run);
run_op_futures_.emplace_back(pool_->enqueue(op_run));
} else {
op_run();
}
......
paddle/fluid/framework/details/threaded_ssa_graph_executor.h
浏览文件 @ f85e16f1
......@@ -15,6 +15,7 @@
#pragma once
#include <deque>
#include <list>
#include <string>
#include <unordered_set>
#include <utility>
......@@ -77,6 +78,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
private:
ExecutionStrategy strategy_;
std::list<std::future<void>> run_op_futures_;
};
} // namespace details
......
paddle/fluid/framework/lod_tensor_test.cu
浏览文件 @ f85e16f1
......@@ -17,9 +17,9 @@
#include <stdio.h>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/init.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/platform/assert.h"
#include "paddle/fluid/platform/init.h"
#include "paddle/fluid/platform/place.h"
__global__ void test(size_t* a, int size) {
......
paddle/fluid/framework/op_info.cc
浏览文件 @ f85e16f1
......@@ -21,8 +21,8 @@ namespace framework {
// a static local variable is already being initialized.
// https://stackoverflow.com/questions/11711920/how-to-implement-multithread-safe-singleton-in-c11-without-using-mutex
OpInfoMap& OpInfoMap::Instance() {
static OpInfoMap* g_op_info_map = new OpInfoMap();
return *g_op_info_map;
static OpInfoMap g_op_info_map;
return g_op_info_map;
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/op_registry.h
浏览文件 @ f85e16f1
......@@ -182,21 +182,15 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
VarTypeInference
InferShapeBase
*/
#define REGISTER_OPERATOR(op_type, op_class, ...) \
STATIC_ASSERT_GLOBAL_NAMESPACE( \
__reg_op__##op_type, \
"REGISTER_OPERATOR must be called in global namespace"); \
class _OpClass_##op_type##_ : public op_class { \
public: \
DEFINE_OP_CLONE_METHOD(_OpClass_##op_type##_); \
DEFINE_OP_CONSTRUCTOR(_OpClass_##op_type##_, op_class); \
}; \
static ::paddle::framework::OperatorRegistrar<_OpClass_##op_type##_, \
##__VA_ARGS__> \
__op_registrar_##op_type##__(#op_type); \
int TouchOpRegistrar_##op_type() { \
__op_registrar_##op_type##__.Touch(); \
return 0; \
#define REGISTER_OPERATOR(op_type, op_class, ...) \
STATIC_ASSERT_GLOBAL_NAMESPACE( \
__reg_op__##op_type, \
"REGISTER_OPERATOR must be called in global namespace"); \
static ::paddle::framework::OperatorRegistrar<op_class, ##__VA_ARGS__> \
__op_registrar_##op_type##__(#op_type); \
int TouchOpRegistrar_##op_type() { \
__op_registrar_##op_type##__.Touch(); \
return 0; \
}
#define REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) \
......
paddle/fluid/framework/op_registry_test.cc
浏览文件 @ f85e16f1
......@@ -193,15 +193,10 @@ TEST(OpRegistry, CustomChecker) {
ASSERT_EQ(test_attr, 4);
}
class CosineOpComplete : public paddle::framework::CosineOp {
public:
DEFINE_OP_CONSTRUCTOR(CosineOpComplete, paddle::framework::CosineOp);
DEFINE_OP_CLONE_METHOD(CosineOpComplete);
};
TEST(OperatorRegistrar, Test) {
paddle::framework::OperatorRegistrar<
CosineOpComplete, paddle::framework::CosineOpProtoAndCheckerMaker>
paddle::framework::CosineOp,
paddle::framework::CosineOpProtoAndCheckerMaker>
reg("cos");
}
......
paddle/fluid/framework/operator.cc
浏览文件 @ f85e16f1
......@@ -633,6 +633,16 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
VLOG(3) << "expected_kernel_key:" << expected_kernel_key;
auto kernel_iter = kernels.find(expected_kernel_key);
#ifdef PADDLE_WITH_MKLDNN
// workaround for missing MKLDNN kernel when FLAGS_use_mkldnn env var is set
if (kernel_iter == kernels.end() &&
expected_kernel_key.library_type_ == LibraryType::kMKLDNN) {
VLOG(3) << "missing MKLDNN kernel: fallbacking to PLAIN one";
expected_kernel_key.library_type_ = LibraryType::kPlain;
expected_kernel_key.data_layout_ = DataLayout::kAnyLayout;
kernel_iter = kernels.find(expected_kernel_key);
}
#endif
if (kernel_iter == kernels.end()) {
PADDLE_THROW("op %s does not have kernel for %s", type_,
KernelTypeToString(expected_kernel_key));
......
paddle/fluid/framework/operator.h
浏览文件 @ f85e16f1
......@@ -121,10 +121,6 @@ class OperatorBase {
//! Get all outputs variable names
virtual std::vector<std::string> OutputVars(bool has_intermediate) const;
// Return a new operator instance, which is as same as this.
// Use unique_ptr to prevent caller forget to delete this pointer.
virtual std::unique_ptr<OperatorBase> Clone() const = 0;
protected:
std::string type_;
// NOTE: in case of OpGrad, inputs_ contains:
......@@ -145,37 +141,6 @@ class OperatorBase {
const platform::Place& place) const = 0;
};
// Macro for define a clone method.
// If you are writing an kernel operator, `Clone` will be defined when you
// register it. i.e. `Clone` method is not needed to define by yourself.
#define DEFINE_OP_CLONE_METHOD(cls) \
std::unique_ptr<::paddle::framework::OperatorBase> Clone() const final { \
return std::unique_ptr<::paddle::framework::OperatorBase>(new cls(*this)); \
}
// Macro for define a default constructor for Operator.
// You can also use
// using PARENT_CLASS::PARENT_CLASS;
// to use parent's constructor.
#define DEFINE_OP_CONSTRUCTOR(cls, parent_cls) \
cls(const std::string& type, \
const ::paddle::framework::VariableNameMap& inputs, \
const ::paddle::framework::VariableNameMap& outputs, \
const paddle::framework::AttributeMap& attrs) \
: parent_cls(type, inputs, outputs, attrs) {}
class NOP : public OperatorBase {
public:
using OperatorBase::OperatorBase;
std::unique_ptr<OperatorBase> Clone() const override {
return std::unique_ptr<OperatorBase>(new NOP(*this));
}
private:
void RunImpl(const Scope& scope,
const platform::Place& place) const override {}
};
class ExecutionContext {
public:
ExecutionContext(const OperatorBase& op, const Scope& scope,
......
paddle/fluid/framework/operator_test.cc
浏览文件 @ f85e16f1
......@@ -13,10 +13,10 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "gtest/gtest.h"
#include "paddle/fluid/framework/init.h"
#include "paddle/fluid/framework/op_info.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/platform/init.h"
namespace paddle {
namespace framework {
......@@ -247,26 +247,3 @@ TEST(OpKernel, multi_inputs) {
auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
op->Run(scope, cpu_place);
}
class OperatorClone : public paddle::framework::OperatorBase {
public:
DEFINE_OP_CLONE_METHOD(OperatorClone);
OperatorClone(const std::string& type,
const paddle::framework::VariableNameMap& inputs,
const paddle::framework::VariableNameMap& outputs,
const paddle::framework::AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs) {}
private:
void RunImpl(const paddle::framework::Scope& scope,
const paddle::platform::Place& place) const override {}
};
TEST(Operator, Clone) {
paddle::framework::InitDevices(true);
OperatorClone a("ABC", paddle::framework::VariableNameMap{},
paddle::framework::VariableNameMap{},
paddle::framework::AttributeMap{});
auto b = a.Clone();
ASSERT_EQ(a.Type(), b->Type());
}
paddle/fluid/framework/reader.cc
浏览文件 @ f85e16f1
......@@ -13,29 +13,61 @@
// limitations under the License.
#include "paddle/fluid/framework/reader.h"
#include <deque>
namespace paddle {
namespace framework {
ReaderBase::~ReaderBase() {}
FileReader::FileReader(const std::vector<DDim> &dims) : dims_(dims) {}
void FileReader::ReadNext(std::vector<LoDTensor> *out) {
void ReaderBase::ReadNext(std::vector<LoDTensor> *out) {
std::lock_guard<std::mutex> lock(mu_);
PADDLE_ENFORCE_EQ(status_, ReaderStatus::kRunning);
ReadNextImpl(out);
if (out->empty()) {
return;
}
}
PADDLE_ENFORCE_EQ(out->size(), dims_.size());
for (size_t i = 0; i < dims_.size(); ++i) {
auto &actual = (*out)[i].dims();
auto &expect = dims_[i];
void ReaderBase::InsertDecoratedReader(
const std::shared_ptr<ReaderBase> &decorated_reader) {
std::lock_guard<std::mutex> guard(mu_);
decorated_readers_.emplace_back(decorated_reader);
}
PADDLE_ENFORCE_EQ(actual.size(), expect.size());
for (int j = 0; j < actual.size(); ++j) {
// PADDLE_ENFORCE(actual[i] == expect[i] || expect[i] == -1);
std::unordered_set<ReaderBase *> ReaderBase::GetEndPoints() {
std::unordered_set<ReaderBase *> result;
std::deque<ReaderBase *> queue;
queue.emplace_back(this);
while (!queue.empty()) { // BFS search
auto *front = queue.front();
queue.pop_front();
if (front->decorated_readers_.empty()) {
result.emplace(front);
} else {
for (auto &reader : front->decorated_readers_) {
if (auto *reader_ptr = reader.lock().get()) {
queue.emplace_back(reader_ptr);
}
}
}
}
return result;
}
void ReaderBase::Shutdown() {
std::lock_guard<std::mutex> lock(mu_);
if (status_ != ReaderStatus::kStopped) {
ShutdownImpl();
status_ = ReaderStatus::kStopped;
}
}
void ReaderBase::Start() {
std::lock_guard<std::mutex> lock(mu_);
if (status_ != ReaderStatus::kRunning) {
StartImpl();
status_ = ReaderStatus::kRunning;
}
}
ReaderBase::~ReaderBase() { Shutdown(); }
} // namespace framework
} // namespace paddle
paddle/fluid/framework/reader.h
浏览文件 @ f85e16f1
......@@ -15,6 +15,7 @@
#pragma once
#include <memory>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/ddim.h"
......@@ -24,61 +25,116 @@
namespace paddle {
namespace framework {
enum ReaderStatus { kRunning, kStopped };
class ReaderBase {
public:
virtual void ReadNext(std::vector<LoDTensor>* out) = 0;
virtual void ReadNext(std::vector<LoDTensor>* out);
virtual void Shutdown();
virtual void ReInit() = 0;
virtual void Start();
// Return the readers which are the end of decorating chain. Basically
// they are readers just before read op.
std::unordered_set<ReaderBase*> GetEndPoints();
virtual ~ReaderBase();
protected:
virtual void ReadNextImpl(std::vector<LoDTensor>* out) {}
virtual void ShutdownImpl() {}
virtual void StartImpl() {}
ReaderStatus status_{kRunning};
mutable std::mutex mu_;
private:
friend class DecoratedReader;
// These methods can be only invoked inside DecoratedReader to record the
// decorating chain.
void InsertDecoratedReader(
const std::shared_ptr<ReaderBase>& decorated_reader);
// A set of which readers that decorated this reader.
std::vector<std::weak_ptr<ReaderBase>> decorated_readers_;
};
class DecoratedReader : public ReaderBase {
class DecoratedReader : public ReaderBase,
public std::enable_shared_from_this<DecoratedReader> {
public:
explicit DecoratedReader(const std::shared_ptr<ReaderBase>& reader)
: ReaderBase(), reader_(reader) {
PADDLE_ENFORCE_NOT_NULL(reader_);
}
void ReInit() override { reader_->ReInit(); }
void RegisterDecorateChain() {
reader_->InsertDecoratedReader(shared_from_this());
}
protected:
std::shared_ptr<ReaderBase> reader_;
};
class FileReader : public ReaderBase {
public:
explicit FileReader(const std::vector<DDim>& dims);
void ReadNext(std::vector<LoDTensor>* out) override;
void ShutdownImpl() override { reader_->Shutdown(); }
protected:
virtual void ReadNextImpl(std::vector<LoDTensor>* out) = 0;
void StartImpl() override { reader_->Start(); }
private:
std::vector<DDim> dims_;
std::shared_ptr<ReaderBase> reader_;
};
// FileReader is just a conceptual class.
class FileReader : public ReaderBase {};
// The ReaderHolder is used as reader' unified wrapper,
// making it easier to access different type reader in Variables.
class ReaderHolder {
public:
void Reset(ReaderBase* reader) { reader_.reset(reader); }
template <typename T>
void Reset(const std::shared_ptr<T>& reader) {
auto reader_base = std::dynamic_pointer_cast<ReaderBase>(reader);
PADDLE_ENFORCE_NOT_NULL(reader_base);
reader_ = reader_base;
}
std::shared_ptr<ReaderBase> Get() const { return reader_; }
const std::shared_ptr<ReaderBase>& Get() const { return reader_; }
void ReadNext(std::vector<LoDTensor>* out) {
PADDLE_ENFORCE_NOT_NULL(reader_);
reader_->ReadNext(out);
}
void ReInit() {
void ResetAll() {
auto end_readers = reader_->GetEndPoints();
for (auto* reader : end_readers) {
reader->Shutdown();
}
for (auto* reader : end_readers) {
reader->Start();
}
}
void Shutdown() {
PADDLE_ENFORCE_NOT_NULL(reader_);
reader_->ReInit();
reader_->Shutdown();
}
void Start() {
PADDLE_ENFORCE_NOT_NULL(reader_);
reader_->Start();
}
operator const std::shared_ptr<ReaderBase>&() const { return this->reader_; }
private:
std::shared_ptr<ReaderBase> reader_;
};
template <typename T, typename... ARGS>
inline std::shared_ptr<DecoratedReader> MakeDecoratedReader(ARGS&&... args) {
std::shared_ptr<DecoratedReader> reader(new T(std::forward<ARGS>(args)...));
reader->RegisterDecorateChain();
return reader;
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/reader_test.cc 0 → 100644
浏览文件 @ f85e16f1
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/reader.h"
#include <memory>
#include "gtest/gtest.h"
class StubDecoratedReader : public paddle::framework::DecoratedReader {
public:
explicit StubDecoratedReader(const std::shared_ptr<ReaderBase> &reader)
: DecoratedReader(reader) {}
void ReadNextImpl(std::vector<paddle::framework::LoDTensor> *out) override {}
};
class StubRootReader : public paddle::framework::ReaderBase {
public:
void ReadNextImpl(std::vector<paddle::framework::LoDTensor> *out) override {}
};
TEST(READER, decorate_chain) {
auto root = std::make_shared<StubRootReader>();
auto end_point1 =
paddle::framework::MakeDecoratedReader<StubDecoratedReader>(root);
auto end_point2 =
paddle::framework::MakeDecoratedReader<StubDecoratedReader>(root);
{
auto endpoints = root->GetEndPoints();
ASSERT_EQ(endpoints.size(), 2U);
ASSERT_NE(endpoints.count(end_point1.get()), 0);
ASSERT_NE(endpoints.count(end_point2.get()), 0);
}
{
auto end_point3 =
paddle::framework::MakeDecoratedReader<StubDecoratedReader>(root);
ASSERT_EQ(root->GetEndPoints().size(), 3U);
}
{ ASSERT_EQ(root->GetEndPoints().size(), 2U); }
}
paddle/fluid/framework/var_type_inference_test.cc
浏览文件 @ f85e16f1
......@@ -22,6 +22,17 @@ limitations under the License. */
namespace paddle {
namespace framework {
class NOP : public OperatorBase {
public:
NOP(const std::string &type, const VariableNameMap &inputs,
const VariableNameMap &outputs, const AttributeMap &attrs)
: OperatorBase(type, inputs, outputs, attrs) {}
private:
void RunImpl(const Scope &scope,
const platform::Place &place) const override {}
};
class SumOpMaker : public OpProtoAndCheckerMaker {
public:
void Make() {
......
paddle/fluid/inference/CMakeLists.txt
浏览文件 @ f85e16f1
set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor init)
set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor )
# TODO(panyx0718): Should this be called paddle_fluid_inference_api_internal?
cc_library(paddle_fluid_api
......
paddle/fluid/inference/analysis/README.md
浏览文件 @ f85e16f1
......@@ -54,4 +54,5 @@ It can be used as a helper class that draws the modified graph after each pass.
There is some helper legacy/function/class for analysis.
- [dot.h](./dot.h) give a easy to use interface for generating `DOT` codes,
- [graph_traits.h](./graph_traits.h) contains the graph traversal algorithms, it uses `iterator` to make the algorithms easy to share across different passes.
- [graph_traits.h](./graph_traits.h) contains the interfaces of the graph traversal algorithms, it uses `iterator`to make the algorithms easy to share across different passes,
there are some implementations in [data_flow_graph.cc](./data_flow_graph.cc) , such as BFS and DFS..
paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc
浏览文件 @ f85e16f1
......@@ -27,7 +27,7 @@ TEST_F(DFG_Tester, Init) {
DataFlowGraph graph;
pass.Run(&graph);
// Analysis is sensitive to ProgramDesc, careful to change the original model.
ASSERT_EQ(graph.nodes.size(), 37);
ASSERT_EQ(graph.nodes.size(), 37UL);
pass.Finalize();
LOG(INFO) << '\n' << graph.DotString();
}
......
paddle/fluid/inference/analysis/pass.h
浏览文件 @ f85e16f1
......@@ -32,19 +32,6 @@ class Pass {
public:
Pass() = default;
virtual ~Pass() = default;
// Virtual method overridden by subclasses to do only necessary initialization
// before any pass is run.
// virtual bool Initialize() { return false; }
// There is some passes such as FlowToDataFlowGraphPass that needs a
// ProgramDesc. Here use the native ProgramDesc ProtoBuf message, so that it
// only couple with the proto file.
// virtual bool Initialize(const framework::proto::ProgramDesc &desc) { return
// false; }
// There are some Passes such as DataFlowGraphToFluidPass that will output a
// ProgramDesc.
// virtual bool Initialize(framework::proto::ProgramDesc *desc) { return
// false; }
// Mutable Pass.
virtual bool Initialize(Argument *argument) { return false; }
// Readonly Pass.
......
paddle/fluid/inference/analysis/subgraph_splitter_tester.cc
浏览文件 @ f85e16f1
......@@ -82,7 +82,7 @@ TEST_F(DFG_Tester, Fuse) {
// At least one nodes should be deleted.
ASSERT_EQ(dfg.nodes.size(), count0 + 1); // added a new FunctionBlock
ASSERT_EQ(6UL, count1);
ASSERT_EQ(6, count1);
}
} // namespace analysis
......
paddle/fluid/inference/io.cc
浏览文件 @ f85e16f1
......@@ -20,7 +20,7 @@ limitations under the License. */
#include "paddle/fluid/framework/block_desc.h"
#include "paddle/fluid/framework/feed_fetch_type.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/platform/cpu_helper.h"
#include "paddle/fluid/pybind/pybind.h"
DEFINE_string(devices, "", "The devices to be used which is joined by comma.");
......@@ -33,7 +33,7 @@ namespace inference {
void Init(const std::vector<std::string> argv) {
framework::InitGflags(argv);
operators::math::SetNumThreads(FLAGS_math_num_threads);
platform::SetNumThreads(FLAGS_math_num_threads);
// init devices
std::vector<int> devices;
std::string token;
......
paddle/fluid/inference/io.h
浏览文件 @ f85e16f1
......@@ -18,9 +18,9 @@ limitations under the License. */
#include <string>
#include <vector>
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/init.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/platform/init.h"
namespace paddle {
namespace inference {
......
paddle/fluid/inference/tests/book/test_inference_nlp.cc
浏览文件 @ f85e16f1
......@@ -19,7 +19,7 @@ limitations under the License. */
#include "gflags/gflags.h"
#include "gtest/gtest.h"
#include "paddle/fluid/inference/tests/test_helper.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/platform/cpu_helper.h"
#ifdef PADDLE_WITH_MKLML
#include <omp.h>
#endif
......@@ -164,7 +164,7 @@ TEST(inference, nlp) {
// only use 1 thread number per std::thread
omp_set_dynamic(0);
omp_set_num_threads(1);
paddle::operators::math::SetNumThreads(1);
paddle::platform::SetNumThreads(1);
#endif
double start_ms = 0, stop_ms = 0;
......
paddle/fluid/memory/detail/buddy_allocator.cc
浏览文件 @ f85e16f1
......@@ -19,8 +19,9 @@ namespace paddle {
namespace memory {
namespace detail {
BuddyAllocator::BuddyAllocator(SystemAllocator* system_allocator,
size_t min_chunk_size, size_t max_chunk_size)
BuddyAllocator::BuddyAllocator(
std::unique_ptr<SystemAllocator> system_allocator, size_t min_chunk_size,
size_t max_chunk_size)
: min_chunk_size_(min_chunk_size),
max_chunk_size_(max_chunk_size),
cache_(system_allocator->UseGpu()),
......
paddle/fluid/memory/detail/buddy_allocator.h
浏览文件 @ f85e16f1
......@@ -14,6 +14,7 @@ limitations under the License. */
#pragma once
#include <memory>
#include <mutex> // NOLINT
#include <set>
#include <tuple>
......@@ -32,8 +33,8 @@ namespace detail {
class BuddyAllocator {
public:
BuddyAllocator(SystemAllocator* system_allocator, size_t min_chunk_size,
size_t max_chunk_size);
BuddyAllocator(std::unique_ptr<SystemAllocator> system_allocator,
size_t min_chunk_size, size_t max_chunk_size);
~BuddyAllocator();
......@@ -103,7 +104,7 @@ class BuddyAllocator {
private:
/*! Allocate CPU/GPU memory from system */
SystemAllocator* system_allocator_;
std::unique_ptr<SystemAllocator> system_allocator_;
std::mutex mutex_;
};
......
paddle/fluid/memory/malloc.cc
浏览文件 @ f85e16f1
......@@ -12,6 +12,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include <vector>
#include "paddle/fluid/memory/malloc.h"
#include "glog/logging.h"
......@@ -34,12 +36,15 @@ namespace memory {
using BuddyAllocator = detail::BuddyAllocator;
BuddyAllocator* GetCPUBuddyAllocator() {
static std::once_flag init_flag;
static detail::BuddyAllocator* a = nullptr;
if (a == nullptr) {
a = new detail::BuddyAllocator(new detail::CPUAllocator,
platform::CpuMinChunkSize(),
platform::CpuMaxChunkSize());
}
std::call_once(init_flag, []() {
a = new detail::BuddyAllocator(
std::unique_ptr<detail::SystemAllocator>(new detail::CPUAllocator),
platform::CpuMinChunkSize(), platform::CpuMaxChunkSize());
});
return a;
}
......@@ -68,27 +73,33 @@ size_t Used<platform::CPUPlace>(platform::CPUPlace place) {
#ifdef PADDLE_WITH_CUDA
BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
static BuddyAllocator** as = NULL;
if (as == NULL) {
static std::once_flag init_flag;
static detail::BuddyAllocator** a_arr = nullptr;
std::call_once(init_flag, [gpu_id]() {
int gpu_num = platform::GetCUDADeviceCount();
as = new BuddyAllocator*[gpu_num];
for (int gpu = 0; gpu < gpu_num; gpu++) {
as[gpu] = nullptr;
PADDLE_ENFORCE(gpu_id < gpu_num, "gpu_id:%d should < gpu_num:%d", gpu_id,
gpu_num);
a_arr = new BuddyAllocator*[gpu_num];
for (int i = 0; i < gpu_num; i++) {
a_arr[i] = nullptr;
platform::SetDeviceId(i);
a_arr[i] = new BuddyAllocator(
std::unique_ptr<detail::SystemAllocator>(new detail::GPUAllocator(i)),
platform::GpuMinChunkSize(), platform::GpuMaxChunkSize());
VLOG(10) << "\n\nNOTE: each GPU device use "
<< FLAGS_fraction_of_gpu_memory_to_use * 100
<< "% of GPU memory.\n"
<< "You can set GFlags environment variable '"
<< "FLAGS_fraction_of_gpu_memory_to_use"
<< "' to change the fraction of GPU usage.\n\n";
}
}
});
platform::SetDeviceId(gpu_id);
if (!as[gpu_id]) {
as[gpu_id] = new BuddyAllocator(new detail::GPUAllocator(gpu_id),
platform::GpuMinChunkSize(),
platform::GpuMaxChunkSize());
VLOG(10) << "\n\nNOTE: each GPU device use "
<< FLAGS_fraction_of_gpu_memory_to_use * 100
<< "% of GPU memory.\n"
<< "You can set GFlags environment variable '"
<< "FLAGS_fraction_of_gpu_memory_to_use"
<< "' to change the fraction of GPU usage.\n\n";
}
return as[gpu_id];
return a_arr[gpu_id];
}
template <>
......@@ -125,12 +136,16 @@ void Free<platform::CUDAPlace>(platform::CUDAPlace place, void* p) {
}
BuddyAllocator* GetCUDAPinnedBuddyAllocator() {
static BuddyAllocator* ba = NULL;
if (ba == NULL) {
ba = new BuddyAllocator(new detail::CUDAPinnedAllocator,
static std::once_flag init_flag;
static BuddyAllocator* ba = nullptr;
std::call_once(init_flag, []() {
ba = new BuddyAllocator(std::unique_ptr<detail::SystemAllocator>(
new detail::CUDAPinnedAllocator),
platform::CUDAPinnedMinChunkSize(),
platform::CUDAPinnedMaxChunkSize());
}
});
return ba;
}
......
paddle/fluid/operators/batch_norm_op.cc
浏览文件 @ f85e16f1
......@@ -216,6 +216,18 @@ class BatchNormKernel<platform::CPUDeviceContext, T>
saved_mean_e.setZero();
saved_variance_e.setZero();
EigenVectorArrayMap<T> running_mean_arr(
mean_out->mutable_data<T>(ctx.GetPlace()), C);
EigenVectorArrayMap<T> running_var_arr(
variance_out->mutable_data<T>(ctx.GetPlace()), C);
if ((N * sample_size) == 1) {
LOG(WARNING) << "Only 1 element in normalization dimension, "
<< "we skip the batch norm calculation, let y = x.";
framework::TensorCopySync(*x, ctx.GetPlace(), y);
return;
}
switch (data_layout) {
case DataLayout::kNCHW: {
ConstEigenArrayMap<T> x_arr(x->data<T>(), sample_size, N * C);
......@@ -247,10 +259,6 @@ class BatchNormKernel<platform::CPUDeviceContext, T>
PADDLE_THROW("Unknown storage order: %s", data_layout_str);
}
EigenVectorArrayMap<T> running_mean_arr(
mean_out->mutable_data<T>(ctx.GetPlace()), C);
EigenVectorArrayMap<T> running_var_arr(
variance_out->mutable_data<T>(ctx.GetPlace()), C);
running_mean_arr =
running_mean_arr * momentum + saved_mean_e * (1. - momentum);
running_var_arr =
......@@ -427,6 +435,11 @@ class BatchNormGradKernel<platform::CPUDeviceContext, T>
d_bias_arr.setZero();
d_scale_arr.setZero();
if ((N * sample_size) == 1) {
framework::TensorCopySync(*d_y, ctx.GetPlace(), d_x);
return;
}
const auto scale_inv_var_nhw = scale_arr * inv_var_arr / (N * sample_size);
switch (data_layout) {
......
paddle/fluid/operators/batch_norm_op.cu.cc
浏览文件 @ f85e16f1
......@@ -72,6 +72,9 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
int N, C, H, W, D;
ExtractNCWHD(x_dims, data_layout, &N, &C, &H, &W, &D);
auto *y = ctx.Output<Tensor>("Y");
y->mutable_data<T>(ctx.GetPlace());
// ------------------- cudnn descriptors ---------------------
cudnnTensorDescriptor_t data_desc_;
cudnnTensorDescriptor_t bn_param_desc_;
......@@ -93,7 +96,7 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
mode_ = CUDNN_BATCHNORM_SPATIAL;
#endif
VLOG(1) << "Setting descriptors.";
VLOG(3) << "Setting descriptors.";
std::vector<int> dims;
std::vector<int> strides;
if (data_layout == DataLayout::kNCHW) {
......@@ -113,11 +116,6 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
const auto *scale = ctx.Input<Tensor>("Scale");
const auto *bias = ctx.Input<Tensor>("Bias");
auto *y = ctx.Output<Tensor>("Y");
// alloc memory
y->mutable_data<T>(ctx.GetPlace());
auto &dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
auto handle = dev_ctx.cudnn_handle();
......@@ -162,22 +160,28 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
functor(dev_ctx, saved_mean, static_cast<BatchNormParamType<T>>(0));
functor(dev_ctx, saved_variance, static_cast<BatchNormParamType<T>>(0));
double this_factor = 1. - momentum;
CUDNN_ENFORCE(platform::dynload::cudnnBatchNormalizationForwardTraining(
handle, mode_, CudnnDataType<T>::kOne(), CudnnDataType<T>::kZero(),
data_desc_, x->template data<T>(), data_desc_,
y->template mutable_data<T>(ctx.GetPlace()), bn_param_desc_,
scale->template data<BatchNormParamType<T>>(),
bias->template data<BatchNormParamType<T>>(), this_factor,
mean_out->template mutable_data<BatchNormParamType<T>>(
ctx.GetPlace()),
variance_out->template mutable_data<BatchNormParamType<T>>(
ctx.GetPlace()),
epsilon, saved_mean->template mutable_data<BatchNormParamType<T>>(
ctx.GetPlace()),
saved_variance->template mutable_data<BatchNormParamType<T>>(
ctx.GetPlace())));
if ((N * H * W * D) == 1) {
LOG(WARNING) << "Only 1 element in normalization dimension, "
<< "we skip the batch norm calculation, let y = x.";
framework::TensorCopySync(*x, ctx.GetPlace(), y);
} else {
double this_factor = 1. - momentum;
CUDNN_ENFORCE(platform::dynload::cudnnBatchNormalizationForwardTraining(
handle, mode_, CudnnDataType<T>::kOne(), CudnnDataType<T>::kZero(),
data_desc_, x->template data<T>(), data_desc_,
y->template mutable_data<T>(ctx.GetPlace()), bn_param_desc_,
scale->template data<BatchNormParamType<T>>(),
bias->template data<BatchNormParamType<T>>(), this_factor,
mean_out->template mutable_data<BatchNormParamType<T>>(
ctx.GetPlace()),
variance_out->template mutable_data<BatchNormParamType<T>>(
ctx.GetPlace()),
epsilon, saved_mean->template mutable_data<BatchNormParamType<T>>(
ctx.GetPlace()),
saved_variance->template mutable_data<BatchNormParamType<T>>(
ctx.GetPlace())));
}
}
// clean when exit.
......@@ -209,6 +213,25 @@ class BatchNormGradKernel<platform::CUDADeviceContext, T>
int N, C, H, W, D;
ExtractNCWHD(x_dims, data_layout, &N, &C, &H, &W, &D);
// init output
auto *d_x = ctx.Output<Tensor>(framework::GradVarName("X"));
auto *d_scale = ctx.Output<Tensor>(framework::GradVarName("Scale"));
auto *d_bias = ctx.Output<Tensor>(framework::GradVarName("Bias"));
d_x->mutable_data<T>(ctx.GetPlace());
d_scale->mutable_data<T>(ctx.GetPlace());
d_bias->mutable_data<T>(ctx.GetPlace());
auto &dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
if ((N * H * W * D) == 1) {
framework::TensorCopySync(*d_y, ctx.GetPlace(), d_x);
math::SetConstant<platform::CUDADeviceContext, BatchNormParamType<T>>
functor;
functor(dev_ctx, d_scale, static_cast<BatchNormParamType<T>>(0));
functor(dev_ctx, d_bias, static_cast<BatchNormParamType<T>>(0));
return;
}
PADDLE_ENFORCE_EQ(scale->dims().size(), 1UL);
PADDLE_ENFORCE_EQ(scale->dims()[0], C);
......@@ -247,21 +270,11 @@ class BatchNormGradKernel<platform::CUDADeviceContext, T>
CUDNN_ENFORCE(platform::dynload::cudnnDeriveBNTensorDescriptor(
bn_param_desc_, data_desc_, mode_));
// init output
auto *d_x = ctx.Output<Tensor>(framework::GradVarName("X"));
auto *d_scale = ctx.Output<Tensor>(framework::GradVarName("Scale"));
auto *d_bias = ctx.Output<Tensor>(framework::GradVarName("Bias"));
d_x->mutable_data<T>(ctx.GetPlace());
d_scale->mutable_data<T>(ctx.GetPlace());
d_bias->mutable_data<T>(ctx.GetPlace());
const auto *saved_mean = ctx.Input<Tensor>("SavedMean");
const auto *saved_var = ctx.Input<Tensor>("SavedVariance");
const void *saved_mean_data = saved_mean->template data<T>();
const void *saved_var_data = saved_var->template data<T>();
auto &dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
CUDNN_ENFORCE(platform::dynload::cudnnBatchNormalizationBackward(
dev_ctx.cudnn_handle(), mode_, CudnnDataType<T>::kOne(),
CudnnDataType<T>::kZero(), CudnnDataType<T>::kOne(),
......
paddle/fluid/operators/conditional_block_op.cc
浏览文件 @ f85e16f1
......@@ -205,9 +205,10 @@ class ConditionalBlockGradInferShape : public framework::InferShapeBase {
context->SetOutputsDim(framework::GradVarName("Params"),
context->GetInputsDim("Params"));
}
PADDLE_ENFORCE(context->HasOutputs(framework::GradVarName("X")));
context->SetOutputsDim(framework::GradVarName("X"),
context->GetInputsDim("X"));
if (context->HasOutputs(framework::GradVarName("X"))) {
context->SetOutputsDim(framework::GradVarName("X"),
context->GetInputsDim("X"));
}
}
};
......
paddle/fluid/operators/cross_entropy_op.cc
浏览文件 @ f85e16f1
......@@ -124,8 +124,7 @@ class CrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker {
"Tensor<float/double> with shape [N x D].");
AddOutput("Y",
"(Tensor, default Tensor<float>), a 2-D tensor with shape "
"[N x 1]. The cross entropy loss.")
.Reuse("X");
"[N x 1]. The cross entropy loss.");
AddAttr<bool>("soft_label",
"(bool, default false), a flag indicating whether to "
"interpretate the given labels as soft labels.")
......
paddle/fluid/operators/distributed/CMakeLists.txt
浏览文件 @ f85e16f1
......@@ -5,7 +5,7 @@ if(WITH_GRPC)
set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
set_source_files_properties(grpc_serde_test.cc rpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
cc_test(serde_test SRCS grpc_serde_test.cc variable_response.cc DEPS grpc++_unsecure grpc_unsecure gpr
cares zlib protobuf sendrecvop_grpc SERIAL)
cares zlib protobuf sendrecvop_grpc scope profiler math_function SERIAL)
cc_test(grpc_server_test SRCS rpc_server_test.cc DEPS sendrecvop_grpc
grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor
proto_desc lookup_table_op SERIAL)
......
paddle/fluid/operators/elementwise_add_mkldnn_op.cc
浏览文件 @ f85e16f1
......@@ -85,7 +85,7 @@ class EltwiseAddMKLDNNKernel : public framework::OpKernel<T> {
"Wrong layout/format set for X tensor");
PADDLE_ENFORCE(y->layout() == DataLayout::kMKLDNN &&
y->format() != memory::format::format_undef,
"Wrong layout/format set for X tensor");
"Wrong layout/format set for Y tensor");
std::vector<int> src_x_tz = framework::vectorize2int(x_dims);
std::vector<int> src_y_tz = framework::vectorize2int(y_dims);
......
paddle/fluid/operators/math/CMakeLists.txt
浏览文件 @ f85e16f1
......@@ -54,13 +54,13 @@ math_library(softmax DEPS math_function)
math_library(unpooling)
math_library(vol2col)
cc_test(math_function_test SRCS math_function_test.cc)
cc_test(math_function_test SRCS math_function_test.cc DEPS math_function)
cc_test(selected_rows_functor_test SRCS selected_rows_functor_test.cc DEPS selected_rows_functor)
cc_test(im2col_test SRCS im2col_test.cc DEPS im2col)
cc_test(vol2col_test SRCS vol2col_test.cc DEPS vol2col)
cc_test(sequence_padding_test SRCS sequence_padding_test.cc DEPS sequence_padding)
if(WITH_GPU)
nv_test(math_function_gpu_test SRCS math_function_test.cu)
nv_test(selected_rows_functor_gpu_test SRCS selected_rows_functor_test.cu DEPS selected_rows_functor)
nv_test(math_function_gpu_test SRCS math_function_test.cu DEPS math_function)
nv_test(selected_rows_functor_gpu_test SRCS selected_rows_functor_test.cu DEPS selected_rows_functor math_function)
endif()
cc_test(concat_test SRCS concat_test.cc DEPS concat)
paddle/fluid/operators/math/blas.h
浏览文件 @ f85e16f1
......@@ -23,41 +23,12 @@
#ifdef PADDLE_USE_OPENBLAS
#include <cblas.h>
#ifdef LAPACK_FOUND
#include <lapacke.h>
#endif
#endif
#ifndef LAPACK_FOUND
extern "C" {
#include <cblas.h> // NOLINT
int LAPACKE_sgetrf(int matrix_layout, int m, int n, float* a, int lda,
int* ipiv);
int LAPACKE_dgetrf(int matrix_layout, int m, int n, double* a, int lda,
int* ipiv);
int LAPACKE_sgetri(int matrix_layout, int n, float* a, int lda,
const int* ipiv);
int LAPACKE_dgetri(int matrix_layout, int n, double* a, int lda,
const int* ipiv);
}
#endif
namespace paddle {
namespace operators {
namespace math {
static void SetNumThreads(int num_threads) {
#ifdef PADDLE_USE_OPENBLAS
int real_num_threads = num_threads > 1 ? num_threads : 1;
openblas_set_num_threads(real_num_threads);
#elif defined(PADDLE_WITH_MKLML)
int real_num_threads = num_threads > 1 ? num_threads : 1;
platform::dynload::MKL_Set_Num_Threads(real_num_threads);
#else
PADDLE_ENFORCE(false, "To be implemented.");
#endif
}
/**
* Matrix Descriptor of a memory buffer.
*
......
paddle/fluid/operators/math/math_function.h
浏览文件 @ f85e16f1
......@@ -19,23 +19,6 @@ limitations under the License. */
#ifdef PADDLE_USE_OPENBLAS
#include <cblas.h>
#ifdef LAPACK_FOUND
#include <lapacke.h>
#endif
#endif
#ifndef LAPACK_FOUND
extern "C" {
#include <cblas.h> // NOLINT
int LAPACKE_sgetrf(int matrix_layout, int m, int n, float* a, int lda,
int* ipiv);
int LAPACKE_dgetrf(int matrix_layout, int m, int n, double* a, int lda,
int* ipiv);
int LAPACKE_sgetri(int matrix_layout, int n, float* a, int lda,
const int* ipiv);
int LAPACKE_dgetri(int matrix_layout, int n, double* a, int lda,
const int* ipiv);
}
#endif
#include <cmath>
......
paddle/fluid/operators/merge_lod_tensor_op.cc
浏览文件 @ f85e16f1
......@@ -44,8 +44,10 @@ class MergeLoDTensorOp : public framework::OperatorBase {
scope.FindVar(Output("Out"))->GetMutable<framework::LoDTensor>();
auto level = static_cast<size_t>(Attr<int>("level"));
auto &mask_dim = mask.dims();
PADDLE_ENFORCE(in_true.numel() || in_false.numel(),
"Input(InTrue) or Input(InFalse) should be initialized.");
auto &mask_dim = mask.dims();
std::unique_ptr<framework::LoDTensor> cpu_mask{new framework::LoDTensor()};
if (platform::is_cpu_place(mask.place())) {
cpu_mask->ShareDataWith(mask);
......@@ -59,19 +61,27 @@ class MergeLoDTensorOp : public framework::OperatorBase {
}
auto *mask_data = cpu_mask->data<bool>();
int rank = in_true.dims().size();
platform::Place place = in_true.place();
std::type_index data_type = in_true.type();
framework::DDim in_true_dims =
framework::slice_ddim(in_true.dims(), 1, rank);
platform::Place place = dev_place;
int64_t batch_size = in_true.dims()[0] + in_false.dims()[0];
auto in_true_dim_vec = framework::vectorize(in_true_dims);
in_true_dim_vec.insert(in_true_dim_vec.begin(), batch_size);
std::type_index data_type =
in_true.IsInitialized() ? in_true.type() : in_false.type();
int rank;
framework::DDim in_dims;
if (in_true.IsInitialized()) {
rank = in_true.dims().size();
in_dims = framework::slice_ddim(in_true.dims(), 1, rank);
} else {
rank = in_false.dims().size();
in_dims = framework::slice_ddim(in_false.dims(), 1, rank);
}
auto in_dim_vec = framework::vectorize(in_dims);
in_dim_vec.insert(in_dim_vec.begin(), batch_size);
framework::DDim out_dims = framework::make_ddim(in_true_dim_vec);
framework::DDim out_dims = framework::make_ddim(in_dim_vec);
out->Resize(out_dims);
out->mutable_data(place, data_type);
auto *out_lod = out->mutable_lod();
......
paddle/fluid/operators/nccl_op_test.cu.cc
浏览文件 @ f85e16f1
......@@ -19,7 +19,6 @@ limitations under the License. */
#include <thread> // NOLINT
#include <vector>
#include "paddle/fluid/framework/init.h"
#include "paddle/fluid/framework/op_desc.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/program_desc.h"
......@@ -27,6 +26,7 @@ limitations under the License. */
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/platform/gpu_info.h"
#include "paddle/fluid/platform/init.h"
#include "paddle/fluid/platform/place.h"
USE_NO_KERNEL_OP(ncclInit);
......
paddle/fluid/operators/read_op.cc
浏览文件 @ f85e16f1
......@@ -92,9 +92,13 @@ class ReadOpMaker : public framework::OpProtoAndCheckerMaker {
void Make() override {
AddInput("Reader", "(ReaderHolder) The executed reader.");
AddOutput("Out", "(LoDTensor) The output data.").AsDuplicable();
AddAttr<bool>("throw_eof_exp",
"If set true, an exception will be thrown when the Reader "
"yields empty (which means there is no next data).")
AddAttr<bool>(
"throw_eof_exp",
"If set true, an exception will be thrown when the Reader "
"yields empty (which means there is no next data).\n"
"NOTES: This flag must be true always. It will be set to false"
" only when the data-balance is enabled in ParallelExecutor"
" and it is set by ParallelExecutor instance, not users.")
.SetDefault(true);
AddComment(R"DOC(
Read Operator
......
paddle/fluid/operators/reader/CMakeLists.txt
浏览文件 @ f85e16f1
......@@ -22,7 +22,6 @@ reader_library(create_batch_reader_op SRCS create_batch_reader_op.cc)
reader_library(create_recordio_file_reader_op SRCS create_recordio_file_reader_op.cc)
reader_library(create_double_buffer_reader_op SRCS create_double_buffer_reader_op.cc)
reader_library(create_multi_pass_reader_op SRCS create_multi_pass_reader_op.cc)
reader_library(create_threaded_reader_op SRCS create_threaded_reader_op.cc)
reader_library(create_custom_reader_op SRCS create_custom_reader_op.cc)
reader_library(create_py_reader_op SRCS create_py_reader_op.cc)
......
paddle/fluid/operators/reader/create_batch_reader_op.cc
浏览文件 @ f85e16f1
......@@ -20,15 +20,19 @@ namespace reader {
class BatchReader : public framework::DecoratedReader {
public:
BatchReader(const std::shared_ptr<ReaderBase>& reader, int batch_size)
: DecoratedReader(reader), batch_size_(batch_size) {
BatchReader(const std::shared_ptr<ReaderBase>& reader, int batch_size,
bool discard_leftover)
: DecoratedReader(reader),
batch_size_(batch_size),
discard_leftover_(discard_leftover) {
buffer_.reserve(batch_size_);
}
void ReadNext(std::vector<framework::LoDTensor>* out) override;
void ReadNextImpl(std::vector<framework::LoDTensor>* out) override;
private:
int batch_size_;
bool discard_leftover_;
std::vector<std::vector<framework::LoDTensor>> buffer_;
};
......@@ -46,8 +50,9 @@ class CreateBatchReaderOp : public framework::OperatorBase {
}
const auto& underlying_reader = scope.FindVar(Input("UnderlyingReader"))
->Get<framework::ReaderHolder>();
out->Reset(
new BatchReader(underlying_reader.Get(), Attr<int>("batch_size")));
out->Reset(framework::MakeDecoratedReader<BatchReader>(
underlying_reader, Attr<int>("batch_size"),
Attr<bool>("discard_leftover")));
}
};
......@@ -57,6 +62,10 @@ class CreateBatchReaderOpMaker : public DecoratedReaderMakerBase {
AddAttr<int>("batch_size",
"How many instances the batch reader yields each time.")
.GreaterThan(0);
AddAttr<bool>("discard_leftover",
"If true, the leftover instances that are not enough for a "
"new batch will be discarded.")
.SetDefault(true);
AddComment(R"DOC(
CreateBatchReader Operator
......@@ -66,7 +75,7 @@ class CreateBatchReaderOpMaker : public DecoratedReaderMakerBase {
}
};
void BatchReader::ReadNext(std::vector<framework::LoDTensor>* out) {
void BatchReader::ReadNextImpl(std::vector<framework::LoDTensor>* out) {
buffer_.clear();
buffer_.reserve(batch_size_);
for (int i = 0; i < batch_size_; ++i) {
......@@ -77,6 +86,9 @@ void BatchReader::ReadNext(std::vector<framework::LoDTensor>* out) {
break;
}
}
if (discard_leftover_ && buffer_.size() < batch_size_) {
buffer_.clear();
}
// Concat instances
out->clear();
if (buffer_.empty()) {
......
paddle/fluid/operators/reader/create_custom_reader_op.cc
浏览文件 @ f85e16f1
......@@ -33,7 +33,7 @@ class CustomReader : public framework::DecoratedReader {
source_var_names_(source_var_names),
sink_var_names_(sink_var_names) {}
void ReadNext(std::vector<framework::LoDTensor>* out) override;
void ReadNextImpl(std::vector<framework::LoDTensor>* out) override;
private:
const framework::ProgramDesc program_;
......@@ -60,10 +60,10 @@ class CreateCustomReaderOp : public framework::OperatorBase {
}
const auto& underlying_reader = scope.FindVar(Input("UnderlyingReader"))
->Get<framework::ReaderHolder>();
out->Reset(
new CustomReader(underlying_reader.Get(), *sub_block,
Attr<std::vector<std::string>>("source_var_names"),
Attr<std::vector<std::string>>("sink_var_names")));
out->Reset(framework::MakeDecoratedReader<CustomReader>(
underlying_reader, *sub_block,
Attr<std::vector<std::string>>("source_var_names"),
Attr<std::vector<std::string>>("sink_var_names")));
}
};
......@@ -143,7 +143,7 @@ class CustomReaderInferVarType : public framework::VarTypeInference {
}
};
void CustomReader::ReadNext(std::vector<framework::LoDTensor>* out) {
void CustomReader::ReadNextImpl(std::vector<framework::LoDTensor>* out) {
out->clear();
std::vector<framework::LoDTensor> underlying_outs;
reader_->ReadNext(&underlying_outs);
......
paddle/fluid/operators/reader/create_double_buffer_reader_op.cc
浏览文件 @ f85e16f1
......@@ -23,13 +23,13 @@ namespace reader {
// 'Double buffer' means we shall maintain two batches of input data at the same
// time. So the kCacheSize shoul be at least 2.
static constexpr size_t kCacheSize = 5;
static constexpr size_t kCacheSize = 3;
// There will be two bacthes out of the channel during training:
// 1. the one waiting to be sent to the channel
// 2. the one just be received from the channel, which is also being used by
// subsequent operators.
// So the channel size should be kChacheSize - 2
static constexpr size_t kChannelSize = 3; // kCacheSize - 2
static constexpr size_t kChannelSize = 1; // kCacheSize - 2
class DoubleBufferReader : public framework::DecoratedReader {
public:
......@@ -50,12 +50,21 @@ class DoubleBufferReader : public framework::DecoratedReader {
StartPrefetcher();
}
void ReadNext(std::vector<framework::LoDTensor>* out) override;
void ReInit() override;
void ReadNextImpl(std::vector<framework::LoDTensor>* out) override;
~DoubleBufferReader() { EndPrefetcher(); }
private:
void ShutdownImpl() override {
EndPrefetcher();
reader_->Shutdown();
}
void StartImpl() override {
reader_->Start();
StartPrefetcher();
}
void StartPrefetcher() {
channel_ = new reader::BlockingQueue<size_t>(kChannelSize);
prefetcher_ = std::thread([this] { PrefetchThreadFunc(); });
......@@ -109,7 +118,8 @@ class CreateDoubleBufferReaderOp : public framework::OperatorBase {
place = platform::CUDAPlace(static_cast<int>(num));
}
out->Reset(new DoubleBufferReader(underlying_reader.Get(), place));
out->Reset(framework::MakeDecoratedReader<DoubleBufferReader>(
underlying_reader, place));
}
};
......@@ -136,7 +146,7 @@ class CreateDoubleBufferReaderOpMaker : public DecoratedReaderMakerBase {
}
};
void DoubleBufferReader::ReadNext(std::vector<framework::LoDTensor>* out) {
void DoubleBufferReader::ReadNextImpl(std::vector<framework::LoDTensor>* out) {
size_t cached_tensor_id;
if (channel_->Receive(&cached_tensor_id)) {
if (platform::is_gpu_place(place_)) {
......@@ -150,12 +160,6 @@ void DoubleBufferReader::ReadNext(std::vector<framework::LoDTensor>* out) {
}
}
void DoubleBufferReader::ReInit() {
reader_->ReInit();
EndPrefetcher();
StartPrefetcher();
}
void DoubleBufferReader::PrefetchThreadFunc() {
VLOG(5) << "A new prefetch thread starts.";
size_t cached_tensor_id = 0;
......
paddle/fluid/operators/reader/create_multi_pass_reader_op.cc
浏览文件 @ f85e16f1
......@@ -24,23 +24,22 @@ class MultiPassReader : public framework::DecoratedReader {
MultiPassReader(const std::shared_ptr<ReaderBase>& reader, int pass_num)
: DecoratedReader(reader), pass_num_(pass_num), pass_count_(0) {}
void ReadNext(std::vector<framework::LoDTensor>* out) override {
void ReadNextImpl(std::vector<framework::LoDTensor>* out) override {
reader_->ReadNext(out);
if (out->empty()) {
if (out->empty() && pass_count_ < pass_num_ - 1) {
reader_->Shutdown();
reader_->Start();
reader_->ReadNext(out);
++pass_count_;
if (pass_count_ < pass_num_) {
reader_->ReInit();
reader_->ReadNext(out);
}
}
}
void ReInit() override {
private:
void StartImpl() override {
pass_count_ = 0;
reader_->ReInit();
reader_->Start();
}
private:
int pass_num_;
mutable int pass_count_;
};
......@@ -60,7 +59,8 @@ class CreateMultiPassReaderOp : public framework::OperatorBase {
const auto& underlying_reader = scope.FindVar(Input("UnderlyingReader"))
->Get<framework::ReaderHolder>();
int pass_num = Attr<int>("pass_num");
out->Reset(new MultiPassReader(underlying_reader.Get(), pass_num));
out->Reset(framework::MakeDecoratedReader<MultiPassReader>(
underlying_reader, pass_num));
}
};
......
paddle/fluid/operators/reader/create_py_reader_op.cc
浏览文件 @ f85e16f1
......@@ -19,9 +19,10 @@ namespace paddle {
namespace operators {
namespace reader {
class PyReader : public framework::ReaderBase {
class PyReader : public framework::FileReader {
public:
explicit PyReader(const std::shared_ptr<LoDTensorBlockingQueue>& queue) {
explicit PyReader(const std::shared_ptr<LoDTensorBlockingQueue>& queue)
: framework::FileReader() {
PADDLE_ENFORCE(queue != nullptr, "LoDTensorBlockingQueue must not be null");
queue_ = queue;
}
......@@ -32,7 +33,9 @@ class PyReader : public framework::ReaderBase {
if (!success) out->clear();
}
void ReInit() override { queue_->ReOpen(); }
void Shutdown() override { queue_->Close(); }
void Start() override { queue_->ReOpen(); }
private:
std::shared_ptr<LoDTensorBlockingQueue> queue_;
......@@ -51,14 +54,14 @@ class CreatePyReaderOp : public framework::OperatorBase {
const std::string& queue_name = Input("blocking_queue");
auto* queue_holder_var = scope.FindVar(queue_name);
PADDLE_ENFORCE(
queue_holder_var != nullptr,
PADDLE_ENFORCE_NOT_NULL(
queue_holder_var,
"No LoDTensorBlockingQueueHolder variable with name %s found",
queue_name);
auto* queue_holder =
queue_holder_var->template GetMutable<LoDTensorBlockingQueueHolder>();
out->Reset(new PyReader(queue_holder->GetQueue()));
out->Reset(std::make_shared<PyReader>(queue_holder->GetQueue()));
}
};
......
paddle/fluid/operators/reader/create_random_data_generator_op.cc
浏览文件 @ f85e16f1
......@@ -19,11 +19,11 @@ namespace operators {
namespace reader {
template <typename T>
class RandomDataGenerator : public framework::ReaderBase {
class RandomDataGenerator : public framework::FileReader {
public:
RandomDataGenerator(const std::vector<framework::DDim>& shapes, float low,
float high)
: framework::ReaderBase(), low_(low), high_(high), shapes_(shapes) {
: framework::FileReader(), low_(low), high_(high), shapes_(shapes) {
PADDLE_ENFORCE_LE(low, high,
"'low' shouldn't be greater than 'high'.(%f vs %f)", low,
high);
......@@ -32,7 +32,7 @@ class RandomDataGenerator : public framework::ReaderBase {
dist_ = std::uniform_real_distribution<float>(low_, high_);
}
void ReadNext(std::vector<framework::LoDTensor>* out) override {
void ReadNextImpl(std::vector<framework::LoDTensor>* out) override {
out->clear();
out->reserve(shapes_.size());
for (const framework::DDim& shape : shapes_) {
......@@ -51,8 +51,6 @@ class RandomDataGenerator : public framework::ReaderBase {
}
}
void ReInit() override { return; }
private:
float low_;
float high_;
......@@ -79,8 +77,8 @@ class CreateRandomDataGeneratorOp : public framework::OperatorBase {
std::vector<framework::DDim> shapes = RestoreShapes(shape_concat, ranks);
auto* out = scope.FindVar(Output("Out"))
->template GetMutable<framework::ReaderHolder>();
out->Reset(new RandomDataGenerator<T>(shapes, Attr<float>("low"),
Attr<float>("high")));
out->Reset(std::make_shared<RandomDataGenerator<T>>(
shapes, Attr<float>("low"), Attr<float>("high")));
}
};
......
paddle/fluid/operators/reader/create_recordio_file_reader_op.cc
浏览文件 @ f85e16f1
......@@ -21,10 +21,8 @@ namespace reader {
template <bool ThreadSafe>
class RecordIOFileReader : public framework::FileReader {
public:
explicit RecordIOFileReader(const std::string& filename,
const std::vector<framework::DDim>& dims)
: FileReader(dims),
scanner_(filename),
explicit RecordIOFileReader(const std::string& filename)
: scanner_(filename),
dev_ctx_(*platform::DeviceContextPool::Instance().Get(
platform::CPUPlace())) {
if (ThreadSafe) {
......@@ -33,8 +31,6 @@ class RecordIOFileReader : public framework::FileReader {
LOG(INFO) << "Creating file reader" << filename;
}
void ReInit() override { scanner_.Reset(); }
protected:
void ReadNextImpl(std::vector<framework::LoDTensor>* out) override {
if (ThreadSafe) {
......@@ -45,6 +41,8 @@ class RecordIOFileReader : public framework::FileReader {
}
}
void StartImpl() override { scanner_.Reset(); }
private:
std::unique_ptr<std::mutex> mutex_;
recordio::Scanner scanner_;
......@@ -58,20 +56,11 @@ class CreateRecordIOReaderOp : public framework::OperatorBase {
private:
void RunImpl(const framework::Scope& scope,
const platform::Place& dev_place) const override {
const auto& shape_concat = Attr<std::vector<int>>("shape_concat");
const auto& ranks = Attr<std::vector<int>>("ranks");
PADDLE_ENFORCE(!shape_concat.empty() && !ranks.empty());
PADDLE_ENFORCE_EQ(std::accumulate(ranks.begin(), ranks.end(), 0),
static_cast<int>(shape_concat.size()),
"The accumulate of all ranks should be equal to the "
"shape concat's length.");
std::string filename = Attr<std::string>("filename");
auto* out = scope.FindVar(Output("Out"))
->template GetMutable<framework::ReaderHolder>();
out->Reset(new RecordIOFileReader<true>(
filename, RestoreShapes(shape_concat, ranks)));
out->Reset(std::make_shared<RecordIOFileReader<true>>(filename));
}
};
......
paddle/fluid/operators/reader/create_shuffle_reader_op.cc
浏览文件 @ f85e16f1
......@@ -34,7 +34,7 @@ class ShuffleReader : public framework::DecoratedReader {
ReloadBuffer();
}
void ReadNext(std::vector<framework::LoDTensor>* out) override {
void ReadNextImpl(std::vector<framework::LoDTensor>* out) override {
out->clear();
if (iteration_pos_ >= buffer_.size()) {
VLOG(10) << "Resetting shuffle buffer";
......@@ -47,6 +47,17 @@ class ShuffleReader : public framework::DecoratedReader {
}
private:
void ShutdownImpl() override {
buffer_.clear();
iteration_pos_ = 0;
reader_->Shutdown();
}
void StartImpl() override {
reader_->Start();
ReloadBuffer();
}
void ReloadBuffer() {
buffer_.clear();
buffer_.reserve(buffer_size_);
......@@ -86,9 +97,8 @@ class CreateShuffleReaderOp : public framework::OperatorBase {
}
const auto& underlying_reader = scope.FindVar(Input("UnderlyingReader"))
->Get<framework::ReaderHolder>();
out->Reset(
new ShuffleReader(underlying_reader.Get(),
static_cast<size_t>(Attr<int>("buffer_size"))));
out->Reset(framework::MakeDecoratedReader<ShuffleReader>(
underlying_reader, static_cast<size_t>(Attr<int>("buffer_size"))));
}
};
......
paddle/fluid/operators/reader/create_threaded_reader_op.cc 已删除 100644 → 0
浏览文件 @ c53c5d45
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/operators/detail/safe_ref.h"
#include "paddle/fluid/operators/reader/reader_op_registry.h"
namespace paddle {
namespace operators {
namespace reader {
class ThreadedReader : public framework::DecoratedReader {
public:
explicit ThreadedReader(const std::shared_ptr<ReaderBase>& reader)
: DecoratedReader(reader) {}
void ReadNext(std::vector<framework::LoDTensor>* out) override {
std::lock_guard<std::mutex> lock(mutex_);
reader_->ReadNext(out);
}
void ReInit() override { reader_->ReInit(); }
private:
std::mutex mutex_;
};
class CreateThreadedReaderOp : public framework::OperatorBase {
public:
using framework::OperatorBase::OperatorBase;
private:
void RunImpl(const framework::Scope& scope,
const platform::Place& dev_place) const override {
auto* out = detail::Ref(scope.FindVar(Output("Out")))
.GetMutable<framework::ReaderHolder>();
if (out->Get() != nullptr) {
return;
}
const auto& underlying_reader = scope.FindVar(Input("UnderlyingReader"))
->Get<framework::ReaderHolder>();
out->Reset(new ThreadedReader(underlying_reader.Get()));
}
};
class CreateThreadedReaderOpMaker : public DecoratedReaderMakerBase {
protected:
void Apply() override {
AddComment(R"DOC(
CreateThreadedReader Operator
This operator creates a threaded reader. A threaded reader's
'ReadNext()' can be invoked by several threads at the same
time.
When the attribute 'safe_mode' is true, the threaded reader's
'ReInit()' is disabled to avoid unexpected bugs in multi-thread
environment.
)DOC");
}
};
} // namespace reader
} // namespace operators
} // namespace paddle
namespace reader = paddle::operators::reader;
REGISTER_DECORATED_READER_OPERATOR(create_threaded_reader,
reader::CreateThreadedReaderOp,
reader::CreateThreadedReaderOpMaker);
paddle/fluid/operators/reader/open_files_op.cc
浏览文件 @ f85e16f1
......@@ -23,24 +23,26 @@ namespace reader {
class MultiFileReader : public framework::ReaderBase {
public:
MultiFileReader(const std::vector<std::string>& file_names,
const std::vector<framework::DDim>& dims, size_t thread_num,
MultiFileReader(const std::vector<std::string>& file_names, size_t thread_num,
size_t buffer_size)
: buffer_size_(buffer_size) {
readers_.reserve(file_names.size());
for (const std::string& f_name : file_names) {
readers_.emplace_back(CreateReaderByFileName(f_name, dims));
readers_.emplace_back(CreateReaderByFileName(f_name));
}
prefetchers_.resize(thread_num);
StartNewScheduler();
}
void ReadNext(std::vector<framework::LoDTensor>* out) override;
void ReInit() override;
void ReadNextImpl(std::vector<framework::LoDTensor>* out) override;
~MultiFileReader() { EndScheduler(); }
private:
void ShutdownImpl() override { EndScheduler(); }
void StartImpl() override { StartNewScheduler(); }
void StartNewScheduler();
void EndScheduler();
void ScheduleThreadFunc();
......@@ -55,17 +57,12 @@ class MultiFileReader : public framework::ReaderBase {
reader::BlockingQueue<std::vector<framework::LoDTensor>>* buffer_;
};
void MultiFileReader::ReadNext(std::vector<framework::LoDTensor>* out) {
void MultiFileReader::ReadNextImpl(std::vector<framework::LoDTensor>* out) {
if (!buffer_->Receive(out)) {
out->clear();
}
}
void MultiFileReader::ReInit() {
EndScheduler();
StartNewScheduler();
}
void MultiFileReader::StartNewScheduler() {
size_t thread_num = prefetchers_.size();
waiting_reader_idx_ = new reader::BlockingQueue<size_t>(readers_.size());
......@@ -120,7 +117,7 @@ void MultiFileReader::ScheduleThreadFunc() {
}
}
}
// If users invoke ReInit() when scheduler is running, it will close the
// If users invoke Shutdown() when scheduler is running, it will close the
// 'avaiable_thread_idx_' and prefecther threads have no way to tell scheduler
// to release their resource. So a check is needed before scheduler ends.
for (auto& p : prefetchers_) {
......@@ -138,7 +135,8 @@ void MultiFileReader::PrefetchThreadFunc(size_t reader_idx, size_t thread_idx) {
std::vector<framework::LoDTensor> ins;
reader->ReadNext(&ins);
if (ins.empty()) {
reader->ReInit();
reader->Shutdown();
reader->Start();
break;
}
try {
......@@ -180,9 +178,8 @@ class OpenFilesOp : public framework::OperatorBase {
auto* out = scope.FindVar(Output("Out"))
->template GetMutable<framework::ReaderHolder>();
out->Reset(new MultiFileReader(file_names,
RestoreShapes(shape_concat, ranks),
thread_num, buffer_size));
out->Reset(
std::make_shared<MultiFileReader>(file_names, thread_num, buffer_size));
}
};
......
paddle/fluid/operators/reader/reader_op_registry.cc
浏览文件 @ f85e16f1
......@@ -39,7 +39,7 @@ std::unordered_map<std::string, FileReaderCreator>& FileReaderRegistry() {
}
std::unique_ptr<framework::ReaderBase> CreateReaderByFileName(
const std::string& file_name, const std::vector<framework::DDim>& dims) {
const std::string& file_name) {
size_t separator_pos = file_name.find_last_of(kFileFormatSeparator);
PADDLE_ENFORCE_NE(separator_pos, std::string::npos,
"File name illegal! A legal file name should be like: "
......@@ -49,7 +49,7 @@ std::unique_ptr<framework::ReaderBase> CreateReaderByFileName(
auto itor = FileReaderRegistry().find(filetype);
PADDLE_ENFORCE(itor != FileReaderRegistry().end(),
"No file reader registered for '%s' format.", filetype);
framework::ReaderBase* reader = (itor->second)(file_name, dims);
framework::ReaderBase* reader = (itor->second)(file_name);
return std::unique_ptr<framework::ReaderBase>(reader);
}
......
paddle/fluid/operators/reader/reader_op_registry.h
浏览文件 @ f85e16f1
......@@ -25,22 +25,21 @@ namespace reader {
static constexpr char kFileFormatSeparator[] = ".";
using FileReaderCreator = std::function<framework::ReaderBase*(
const std::string&, const std::vector<framework::DDim>&)>;
using FileReaderCreator =
std::function<framework::ReaderBase*(const std::string&)>;
std::unordered_map<std::string, FileReaderCreator>& FileReaderRegistry();
template <typename Reader>
int RegisterFileReader(const std::string& filetype) {
FileReaderRegistry()[filetype] = [](
const std::string& fn, const std::vector<framework::DDim>& dims) {
return new Reader(fn, dims);
FileReaderRegistry()[filetype] = [](const std::string& fn) {
return new Reader(fn);
};
return 0;
}
std::unique_ptr<framework::ReaderBase> CreateReaderByFileName(
const std::string& file_name, const std::vector<framework::DDim>& dims);
const std::string& file_name);
extern std::vector<framework::DDim> RestoreShapes(
const std::vector<int>& shape_concat, const std::vector<int>& ranks);
......
paddle/fluid/platform/CMakeLists.txt
浏览文件 @ f85e16f1
......@@ -28,6 +28,9 @@ cc_test(place_test SRCS place_test.cc DEPS place glog gflags)
add_subdirectory(dynload)
cc_library(cpu_helper SRCS cpu_helper.cc DEPS cblas enforce)
cc_test(cpu_helper_test SRCS cpu_helper_test.cc DEPS cpu_helper)
IF(WITH_GPU)
set(GPU_CTX_DEPS dynload_cuda dynamic_loader)
ELSE()
......@@ -42,10 +45,12 @@ ENDIF()
# memcpy depends on device_context, here add deps individually for
# avoiding cycle dependencies
cc_library(device_context SRCS device_context.cc DEPS malloc
place eigen3 ${GPU_CTX_DEPS} ${MKLDNN_CTX_DEPS})
cc_library(device_context SRCS device_context.cc init.cc DEPS malloc
place eigen3 stringpiece cpu_helper ${GPU_CTX_DEPS} ${MKLDNN_CTX_DEPS})
nv_test(device_context_test SRCS device_context_test.cu DEPS device_context gpu_info)
cc_test(init_test SRCS init_test.cc DEPS device_context)
nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda)
nv_test(transform_test SRCS transform_test.cu DEPS memory place device_context)
......@@ -53,5 +58,5 @@ cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto framewo
cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer)
cc_test(profiler_test SRCS profiler_test.cc DEPS profiler)
nv_test(float16_gpu_test SRCS float16_test.cu)
cc_test(float16_test SRCS float16_test.cc)
nv_test(float16_gpu_test SRCS float16_test.cu DEPS lod_tensor)
cc_test(float16_test SRCS float16_test.cc DEPS lod_tensor)
paddle/fluid/platform/cpu_helper.cc 0 → 100644
浏览文件 @ f85e16f1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/platform/cpu_helper.h"
#include "paddle/fluid/platform/enforce.h"
#ifdef PADDLE_WITH_MKLML
#include "paddle/fluid/platform/dynload/mklml.h"
#endif
#ifdef PADDLE_USE_OPENBLAS
#include <cblas.h>
#endif
namespace paddle {
namespace platform {
void SetNumThreads(int num_threads) {
#ifdef PADDLE_USE_OPENBLAS
int real_num_threads = num_threads > 1 ? num_threads : 1;
openblas_set_num_threads(real_num_threads);
#elif defined(PADDLE_WITH_MKLML)
int real_num_threads = num_threads > 1 ? num_threads : 1;
platform::dynload::MKL_Set_Num_Threads(real_num_threads);
#else
PADDLE_ENFORCE(false, "To be implemented.");
#endif
}
} // namespace platform
} // namespace paddle
paddle/fluid/platform/cpu_helper.h 0 → 100644
浏览文件 @ f85e16f1
/* 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 <stddef.h>
namespace paddle {
namespace platform {
//! Set the number of threads in use.
void SetNumThreads(int num_threads);
} // namespace platform
} // namespace paddle
paddle/fluid/platform/cpu_helper_test.cc 0 → 100644
浏览文件 @ f85e16f1
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/platform/cpu_helper.h"
#include "gtest/gtest.h"
TEST(CpuHelper, SetNumThread) {
paddle::platform::SetNumThreads(1);
paddle::platform::SetNumThreads(4);
}
paddle/fluid/platform/device_context_test.cu
浏览文件 @ f85e16f1
......@@ -69,19 +69,3 @@ TEST(Device, DeviceContextPool) {
ASSERT_NE(dev_ctx, nullptr);
}
}
int main(int argc, char** argv) {
std::vector<paddle::platform::Place> places;
places.emplace_back(paddle::platform::CPUPlace());
int count = paddle::platform::GetCUDADeviceCount();
for (int i = 0; i < count; ++i) {
places.emplace_back(paddle::platform::CUDAPlace(i));
}
VLOG(0) << " DeviceCount " << count;
paddle::platform::DeviceContextPool::Init(places);
testing::InitGoogleTest(&argc, argv);
return RUN_ALL_TESTS();
}
paddle/fluid/platform/dynload/dynamic_loader.cc
浏览文件 @ f85e16f1
......@@ -36,8 +36,6 @@ DEFINE_string(cuda_dir, "",
DEFINE_string(warpctc_dir, "", "Specify path for loading libwarpctc.so.");
DEFINE_string(lapack_dir, "", "Specify path for loading liblapack.so.");
DEFINE_string(nccl_dir, "",
"Specify path for loading nccl library, such as libcublas, "
"libcurand. For instance, /usr/local/cuda/lib64. If default, "
......@@ -189,14 +187,6 @@ void* GetWarpCTCDsoHandle() {
#endif
}
void* GetLapackDsoHandle() {
#if defined(__APPLE__) || defined(__OSX__)
return GetDsoHandleFromSearchPath(FLAGS_lapack_dir, "liblapacke.dylib");
#else
return GetDsoHandleFromSearchPath(FLAGS_lapack_dir, "liblapacke.so");
#endif
}
void* GetNCCLDsoHandle() {
#if defined(__APPLE__) || defined(__OSX__)
return GetDsoHandleFromSearchPath(FLAGS_nccl_dir, "libnccl.dylib");
......
paddle/fluid/platform/dynload/dynamic_loader.h
浏览文件 @ f85e16f1
......@@ -23,7 +23,6 @@ void* GetCUDNNDsoHandle();
void* GetCUPTIDsoHandle();
void* GetCurandDsoHandle();
void* GetWarpCTCDsoHandle();
void* GetLapackDsoHandle();
void* GetNCCLDsoHandle();
void* GetTensorRtDsoHandle();
void* GetMKLMLDsoHandle();
......
paddle/fluid/platform/float16_test.cc
浏览文件 @ f85e16f1
......@@ -13,8 +13,8 @@ limitations under the License. */
#include <vector>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/init.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/platform/init.h"
namespace paddle {
namespace platform {
......
paddle/fluid/framework/init.cc → paddle/fluid/platform/init.cc
浏览文件 @ f85e16f1
......@@ -16,10 +16,10 @@ limitations under the License. */
#include <stdexcept>
#include <string>
#include "paddle/fluid/framework/init.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/platform/cpu_helper.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/init.h"
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/string/piece.h"
......@@ -115,7 +115,7 @@ void InitDevices(bool init_p2p, const std::vector<int> devices) {
places.emplace_back(platform::CPUPlace());
platform::DeviceContextPool::Init(places);
#ifndef PADDLE_WITH_MKLDNN
operators::math::SetNumThreads(1);
platform::SetNumThreads(1);
#endif
}
......
paddle/fluid/framework/init_test.cc → paddle/fluid/platform/init_test.cc
浏览文件 @ f85e16f1
......@@ -13,8 +13,8 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "gtest/gtest.h"
#include "paddle/fluid/framework/init.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/init.h"
TEST(InitDevices, CPU) {
using paddle::framework::InitDevices;
......
paddle/fluid/pybind/CMakeLists.txt
浏览文件 @ f85e16f1
......@@ -2,13 +2,13 @@ if(WITH_PYTHON)
if(WITH_AMD_GPU)
hip_library(paddle_pybind SHARED
SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc
DEPS pybind python proto_desc memory executor prune init profiler feed_fetch_method
DEPS pybind python proto_desc memory executor prune profiler feed_fetch_method
parallel_executor
${GLOB_OP_LIB})
else()
cc_library(paddle_pybind SHARED
SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc
DEPS pybind python proto_desc memory executor prune init profiler feed_fetch_method
DEPS pybind python proto_desc memory executor prune profiler feed_fetch_method
parallel_executor
${GLOB_OP_LIB})
if(NOT APPLE AND NOT ANDROID)
......
paddle/fluid/pybind/pybind.cc
浏览文件 @ f85e16f1
......@@ -25,7 +25,6 @@ limitations under the License. */
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/framework/framework.pb.h"
#include "paddle/fluid/framework/init.h"
#include "paddle/fluid/framework/lod_rank_table.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
......@@ -37,6 +36,7 @@ limitations under the License. */
#include "paddle/fluid/operators/activation_op.h"
#include "paddle/fluid/operators/reader/lod_tensor_blocking_queue.h"
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/platform/init.h"
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/platform/profiler.h"
#include "paddle/fluid/pybind/const_value.h"
......@@ -297,7 +297,7 @@ All parameter, weight, gradient are variables in Paddle.
py::return_value_policy::reference);
py::class_<framework::ReaderHolder>(m, "Reader", "")
.def("reset", &framework::ReaderHolder::ReInit);
.def("reset", &framework::ReaderHolder::ResetAll);
using LoDTensorBlockingQueue =
::paddle::operators::reader::LoDTensorBlockingQueue;
......
paddle/fluid/string/printf.h
浏览文件 @ f85e16f1
......@@ -83,6 +83,13 @@ void Fprintf(std::ostream& out, const char* fmt, const Args&... args) {
tinyformat::vformat(out, fmt, tinyformat::makeFormatList(args...));
}
template <typename... Args>
std::string Sprintf(const Args&... args) {
std::ostringstream oss;
Fprintf(oss, "");
return oss.str();
}
template <typename... Args>
std::string Sprintf(const char* fmt, const Args&... args) {
std::ostringstream oss;
......
paddle/fluid/string/printf_test.cc
浏览文件 @ f85e16f1
......@@ -27,4 +27,5 @@ TEST(StringPrintf, StringPrintf) {
EXPECT_EQ(std::string("Wednesday, July 27, 14:44"),
paddle::string::Sprintf("%s, %s %d, %.2d:%.2d", weekday, month, day,
hour, min));
EXPECT_EQ(std::string(""), paddle::string::Sprintf());
}
paddle/fluid/train/demo/demo_trainer.cc
浏览文件 @ f85e16f1
......@@ -15,11 +15,11 @@
#include <fstream>
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/init.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/tensor_util.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/init.h"
#include "paddle/fluid/platform/place.h"
namespace paddle {
......
paddle/legacy/gserver/gradientmachines/MultiGradientMachine.cpp
浏览文件 @ f85e16f1
......@@ -532,6 +532,7 @@ void TrainerThread::computeThread() {
break;
}
}
hl_fini();
}
void TrainerThread::prefetch() {
......@@ -651,6 +652,7 @@ void TrainerThread::copyGradToBufferThread() {
}
partnerThread->notifyGradientCollect(pid);
}
hl_fini();
}
void TrainerThread::gradCollectThread() {
......@@ -693,6 +695,7 @@ void TrainerThread::gradCollectThread() {
notifyCopyGradToBuffer(pid);
}
}
hl_fini();
}
void TrainerThread::doCallback(int pid) {
......@@ -741,6 +744,7 @@ void TrainerThread::valueDispatchThread() {
thread->notifyValueReady(pid);
}
hl_fini();
}
void TrainerThread::notifyValueReady(int paramId) {
......
paddle/legacy/gserver/gradientmachines/ParallelNeuralNetwork.cpp
浏览文件 @ f85e16f1
......@@ -197,6 +197,7 @@ void ParallelThread::computeThread() {
job_work.layer_->markAllInputGrad();
}
}
hl_fini();
}
void ParallelThread::start() {
......
paddle/scripts/paddle_build.sh
浏览文件 @ f85e16f1
......@@ -312,6 +312,20 @@ EOF
fi
}
function assert_api_not_changed() {
mkdir -p ${PADDLE_ROOT}/build/.check_api_workspace
cd ${PADDLE_ROOT}/build/.check_api_workspace
virtualenv .env
source .env/bin/activate
pip install ${PADDLE_ROOT}/build/python/dist/*whl
curl ${PADDLE_API_SPEC_URL:-https://raw.githubusercontent.com/PaddlePaddle/FluidAPISpec/master/API.spec} \
> origin.spec
python ${PADDLE_ROOT}/tools/print_signatures.py paddle.fluid > new.spec
python ${PADDLE_ROOT}/tools/diff_api.py origin.spec new.spec
deactivate
}
function single_test() {
TEST_NAME=$1
if [ -z "${TEST_NAME}" ]; then
......@@ -550,6 +564,7 @@ function main() {
cicheck)
cmake_gen ${PYTHON_ABI:-""}
build
assert_api_not_changed
run_test
gen_capi_package
gen_fluid_inference_lib
......
paddle/testing/CMakeLists.txt
浏览文件 @ f85e16f1
......@@ -6,6 +6,6 @@ if(WITH_TESTING)
add_library(paddle_test_util STATIC TestUtil.cpp)
add_dependencies(paddle_test_util paddle_proto ${external_project_dependencies})
if(NOT MOBILE_INFERENCE)
cc_library(paddle_gtest_main SRCS paddle_gtest_main.cc DEPS init memory gtest gflags)
cc_library(paddle_gtest_main SRCS paddle_gtest_main.cc DEPS device_context memory gtest gflags)
endif()
endif()
paddle/testing/paddle_gtest_main.cc
浏览文件 @ f85e16f1
......@@ -16,8 +16,8 @@ limitations under the License. */
#include "gflags/gflags.h"
#include "gtest/gtest.h"
#include "paddle/fluid/framework/init.h"
#include "paddle/fluid/memory/memory.h"
#include "paddle/fluid/platform/init.h"
int main(int argc, char** argv) {
testing::InitGoogleTest(&argc, argv);
......
python/paddle/fluid/framework.py
浏览文件 @ f85e16f1
......@@ -26,7 +26,7 @@ except ImportError, e:
"""NOTE: You may need to run \"export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH\"
if you encounters \"libmkldnn.so not found\" errors. If you have python
installed in other directory, replace \"/usr/local/lib\" with your own
directory. The original error is: """ % str(e))
directory. The original error is: \n""" + e.message)
except Exception, e:
raise e
import unique_name
......
python/paddle/fluid/layers/io.py
浏览文件 @ f85e16f1
......@@ -376,9 +376,6 @@ def open_recordio_file(filename,
if pass_num > 1:
main_prog_var = multi_pass(reader=main_prog_var, pass_num=pass_num)
if for_parallel:
main_prog_var = parallel(reader=main_prog_var)
return monkey_patch_reader_methods(main_prog_var)
......@@ -610,9 +607,6 @@ def open_files(filenames,
main_prog_reader = multi_pass(
reader=main_prog_reader, pass_num=pass_num)
if for_parallel:
main_prog_reader = parallel(reader=main_prog_reader)
return monkey_patch_reader_methods(main_prog_reader)
......@@ -728,11 +722,6 @@ def multi_pass(reader, pass_num):
'create_multi_pass_reader', reader, {'pass_num': int(pass_num)})
def parallel(reader):
return __create_shared_decorated_reader__('create_threaded_reader', reader,
{})
def read_file(reader):
"""
Execute the given reader and get data via it.
......
python/paddle/fluid/tests/test_mnist_if_else_op.py → python/paddle/fluid/tests/test_if_else_op.py
浏览文件 @ f85e16f1
......@@ -14,10 +14,11 @@
import paddle
import paddle.fluid.layers as layers
from paddle.fluid.framework import Program, program_guard, default_main_program, default_startup_program
from paddle.fluid.framework import Program, program_guard
from paddle.fluid.executor import Executor
from paddle.fluid.optimizer import MomentumOptimizer
import paddle.fluid.core as core
import paddle.fluid as fluid
import unittest
import numpy as np
......@@ -31,14 +32,13 @@ class TestMNISTIfElseOp(unittest.TestCase):
label = layers.data(name='y', shape=[1], dtype='int64')
limit = layers.fill_constant_batch_size_like(
input=label, dtype='int64', shape=[1], value=5.0)
limit = layers.fill_constant(shape=[1], dtype='int64', value=5)
cond = layers.less_than(x=label, y=limit)
true_image, false_image = layers.split_lod_tensor(
input=image, mask=cond)
true_out = layers.create_tensor(dtype='float32')
true_cond = layers.ConditionalBlock([true_image])
true_cond = layers.ConditionalBlock([cond])
with true_cond.block():
hidden = layers.fc(input=true_image, size=100, act='tanh')
......@@ -46,7 +46,7 @@ class TestMNISTIfElseOp(unittest.TestCase):
layers.assign(input=prob, output=true_out)
false_out = layers.create_tensor(dtype='float32')
false_cond = layers.ConditionalBlock([false_image])
false_cond = layers.ConditionalBlock([cond])
with false_cond.block():
hidden = layers.fc(input=false_image, size=200, act='tanh')
......@@ -64,7 +64,7 @@ class TestMNISTIfElseOp(unittest.TestCase):
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.mnist.train(), buf_size=8192),
batch_size=200)
batch_size=10)
place = core.CPUPlace()
exe = Executor(place)
......@@ -94,8 +94,7 @@ class TestMNISTIfElseOp(unittest.TestCase):
label = layers.data(name='y', shape=[1], dtype='int64')
limit = layers.fill_constant_batch_size_like(
input=label, dtype='int64', shape=[1], value=5.0)
limit = layers.fill_constant(shape=[1], dtype='int64', value=5)
cond = layers.less_than(x=label, y=limit)
ie = layers.IfElse(cond)
......@@ -125,7 +124,7 @@ class TestMNISTIfElseOp(unittest.TestCase):
place = core.CPUPlace()
exe = Executor(place)
exe.run(kwargs['startup_program'])
exe.run(startup_prog)
PASS_NUM = 100
for pass_id in range(PASS_NUM):
for data in train_reader():
......@@ -133,7 +132,7 @@ class TestMNISTIfElseOp(unittest.TestCase):
y_data = np.array(map(lambda x: x[1], data)).astype("int64")
y_data = y_data.reshape((y_data.shape[0], 1))
outs = exe.run(kwargs['main_program'],
outs = exe.run(prog,
feed={'x': x_data,
'y': y_data},
fetch_list=[avg_loss])
......@@ -143,6 +142,67 @@ class TestMNISTIfElseOp(unittest.TestCase):
self.assertFalse(True)
class TestIfElse(unittest.TestCase):
def set_test_case(self):
# condiction is: self.data < self.cond_value
self.cond_value = 0.5
self.data = np.random.rand(25, 1).astype(np.float32)
def compare_ifelse_op_and_numpy(self, place):
self.set_test_case()
prog = Program()
startup_prog = Program()
with program_guard(prog, startup_prog):
src = layers.data(name='data', shape=[1], dtype='float32')
cond = layers.fill_constant(
[1], dtype='float32', value=self.cond_value)
ifcond = layers.less_than(x=src, y=cond)
ie = layers.IfElse(ifcond)
with ie.true_block():
true_target = ie.input(src)
ie.output(true_target)
with ie.false_block():
false_target = ie.input(src)
ie.output(false_target)
if_out = ie()
out = layers.reduce_sum(if_out)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
fetch_list = [out]
o1, = exe.run(fluid.default_main_program(),
feed={'data': self.data},
fetch_list=[out])
o2 = np.sum(self.data)
self.assertTrue(
np.allclose(
o1, o2, atol=1e-8),
"IfElse result : " + str(o1) + "\n Numpy result :" + str(o2))
def test_cpu(self):
self.compare_ifelse_op_and_numpy(fluid.CPUPlace())
def test_cuda(self):
if not core.is_compiled_with_cuda():
return
self.compare_ifelse_op_and_numpy(fluid.CUDAPlace(0))
class TestIfElseTrueBranch(TestIfElse):
def set_test_case(self):
# condiction is: self.data < self.cond_value
self.cond_value = 10.
self.data = np.random.rand(25, 1).astype(np.float32)
class TestIfElseFalseBranch(TestIfElse):
def set_test_case(self):
# condiction is: self.data < self.cond_value
self.cond_value = -10.
self.data = np.random.rand(25, 1).astype(np.float32)
if __name__ == '__main__':
# temp disable if else unittest since it could be buggy.
exit(0)
unittest.main()
python/paddle/fluid/tests/unittests/test_data_balance.py
浏览文件 @ f85e16f1
......@@ -103,8 +103,12 @@ class TestDataBalance(unittest.TestCase):
exe = fluid.Executor(place)
exe.run(startup_prog)
build_strategy = fluid.BuildStrategy()
build_strategy.enable_data_balance = True
parallel_exe = fluid.ParallelExecutor(
use_cuda=self.use_cuda, main_program=main_prog)
use_cuda=self.use_cuda,
main_program=main_prog,
build_strategy=build_strategy)
if (parallel_exe.device_count > self.batch_size):
print("WARNING: Unittest TestDataBalance skipped. \
......@@ -145,9 +149,12 @@ class TestDataBalance(unittest.TestCase):
place = fluid.CUDAPlace(0) if self.use_cuda else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(startup_prog)
build_strategy = fluid.BuildStrategy()
build_strategy.enable_data_balance = True
parallel_exe = fluid.ParallelExecutor(
use_cuda=self.use_cuda, main_program=main_prog)
use_cuda=self.use_cuda,
main_program=main_prog,
build_strategy=build_strategy)
if (parallel_exe.device_count > self.batch_size):
print("WARNING: Unittest TestDataBalance skipped. \
......
python/paddle/fluid/tests/unittests/test_fake_dequantize_op.py
浏览文件 @ f85e16f1
......@@ -40,7 +40,6 @@ class TestFakeDequantizeMaxAbsOp(OpTest):
self.op_type = "fake_dequantize_max_abs"
x = np.random.randn(31, 65).astype("float32")
yq, scale = quantize_max_abs(x, self.num_bits)
print 'scale ', scale
ydq = dequantize_max_abs(yq, self.num_bits, scale)
self.inputs = {'X': yq}
......
python/paddle/fluid/tests/unittests/test_parallel_op.py
浏览文件 @ f85e16f1
......@@ -113,7 +113,9 @@ class BaseParallelForTest(unittest.TestCase):
generator = callback()
# Automatically insert parallel do if use_parallel = True
if use_parallel:
places = fluid.layers.get_places()
thread_num = fluid.core.get_cuda_device_count(
) if use_gpu else 8
places = fluid.layers.get_places(thread_num)
pd = fluid.layers.ParallelDo(places, use_nccl=use_nccl)
data = next(generator)
......
python/paddle/fluid/transpiler/distribute_transpiler.py
浏览文件 @ f85e16f1
......@@ -309,10 +309,10 @@ class DistributeTranspiler(object):
def get_pserver_program(self, endpoint):
"""
Get parameter server side program.
Args:
endpoint (str): current parameter server endpoint.
Returns:
Program: the program for current parameter server to run.
"""
......@@ -516,7 +516,7 @@ class DistributeTranspiler(object):
endpoint (str): current pserver endpoint.
pserver_program (Program): call get_pserver_program first and
pass the result here.
Returns:
Program: parameter server side startup program.
"""
......@@ -552,10 +552,10 @@ class DistributeTranspiler(object):
op_on_pserver = True
new_outputs[key] = pserver_vars[op.output(key)[0]]
# most startup program ops have no inputs
new_inputs = self._get_input_map_from_op(pserver_vars, op)
if op_on_pserver:
# most startup program ops have no inputs
new_inputs = self._get_input_map_from_op(pserver_vars, op)
if op.type in [
"gaussian_random", "fill_constant", "uniform_random"
]:
......
python/paddle/fluid/transpiler/inference_transpiler.py
浏览文件 @ f85e16f1
......@@ -19,7 +19,7 @@ from ..framework import Program
from ..executor import global_scope
class InferenceTranspiler:
class InferenceTranspiler(object):
'''
Convert the fluid program to optimized inference program.
......
python/paddle/libs/__init__.py 0 → 100644
浏览文件 @ f85e16f1
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# used for setup.py.in to store the thirdparty shared libraries
python/setup.py.in
浏览文件 @ f85e16f1
from setuptools import setup, Distribution, Extension
import subprocess
import os
import re
import shutil
class BinaryDistribution(Distribution):
def has_ext_modules(foo):
return True
MAJOR = 0
MINOR = 14
PATCH = 0
RC = 0
ISTAGED = False
......@@ -20,14 +19,47 @@ def git_commit():
git_commit = 'Unknown'
return git_commit
def _get_version_detail(idx):
assert idx < 3, "vesion info consists of %(major)d.%(minor)d.%(patch)d, \
so detail index must less than 3"
if re.match('@TAG_VERSION_REGEX@', '@PADDLE_VERSION@'):
version_details = '@PADDLE_VERSION@'.split('.')
if len(version_details) == 3:
return version_details[idx]
return 0
def get_major():
return int(_get_version_detail(0))
def get_minor():
return int(_get_version_detail(1))
def get_patch():
return str(_get_version_detail(2))
def is_taged():
try:
cmd = ['git', 'describe', '--exact-match', '--tags']
git_tag = subprocess.Popen(cmd, stdout = subprocess.PIPE).communicate()[0].strip()
except:
return False
if git_tag.replace('v', '') == '@PADDLE_VERSION@':
return True
else:
return False
def write_version_py(filename='paddle/version.py'):
cnt = '''
# THIS FILE IS GENERATED FROM PADDLEPADDLE SETUP.PY
#
full_version = '%(major)d.%(minor)d.%(patch)d'
full_version = '%(major)d.%(minor)d.%(patch)s'
major = '%(major)d'
minor = '%(minor)d'
patch = '%(patch)d'
patch = '%(patch)s'
rc = '%(rc)d'
istaged = %(istaged)s
commit = '%(commit)s'
......@@ -49,19 +81,20 @@ def mkl():
commit = git_commit()
with open(filename, 'w') as f:
f.write(cnt % {
'major': MAJOR,
'minor': MINOR,
'patch': PATCH,
'major': get_major(),
'minor': get_minor(),
'patch': get_patch(),
'rc': RC,
'version': '${PADDLE_VERSION}',
'commit': commit,
'istaged': ISTAGED,
'istaged': is_taged(),
'with_mkl': '@WITH_MKL@'})
write_version_py(filename='@PADDLE_BINARY_DIR@/python/paddle/version.py')
packages=['paddle',
'paddle.libs',
'paddle.utils',
'paddle.dataset',
'paddle.reader',
......@@ -113,12 +146,35 @@ package_dir={
}
if '${WITH_FLUID_ONLY}'== 'OFF':
package_dir['py_paddle']='${PADDLE_BINARY_DIR}/python/py_paddle'
paddle_rt_lib_dir = 'lib'
paddle_rt_libs = ['${WARPCTC_LIBRARIES}']
if '${MKL_SHARED_LIBS}'!= '':
paddle_rt_libs += '${MKL_SHARED_LIBS}'.split(';')
# put all thirdparty libraries in paddle.libs
package_data['paddle.libs']=['libwarpctc.so']
libs_path='${PADDLE_BINARY_DIR}/python/paddle/libs'
shutil.copy('${WARPCTC_LIBRARIES}', libs_path)
if '${WITH_MKL}' == 'ON':
shutil.copy('${MKLML_LIB}', libs_path)
shutil.copy('${MKLML_IOMP_LIB}', libs_path)
package_data['paddle.libs']+=['libmklml_intel.so','libiomp5.so']
if '${WITH_MKLDNN}' == 'ON':
# change rpath of libmkldnn.so.0, add $ORIGIN/ to it.
# The reason is that all thirdparty libraries in the same directory,
# thus, libmkldnn.so.0 will find libmklml_intel.so and libiomp5.so.
command = "patchelf --set-rpath '$ORIGIN/' ${MKLDNN_SHARED_LIB}"
if os.system(command) != 0:
raise Exception("patchelf --set-rpath for libmkldnn.so.0 fails")
package_data['paddle.libs']+=['libmkldnn.so.0']
shutil.copy('${MKLDNN_SHARED_LIB}', libs_path)
# remove unused paddle/libs/__init__.py
os.remove(libs_path+'/__init__.py')
package_dir['paddle.libs']=libs_path
# change rpath of core.so, add $ORIGIN/../libs/ to it.
# The reason is that libwarpctc.so, libiomp5.so etc are in paddle.libs, and
# core.so is in paddle.fluid, thus paddle/fluid/../libs will pointer to above libraries.
# This operation will fix https://github.com/PaddlePaddle/Paddle/issues/3213
command = "patchelf --set-rpath '$ORIGIN/../libs/' ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so"
if os.system(command) != 0:
raise Exception("patchelf --set-rpath for core.so fails")
setup(name='${PACKAGE_NAME}',
version='${PADDLE_VERSION}',
......@@ -128,6 +184,5 @@ setup(name='${PACKAGE_NAME}',
ext_modules=[Extension('_foo', ['stub.cc'])],
package_data=package_data,
package_dir=package_dir,
scripts=paddle_bins,
data_files=[(paddle_rt_lib_dir, paddle_rt_libs)]
scripts=paddle_bins
)
tools/diff_api.py 0 → 100644
浏览文件 @ f85e16f1
#!/usr/bin/env python
from __future__ import print_function
import difflib
import sys
with open(sys.argv[1], 'r') as f:
origin = f.read()
origin = origin.splitlines()
with open(sys.argv[2], 'r') as f:
new = f.read()
new = new.splitlines()
differ = difflib.Differ()
result = differ.compare(origin, new)
error = False
print('API Difference is: ')
for each_diff in result:
if each_diff[0] in ['-', '?']: # delete or change API is not allowed
error = True
elif each_diff[0] == '+':
# only new layers is allowed.
if not each_diff.startswith('+ paddle.fluid.layers.'):
error = True
if each_diff[0] != ' ':
print(each_diff)
if error:
sys.exit(1)
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