Skip to content

P
Paddle

PaddlePaddle / Paddle
大约 1 年 前同步成功

通知 2299
Star 20931
Fork 5422
P
Paddle
提交 aece2905 编写于 作者: X xzl
浏览文件
操作

add depthwise gpu forward, backward, test, interface

上级 06db7038 88a95a02
显示空白变更内容
内联 并排
Showing with 1752 addition and 1408 deletion
CMakeLists.txt
浏览文件 @ aece2905
......@@ -55,6 +55,8 @@ option(WITH_COVERAGE "Compile PaddlePaddle with code coverage" OFF)
option(COVERALLS_UPLOAD "Package code coverage data to coveralls" OFF)
option(ON_TRAVIS "Exclude special unit test on Travis CI" OFF)
option(WITH_C_API "Compile PaddlePaddle with C-API(Prediction)" OFF)
# TODO: Only compile PaddlePaddle fluid version by WITH_FLUID option.
option(WITH_FLUID "Compile PaddlePaddle fluid only(TODO)" ON)
option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF)
option(GLIDE_INSTALL "Download and install go dependencies " ON)
option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF)
......@@ -107,6 +109,10 @@ if (WITH_C_API AND WITH_PYTHON)
"different Python interpreter from compiling.")
endif()
if (WITH_C_API)
set(WITH_FLUID OFF CACHE STRING "Disable install fluid when compile the C_API" FORCE)
endif()
if(MOBILE_INFERENCE)
set(THIRD_PARTY_BUILD_TYPE MinSizeRel)
else()
......
cmake/external/eigen.cmake
浏览文件 @ aece2905
INCLUDE(ExternalProject)
SET(EIGEN_SOURCE_DIR ${THIRD_PARTY_PATH}/eigen3)
INCLUDE_DIRECTORIES(${EIGEN_SOURCE_DIR}/src/extern_eigen3)
SET(EIGEN_INCLUDE_DIR ${EIGEN_SOURCE_DIR}/src/extern_eigen3)
INCLUDE_DIRECTORIES(${EIGEN_INCLUDE_DIR})
ExternalProject_Add(
extern_eigen3
......@@ -28,3 +28,9 @@ endif()
add_dependencies(eigen3 extern_eigen3)
LIST(APPEND external_project_dependencies eigen3)
IF(NOT WITH_C_API AND WITH_FLUID)
INSTALL(FILES ${EIGEN_INCLUDE_DIR}/Eigen/Core DESTINATION third_party/eigen3/Eigen)
INSTALL(DIRECTORY ${EIGEN_INCLUDE_DIR}/Eigen/src DESTINATION third_party/eigen3/Eigen)
INSTALL(DIRECTORY ${EIGEN_INCLUDE_DIR}/unsupported/Eigen DESTINATION third_party/eigen3/unsupported)
ENDIF()
cmake/external/gflags.cmake
浏览文件 @ aece2905
......@@ -52,7 +52,7 @@ ADD_DEPENDENCIES(gflags extern_gflags)
LIST(APPEND external_project_dependencies gflags)
IF(WITH_C_API)
IF(WITH_C_API OR WITH_FLUID)
INSTALL(DIRECTORY ${GFLAGS_INCLUDE_DIR} DESTINATION third_party/gflags)
IF(ANDROID)
INSTALL(FILES ${GFLAGS_LIBRARIES} DESTINATION third_party/gflags/lib/${ANDROID_ABI})
......
cmake/external/glog.cmake
浏览文件 @ aece2905
......@@ -68,7 +68,7 @@ LINK_LIBRARIES(glog gflags)
LIST(APPEND external_project_dependencies glog)
IF(WITH_C_API)
IF(WITH_C_API OR WITH_FLUID)
INSTALL(DIRECTORY ${GLOG_INCLUDE_DIR} DESTINATION third_party/glog)
IF(ANDROID)
INSTALL(FILES ${GLOG_LIBRARIES} DESTINATION third_party/glog/lib/${ANDROID_ABI})
......
cmake/external/protobuf.cmake
浏览文件 @ aece2905
......@@ -250,7 +250,7 @@ IF(NOT PROTOBUF_FOUND)
SET(PROTOBUF_PROTOC_LIBRARY ${extern_protobuf_PROTOC_LIBRARY}
CACHE FILEPATH "protoc library." FORCE)
IF(WITH_C_API)
IF(WITH_C_API OR WITH_FLUID)
INSTALL(DIRECTORY ${PROTOBUF_INCLUDE_DIR} DESTINATION third_party/protobuf)
IF(ANDROID)
INSTALL(FILES ${PROTOBUF_LITE_LIBRARY} DESTINATION third_party/protobuf/lib/${ANDROID_ABI})
......
doc/api/v1/data_provider/dataprovider_cn.rst 已删除 100644 → 0
浏览文件 @ 06db7038
.. _api_dataprovider:
DataProvider的介绍
==================
DataProvider是PaddlePaddle负责提供数据的模块。其作用是将数据传入内存或显存，让神经网络可以进行训练或预测。用户可以通过简单使用Python接口 :ref:`api_pydataprovider2` ，来自定义传数据的过程。如果有更复杂的使用，或者需要更高的效率，用户也可以在C++端自定义一个 ``DataProvider`` 。
PaddlePaddle需要用户在网络配置（trainer_config.py）中定义使用哪种DataProvider，并且在DataProvider中实现如何访问训练文件列表（train.list）或测试文件列表（test.list）。
- train.list和test.list存放在本地（推荐直接存放到训练目录，以相对路径引用)。一般情况下，两者均为纯文本文件，其中每一行对应一个数据文件地址：
- 如果数据文件存于本地磁盘，这个地址则为它的绝对路径或相对路径(相对于PaddlePaddle程序运行时的路径)。
- 地址也可以为hdfs文件路径，或者数据库连接路径等。
- 由于这个地址会被DataProvider使用，因此，如何解析该地址也是用户自定义DataProvider时需要考虑的地方。
- 如果没有设置test.list，或设置为None，那么在训练过程中不会执行测试操作；否则，会根据命令行参数指定的测试方式，在训练过程中进行测试，从而防止过拟合。
doc/api/v1/data_provider/dataprovider_en.rst 已删除 100644 → 0
浏览文件 @ 06db7038
Introduction
==============
DataProvider is a module that loads training or testing data into cpu or gpu
memory for the following triaining or testing process.
For simple use, users can use Python :code:`PyDataProvider` to dynamically reads
the original data in any format or in any form, and then transfer them into a
data format PaddlePaddle requires. The process is extremly flexible and highly
customized, with sacrificing the efficiency only a little. This is extremly
useful when you have to dynamically generate certain kinds of data according to,
for example, the training performance.
Besides, users also can customize a C++ :code:`DataProvider` for a more
complex usage, or for a higher efficiency.
The following parameters are required to define in the PaddlePaddle network
configuration file (trainer_config.py): which DataProvider is chosen to used,
and specific parameters for DataProvider, including training file list
(train.list) and testing file list (test.list).
Train.list and test.list are simply two plain text files, which defines path
of training or testing data. It is recommended that directly placing them into
the training directory, and reference to them by using a relative path (
relative to the PaddePaddle program).
Testing or evaluating will not be performed during training if the test.list is
not set or set to None. Otherwise, PaddlePaddle will evaluate the trained model
by the specified tesing data while training, every testing period (a user
defined command line parameter in PaddlePaddle) to prevent over-fitting.
Each line of train.list and test.list is an absolute or relative path (relative
to the PaddePaddle program runtime) of data file. Fascinatingly more, each line
can also be a HDFS file path or a SQL connection string. As long as the user
assures how to access each file in DataProvider.
doc/api/v1/data_provider/pydataprovider2_cn.rst 已删除 100644 → 0
浏览文件 @ 06db7038
.. _api_pydataprovider2:
PyDataProvider2的使用
=====================
PyDataProvider2是PaddlePaddle使用Python提供数据的推荐接口。该接口使用多线程读取数据，并提供了简单的Cache功能；同时可以使用户只关注如何从文件中读取每一条数据，而不用关心数据如何传输，如何存储等等。
.. contents::
MNIST的使用场景
---------------
我们以MNIST手写识别为例，来说明PyDataProvider2的简单使用场景。
样例数据
++++++++
MNIST是一个包含有70,000张灰度图片的数字分类数据集。样例数据 ``mnist_train.txt`` 如下：
.. literalinclude:: src/mnist_train.txt
其中每行数据代表一张图片，行内使用 ``;`` 分成两部分。第一部分是图片的标签，为0-9中的一个数字；第二部分是28*28的图片像素灰度值。 对应的 ``train.list`` 即为这个数据文件的名字：
.. literalinclude:: src/train.list
dataprovider的使用
++++++++++++++++++
.. literalinclude:: src/mnist_provider.dict.py
- 首先，引入PaddlePaddle的PyDataProvider2包。
- 其次，定义一个Python的 `Decorator <http://www.learnpython.org/en/Decorators>`_ `@provider`_ 。用于将下一行的数据输入函数标记成一个PyDataProvider2，同时设置它的input_types属性。
- `input_types`_：设置这个PyDataProvider2返回什么样的数据。本例根据网络配置中 ``data_layer`` 的名字，显式指定返回的是一个28*28维的稠密浮点数向量和一个[0-9]的10维整数标签。
.. literalinclude:: src/mnist_config.py
:lines: 9-10
- 注意：如果用户不显示指定返回数据的对应关系，那么PaddlePaddle会根据layer的声明顺序，来确定对应关系。但这个关系可能不正确，所以推荐使用显式指定的方式来设置input_types。
- 最后，实现数据输入函数（如本例的 ``process`` 函数）。
- 该函数的功能是：打开文本文件，读取每一行，将行中的数据转换成与input_types一致的格式，然后返回给PaddlePaddle进程。注意，
- 返回的顺序需要和input_types中定义的顺序一致。
- 返回时，必须使用Python关键词 ``yield`` ，相关概念是 ``generator`` 。
- 一次yield调用，返回一条完整的样本。如果想为一个数据文件返回多条样本，只需要在函数中调用多次yield即可（本例中使用for循环进行多次调用）。
- 该函数具有两个参数：
- settings：在本例中没有使用，具体可以参考 `init_hook`_ 中的说明。
- filename：为 ``train.list`` 或 ``test.list`` 中的一行，即若干数据文件路径的某一个。
网络配置中的调用
++++++++++++++++
在网络配置里，只需要一行代码就可以调用这个PyDataProvider2，如，
.. literalinclude:: src/mnist_config.py
:lines: 1-7
训练数据是 ``train.list`` ，没有测试数据，调用的PyDataProvider2是 ``mnist_provider`` 模块中的 ``process`` 函数。
小结
+++++
至此，简单的PyDataProvider2样例就说明完毕了。对用户来说，仅需要知道如何从 **一个文件** 中读取 **一条样本** ，就可以将数据传送给PaddlePaddle了。而PaddlePaddle则会帮用户做以下工作：
* 将数据组合成Batch进行训练
* 对训练数据进行Shuffle
* 多线程的数据读取
* 缓存训练数据到内存(可选)
* CPU->GPU双缓存
是不是很简单呢？
时序模型的使用场景
------------------
样例数据
++++++++
时序模型是指数据的某一维度是一个序列形式，即包含时间步信息。所谓时间步信息，不一定和时间有关系，只是说明数据的顺序是重要的。例如，文本信息就是一个序列数据。
本例采用英文情感分类的数据，即将一段英文文本数据，分类成正面情绪和负面情绪两类(用0和1表示)。样例数据 ``sentimental_train.txt`` 如下：
.. literalinclude:: src/sentimental_train.txt
dataprovider的使用
++++++++++++++++++
相对MNIST而言，这个dataprovider较复杂，主要原因是增加了初始化机制 `init_hook`_。本例的 ``on_init`` 函数就是根据该机制配置的，它会在dataprovider创建的时候执行。
- 其中 ``input_types`` 和在 `@provider`_ 中配置的效果一致。本例中的输入特征是词ID的序列，因此使用 ``integer_value_sequence`` 类型来设置。
- 将 ``dictionary`` 存入settings对象，在 ``process`` 函数中使用。 dictionary是从网络配置中传入的dict对象，即一个将单词字符串映射到单词ID的字典。
.. literalinclude:: src/sentimental_provider.py
网络配置中的调用
++++++++++++++++
调用这个PyDataProvider2的方法，基本上和MNIST样例一致，除了
* 在配置中需要读取外部字典。
* 在声明DataProvider的时候传入dictionary作为参数。
.. literalinclude:: src/sentimental_config.py
:emphasize-lines: 12-14
参考(Reference)
---------------
@provider
+++++++++
``@provider`` 是一个Python的 `Decorator`_ ，可以将某一个函数标记成一个PyDataProvider2。如果不了解 `Decorator`_ 是什么也没关系，只需知道这是一个标记属性的方法就可以了。它包含的属性参数如下:
* input_types：数据输入格式。具体的格式说明，请参考 `input_types`_ 。
* should_shuffle：是不是要对数据做Shuffle。训练时默认shuffle，测试时默认不shuffle。
* min_pool_size：设置内存中最小暂存的数据条数，也是PaddlePaddle所能够保证的shuffle粒度。如果为-1，则会预先读取全部数据到内存中。
* pool_size： 设置内存中暂存的数据条数。如果为-1（默认），则不在乎内存暂存多少条数据。如果设置，则推荐大于训练时batch size的值，并且在内存足够的情况下越大越好。
* can_over_batch_size：是否允许暂存略微多余pool_size的数据。由于这样做可以避免很多死锁问题，一般推荐设置成True。
* calc_batch_size：可以传入一个函数，用于自定义每条数据的batch size（默认为1）。
* cache： 数据缓存的策略，具体请参考 `cache`_ 。
* init_hook：初始化时调用的函数，具体请参考 `init_hook`_ 。
* check：如果为true，会根据input_types检查数据的合法性。
* check_fail_continue：如果为true，那么当check出数据不合法时，会扔到这条数据，继续训练或预测。（对check=false的情况，没有作用）
input_types
+++++++++++
PaddlePaddle的数据包括四种主要类型，和三种序列模式。
四种数据类型：
* dense_vector：稠密的浮点数向量。
* sparse_binary_vector：稀疏的01向量，即大部分值为0，但有值的地方必须为1。
* sparse_float_vector：稀疏的向量，即大部分值为0，但有值的部分可以是任何浮点数。
* integer：整数标签。
三种序列模式：
* SequenceType.NO_SEQUENCE：不是一条序列
* SequenceType.SEQUENCE：是一条时间序列
* SequenceType.SUB_SEQUENCE： 是一条时间序列，且序列的每一个元素还是一个时间序列。
不同的数据类型和序列模式返回的格式不同，列表如下：
+----------------------+---------------------+-----------------------------------+------------------------------------------------+
| | NO_SEQUENCE | SEQUENCE | SUB_SEQUENCE |
+======================+=====================+===================================+================================================+
| dense_vector | [f, f, ...] | [[f, ...], [f, ...], ...] | [[[f, ...], ...], [[f, ...], ...],...] |
+----------------------+---------------------+-----------------------------------+------------------------------------------------+
| sparse_binary_vector | [i, i, ...] | [[i, ...], [i, ...], ...] | [[[i, ...], ...], [[i, ...], ...],...] |
+----------------------+---------------------+-----------------------------------+------------------------------------------------+
| sparse_float_vector | [(i,f), (i,f), ...] | [[(i,f), ...], [(i,f), ...], ...] | [[[(i,f), ...], ...], [[(i,f), ...], ...],...] |
+----------------------+---------------------+-----------------------------------+------------------------------------------------+
| integer_value | i | [i, i, ...] | [[i, ...], [i, ...], ...] |
+----------------------+---------------------+-----------------------------------+------------------------------------------------+
其中，f代表一个浮点数，i代表一个整数。
注意：对sparse_binary_vector和sparse_float_vector，PaddlePaddle存的是有值位置的索引。例如，
- 对一个5维非序列的稀疏01向量 ``[0, 1, 1, 0, 0]`` ，类型是sparse_binary_vector，返回的是 ``[1, 2]`` 。
- 对一个5维非序列的稀疏浮点向量 ``[0, 0.5, 0.7, 0, 0]`` ，类型是sparse_float_vector，返回的是 ``[(1, 0.5), (2, 0.7)]`` 。
init_hook
+++++++++
init_hook可以传入一个函数。该函数在初始化的时候会被调用，其参数如下:
* 第一个参数是settings对象，它和数据传入函数的第一个参数（如本例中 ``process`` 函数的 ``settings`` 参数）必须一致。该对象具有以下两个属性：
* settings.input_types：数据输入格式，具体请参考 `input_types`_ 。
* settings.logger：一个logging对象。
* 其他参数使用 ``kwargs`` （key word arguments）传入，包括以下两种：
* PaddlePaddle定义的参数: 1）is_train：bool型参数，表示用于训练或预测；2）file_list：所有文件列表。
* 用户定义的参数：使用args在网络配置中设置。
注意：PaddlePaddle保留添加参数的权力，因此init_hook尽量使用 ``**kwargs`` 来接受不使用的函数以保证兼容性。
cache
+++++
PyDataProvider2提供了两种简单的Cache策略：
* CacheType.NO_CACHE：不缓存任何数据，每次都会从python端读取数据
* CacheType.CACHE_PASS_IN_MEM：第一个pass会从python端读取数据，剩下的pass会直接从内存里
读取数据。
注意事项
--------
可能的内存泄露问题
++++++++++++++++++
PaddlePaddle将train.list中的每一行都传递给process函数，从而生成多个generator。当训练数据非常多时，就会生成非常多的generator。
虽然每个generator在没有调用的时候，是几乎不占内存的；但当调用过一次后，generator便会存下当前的上下文(Context)，而这个Context可能会非常大。并且，generator至少需要调用两次才会知道是否停止。所以，即使process函数里面只有一个yield，也需要两次随机选择到相同generator的时候，才会释放该段内存。
.. code-block:: python
def func():
yield 0
f = func() # 创建generator
tmp = next(f) # 调用一次，返回0
tmp = next(f) # 调用第二次的时候，才会Stop Iteration
由于顺序调用这些generator不会出现上述问题，因此有两种解决方案：
1. **最佳推荐**：将样本的地址放入另一个文本文件，train.list写入那个文本文件的地址。即不要将每一个样本都放入train.list。
2. 在generator的上下文中尽量留下非常少的变量引用，例如
.. code-block:: python
def real_process(fn):
# ... read from fn
return result # 当函数返回的时候，python可以解除掉内部变量的引用。
def process(fn):
yield real_process(fn)
注意：这个问题是PyDataProvider读数据时候的逻辑问题，很难整体修正。
内存不够用的情况
++++++++++++++++
PyDataProvider2会尽可能多的使用内存。因此，对于内存较小的机器，推荐使用 ``pool_size`` 变量来设置内存中暂存的数据条。具体请参考 `@provider`_ 中的说明。
doc/api/v1/data_provider/pydataprovider2_en.rst 已删除 100644 → 0
浏览文件 @ 06db7038
.. _api_pydataprovider2:
PyDataProvider2
===============
We highly recommand users to use PyDataProvider2 to provide training or testing
data to PaddlePaddle. The user only needs to focus on how to read a single
sample from the original data file by using PyDataProvider2, leaving all of the
trivial work, including, transfering data into cpu/gpu memory, shuffle, binary
serialization to PyDataProvider2. PyDataProvider2 uses multithreading and a
fanscinating but simple cache strategy to optimize the efficiency of the data
providing process.
DataProvider for the non-sequential model
-----------------------------------------
Here we use the MNIST handwriting recognition data as an example to illustrate
how to write a simple PyDataProvider.
MNIST is a handwriting classification data set. It contains 70,000 digital
grayscale images. Labels of the training sample range from 0 to 9. All the
images have been size-normalized and centered into images with the same size
of 28 x 28 pixels.
A small part of the original data as an example is shown as below:
.. literalinclude:: src/mnist_train.txt
Each line of the data contains two parts, separated by :code:`;`. The first part is
label of an image. The second part contains 28x28 pixel float values.
Just write path of the above data into train.list. It looks like this:
.. literalinclude:: src/train.list
The corresponding dataprovider is shown as below:
.. literalinclude:: src/mnist_provider.dict.py
The first line imports PyDataProvider2 package.
The main function is the process function, that has two parameters.
The first parameter is the settings, which is not used in this example.
The second parameter is the filename, that is exactly each line of train.list.
This parameter is passed to the process function by PaddlePaddle.
:code:`@provider` is a Python
`Decorator <http://www.learnpython.org/en/Decorators>`_ .
It sets some properties to DataProvider, and constructs a real PaddlePaddle
DataProvider from a very simple user implemented python function. It does not
matter if you are not familiar with `Decorator`_. You can keep it simple by
just taking :code:`@provider` as a fixed mark above the provider function you
implemented.
`input_types`_ defines the data format that a DataProvider returns.
In this example, it is set to a 28x28-dimensional dense vector and an integer
scalar, whose value ranges from 0 to 9.
`input_types`_ can be set to several kinds of input formats, please refer to the
document of `input_types`_ for more details.
The process method is the core part to construct a real DataProvider in
PaddlePaddle. It implements how to open the text file, how to read one sample
from the original text file, convert them into `input_types`_, and give them
back to PaddlePaddle process at line 23.
Note that data yielded by the process function must follow the same order that
`input_types`_ are defined.
With the help of PyDataProvider2, user can focus on how to generate ONE traning
sample by using keywords :code:`yield`.
:code:`yield` is a python keyword, and a concept related to it includes
:code:`generator`.
Only a few lines of codes need to be added into the training configuration file,
you can take this as an example.
.. literalinclude:: src/mnist_config.py
Here we specify training data by :code:`train.list`, and no testing data is specified.
The method which actually provide data is :code:`process`.
User also can use another style to provide data, which defines the
:code:`data_layer`'s name explicitly when `yield`. For example,
the :code:`dataprovider` is shown as below.
.. literalinclude:: src/mnist_provider.dict.py
:linenos:
If user did't give the :code:`data_layer`'s name, PaddlePaddle will use
the order of :code:`data_layer` definition roughly to determine which feature to
which :code:`data_layer`. This order may be not correct, so TO DEFINE THE
:code:`data_layer`'s NAMES EXPLICITLY IS THE RECOMMANDED WAY TO PROVIDER DATA.
Now, this simple example of using PyDataProvider is finished.
The only thing that the user should know is how to generte **one sample** from
**one data file**.
And PaddlePadle will do all of the rest things\:
* Form a training batch
* Shuffle the training data
* Read data with multithreading
* Cache the training data (Optional)
* CPU-> GPU double buffering.
Is this cool?
.. _api_pydataprovider2_sequential_model:
DataProvider for the sequential model
-------------------------------------
A sequence model takes sequences as its input. A sequence is made up of several
timesteps. The so-called timestep, is not necessary to have something to do
with time. It can also be explained to that the order of data are taken into
consideration into model design and training.
For example, the sentence can be interpreted as a kind of sequence data in NLP
tasks.
Here is an example on data proivider for English sentiment classification data.
The original input data are simple English text, labeled into positive or
negative sentiment (marked by 0 and 1 respectively).
A small part of the original data as an example can be found in the path below:
.. literalinclude:: src/sentimental_train.txt
The corresponding data provider can be found in the path below:
.. literalinclude:: src/sentimental_provider.py
This data provider for sequential model is a little more complex than that
for MINST dataset.
A new initialization method is introduced here.
The method :code:`on_init` is configured to DataProvider by :code:`@provider`'s
:code:`init_hook` parameter, and it will be invoked once DataProvider is
initialized. The :code:`on_init` function has the following parameters:
* The first parameter is the settings object.
* The rest parameters are passed by key word arguments. Some of them are passed
by PaddlePaddle, see reference for `init_hook`_.
The :code:`dictionary` object is a python dict object passed from the trainer
configuration file, and it maps word string to word id.
To pass these parameters into DataProvider, the following lines should be added
into trainer configuration file.
.. literalinclude:: src/sentimental_config.py
The definition is basically same as MNIST example, except:
* Load dictionary in this configuration
* Pass it as a parameter to the DataProvider
The `input_types` is configured in method :code:`on_init`. It has the same
effect to configure them by :code:`@provider`'s :code:`input_types` parameter.
However, the :code:`input_types` is set at runtime, so we can set it to
different types according to the input data. Input of the neural network is a
sequence of word id, so set :code:`seq_type` to :code:`integer_value_sequence`.
Durning :code:`on_init`, we save :code:`dictionary` variable to
:code:`settings`, and it will be used in :code:`process`. Note the settings
parameter for the process function and for the on_init's function are a same
object.
The basic processing logic is the same as MNIST's :code:`process` method. Each
sample in the data file is given back to PaddlePaddle process.
Thus, the basic usage of PyDataProvider is here.
Please refer to the following section reference for details.
Reference
---------
@provider
+++++++++
.. autofunction:: paddle.trainer.PyDataProvider2.provider
input_types
+++++++++++
PaddlePaddle has four data types, and three sequence types.
The four data types are:
* :code:`dense_vector`: dense float vector.
* :code:`sparse_binary_vector`: sparse binary vector, most of the value is 0, and
the non zero elements are fixed to 1.
* :code:`sparse_float_vector`: sparse float vector, most of the value is 0, and some
non zero elements can be any float value. They are given by the user.
* :code:`integer`: an integer scalar, that is especially used for label or word index.
The three sequence types are:
* :code:`SequenceType.NO_SEQUENCE` means the sample is not a sequence.
* :code:`SequenceType.SEQUENCE` means the sample is a sequence.
* :code:`SequenceType.SUB_SEQUENCE` means it is a nested sequence, that each timestep of
the input sequence is also a sequence.
Different input type has a defferenct input format. Their formats are shown
in the above table.
+----------------------+---------------------+-----------------------------------+------------------------------------------------+
| | NO_SEQUENCE | SEQUENCE | SUB_SEQUENCE |
+======================+=====================+===================================+================================================+
| dense_vector | [f, f, ...] | [[f, ...], [f, ...], ...] | [[[f, ...], ...], [[f, ...], ...],...] |
+----------------------+---------------------+-----------------------------------+------------------------------------------------+
| sparse_binary_vector | [i, i, ...] | [[i, ...], [i, ...], ...] | [[[i, ...], ...], [[i, ...], ...],...] |
+----------------------+---------------------+-----------------------------------+------------------------------------------------+
| sparse_float_vector | [(i,f), (i,f), ...] | [[(i,f), ...], [(i,f), ...], ...] | [[[(i,f), ...], ...], [[(i,f), ...], ...],...] |
+----------------------+---------------------+-----------------------------------+------------------------------------------------+
| integer_value | i | [i, i, ...] | [[i, ...], [i, ...], ...] |
+----------------------+---------------------+-----------------------------------+------------------------------------------------+
where f represents a float value, i represents an integer value.
init_hook
+++++++++
init_hook is a function that is invoked once the data provoder is initialized.
Its parameters lists as follows:
* The first parameter is a settings object, which is the same to :code:`settings`
in :code:`process` method. The object contains several attributes, including:
* :code:`settings.input_types`: the input types. Reference `input_types`_.
* :code:`settings.logger`: a logging object.
* The rest parameters are the key word arguments. It is made up of PaddpePaddle
pre-defined parameters and user defined parameters.
* PaddlePaddle-defined parameters including:
* :code:`is_train` is a bool parameter that indicates the DataProvider is used in
training or testing.
* :code:`file_list` is the list of all files.
* User-defined parameters args can be set in training configuration.
Note, PaddlePaddle reserves the right to add pre-defined parameter, so please
use :code:`**kwargs` in init_hook to ensure compatibility by accepting the
parameters which your init_hook does not use.
cache
+++++
DataProvider provides two simple cache strategy. They are:
* :code:`CacheType.NO_CACHE` means do not cache any data, then data is read at runtime by
the user implemented python module every pass.
* :code:`CacheType.CACHE_PASS_IN_MEM` means the first pass reads data by the user
implemented python module, and the rest passes will directly read data from
memory.
doc/api/v1/data_provider/src/mnist_config.py 已删除 100644 → 0
浏览文件 @ 06db7038
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle.trainer_config_helpers import *
define_py_data_sources2(
train_list='train.list',
test_list=None,
module='mnist_provider',
obj='process')
img = data_layer(name='pixel', size=784)
label = data_layer(name='label', size=10)
doc/api/v1/data_provider/src/mnist_provider.dict.py 已删除 100644 → 0
浏览文件 @ 06db7038
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle.trainer.PyDataProvider2 import *
# Define a py data provider
@provider(
input_types={'pixel': dense_vector(28 * 28),
'label': integer_value(10)})
def process(settings, filename): # settings is not used currently.
f = open(filename, 'r') # open one of training file
for line in f: # read each line
label, pixel = line.split(';')
# get features and label
pixels_str = pixel.split(' ')
pixels_float = []
for each_pixel_str in pixels_str:
pixels_float.append(float(each_pixel_str))
# give data to paddle.
yield {"pixel": pixels_float, 'label': int(label)}
f.close() # close file
doc/api/v1/data_provider/src/mnist_train.txt 已删除 100644 → 0
浏览文件 @ 06db7038
5;0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.215686 0.533333 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.67451 0.992157 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.070588 0.886275 0.992157 0 0 0 0 0 0 0 0 0 0 0.192157 0.070588 0 0 0 0 0 0 0 0 0 0 0 0 0 0.670588 0.992157 0.992157 0 0 0 0 0 0 0 0 0 0.117647 0.933333 0.858824 0.313725 0 0 0 0 0 0 0 0 0 0 0 0.090196 0.858824 0.992157 0.831373 0 0 0 0 0 0 0 0 0 0.141176 0.992157 0.992157 0.611765 0.054902 0 0 0 0 0 0 0 0 0 0 0.258824 0.992157 0.992157 0.529412 0 0 0 0 0 0 0 0 0 0.368627 0.992157 0.992157 0.419608 0.003922 0 0 0 0 0 0 0 0 0 0.094118 0.835294 0.992157 0.992157 0.517647 0 0 0 0 0 0 0 0 0 0.603922 0.992157 0.992157 0.992157 0.603922 0.545098 0.043137 0 0 0 0 0 0 0 0.447059 0.992157 0.992157 0.956863 0.062745 0 0 0 0 0 0 0 0 0.011765 0.666667 0.992157 0.992157 0.992157 0.992157 0.992157 0.745098 0.137255 0 0 0 0 0 0.152941 0.866667 0.992157 0.992157 0.521569 0 0 0 0 0 0 0 0 0 0.070588 0.992157 0.992157 0.992157 0.803922 0.352941 0.745098 0.992157 0.945098 0.317647 0 0 0 0 0.580392 0.992157 0.992157 0.764706 0.043137 0 0 0 0 0 0 0 0 0 0.070588 0.992157 0.992157 0.776471 0.043137 0 0.007843 0.27451 0.882353 0.941176 0.176471 0 0 0.180392 0.898039 0.992157 0.992157 0.313725 0 0 0 0 0 0 0 0 0 0 0.070588 0.992157 0.992157 0.713725 0 0 0 0 0.627451 0.992157 0.729412 0.062745 0 0.509804 0.992157 0.992157 0.776471 0.035294 0 0 0 0 0 0 0 0 0 0 0.494118 0.992157 0.992157 0.968627 0.168627 0 0 0 0.423529 0.992157 0.992157 0.364706 0 0.717647 0.992157 0.992157 0.317647 0 0 0 0 0 0 0 0 0 0 0 0.533333 0.992157 0.984314 0.945098 0.603922 0 0 0 0.003922 0.466667 0.992157 0.988235 0.976471 0.992157 0.992157 0.788235 0.007843 0 0 0 0 0 0 0 0 0 0 0 0.686275 0.882353 0.364706 0 0 0 0 0 0 0.098039 0.588235 0.992157 0.992157 0.992157 0.980392 0.305882 0 0 0 0 0 0 0 0 0 0 0 0 0.101961 0.67451 0.321569 0 0 0 0 0 0 0 0.105882 0.733333 0.976471 0.811765 0.713725 0 0 0 0 0 0 0 0 0 0 0 0 0 0.65098 0.992157 0.321569 0 0 0 0 0 0 0 0 0 0.25098 0.007843 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0.94902 0.219608 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.968627 0.764706 0.152941 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.498039 0.25098 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0;
0;0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.298039 0.333333 0.333333 0.333333 0.337255 0.333333 0.333333 0.109804 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.027451 0.223529 0.776471 0.964706 0.988235 0.988235 0.988235 0.992157 0.988235 0.988235 0.780392 0.098039 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.14902 0.698039 0.988235 0.992157 0.988235 0.901961 0.87451 0.568627 0.882353 0.976471 0.988235 0.988235 0.501961 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.188235 0.647059 0.988235 0.988235 0.745098 0.439216 0.098039 0 0 0 0.572549 0.988235 0.988235 0.988235 0 0 0 0 0 0 0 0 0 0 0 0 0 0.2 0.933333 0.992157 0.941176 0.247059 0 0 0 0 0 0 0.188235 0.898039 0.992157 0.992157 0 0 0 0 0 0 0 0 0 0 0 0.039216 0.639216 0.933333 0.988235 0.913725 0.278431 0 0 0 0 0 0 0 0.113725 0.843137 0.988235 0.988235 0 0 0 0 0 0 0 0 0 0 0 0.235294 0.988235 0.992157 0.988235 0.815686 0.07451 0 0 0 0 0 0 0 0.333333 0.988235 0.988235 0.552941 0 0 0 0 0 0 0 0 0 0 0.211765 0.878431 0.988235 0.992157 0.701961 0.329412 0.109804 0 0 0 0 0 0 0 0.698039 0.988235 0.913725 0.145098 0 0 0 0 0 0 0 0 0 0.188235 0.890196 0.988235 0.988235 0.745098 0.047059 0 0 0 0 0 0 0 0 0 0.882353 0.988235 0.568627 0 0 0 0 0 0 0 0 0 0.2 0.933333 0.992157 0.992157 0.992157 0.447059 0.294118 0 0 0 0 0 0 0 0 0.447059 0.992157 0.768627 0 0 0 0 0 0 0 0 0 0 0.623529 0.988235 0.988235 0.988235 0.988235 0.992157 0.47451 0 0 0 0 0 0 0 0.188235 0.933333 0.87451 0.509804 0 0 0 0 0 0 0 0 0 0 0.992157 0.988235 0.937255 0.792157 0.988235 0.894118 0.082353 0 0 0 0 0 0 0.027451 0.647059 0.992157 0.654902 0 0 0 0 0 0 0 0 0 0 0 0.623529 0.988235 0.913725 0.329412 0.376471 0.184314 0 0 0 0 0 0 0.027451 0.513725 0.988235 0.635294 0.219608 0 0 0 0 0 0 0 0 0 0 0 0.196078 0.929412 0.988235 0.988235 0.741176 0.309804 0 0 0 0 0 0 0.529412 0.988235 0.678431 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.223529 0.992157 0.992157 1 0.992157 0.992157 0.992157 0.992157 1 0.992157 0.992157 0.882353 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.023529 0.478431 0.654902 0.658824 0.952941 0.988235 0.988235 0.988235 0.992157 0.988235 0.729412 0.278431 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.196078 0.647059 0.764706 0.764706 0.768627 0.580392 0.047059 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0;
4;0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.180392 0.470588 0.623529 0.623529 0.623529 0.588235 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.243137 0.494118 0.862745 0.870588 0.960784 0.996078 0.996078 0.996078 0.996078 0.992157 0.466667 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.317647 0.639216 0.639216 0.639216 0.639216 0.639216 0.470588 0.262745 0.333333 0.929412 0.694118 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.811765 0.694118 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.811765 0.694118 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.811765 0.694118 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.184314 0.992157 0.694118 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.192157 0.996078 0.384314 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.454902 0.980392 0.219608 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.564706 0.941176 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.588235 0.776471 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.945098 0.560784 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.054902 0.952941 0.356863 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.337255 0.917647 0.109804 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.698039 0.701961 0.019608 0.4 0.662745 0.662745 0.662745 0.662745 0.662745 0.662745 0.662745 0.376471 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.090196 0.639216 0.972549 0.945098 0.913725 0.996078 0.996078 0.996078 0.996078 1 0.996078 0.996078 1 0.996078 0 0 0 0 0 0 0 0 0 0 0.007843 0.105882 0.717647 0.776471 0.905882 0.996078 0.996078 0.988235 0.980392 0.862745 0.537255 0.223529 0.223529 0.368627 0.376471 0.6 0.6 0.6 0 0 0 0 0 0 0 0 0.262745 0.470588 0.6 0.996078 0.996078 0.996078 0.996078 0.847059 0.356863 0.156863 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.909804 0.705882 0.823529 0.635294 0.490196 0.219608 0.113725 0.062745 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.152941 0.152941 0.156863 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0;
doc/api/v1/data_provider/src/sentimental_config.py 已删除 100644 → 0
浏览文件 @ 06db7038
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle.trainer_config_helpers import *
dictionary = dict()
... # read dictionary from outside
define_py_data_sources2(
train_list='train.list',
test_list=None,
module='sentimental_provider',
obj='process',
# above codes same as mnist sample.
args={ # pass to provider.
'dictionary': dictionary
})
doc/api/v1/data_provider/src/sentimental_provider.py 已删除 100644 → 0
浏览文件 @ 06db7038
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle.trainer.PyDataProvider2 import *
def on_init(settings, dictionary, **kwargs):
# on_init will invoke when data provider is initialized. The dictionary
# is passed from trainer_config, and is a dict object with type
# (word string => word id).
# set input types in runtime. It will do the same thing as
# @provider(input_types) will do, but it is set dynamically during runtime.
settings.input_types = {
# The text is a sequence of integer values, and each value is a word id.
# The whole sequence is the sentences that we want to predict its
# sentimental.
'data': integer_value_sequence(len(dictionary)), # text input
'label': integer_value(2) # label positive/negative
}
# save dictionary as settings.dictionary.
# It will be used in process method.
settings.dictionary = dictionary
@provider(init_hook=on_init)
def process(settings, filename):
f = open(filename, 'r')
for line in f: # read each line of file
label, sentence = line.split('\t') # get label and sentence
words = sentence.split(' ') # get words
# convert word string to word id
# the word not in dictionary will be ignored.
word_ids = []
for each_word in words:
if each_word in settings.dictionary:
word_ids.append(settings.dictionary[each_word])
# give data to paddle.
yield word_ids, int(label)
f.close()
doc/api/v1/data_provider/src/sentimental_train.txt 已删除 100644 → 0
浏览文件 @ 06db7038
0 I saw this movie at the AFI Dallas festival . It all takes place at a lake house and it looks wonderful .
1 This documentary makes you travel all around the globe . It contains rare and stunning sequels from the wilderness .
...
doc/api/v1/index_cn.rst 已删除 100644 → 0
浏览文件 @ 06db7038
API中文手册
============
DataProvider API
----------------
.. toctree::
:maxdepth: 1
data_provider/dataprovider_cn.rst
data_provider/pydataprovider2_cn.rst
.. _api_trainer_config:
Model Config API
----------------
.. toctree::
:maxdepth: 1
trainer_config_helpers/optimizers.rst
trainer_config_helpers/data_sources.rst
trainer_config_helpers/layers.rst
trainer_config_helpers/activations.rst
trainer_config_helpers/poolings.rst
trainer_config_helpers/networks.rst
trainer_config_helpers/evaluators.rst
trainer_config_helpers/attrs.rst
Applications API
----------------
.. toctree::
:maxdepth: 1
predict/swig_py_paddle_cn.rst
doc/api/v1/index_en.rst 已删除 100644 → 0
浏览文件 @ 06db7038
API
===
DataProvider API
----------------
.. toctree::
:maxdepth: 1
data_provider/dataprovider_en.rst
data_provider/pydataprovider2_en.rst
.. _api_trainer_config:
Model Config API
----------------
.. toctree::
:maxdepth: 1
trainer_config_helpers/optimizers.rst
trainer_config_helpers/data_sources.rst
trainer_config_helpers/layers.rst
trainer_config_helpers/activations.rst
trainer_config_helpers/poolings.rst
trainer_config_helpers/networks.rst
trainer_config_helpers/evaluators.rst
trainer_config_helpers/attrs.rst
Applications API
----------------
.. toctree::
:maxdepth: 1
predict/swig_py_paddle_en.rst
doc/api/v1/predict/src/predict_sample.py 已删除 100644 → 0
浏览文件 @ 06db7038
# 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.
from py_paddle import swig_paddle, DataProviderConverter
from paddle.trainer.PyDataProvider2 import dense_vector
from paddle.trainer.config_parser import parse_config
TEST_DATA = [[[
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.215686, 0.533333, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.67451, 0.992157, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0.070588, 0.886275, 0.992157, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.192157,
0.070588, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.670588, 0.992157,
0.992157, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.117647, 0.933333, 0.858824, 0.313725,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.090196, 0.858824, 0.992157, 0.831373, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0.141176, 0.992157, 0.992157, 0.611765, 0.054902, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0.258824, 0.992157, 0.992157, 0.529412, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0.368627, 0.992157, 0.992157, 0.419608, 0.003922, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0.094118, 0.835294, 0.992157, 0.992157, 0.517647, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0.603922, 0.992157, 0.992157, 0.992157, 0.603922,
0.545098, 0.043137, 0, 0, 0, 0, 0, 0, 0, 0.447059, 0.992157, 0.992157,
0.956863, 0.062745, 0, 0, 0, 0, 0, 0, 0, 0, 0.011765, 0.666667, 0.992157,
0.992157, 0.992157, 0.992157, 0.992157, 0.745098, 0.137255, 0, 0, 0, 0, 0,
0.152941, 0.866667, 0.992157, 0.992157, 0.521569, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0.070588, 0.992157, 0.992157, 0.992157, 0.803922, 0.352941, 0.745098,
0.992157, 0.945098, 0.317647, 0, 0, 0, 0, 0.580392, 0.992157, 0.992157,
0.764706, 0.043137, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.070588, 0.992157, 0.992157,
0.776471, 0.043137, 0, 0.007843, 0.27451, 0.882353, 0.941176, 0.176471, 0,
0, 0.180392, 0.898039, 0.992157, 0.992157, 0.313725, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0.070588, 0.992157, 0.992157, 0.713725, 0, 0, 0, 0, 0.627451,
0.992157, 0.729412, 0.062745, 0, 0.509804, 0.992157, 0.992157, 0.776471,
0.035294, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.494118, 0.992157, 0.992157,
0.968627, 0.168627, 0, 0, 0, 0.423529, 0.992157, 0.992157, 0.364706, 0,
0.717647, 0.992157, 0.992157, 0.317647, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0.533333, 0.992157, 0.984314, 0.945098, 0.603922, 0, 0, 0, 0.003922,
0.466667, 0.992157, 0.988235, 0.976471, 0.992157, 0.992157, 0.788235,
0.007843, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.686275, 0.882353, 0.364706, 0,
0, 0, 0, 0, 0, 0.098039, 0.588235, 0.992157, 0.992157, 0.992157, 0.980392,
0.305882, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.101961, 0.67451, 0.321569,
0, 0, 0, 0, 0, 0, 0, 0.105882, 0.733333, 0.976471, 0.811765, 0.713725, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.65098, 0.992157, 0.321569, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0.25098, 0.007843, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1,
0.94902, 0.219608, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0.968627, 0.764706, 0.152941, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.498039, 0.25098, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0
]], [[
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0.298039, 0.333333, 0.333333, 0.333333, 0.337255,
0.333333, 0.333333, 0.109804, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0.027451, 0.223529, 0.776471, 0.964706, 0.988235, 0.988235, 0.988235,
0.992157, 0.988235, 0.988235, 0.780392, 0.098039, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0.14902, 0.698039, 0.988235, 0.992157, 0.988235, 0.901961,
0.87451, 0.568627, 0.882353, 0.976471, 0.988235, 0.988235, 0.501961, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.188235, 0.647059, 0.988235, 0.988235,
0.745098, 0.439216, 0.098039, 0, 0, 0, 0.572549, 0.988235, 0.988235,
0.988235, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.2, 0.933333, 0.992157,
0.941176, 0.247059, 0, 0, 0, 0, 0, 0, 0.188235, 0.898039, 0.992157,
0.992157, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.039216, 0.639216, 0.933333,
0.988235, 0.913725, 0.278431, 0, 0, 0, 0, 0, 0, 0, 0.113725, 0.843137,
0.988235, 0.988235, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.235294, 0.988235,
0.992157, 0.988235, 0.815686, 0.07451, 0, 0, 0, 0, 0, 0, 0, 0.333333,
0.988235, 0.988235, 0.552941, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.211765,
0.878431, 0.988235, 0.992157, 0.701961, 0.329412, 0.109804, 0, 0, 0, 0, 0,
0, 0, 0.698039, 0.988235, 0.913725, 0.145098, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0.188235, 0.890196, 0.988235, 0.988235, 0.745098, 0.047059, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0.882353, 0.988235, 0.568627, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.2,
0.933333, 0.992157, 0.992157, 0.992157, 0.447059, 0.294118, 0, 0, 0, 0, 0,
0, 0, 0, 0.447059, 0.992157, 0.768627, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0.623529, 0.988235, 0.988235, 0.988235, 0.988235, 0.992157, 0.47451, 0, 0,
0, 0, 0, 0, 0, 0.188235, 0.933333, 0.87451, 0.509804, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0.992157, 0.988235, 0.937255, 0.792157, 0.988235, 0.894118,
0.082353, 0, 0, 0, 0, 0, 0, 0.027451, 0.647059, 0.992157, 0.654902, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0.623529, 0.988235, 0.913725, 0.329412, 0.376471,
0.184314, 0, 0, 0, 0, 0, 0, 0.027451, 0.513725, 0.988235, 0.635294,
0.219608, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.196078, 0.929412, 0.988235,
0.988235, 0.741176, 0.309804, 0, 0, 0, 0, 0, 0, 0.529412, 0.988235,
0.678431, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.223529, 0.992157,
0.992157, 1, 0.992157, 0.992157, 0.992157, 0.992157, 1, 0.992157, 0.992157,
0.882353, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.023529,
0.478431, 0.654902, 0.658824, 0.952941, 0.988235, 0.988235, 0.988235,
0.992157, 0.988235, 0.729412, 0.278431, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0.196078, 0.647059, 0.764706, 0.764706, 0.768627,
0.580392, 0.047059, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0
]]]
def main():
conf = parse_config("./mnist_model/trainer_config.py", "")
print conf.data_config.load_data_args
network = swig_paddle.GradientMachine.createFromConfigProto(
conf.model_config)
assert isinstance(network, swig_paddle.GradientMachine) # For code hint.
network.loadParameters("./mnist_model/")
converter = DataProviderConverter([dense_vector(784)])
inArg = converter(TEST_DATA)
print network.forwardTest(inArg)
if __name__ == '__main__':
swig_paddle.initPaddle("--use_gpu=0")
main()
doc/api/v1/predict/swig_py_paddle_cn.rst 已删除 100644 → 0
浏览文件 @ 06db7038
.. _api_swig_py_paddle:
基于Python的预测
================
预测流程
--------
PaddlePaddle使用swig对常用的预测接口进行了封装，通过编译会生成py_paddle软件包，安装该软件包就可以在python环境下实现模型预测。可以使用python的 ``help()`` 函数查询软件包相关API说明。
基于Python的模型预测，主要包括以下五个步骤。
1. 初始化PaddlePaddle环境
在程序开始阶段，通过调用 ``swig_paddle.initPaddle()`` 并传入相应的命令行参数初始化PaddlePaddle。
2. 解析模型配置文件
初始化之后，可以通过调用 ``parse_config()`` 解析训练模型时用的配置文件。注意预测数据通常不包含label, 同时预测网络通常直接输出最后一层的结果而不是像训练网络一样再接一层cost layer，所以一般需要对训练用的模型配置文件稍作相应修改才能在预测时使用。
3. 构造paddle.GradientMachine
通过调用 ``swig_paddle.GradientMachine.createFromConfigproto()`` 传入上一步解析出来的模型配置就可以创建一个 ``GradientMachine``。
4. 准备预测数据
swig_paddle中的预测接口的参数是自定义的C++数据类型，py_paddle里面提供了一个工具类 ``DataProviderConverter`` 可以用于接收和PyDataProvider2一样的输入数据并转换成预测接口所需的数据类型。
5. 模型预测
通过调用 ``forwardTest()`` 传入预测数据，直接返回计算结果。
预测Demo
--------
如下是一段使用mnist model来实现手写识别的预测代码。完整的代码见 ``src_root/doc/ui/predict/predict_sample.py`` 。mnist model可以通过 ``src_root\demo\mnist`` 目录下的demo训练出来。
.. literalinclude:: src/predict_sample.py
:language: python
:lines: 15-18,121-136
Demo预测输出如下，其中value即为softmax层的输出。由于TEST_DATA包含两条预测数据，所以输出的value包含两个向量 。
.. code-block:: text
[{'id': None, 'value': array(
[[ 5.53018653e-09, 1.12194102e-05, 1.96644767e-09,
1.43630644e-02, 1.51111044e-13, 9.85625684e-01,
2.08823112e-10, 2.32777140e-08, 2.00186201e-09,
1.15501715e-08],
[ 9.99982715e-01, 1.27787406e-10, 1.72296313e-05,
1.49316648e-09, 1.36540484e-11, 6.93137714e-10,
2.70634608e-08, 3.48565123e-08, 5.25639710e-09,
4.48684503e-08]], dtype=float32)}]
doc/api/v1/predict/swig_py_paddle_en.rst 已删除 100644 → 0
浏览文件 @ 06db7038
Python Prediction
==================
PaddlePaddle offers a set of clean prediction interfaces for python with the help of
SWIG. The main steps of predict values in python are:
* Parse training configurations
* Construct GradientMachine
* Prepare data
* Predict
Here is a sample python script that shows the typical prediction process for the
MNIST classification problem. A complete sample code could be found at
:code:`src_root/doc/ui/predict/predict_sample.py`.
.. literalinclude:: src/predict_sample.py
:language: python
:lines: 15-18,90-100,101-104
The module that does the most of the job is py_paddle.swig_paddle, it's
generated by SWIG and has complete documents, for more details you can use
python's :code:`help()` function. Let's walk through the above python script:
* At the beginning, use :code:`swig_paddle.initPaddle()` to initialize
PaddlePaddle with command line arguments, for more about command line arguments
see :ref:`cmd_detail_introduction` .
* Parse the configuration file that is used in training with :code:`parse_config()`.
Because data to predict with always have no label, and output of prediction work
normally is the output layer rather than the cost layer, so you should modify
the configuration file accordingly before using it in the prediction work.
* Create a neural network with
:code:`swig_paddle.GradientMachine.createFromConfigproto()`, which takes the
parsed configuration :code:`conf.model_config` as argument. Then load the
trained parameters from the model with :code:`network.loadParameters()`.
* Create a data converter object of utility class :code:`DataProviderConverter`.
- Note: As swig_paddle can only accept C++ matrices, we offer a utility
class DataProviderConverter that can accept the same input data with
PyDataProvider2, for more information please refer to document
of :ref:`api_pydataprovider2` .
* Do the prediction with :code:`forwardTest()`, which takes the converted
input data and outputs the activations of the output layer.
Here is a typical output:
.. code-block:: text
[{'id': None, 'value': array([[ 5.53018653e-09, 1.12194102e-05, 1.96644767e-09,
1.43630644e-02, 1.51111044e-13, 9.85625684e-01,
2.08823112e-10, 2.32777140e-08, 2.00186201e-09,
1.15501715e-08],
[ 9.99982715e-01, 1.27787406e-10, 1.72296313e-05,
1.49316648e-09, 1.36540484e-11, 6.93137714e-10,
2.70634608e-08, 3.48565123e-08, 5.25639710e-09,
4.48684503e-08]], dtype=float32)}]
:code:`value` is the output of the output layer, each row represents result of
the corresponding row in the input data, each element represents activation of
the corresponding neuron in the output layer.
doc/api/v2/fluid/layers.rst
浏览文件 @ aece2905
......@@ -500,6 +500,16 @@ swish
.. autofunction:: paddle.v2.fluid.layers.swish
:noindex:
edit_distance
---------------
.. autofunction:: paddle.v2.fluid.layers.edit_distance_error
:noindex:
ctc_greedy_decoder
---------------
.. autofunction:: paddle.v2.fluid.layers.ctc_greedy_decoder
:noindex:
l2_normalize
------------
.. autofunction:: paddle.v2.fluid.layers.l2_normalize
......
doc/faq/local/index_cn.rst
浏览文件 @ aece2905
......@@ -211,3 +211,49 @@ decoder_inputs = paddle.layer.fc(
* list 中元素的个数等于网络中输出层的个数；
* list 中每个元素是一个layer的输出结果矩阵，类型是numpy的ndarray；
* 每一个layer输出矩阵的高度，在非序列输入时：等于样本数；序列输入时等于：输入序列中元素的总数；宽度等于配置中layer的size；
6. 如何在训练过程中获得某一个layer的output
-----------------------------------------------
可以在event_handler中，通过 :code:`event.gm.getLayerOutputs("layer_name")` 获得在模型配置中某一层的name :code:`layer_name` 在当前
mini-batch forward的output的值。获得的值类型均为 :code:`numpy.ndarray` ，可以通过这个输出来完成自定义的评估指标计算等功能。例如下面代码：
.. code-block:: python
def score_diff(right_score, left_score):
return np.average(np.abs(right_score - left_score))
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 25 == 0:
diff = score_diff(
event.gm.getLayerOutputs("right_score")["right_score"][
"value"],
event.gm.getLayerOutputs("left_score")["left_score"][
"value"])
logger.info(("Pass %d Batch %d : Cost %.6f, "
"average absolute diff scores: %.6f") %
(event.pass_id, event.batch_id, event.cost, diff))
注意：此方法不能获取 :code:`paddle.layer.recurrent_group` 里step的内容，但可以获取 :code:`paddle.layer.recurrent_group` 的输出。
7. 如何在训练过程中获得参数的权重和梯度
-----------------------------------------------
在某些情况下，获得当前mini-batch的权重（或称作weights, parameters）有助于在训练时观察具体数值，方便排查以及快速定位问题。
可以通过在 :code:`event_handler` 中打印其值（注意，需要使用 :code:`paddle.event.EndForwardBackward` 保证使用GPU训练时也可以获得），
示例代码如下：
.. code-block:: python
...
parameters = paddle.parameters.create(cost)
...
def event_handler(event):
if isinstance(event, paddle.event.EndForwardBackward):
if event.batch_id % 25 == 0:
for p in parameters.keys():
logger.info("Param %s, Grad %s",
parameters.get(p), parameters.get_grad(p))
注意：“在训练过程中获得某一个layer的output”和“在训练过程中获得参数的权重和梯度”都会造成训练中的数据从C++拷贝到numpy，会对训练性能造成影响。不要在注重性能的训练场景下使用。
\ No newline at end of file
doc/getstarted/build_and_install/docker_install_cn.rst
浏览文件 @ aece2905
......@@ -25,14 +25,14 @@
.. code-block:: bash
docker pull docker.paddlepaddle.org/paddle
docker pull docker.paddlepaddlehub.com/paddle
下载GPU版本（cuda8.0_cudnn5_avx_mkl）的Docker镜像：
.. code-block:: bash
docker pull paddlepaddle/paddle:latest-gpu
docker pull docker.paddlepaddle.org/paddle:latest-gpu
docker pull docker.paddlepaddlehub.com/paddle:latest-gpu
选择下载使用不同的BLAS库的Docker镜像：
......@@ -49,7 +49,7 @@
docker pull paddlepaddle/paddle:[tag]
# 比如：
docker pull docker.paddlepaddle.org/paddle:0.10.0-gpu
docker pull docker.paddlepaddlehub.com/paddle:0.11.0-gpu
.. _docker_run:
......
doc/getstarted/build_and_install/docker_install_en.rst
浏览文件 @ aece2905
......@@ -26,14 +26,14 @@ For users in China, we provide a faster mirror:
.. code-block:: bash
docker pull docker.paddlepaddle.org/paddle
docker pull docker.paddlepaddlehub.com/paddle
Download GPU version (cuda8.0_cudnn5_avx_mkl) images:
.. code-block:: bash
docker pull paddlepaddle/paddle:latest-gpu
docker pull docker.paddlepaddle.org/paddle:latest-gpu
docker pull docker.paddlepaddlehub.com/paddle:latest-gpu
Choose between different BLAS version:
......@@ -53,7 +53,7 @@ and run:
docker pull paddlepaddle/paddle:[tag]
# i.e.
docker pull docker.paddlepaddle.org/paddle:0.10.0-gpu
docker pull docker.paddlepaddlehub.com/paddle:0.11.0-gpu
.. _docker_run:
......
doc/howto/usage/cluster/fluid_cluster_train_en.md
浏览文件 @ aece2905
......@@ -2,27 +2,27 @@
## Introduction
In this article, we'll explain how to config and run distributed training jobs with PaddlePaddle Fluid in a bare metal cluster.
In this article, we'll explain how to configure and run distributed training jobs with PaddlePaddle Fluid in a bare metal cluster.
## Preparations
### Get your cluster ready
### Getting the cluster ready
Prepare your computer nodes in the cluster. Nodes in this cluster can be of any specification that runs PaddlePaddle, and with a unique IP address assigned to it. Make sure they can communicate with each other.
Prepare the compute nodes in the cluster. Nodes in this cluster can be of any specification that runs PaddlePaddle, and with a unique IP address assigned to it. Make sure they can communicate to each other.
### Have PaddlePaddle installed
PaddlePaddle must be installed on all nodes. If you have GPU cards on your nodes, be sure to properly install drivers and CUDA libraries.
PaddlePaddle build and installation guide can be found from [here](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/index_en.html).
PaddlePaddle build and installation guide can be found [here](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/index_en.html).
### Update training script
### Update the training script
#### Non-cluster training script
Let's take [Deep Learning 101](http://www.paddlepaddle.org/docs/develop/book/01.fit_a_line/index.html)'s first chapter: "fit a line" as an example.
This demo's non-cluster version with fluid API is as follows:
The non-cluster version of this demo with fluid API is as follows:
``` python
import paddle.v2 as paddle
......@@ -65,25 +65,25 @@ for pass_id in range(PASS_NUM):
exit(1)
```
We created a simple fully connected neural networks training program and handed it to the fluid executor to run for 100 passes.
We created a simple fully-connected neural network training program and handed it to the fluid executor to run for 100 passes.
Now let's try to convert it to a distributed version to run in a cluster.
Now let's try to convert it to a distributed version to run on a cluster.
#### Introducing parameter server
As you see from the non-cluster version of training script, there is only one role in it: the trainer, who does the computing as well as holding parameters. In cluster training, since multi-trainers are working on the same task, they need one centralized place to hold and distribute parameters. This centralized place is called the Parameter Server in PaddlePaddle.
As we can see from the non-cluster version of training script, there is only one role in the script: the trainer, that performs the computing as well as holds the parameters. In cluster training, since multi-trainers are working on the same task, they need one centralized place to hold and distribute parameters. This centralized place is called the Parameter Server in PaddlePaddle.
![parameter server architect](src/trainer.png)
![parameter server architecture](src/trainer.png)
Parameter Server in fluid does not only hold parameters but is also assigned with a part of the program. Trainers communicate with parameter servers via send/receive OPs. For more tech detail, please refer to this [document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/dist_refactor/distributed_architecture.md).
Parameter Server in fluid not only holds the parameters but is also assigned with a part of the program. Trainers communicate with parameter servers via send/receive OPs. For more technical details, please refer to [this document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/dist_refactor/distributed_architecture.md).
Now we need to create program for both trainers and parameter servers, the question is how?
Now we need to create programs for both: trainers and parameter servers, the question is how?
#### Slice the program
Fluid provides a tool called "Distribute Transpiler" to automatically convert the non-cluster program into cluster program.
Fluid provides a tool called "Distributed Transpiler" that automatically converts the non-cluster program into cluster program.
The idea behind this tool is to find optimize OPs and gradient parameters, slice the program into 2 pieces and connect them with send/receive OP.
The idea behind this tool is to find the optimize OPs and gradient parameters, slice the program into 2 pieces and connect them with send/receive OP.
Optimize OPs and gradient parameters can be found from the return values of optimizer's minimize function.
......@@ -94,7 +94,7 @@ To put them together:
optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) #get optimize OPs and gradient parameters
t = fluid.DistributeTranspiler() # create transpiler instance
t = fluid.DistributeTranspiler() # create the transpiler instance
# slice the program into 2 pieces with optimizer_ops and gradient parameters list, as well as pserver_endpoints, which is a comma separated list of [IP:PORT] and number of trainers
t.transpile(optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2)
......@@ -119,7 +119,7 @@ for pass_id in range(100):
### E2E demo
Please find the complete demo from [here](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/tests/book_distribute/notest_dist_fit_a_line.py). In parameter server node run this in the command line:
Please find the complete demo from [here](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/tests/book_distribute/notest_dist_fit_a_line.py). In parameter server node run the following in the command line:
``` bash
PSERVERS=192.168.1.2:6174 SERVER_ENDPOINT=192.168.1.2:6174 TRAINING_ROLE=PSERVER python notest_dist_fit_a_line.py
......@@ -129,12 +129,12 @@ PSERVERS=192.168.1.2:6174 SERVER_ENDPOINT=192.168.1.2:6174 TRAINING_ROLE=PSERVER
Wait until the prompt `Server listening on 192.168.1.2:6174`
Then in 2 of your trainer node run this:
Then in 2 of your trainer nodes run this:
``` bash
PSERVERS=192.168.1.2:6174 SERVER_ENDPOINT=192.168.1.2:6174 TRAINING_ROLE=TRAINER python notest_dist_fit_a_line.py
```
*the reason you need to run this command twice in 2 nodes is: in the script we set the trainer count to be 2. You can change this setting on line 50*
*the reason you need to run this command twice in 2 nodes is because: in the script we set the trainer count to be 2. You can change this setting on line 50*
Now you have 2 trainers and 1 parameter server up and running.
paddle/framework/CMakeLists.txt
浏览文件 @ aece2905
......@@ -88,3 +88,10 @@ 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)
if(NOT WITH_C_API AND WITH_FLUID)
file(GLOB FRAMEWORK_HEADERS *.h)
install(FILES ${FRAMEWORK_HEADERS} DESTINATION include/paddle/framework)
install(FILES ${CMAKE_CURRENT_BINARY_DIR}/framework.pb.h DESTINATION include/paddle/framework)
install(FILES details/cow_ptr.h details/op_registry.h DESTINATION include/paddle/framework/details)
endif()
paddle/framework/variable_test.cc
浏览文件 @ aece2905
......@@ -12,19 +12,6 @@
// See the License for the specific language governing permissions and
// limitations under the License.
/*
Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
*/
#include <memory>
#include <string>
......
paddle/inference/CMakeLists.txt
浏览文件 @ aece2905
set(FLUID_CORE_MODULES
backward proto_desc paddle_memory executor prune init ${GLOB_OP_LIB})
set(FLUID_CORE_MODULES proto_desc paddle_memory executor prune init)
cc_library(paddle_fluid_api
SRCS inference.cc
DEPS ${FLUID_CORE_MODULES})
DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
# Merge all modules into a simgle static library
cc_library(paddle_fluid DEPS paddle_fluid_api ${FLUID_CORE_MODULES})
# Merge all modules into a single static library
cc_library(paddle_fluid DEPS paddle_fluid_api ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
# Create shared library
add_library(paddle_fluid_shared SHARED inference.cc)
target_circle_link_libraries(paddle_fluid_shared
ARCHIVE_START
${GLOB_OP_LIB}
ARCHIVE_END
${FLUID_CORE_MODULES})
SET_TARGET_PROPERTIES(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)
# install library & headers
if(NOT WITH_C_API AND WITH_FLUID)
install(FILES inference.h DESTINATION include/paddle/inference)
install(TARGETS paddle_fluid_shared DESTINATION lib)
endif()
add_executable(example example.cc)
if(APPLE)
......
paddle/memory/CMakeLists.txt
浏览文件 @ aece2905
......@@ -14,3 +14,10 @@ cc_library(paddle_memory
system_allocator)
cc_test(memory_test SRCS memory_test.cc DEPS place paddle_memory)
if(NOT WITH_C_API AND WITH_FLUID)
file(GLOB MEMORY_HEADERS *.h)
file(GLOB MEMORY_DETAIL_HEADERS detail/*.h)
install(FILES ${MEMORY_HEADERS} DESTINATION include/paddle/memory)
install(FILES ${MEMORY_DETAIL_HEADERS} DESTINATION include/paddle/memory/detail)
endif()
paddle/operators/CMakeLists.txt
浏览文件 @ aece2905
......@@ -155,8 +155,12 @@ op_library(parallel_do_op DEPS executor)
# Regist multiple Kernel to pybind
if (WITH_GPU)
op_library(conv_op SRCS conv_op.cc conv_op.cu.cc conv_cudnn_op.cu.cc DEPS
vol2col depthwise_conv)
# op_library(conv_op SRCS conv_op.cc conv_op.cu.cc conv_cudnn_op.cu.cc DEPS vol2col)
op_library(edit_distance_op SRCS edit_distance_op.cc edit_distance_op.cu DEPS math_function)
op_library(pool_op SRCS pool_op.cc pool_op.cu.cc pool_cudnn_op.cu.cc DEPS pooling)
op_library(conv_transpose_op SRCS conv_transpose_op.cc conv_transpose_op.cu.cc
conv_transpose_cudnn_op.cu.cc DEPS vol2col)
......
paddle/operators/bipartite_match_op.cc 0 → 100644
浏览文件 @ aece2905
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/framework/op_registry.h"
#include "paddle/operators/math/math_function.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
using LoDTensor = framework::LoDTensor;
class BipartiteMatchOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("DistMat"),
"Input(DistMat) of BipartiteMatch should not be null.");
auto dims = ctx->GetInputDim("DistMat");
PADDLE_ENFORCE_EQ(dims.size(), 2, "The rank of Input(DistMat) must be 2.");
ctx->SetOutputDim("ColToRowMatchIndices", dims);
ctx->SetOutputDim("ColToRowMatchDis", dims);
}
};
template <typename T>
class BipartiteMatchKernel : public framework::OpKernel<T> {
public:
// The match_indices must be initialized to -1 at first.
// The match_dist must be initialized to 0 at first.
void BipartiteMatch(const Tensor& dist, int* match_indices,
T* match_dist) const {
constexpr T kEPS = static_cast<T>(1e-6);
PADDLE_ENFORCE_EQ(dist.dims().size(), 2, "The rank of dist must be 2.");
int64_t row = dist.dims()[0];
int64_t col = dist.dims()[1];
auto* dist_data = dist.data<T>();
std::vector<int> row_pool;
for (int i = 0; i < row; ++i) {
row_pool.push_back(i);
}
while (row_pool.size() > 0) {
int max_idx = -1;
int max_row_idx = -1;
T max_dist = -1;
for (int64_t j = 0; j < col; ++j) {
if (match_indices[j] != -1) {
continue;
}
for (size_t k = 0; k < row_pool.size(); ++k) {
int m = row_pool[k];
// distance is 0 between m-th row and j-th column
if (dist_data[m * col + j] < kEPS) {
continue;
}
if (dist_data[m * col + j] > max_dist) {
max_idx = j;
max_row_idx = m;
max_dist = dist_data[m * col + j];
}
}
}
if (max_idx == -1) {
// Cannot find good match.
break;
} else {
PADDLE_ENFORCE_EQ(match_indices[max_idx], -1);
match_indices[max_idx] = max_row_idx;
match_dist[max_idx] = max_dist;
// Erase the row index.
row_pool.erase(
std::find(row_pool.begin(), row_pool.end(), max_row_idx));
}
}
}
void Compute(const framework::ExecutionContext& context) const override {
auto* dist_mat = context.Input<LoDTensor>("DistMat");
auto* match_indices = context.Output<Tensor>("ColToRowMatchIndices");
auto* match_dist = context.Output<Tensor>("ColToRowMatchDis");
auto& dev_ctx = context.device_context<platform::CPUDeviceContext>();
auto col = dist_mat->dims()[1];
int64_t n = dist_mat->lod().size() == 0UL
? 1
: static_cast<int64_t>(dist_mat->lod().back().size() - 1);
if (dist_mat->lod().size()) {
PADDLE_ENFORCE_EQ(dist_mat->lod().size(), 1UL,
"Only support 1 level of LoD.");
}
match_indices->mutable_data<int>({n, col}, context.GetPlace());
match_dist->mutable_data<T>({n, col}, context.GetPlace());
math::SetConstant<platform::CPUDeviceContext, int> iset;
iset(dev_ctx, match_indices, static_cast<int>(-1));
math::SetConstant<platform::CPUDeviceContext, T> tset;
tset(dev_ctx, match_dist, static_cast<T>(0));
int* indices = match_indices->data<int>();
T* dist = match_dist->data<T>();
if (n == 1) {
BipartiteMatch(*dist_mat, indices, dist);
} else {
auto lod = dist_mat->lod().back();
for (size_t i = 0; i < lod.size() - 1; ++i) {
Tensor one_ins = dist_mat->Slice(lod[i], lod[i + 1]);
BipartiteMatch(one_ins, indices + i * col, dist + i * col);
}
}
}
};
class BipartiteMatchOpMaker : public framework::OpProtoAndCheckerMaker {
public:
BipartiteMatchOpMaker(OpProto* proto, OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput(
"DistMat",
"(LoDTensor or Tensor) this input is a 2-D LoDTensor with shape "
"[K, M]. It is pair-wise distance matrix between the entities "
"represented by each row and each column. For example, assumed one "
"entity is A with shape [K], another entity is B with shape [M]. The "
"DistMat[i][j] is the distance between A[i] and B[j]. The bigger "
"the distance is, the better macthing the pairs are. Please note, "
"This tensor can contain LoD information to represent a batch of "
"inputs. One instance of this batch can contain different numbers of "
"entities.");
AddOutput("ColToRowMatchIndices",
"(Tensor) A 2-D Tensor with shape [N, M] in int type. "
"N is the batch size. If ColToRowMatchIndices[i][j] is -1, it "
"means B[j] does not match any entity in i-th instance. "
"Otherwise, it means B[j] is matched to row "
"ColToRowMatchIndices[i][j] in i-th instance. The row number of "
"i-th instance is saved in ColToRowMatchIndices[i][j].");
AddOutput("ColToRowMatchDis",
"(Tensor) A 2-D Tensor with shape [N, M] in float type. "
"N is batch size. If ColToRowMatchIndices[i][j] is -1, "
"ColToRowMatchDis[i][j] is also -1.0. Otherwise, assumed "
"ColToRowMatchIndices[i][j] = d, and the row offsets of each "
"instance are called LoD. Then "
"ColToRowMatchDis[i][j] = DistMat[d+LoD[i]][j]");
AddComment(R"DOC(
This operator is a greedy bipartite matching algorithm, which is used to
obtain the matching with the maximum distance based on the input
distance matrix. For input 2D matrix, the bipartite matching algorithm can
find the matched column for each row, also can find the matched row for
each column. And this operator only calculate matched indices from column
to row. For each instance, the number of matched indices is the number of
of columns of the input ditance matrix.
There are two outputs to save matched indices and distance.
A simple description, this algothrim matched the best (maximum distance)
row entity to the column entity and the matched indices are not duplicated
in each row of ColToRowMatchIndices. If the column entity is not matched
any row entity, set -1 in ColToRowMatchIndices.
Please note that the input DistMat can be LoDTensor (with LoD) or Tensor.
If LoDTensor with LoD, the height of ColToRowMatchIndices is batch size.
If Tensor, the height of ColToRowMatchIndices is 1.
)DOC");
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(bipartite_match, ops::BipartiteMatchOp,
ops::BipartiteMatchOpMaker,
paddle::framework::EmptyGradOpMaker);
REGISTER_OP_CPU_KERNEL(bipartite_match, ops::BipartiteMatchKernel<float>,
ops::BipartiteMatchKernel<double>);
paddle/operators/conv_op.cc
浏览文件 @ aece2905
......@@ -318,18 +318,22 @@ framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
namespace ops = paddle::operators;
REGISTER_OP(conv2d, ops::ConvOp, ops::Conv2DOpMaker, conv2d_grad,
ops::ConvOpGrad);
REGISTER_OP(depthwiseConv, ops::ConvOp, ops::Conv2DOpMaker, depthwiseConv_grad,
ops::ConvOpGrad);
// depthwise convolution op
REGISTER_OP(depthwise_conv, ops::ConvOp, ops::Conv2DOpMaker,
depthwise_conv_grad, ops::ConvOpGrad);
REGISTER_OP(conv3d, ops::ConvOp, ops::Conv3DOpMaker, conv3d_grad,
ops::ConvOpGrad);
// depthwise conv kernel
// TODO(xingzhaolong): neon kernel for mobile
REGISTER_OP_CPU_KERNEL(
depthwiseConv,
depthwise_conv,
ops::GemmConvKernel<paddle::platform::CPUDeviceContext, float>,
ops::GemmConvKernel<paddle::platform::CPUDeviceContext, double>);
REGISTER_OP_CPU_KERNEL(
depthwiseConv_grad,
depthwise_conv_grad,
ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, float>,
ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, double>);
......
paddle/operators/conv_op.cu.cc
浏览文件 @ aece2905
......@@ -17,10 +17,15 @@ limitations under the License. */
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(
depthwiseConv,
depthwise_conv,
ops::DepthwiseConvKernel<paddle::platform::CUDADeviceContext, float>,
ops::DepthwiseConvKernel<paddle::platform::CUDADeviceContext, double>);
REGISTER_OP_CUDA_KERNEL(
depthwise_conv_grad,
ops::DepthwiseConvGradKernel<paddle::platform::CUDADeviceContext, float>,
ops::DepthwiseConvGradKernel<paddle::platform::CUDADeviceContext, double>);
REGISTER_OP_CUDA_KERNEL(
conv2d, ops::GemmConvKernel<paddle::platform::CUDADeviceContext, float>,
ops::GemmConvKernel<paddle::platform::CUDADeviceContext, double>);
......
paddle/operators/conv_op.h
浏览文件 @ aece2905
......@@ -357,14 +357,10 @@ class DepthwiseConvKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
const Tensor* input = context.Input<Tensor>("Input");
// The filter will be reshaped in the calculations,
// so here use an assignment operation,
// that avoids modifying the variable in the Scope.
Tensor filter = *context.Input<Tensor>("Filter");
Tensor* output = context.Output<Tensor>("Output");
output->mutable_data<T>(context.GetPlace());
std::vector<int> ksize = context.Attr<std::vector<int>>("ksize");
std::vector<int> strides = context.Attr<std::vector<int>>("strides");
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
std::vector<int> dilations = context.Attr<std::vector<int>>("dilations");
......@@ -372,8 +368,50 @@ class DepthwiseConvKernel : public framework::OpKernel<T> {
math::DepthwiseConvFunctor<DeviceContext, T> depthwiseConv;
auto& dev_ctx = context.template device_context<DeviceContext>();
depthwiseConv(dev_ctx, *input, filter, ksize, strides, paddings,
output);
depthwiseConv(dev_ctx, *input, filter, strides, paddings, output);
}
};
template <typename DeviceContext, typename T>
class DepthwiseConvGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
const Tensor* input = context.Input<Tensor>("Input");
const Tensor* output_grad =
context.Input<Tensor>(framework::GradVarName("Output"));
Tensor* input_grad =
context.Output<Tensor>(framework::GradVarName("Input"));
Tensor* filter_grad =
context.Output<Tensor>(framework::GradVarName("Filter"));
Tensor filter = *context.Input<Tensor>("Filter");
if (!input_grad && !filter_grad) return;
std::vector<int> strides = context.Attr<std::vector<int>>("strides");
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
std::vector<int> dilations = context.Attr<std::vector<int>>("dilations");
math::SetConstant<DeviceContext, T> set_zero;
auto& dev_ctx = context.template device_context<DeviceContext>();
math::DepthwiseConvInputGradFunctor<DeviceContext, T>
depthwiseConvInputGrad;
math::DepthwiseConvFilterGradFunctor<DeviceContext, T>
depthwiseConvFilterGrad;
if (input_grad) {
input_grad->mutable_data<T>(context.GetPlace());
set_zero(dev_ctx, input_grad, static_cast<T>(0));
depthwiseConvInputGrad(dev_ctx, *input, filter, *output_grad, strides,
paddings, input_grad);
}
if (filter_grad) {
filter_grad->mutable_data<T>(context.GetPlace());
set_zero(dev_ctx, filter_grad, static_cast<T>(0));
depthwiseConvFilterGrad(dev_ctx, *input, *output_grad, strides, paddings,
filter_grad);
}
}
};
......
paddle/operators/conv_transpose_op.cc
浏览文件 @ aece2905
......@@ -160,8 +160,8 @@ Example:
Output shape: $(N, C_{out}, H_{out}, W_{out})$
Where
$$
H_{out} = (H_{in} - 1) * strides[0] - 2 * paddings[0] + H_f \\
W_{out} = (W_{in} - 1) * strides[1] - 2 * paddings[1] + W_f
H_{out} = (H_{in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (H_f - 1) + 1 \\
W_{out} = (W_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (W_f - 1) + 1
$$
)DOC");
}
......@@ -249,9 +249,9 @@ Example:
Output shape: $(N, C_{out}, D_{out}, H_{out}, W_{out})$
Where
$$
D_{out} = (D_{in} - 1) * strides[0] - 2 * paddings[0] + D_f \\
H_{out} = (H_{in} - 1) * strides[1] - 2 * paddings[1] + H_f \\
W_{out} = (W_{in} - 1) * strides[2] - 2 * paddings[2] + W_f
D_{out} = (D_{in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (D_f - 1) + 1 \\
H_{out} = (H_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (H_f - 1) + 1 \\
W_{out} = (W_{in} - 1) * strides[2] - 2 * paddings[2] + dilations[2] * (W_f - 1) + 1
$$
)DOC");
}
......
paddle/operators/conv_transpose_op.h
浏览文件 @ aece2905
......@@ -141,8 +141,8 @@ class GemmConvTransposeKernel : public framework::OpKernel<T> {
if (data_dim == 2U) {
// col2im: col_matrix -> dy
// from (c * k_h * k_w, h * w) to (c, o_h, o_w)
col2im(dev_ctx, col, std::vector<int>{dilations[0], dilations[1]},
strides, std::vector<int>{paddings[0], paddings[1], paddings[0],
col2im(dev_ctx, col, dilations, strides,
std::vector<int>{paddings[0], paddings[1], paddings[0],
paddings[1]},
&output_batch);
} else if (data_dim == 3U) {
......@@ -247,8 +247,7 @@ class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
if (data_dim == 2U) {
// im2col: dy -> col matrix
// from (c, o_h, o_w) to (c * k_h * k_w, h * w)
im2col(dev_ctx, output_grad_batch,
std::vector<int>{dilations[0], dilations[1]}, strides,
im2col(dev_ctx, output_grad_batch, dilations, strides,
std::vector<int>{paddings[0], paddings[1], paddings[0],
paddings[1]},
&col);
......
paddle/operators/edit_distance_op.cc
浏览文件 @ aece2905
......@@ -25,6 +25,8 @@ class EditDistanceOp : public framework::OperatorWithKernel {
PADDLE_ENFORCE(ctx->HasInput("Hyps"), "Input(Hyps) shouldn't be null.");
PADDLE_ENFORCE(ctx->HasInput("Refs"), "Input(Refs) shouldn't be null.");
PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) shouldn't be null.");
PADDLE_ENFORCE(ctx->HasOutput("SequenceNum"),
"Output(SequenceNum) shouldn't be null.");
auto hyp_dims = ctx->GetInputDim("Hyps");
auto ref_dims = ctx->GetInputDim("Refs");
PADDLE_ENFORCE(hyp_dims.size() == 2 && hyp_dims[1] == 1,
......@@ -34,6 +36,7 @@ class EditDistanceOp : public framework::OperatorWithKernel {
"Input(Refs) must be a 2-D LoDTensor with the 2nd dimension "
"equal to 1.");
ctx->SetOutputDim("Out", ctx->GetInputDim("Refs"));
ctx->SetOutputDim("SequenceNum", {1});
}
protected:
......@@ -54,6 +57,7 @@ class EditDistanceOpMaker : public framework::OpProtoAndCheckerMaker {
AddInput("Refs",
"(2-D LoDTensor<int64_t>, 2nd dim. equal to 1) "
"The indices for reference strings.");
AddOutput("SequenceNum", "The sequence count of current batch");
AddAttr<bool>("normalized",
"(bool, default false) Indicated whether to normalize "
"the edit distance by the length of reference string.")
......
paddle/operators/edit_distance_op.cu
浏览文件 @ aece2905
......@@ -14,6 +14,7 @@ limitations under the License. */
#include <algorithm>
#include "paddle/framework/op_registry.h"
#include "paddle/operators/math/math_function.h"
#include "paddle/platform/cuda_helper.h"
#include "paddle/platform/gpu_info.h"
......@@ -72,6 +73,8 @@ class EditDistanceGPUKernel : public framework::OpKernel<T> {
auto* x1_t = ctx.Input<framework::LoDTensor>("Hyps");
auto* x2_t = ctx.Input<framework::LoDTensor>("Refs");
auto* sequence_num = ctx.Output<framework::Tensor>("SequenceNum");
sequence_num->mutable_data<int64_t>(ctx.GetPlace());
auto normalized = ctx.Attr<bool>("normalized");
auto stream = reinterpret_cast<const platform::CUDADeviceContext&>(
......@@ -88,7 +91,11 @@ class EditDistanceGPUKernel : public framework::OpKernel<T> {
"Reference string %d is empty.", i);
}
auto num_strs = hyp_lod.size() - 1;
const size_t num_strs = hyp_lod.size() - 1;
math::SetConstant<platform::CUDADeviceContext, int64_t> set_constant;
set_constant(ctx.template device_context<platform::CUDADeviceContext>(),
sequence_num, static_cast<int64_t>(num_strs));
out_t->Resize({static_cast<int64_t>(num_strs), 1});
out_t->mutable_data<T>(ctx.GetPlace());
auto out = out_t->data<T>();
......
paddle/operators/edit_distance_op.h
浏览文件 @ aece2905
......@@ -16,7 +16,6 @@ limitations under the License. */
#include <algorithm>
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
......@@ -28,6 +27,8 @@ class EditDistanceKernel : public framework::OpKernel<T> {
auto* x1_t = ctx.Input<framework::LoDTensor>("Hyps");
auto* x2_t = ctx.Input<framework::LoDTensor>("Refs");
auto* sequence_num = ctx.Output<framework::Tensor>("SequenceNum");
int64_t* seq_num_data = sequence_num->mutable_data<int64_t>(ctx.GetPlace());
auto normalized = ctx.Attr<bool>("normalized");
......@@ -41,6 +42,7 @@ class EditDistanceKernel : public framework::OpKernel<T> {
"Reference string %d is empty.", i);
}
auto num_strs = hyp_lod.size() - 1;
*seq_num_data = static_cast<int64_t>(num_strs);
out_t->Resize({static_cast<int64_t>(num_strs), 1});
out_t->mutable_data<float>(ctx.GetPlace());
......
paddle/operators/im2sequence_op.cc 0 → 100644
浏览文件 @ aece2905
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/operators/im2sequence_op.h"
namespace paddle {
namespace operators {
class Im2SequenceOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"),
"Input(X) of Im2SequenceOp should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("Out"),
"Output(Out) of Im2SequenceOp op should not be null.");
auto in_dim = ctx->GetInputDim("X");
PADDLE_ENFORCE_EQ(in_dim.size(), 4,
"Input(X) format must be 4D tensor, eg., NCHW.");
auto kernels = ctx->Attrs().Get<std::vector<int>>("kernels");
auto strides = ctx->Attrs().Get<std::vector<int>>("strides");
auto paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
int batch_size = in_dim[0];
int img_channels = in_dim[1];
int img_height = in_dim[2];
int img_width = in_dim[3];
int output_height = OutputSize(img_height, kernels[0], paddings[0],
paddings[2], strides[0]);
int output_width =
OutputSize(img_width, kernels[1], paddings[1], paddings[3], strides[1]);
ctx->SetOutputDim("Out", {batch_size * output_height * output_width,
img_channels * kernels[0] * kernels[1]});
}
};
class Im2SequenceOpMaker : public framework::OpProtoAndCheckerMaker {
public:
Im2SequenceOpMaker(OpProto* proto, OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X",
"(Tensor) The input tensor has NCHW format."
"N: batch size"
"C: channels"
"H: height"
"W: width");
AddOutput("Out", "(LodTensor) The output data of im2sequence op,");
AddAttr<std::vector<int>>("kernels",
"(vector<int>), the "
"kernels(kernel_height, kernel_width)");
AddAttr<std::vector<int>>("strides",
"(vector<int> default:{1, 1}), the "
"strides(h_stride, w_stride)")
.SetDefault({1, 1});
AddAttr<std::vector<int>>("paddings",
"(vector<int> default:{0, 0, 0, 0}), the "
"paddings(up_pad, left_pad, down_pad, right_pad)")
.SetDefault({0, 0, 0, 0});
AddComment(R"DOC(
This op uses kernels to scan images and converts these images to sequences.
After expanding, The number of time steps are output_height * output_width
and the dimension of each time step is kernel_height * kernel_width * channels,
in which:
output_height =
1 + (padding_height + padding_down + img_height - kernel_height + stride_height - 1) /
stride_height;
output_width =
1 + (padding_left + padding+right + img_width - kernel_width + stride_width - 1) /
stride_width;
This op can be used after convolution neural network, and before recurrent neural network.
Given:
x = [[[[ 6. 2. 1.]
[ 8. 3. 5.]
[ 0. 2. 6.]]
[[ 2. 4. 4.]
[ 6. 3. 0.]
[ 6. 4. 7.]]]
[[[ 6. 7. 1.]
[ 5. 7. 9.]
[ 2. 4. 8.]]
[[ 1. 2. 1.]
[ 1. 3. 5.]
[ 9. 0. 8.]]]]
x.dims = {2, 2, 3, 3}
And:
kernels = [2, 2]
strides = [1, 1]
paddings = [0, 0, 0, 0]
Then:
output.data = [[ 6. 2. 8. 3. 2. 4. 6. 3.]
[ 2. 1. 3. 5. 4. 4. 3. 0.]
[ 8. 3. 0. 2. 6. 3. 6. 4.]
[ 3. 5. 2. 6. 3. 0. 4. 7.]
[ 6. 7. 5. 7. 1. 2. 1. 3.]
[ 7. 1. 7. 9. 2. 1. 3. 5.]
[ 5. 7. 2. 4. 1. 3. 9. 0.]
[ 7. 9. 4. 8. 3. 5. 0. 8.]]
output.dims = {8, 9}
output.lod = [[0, 4, 8]]
)DOC");
}
};
class Im2SequenceGradOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null");
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
"Input(Out@GRAD) shouldn't be null.");
ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(im2sequence, ops::Im2SequenceOp, ops::Im2SequenceOpMaker,
im2sequence_grad, ops::Im2SequenceGradOp);
REGISTER_OP_CPU_KERNEL(
im2sequence,
ops::Im2SequenceKernel<paddle::platform::CPUDeviceContext, float>);
REGISTER_OP_CPU_KERNEL(
im2sequence_grad,
ops::Im2SequenceGradKernel<paddle::platform::CPUDeviceContext, float>);
paddle/operators/im2sequence_op.cu 0 → 100644
浏览文件 @ aece2905
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#define EIGEN_USE_GPU
#include "paddle/operators/im2sequence_op.h"
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(
im2sequence,
ops::Im2SequenceKernel<paddle::platform::CUDADeviceContext, float>);
REGISTER_OP_CUDA_KERNEL(
im2sequence_grad,
ops::Im2SequenceGradKernel<paddle::platform::CUDADeviceContext, float>);
paddle/operators/im2sequence_op.h 0 → 100644
浏览文件 @ aece2905
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
You may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "paddle/framework/data_layout.h"
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
#include "paddle/operators/math/im2col.h"
#include "paddle/operators/math/math_function.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
using LoDTensor = framework::LoDTensor;
inline int OutputSize(int input_size, int filter_size, int padding_0,
int padding_1, int stride) {
const int output_size =
(input_size + padding_0 + padding_1 - filter_size) / stride + 1;
return output_size;
}
template <typename DeviceContext, typename T>
class Im2SequenceKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
const Tensor* in = ctx.Input<Tensor>("X");
LoDTensor* out = ctx.Output<LoDTensor>("Out");
out->mutable_data<T>(ctx.GetPlace());
// TODO(wanghaoshuang): Add layout checker after 'set_layout'
// being available for python API
// PADDLE_ENFORCE_EQ(in->layout(), framework::DataLayout::kNCHW,
// "Input(X) layout must be NCHW");
auto in_dim = in->dims();
int batch_size = in_dim[0];
int img_channels = in_dim[1];
int img_height = in_dim[2];
int img_width = in_dim[3];
auto kernels = ctx.Attr<std::vector<int>>("kernels");
auto strides = ctx.Attr<std::vector<int>>("strides");
auto paddings = ctx.Attr<std::vector<int>>("paddings");
int output_height = OutputSize(img_height, kernels[0], paddings[0],
paddings[2], strides[0]);
int output_width =
OutputSize(img_width, kernels[1], paddings[1], paddings[3], strides[1]);
const std::vector<int> dilations({1, 1});
auto out_dims = out->dims();
out->Resize({batch_size, out->numel() / batch_size});
for (int i = 0; i < batch_size; i++) {
const Tensor src =
in->Slice(i, i + 1).Resize({img_channels, img_height, img_width});
Tensor dst = out->Slice(i, i + 1).Resize(
{output_height, output_width, img_channels, kernels[0], kernels[1]});
math::Im2ColFunctor<math::ColFormat::kOCF, DeviceContext, T> f;
auto& dev_ctx = ctx.template device_context<DeviceContext>();
f(dev_ctx, src, dilations, strides, paddings, &dst);
}
out->Resize(out_dims);
// set lod information
// TODO(wanghaoshuang): Move this to InferShape
framework::LoD lod(1);
lod[0].reserve(batch_size + 1);
for (int i = 0, offset = 0; i < batch_size + 1; ++i) {
lod[0][i] = offset;
offset += output_height * output_width;
}
out->set_lod(lod);
}
};
template <typename DeviceContext, typename T>
class Im2SequenceGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* in = ctx.Input<Tensor>("X");
Tensor* d_out =
const_cast<Tensor*>(ctx.Input<Tensor>(framework::GradVarName("Out")));
auto* d_x = ctx.Output<Tensor>(framework::GradVarName("X"));
d_x->mutable_data<T>(ctx.GetPlace());
auto x_v = framework::EigenVector<T>::Flatten(*d_x);
auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
x_v.device(place) = x_v.constant(0.0);
auto in_dim = in->dims();
int batch_size = in_dim[0];
int img_channels = in_dim[1];
int img_height = in_dim[2];
int img_width = in_dim[3];
auto kernels = ctx.Attr<std::vector<int>>("kernels");
auto strides = ctx.Attr<std::vector<int>>("strides");
auto paddings = ctx.Attr<std::vector<int>>("paddings");
int output_height = OutputSize(img_height, kernels[0], paddings[0],
paddings[2], strides[0]);
int output_width =
OutputSize(img_width, kernels[1], paddings[1], paddings[3], strides[1]);
const std::vector<int> dilations({1, 1});
auto d_out_dims = d_out->dims();
d_out->Resize({batch_size, d_out->numel() / batch_size});
for (int i = 0; i < batch_size; i++) {
Tensor dst =
d_x->Slice(i, i + 1).Resize({img_channels, img_height, img_width});
const Tensor src = d_out->Slice(i, i + 1).Resize(
{output_height, output_width, img_channels, kernels[0], kernels[1]});
math::Col2ImFunctor<math::ColFormat::kOCF, DeviceContext, T> f;
auto& dev_ctx = ctx.template device_context<DeviceContext>();
f(dev_ctx, src, dilations, strides, paddings, &dst);
}
d_out->Resize(d_out_dims);
}
};
} // namespace operators
} // namespace paddle
paddle/operators/math/depthwise_conv.cu
浏览文件 @ aece2905
......@@ -19,7 +19,8 @@ namespace paddle {
namespace operators {
namespace math {
// CUDA kernel to compute the depthwise convolution forward pass
// A Cuda kernel to compute the depthwise convolution forward pass
// in NCHW format.
template <typename T>
__global__ void KernelDepthwiseConv(
const int nthreads, const T* const input_data, const T* const filter_data,
......@@ -79,116 +80,122 @@ __global__ void KernelDepthwiseConv(
output_data[index] = value;
}
}
/*
// CUDA kernel to compute the depthwise convolution backprop w.r.t input.
template <typename T>
__global__ void KernelDepthwiseConvInputGrad(const int nthreads,
const T* const top_diff,
const T* const weight_data,
const int num,
const int outputChannels,
const int outputHeight,
const int outputWidth,
const int inputChannels,
const int inputHeight,
const int inputWidth,
const int filterMultiplier,
const int filterHeight,
const int filterWidth,
const int strideH,
const int strideW,
const int paddingH,
const int paddingW,
T* const bottom_diff) {
__global__ void KernelDepthwiseConvInputGrad(
const int nthreads, const T* const output_grad_data,
const T* const filter_data, const int batch_size, const int output_channels,
const int output_height, const int output_width, const int input_channels,
const int input_height, const int input_width, const int filter_multiplier,
const int filter_height, const int filter_width, const int stride_height,
const int stride_width, const int padding_height, const int padding_width,
T* const input_grad_data) {
int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x;
if (index < nthreads) {
const int batch = index / inputChannels / inputHeight / inputWidth;
const int c_in = (index / inputHeight / inputWidth) % inputChannels;
const int h_in = (index / inputWidth) % inputHeight;
const int w_in = index % inputWidth;
const int batch = index / input_channels / input_height / input_width;
const int c_in = (index / input_height / input_width) % input_channels;
const int h_in = (index / input_width) % input_height;
const int w_in = index % input_width;
const int c_out_start = c_in * filterMultiplier;
const int c_out_start = c_in * filter_multiplier;
int h_out_start = (h_in - filterHeight + paddingH + strideH) / strideH;
int h_out_start =
(h_in - filter_height + padding_height + stride_height) / stride_height;
h_out_start = 0 > h_out_start ? 0 : h_out_start;
int h_out_end = (h_in + paddingH) / strideH;
h_out_end = outputHeight - 1 < h_out_end ? outputHeight - 1 : h_out_end;
int w_out_start = (w_in - filterWidth + paddingW + strideW) / strideW;
int h_out_end = (h_in + padding_height) / stride_height;
h_out_end = output_height - 1 < h_out_end ? output_height - 1 : h_out_end;
int w_out_start =
(w_in - filter_width + padding_width + stride_width) / stride_width;
w_out_start = 0 > w_out_start ? 0 : w_out_start;
int w_out_end = (w_in + paddingW) / strideW;
w_out_end = outputWidth - 1 < w_out_end ? outputWidth - 1 : w_out_end;
int w_out_end = (w_in + padding_width) / stride_width;
w_out_end = output_width - 1 < w_out_end ? output_width - 1 : w_out_end;
T value = 0;
for (int c_out = c_out_start; c_out < c_out_start + filterMultiplier;
for (int c_out = c_out_start; c_out < c_out_start + filter_multiplier;
c_out++) {
for (int h_out = h_out_start; h_out <= h_out_end; ++h_out) {
const int filter_h = h_in + paddingH - h_out * strideH;
const int filter_h = h_in + padding_height - h_out * stride_height;
for (int w_out = w_out_start; w_out <= w_out_end; ++w_out) {
const int filter_w = w_in + paddingW - w_out * strideW;
const int filter_offset = c_out * filterHeight * filterWidth +
filter_h * filterWidth + filter_w;
const int top_diff_offset =
((batch * outputChannels + c_out) * outputHeight + h_out) *
outputWidth +
const int filter_w = w_in + padding_width - w_out * stride_width;
const int filter_offset = c_out * filter_height * filter_width +
filter_h * filter_width + filter_w;
const int output_grad_offset =
((batch * output_channels + c_out) * output_height + h_out) *
output_width +
w_out;
value += top_diff[top_diff_offset] * weight_data[filter_offset];
value +=
output_grad_data[output_grad_offset] * filter_data[filter_offset];
}
}
}
bottom_diff[index] += value;
input_grad_data[index] += value;
}
}
// CUDA kernel to compute the depthwise convolution backprop w.r.t filter.
// Cuda kernel to compute the depthwise convolution backprop w.r.t. filter.
template <typename T>
__global__ void KernelDepthwiseConvFilterGrad(const int num_i,
const int nthreads,
const T* const top_diff,
const T* const inputData,
const int num,
const int outputChannels,
const int outputHeight,
const int outputWidth,
const int inputChannels,
const int inputHeight,
const int inputWidth,
const int filterMultiplier,
const int filterHeight,
const int filterWidth,
const int strideH,
const int strideW,
const int paddingH,
const int paddingW,
T* const buffer_data) {
__global__ void KernelDepthwiseConvFilterGrad(
const int nthreads, const T* const output_grad_data,
const T* const input_data, const int num, const int output_channels,
const int output_height, const int output_width, const int input_channels,
const int input_height, const int input_width, const int filter_multiplier,
const int filter_height, const int filter_width, const int stride_height,
const int stride_width, const int padding_height, const int padding_width,
T* const filter_grad_data) {
int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x;
if (index < nthreads) {
const int h_out = (index / outputWidth) % outputHeight;
const int w_out = index % outputWidth;
const int kh =
(index / filterWidth / outputHeight / outputWidth) % filterHeight;
const int kw = (index / outputHeight / outputWidth) % filterWidth;
const int h_in = -paddingH + h_out * strideH + kh;
const int w_in = -paddingW + w_out * strideW + kw;
if ((h_in >= 0) && (h_in < inputHeight) && (w_in >= 0) &&
(w_in < inputWidth)) {
const int c_out =
index / (filterHeight * filterWidth * outputHeight * outputWidth);
const int c_in = c_out / filterMultiplier;
const int batch = num_i;
const int top_offset =
((batch * outputChannels + c_out) * outputHeight + h_out) *
outputWidth + w_out;
const int bottom_offset =
((batch * inputChannels + c_in) * inputHeight + h_in) * inputWidth +
const int w_out = index % output_width;
const int h_out = (index / output_width) % output_height;
const int c_out = (index / output_width / output_height) % output_channels;
const int batch = (index / output_width / output_height / output_channels);
const int c_in = c_out / filter_multiplier;
const int h_in_start = -padding_height + h_out * stride_height;
const int w_in_start = -padding_width + w_out * stride_width;
const int h_in_end =
-padding_height + h_out * stride_height + filter_height;
const int w_in_end = -padding_width + w_out * stride_width + filter_width;
if ((h_in_start >= 0) && (h_in_end < input_height) && (w_in_start >= 0) &&
(w_in_end < input_width)) {
for (int kw = 0; kw < filter_width; kw++) {
for (int kh = 0; kh < filter_height; kh++) {
const int h_in = -padding_height + h_out * stride_height + kh;
const int w_in = -padding_width + w_out * stride_width + kw;
const int offset =
((batch * input_channels + c_in) * input_height + h_in) *
input_width +
w_in;
buffer_data[index] = top_diff[top_offset] * inputData[bottom_offset];
const T diff_temp = output_grad_data[index] * input_data[offset];
T* addr = filter_grad_data + c_out * filter_height * filter_width +
kh * filter_width + kw;
paddle::platform::CudaAtomicAdd(addr, diff_temp);
}
}
} else {
buffer_data[index] = 0;
for (int kw = 0; kw < filter_width; kw++) {
for (int kh = 0; kh < filter_height; kh++) {
const int h_in = -padding_height + h_out * stride_height + kh;
const int w_in = -padding_width + w_out * stride_width + kw;
if ((h_in >= 0) && (h_in < input_height) && (w_in >= 0) &&
(w_in < input_width)) {
const int offset =
((batch * input_channels + c_in) * input_height + h_in) *
input_width +
w_in;
const T diff_temp = output_grad_data[index] * input_data[offset];
T* addr = filter_grad_data + c_out * filter_height * filter_width +
kh * filter_width + kw;
paddle::platform::CudaAtomicAdd(addr, diff_temp);
}
}
}
}
}
}
*/
/*
* All tensors are in NCHW format.
......@@ -200,9 +207,8 @@ class DepthwiseConvFunctor<platform::CUDADeviceContext, T> {
public:
void operator()(const platform::CUDADeviceContext& context,
const framework::Tensor& input,
const framework::Tensor& filter, std::vector<int>& ksize,
std::vector<int>& strides, std::vector<int>& paddings,
framework::Tensor* output) {
const framework::Tensor& filter, std::vector<int>& strides,
std::vector<int>& paddings, framework::Tensor* output) {
const int batch_size = input.dims()[0];
const int input_channels = input.dims()[1];
const int input_height = input.dims()[2];
......@@ -210,8 +216,8 @@ class DepthwiseConvFunctor<platform::CUDADeviceContext, T> {
const int output_channels = output->dims()[1];
const int output_height = output->dims()[2];
const int output_width = output->dims()[3];
const int ksize_height = ksize[0];
const int ksize_width = ksize[1];
const int ksize_height = filter.dims()[2];
const int ksize_width = filter.dims()[3];
const int stride_height = strides[0];
const int stride_width = strides[1];
const int padding_height = paddings[0];
......@@ -235,31 +241,30 @@ class DepthwiseConvFunctor<platform::CUDADeviceContext, T> {
}
};
/*
template <typename T>
class DepthwiseConvInputGradFunctor<platform::CUDADeviceContext, PoolProcess, T>
{
class DepthwiseConvInputGradFunctor<platform::CUDADeviceContext, T> {
public:
void operator()(const platform::CUDADeviceContext& context,
const framework::Tensor& input,
const framework::Tensor& output,
const framework::Tensor& output_grad, std::vector<int>& ksize,
const framework::Tensor& filter,
const framework::Tensor& output_grad,
std::vector<int>& strides, std::vector<int>& paddings,
PoolProcess pool_process, framework::Tensor* input_grad) {
framework::Tensor* input_grad) {
const int batch_size = input.dims()[0];
const int input_channels = input.dims()[1];
const int input_height = input.dims()[2];
const int input_width = input.dims()[3];
const int output_height = output.dims()[2];
const int output_width = output.dims()[3];
const int ksize_height = ksize[0];
const int ksize_width = ksize[1]; const int stride_height = strides[0];
const int output_channels = output_grad.dims()[1];
const int output_height = output_grad.dims()[2];
const int output_width = output_grad.dims()[3];
const int ksize_height = filter.dims()[2];
const int ksize_width = filter.dims()[3];
const int stride_height = strides[0];
const int stride_width = strides[1];
const int padding_height = paddings[0];
const int padding_width = paddings[1];
const T* input_data = input.data<T>();
const T* output_data = output.data<T>();
const T* filter_data = filter.data<T>();
const T* output_grad_data = output_grad.data<T>();
T* input_grad_data = input_grad->mutable_data<T>(context.GetPlace());
......@@ -268,73 +273,68 @@ class DepthwiseConvInputGradFunctor<platform::CUDADeviceContext, PoolProcess, T>
dim3 threads(1024, 1);
dim3 grid(blocks, 1);
KernelPool2DGrad<PoolProcess, T><<<grid, threads, 0, context.stream()>>>(
nthreads, input_data, output_data, output_grad_data, input_channels,
input_height, input_width, output_height, output_width, ksize_height,
ksize_width, stride_height, stride_width, padding_height, padding_width,
pool_process, input_grad_data);
KernelDepthwiseConvInputGrad<T><<<grid, threads, 0, context.stream()>>>(
nthreads, output_grad_data, filter_data, batch_size, output_channels,
output_height, output_width, input_channels, input_height, input_width,
output_channels / input_channels, ksize_height, ksize_width,
stride_height, stride_width, padding_height, padding_width,
input_grad_data);
}
};
template <typename T>
class DepthwiseConvdFilterGradFunctor<platform::CUDADeviceContext, T> {
class DepthwiseConvFilterGradFunctor<platform::CUDADeviceContext, T> {
public:
void operator()(const platform::CUDADeviceContext& context,
const framework::Tensor& input,
const framework::Tensor& output,
const framework::Tensor& output_grad, std::vector<int>& ksize,
const framework::Tensor& output_grad,
std::vector<int>& strides, std::vector<int>& paddings,
framework::Tensor* input_grad) {
framework::Tensor* filter_grad) {
const int batch_size = input.dims()[0];
const int input_channels = input.dims()[1];
const int input_height = input.dims()[2];
const int input_width = input.dims()[3];
const int output_channels = output.dims()[1];
const int output_height = output.dims()[2];
const int output_width = output.dims()[3];
const int ksize_height = ksize[0];
const int ksize_width = ksize[1];
const int output_channels = output_grad.dims()[1];
const int output_height = output_grad.dims()[2];
const int output_width = output_grad.dims()[3];
const int ksize_height = filter_grad->dims()[2];
const int ksize_width = filter_grad->dims()[3];
const int stride_height = strides[0];
const int stride_width = strides[1];
const int padding_height = paddings[0];
const int padding_width = paddings[1];
const T* input_data = input.data<T>();
const T* output_data = output.data<T>();
const T* output_grad_data = output_grad.data<T>();
T* input_grad_data = input_grad->mutable_data<T>(context.GetPlace());
T* filter_grad_data = filter_grad->mutable_data<T>(context.GetPlace());
int nthreads = batch_size * output_channels * output_height * output_width;
int blocks = (nthreads + 1024 - 1) / 1024;
dim3 threads(1024, 1);
dim3 grid(blocks, 1);
KernelMaxPool2DGrad<T><<<grid, threads, 0, context.stream()>>>(
nthreads, input_data, output_data, output_grad_data, input_channels,
input_height, input_width, output_height, output_width, ksize_height,
ksize_width, stride_height, stride_width, padding_height, padding_width,
input_grad_data);
KernelDepthwiseConvFilterGrad<T><<<grid, threads, 0, context.stream()>>>(
nthreads, output_grad_data, input_data, batch_size, output_channels,
output_height, output_width, input_channels, input_height, input_width,
output_channels / input_channels, ksize_height, ksize_width,
stride_height, stride_width, padding_height, padding_width,
filter_grad_data);
}
};
*/
template class DepthwiseConvFunctor<platform::CUDADeviceContext,
float>;
template class DepthwiseConvFunctor<platform::CUDADeviceContext,
double>;
template class DepthwiseConvFunctor<platform::CUDADeviceContext, float>;
template class DepthwiseConvFunctor<platform::CUDADeviceContext, double>;
/*
template class DepthwiseConvInputGradFunctor<platform::CUDADeviceContext,
float>;
template class DepthwiseConvFilterGradFunctor<platform::CUDADeviceContext,
float>;
template class DepthwiseConvFunctor<platform::CUDADeviceContext,
template class DepthwiseConvInputGradFunctor<platform::CUDADeviceContext,
double>;
template class DepthwiseConvFilterGradFunctor<platform::CUDADeviceContext,
float>;
template class DepthwiseConvFilterGradFunctor<platform::CUDADeviceContext,
double>;
*/
} // namespace math
} // namespace operators
......
paddle/operators/math/depthwise_conv.h
浏览文件 @ aece2905
......@@ -21,36 +21,36 @@ namespace paddle {
namespace operators {
namespace math {
/*
* \brief Compute the depthwise convolution which include
* forward process and backpropagation process
*/
template <typename DeviceContext, typename T>
class DepthwiseConvFunctor {
public:
void operator()(const DeviceContext& context, const framework::Tensor& input,
const framework::Tensor& filter, std::vector<int>& ksize,
std::vector<int>& strides, std::vector<int>& paddings,
framework::Tensor* output);
const framework::Tensor& filter, std::vector<int>& strides,
std::vector<int>& paddings, framework::Tensor* output);
};
/*
template <typename DeviceContext, typename T>
class DepthwiseConvInputGradFunctor {
public:
void operator()(const DeviceContext& context,
public:
void operator()(const DeviceContext& context, const framework::Tensor& input,
const framework::Tensor& filter,
const framework::Tensor& output_grad, std::vector<int>& ksize,
const framework::Tensor& output_grad,
std::vector<int>& strides, std::vector<int>& paddings,
framework::Tensor* input_grad);
};
template <typename DeviceContext, typename T>
class DepthwiseConvFilterGradFunctor {
public:
void operator()(const DeviceContext& context,
const framework::Tensor& input,
const framework::Tensor& output_grad, std::vector<int>& ksize,
public:
void operator()(const DeviceContext& context, const framework::Tensor& input,
const framework::Tensor& output_grad,
std::vector<int>& strides, std::vector<int>& paddings,
framework::Tensor* filter_grad);
};
*/
} // namespace math
} // namespace operators
......
paddle/operators/recv_op.cc
浏览文件 @ aece2905
......@@ -49,7 +49,7 @@ static void CreateTensorFromMessageType(framework::Variable *var,
var->GetMutable<framework::SelectedRows>();
} else {
PADDLE_THROW(
"VraibleMessage type %d is not in "
"VariableMessage type %d is not in "
"[LoDTensor, SelectedRows]",
var_type);
}
......@@ -121,17 +121,17 @@ class RecvOp : public framework::OperatorBase {
if (it != grad_list.end()) {
param_var_name = param_list[it - grad_list.begin()];
} else {
LOG(ERROR) << "grad have no paired param:" << grad_var_name;
LOG(ERROR) << "grad has no paired param:" << grad_var_name;
}
VLOG(3) << "recved grad: " << grad_var_name
VLOG(3) << "received grad: " << grad_var_name
<< " updating param: " << param_var_name;
if (fan_in > 1) {
grad_var_name = this->GetGradVarNameForTrainer(grad_var_name);
}
auto *var = recv_scope.FindVar(grad_var_name);
if (var == nullptr) {
LOG(ERROR) << "can not find server side var: " << grad_var_name;
PADDLE_THROW("can not find server side var");
LOG(ERROR) << "Can not find server side var: " << grad_var_name;
PADDLE_THROW("Can not find server side var");
}
detail::DeserializeFromMessage(v.second, dev_ctx, var);
}
......@@ -165,7 +165,7 @@ class RecvOpMaker : public framework::OpProtoAndCheckerMaker {
AddComment(R"DOC(
Recv operator
This operator will recv tensor from send_op
This operator will recieve tensor from send_op
)DOC");
AddAttr<std::string>("endpoint",
"(string, default 127.0.0.1:6164)"
......@@ -176,11 +176,11 @@ This operator will recv tensor from send_op
kOptimizeBlock, "Serialized ProgramDesc string for recv to run.");
AddAttr<std::vector<std::string>>(
"ParamList", "type list of string",
"grad->param name mapping to find which param to optimize.")
"grad->param name mapping to find which parameters to optimize.")
.SetDefault({});
AddAttr<std::vector<std::string>>(
"GradList", "type list of string",
"grad->param name mapping to find which param to optimize.")
"grad->param name mapping to find which parameters to optimize.")
.SetDefault({});
AddAttr<int>("Fanin", "type int",
"Number of trainers in the current cluster job")
......
paddle/operators/reduce_op.cc
浏览文件 @ aece2905
......@@ -190,10 +190,22 @@ REGISTER_OP(reduce_min, ops::ReduceOp, ops::ReduceMinOpMaker, reduce_min_grad,
#define REGISTER_REDUCE_CPU_KERNEL(reduce_type, functor, grad_functor) \
REGISTER_OP_CPU_KERNEL(reduce_type, \
ops::ReduceKernel<paddle::platform::CPUDeviceContext, \
float, ops::functor>); \
float, ops::functor>, \
ops::ReduceKernel<paddle::platform::CPUDeviceContext, \
double, ops::functor>, \
ops::ReduceKernel<paddle::platform::CPUDeviceContext, \
int, ops::functor>, \
ops::ReduceKernel<paddle::platform::CPUDeviceContext, \
int64_t, ops::functor>); \
REGISTER_OP_CPU_KERNEL( \
reduce_type##_grad, \
ops::ReduceGradKernel<paddle::platform::CPUDeviceContext, float, \
ops::grad_functor>, \
ops::ReduceGradKernel<paddle::platform::CPUDeviceContext, double, \
ops::grad_functor>, \
ops::ReduceGradKernel<paddle::platform::CPUDeviceContext, int, \
ops::grad_functor>, \
ops::ReduceGradKernel<paddle::platform::CPUDeviceContext, int64_t, \
ops::grad_functor>);
FOR_EACH_KERNEL_FUNCTOR(REGISTER_REDUCE_CPU_KERNEL);
paddle/operators/reduce_op.cu
浏览文件 @ aece2905
......@@ -20,10 +20,22 @@ namespace ops = paddle::operators;
#define REGISTER_REDUCE_GPU_KERNEL(reduce_type, functor, grad_functor) \
REGISTER_OP_CUDA_KERNEL( \
reduce_type, ops::ReduceKernel<paddle::platform::CUDADeviceContext, \
float, ops::functor>); \
float, ops::functor>, \
ops::ReduceKernel<paddle::platform::CUDADeviceContext, double, \
ops::functor>, \
ops::ReduceKernel<paddle::platform::CUDADeviceContext, int, \
ops::functor>, \
ops::ReduceKernel<paddle::platform::CUDADeviceContext, int64_t, \
ops::functor>); \
REGISTER_OP_CUDA_KERNEL( \
reduce_type##_grad, \
ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, float, \
ops::grad_functor>, \
ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, double, \
ops::grad_functor>, \
ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int, \
ops::grad_functor>, \
ops::ReduceGradKernel<paddle::platform::CUDADeviceContext, int64_t, \
ops::grad_functor>);
FOR_EACH_KERNEL_FUNCTOR(REGISTER_REDUCE_GPU_KERNEL);
paddle/operators/send_op.cc
浏览文件 @ aece2905
......@@ -62,13 +62,13 @@ class SendOpMaker : public framework::OpProtoAndCheckerMaker {
public:
SendOpMaker(OpProto* proto, OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "(Tensor) Input tensor to be send").AsDuplicable();
AddOutput("Out", "(Tensor) Output tensor to get from server")
AddInput("X", "(Tensor) Input tensor to be sent").AsDuplicable();
AddOutput("Out", "(Tensor) Output tensor to be received from server")
.AsDuplicable();
AddComment(R"DOC(
Send operator
This operator will send tensor to recv_op.
This operator will send tensor to recv_op at the parameter server.
)DOC");
AddAttr<std::vector<std::string>>("endpoints",
"(string vector, default 127.0.0.1:6164)"
......
paddle/operators/transpose_op.cc
浏览文件 @ aece2905
......@@ -59,44 +59,39 @@ class TransposeOpMaker : public framework::OpProtoAndCheckerMaker {
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput(
"X",
"(Tensor)The input tensor, tensors with rank at most 6 are supported");
AddOutput("Out", "(Tensor)The output tensor");
"(Tensor) The input tensor, tensors with rank up to 6 are supported.");
AddOutput("Out", "(Tensor)The output tensor.");
AddAttr<std::vector<int>>(
"axis",
"(vector<int>)A list of values, and the size of the list should be "
"the same with the input tensor rank, the tensor will "
"permute the axes according the the values given");
"(vector<int>) A list of values, and the size of the list should be "
"the same with the input tensor rank. This operator permutes the input "
"tensor's axes according to the values given.");
AddComment(R"DOC(
Transpose Operator.
The input tensor will be permuted according to the axis values given.
The op functions is similar to how numpy.transpose works in python.
The input tensor will be permuted according to the axes given.
The behavior of this operator is similar to how `numpy.transpose` works.
For example:
- suppose the input `X` is a 2-D tensor:
$$
X = \begin{pmatrix}
0 &1 &2 \\
3 &4 &5
\end{pmatrix}$$
.. code-block:: text
the given `axes` is: $[1, 0]$, and $Y$ = transpose($X$, axis)
input = numpy.arange(6).reshape((2,3))
then the output $Y$ is:
the input is:
$$
Y = \begin{pmatrix}
0 &3 \\
1 &4 \\
2 &5
\end{pmatrix}$$
array([[0, 1, 2],
[3, 4, 5]])
given axis is:
[1, 0]
output = input.transpose(axis)
then the output is:
array([[0, 3],
[1, 4],
[2, 5]])
So, given a input tensor of shape(N, C, H, W) and the axis is {0, 2, 3, 1},
the output tensor shape will be (N, H, W, C)
- Given a input tensor with shape $(N, C, H, W)$ and the `axes` is
$[0, 2, 3, 1]$, then shape of the output tensor will be: $(N, H, W, C)$.
)DOC");
}
......
paddle/platform/CMakeLists.txt
浏览文件 @ aece2905
......@@ -39,3 +39,11 @@ nv_test(nccl_test SRCS nccl_test.cu DEPS dynload_cuda gpu_info device_context)
cc_library(profiler SRCS profiler.cc DEPS device_context)
cc_test(profiler_test SRCS profiler_test.cc DEPS profiler)
if(NOT WITH_C_API AND WITH_FLUID)
file(GLOB PLATFORM_HEADERS *.h)
file(GLOB PLATFORM_dynload_HEADERS dynload/*.h)
install(FILES ${PLATFORM_HEADERS} DESTINATION include/paddle/platform)
install(FILES ${PLATFORM_HEADERS} DESTINATION include/paddle/platform/dynload)
install(FILES details/device_ptr_cast.h DESTINATION include/paddle/platform/details)
endif()
paddle/string/CMakeLists.txt
浏览文件 @ aece2905
cc_library(stringpiece SRCS piece.cc)
cc_test(stringpiece_test SRCS piece_test.cc DEPS stringpiece glog gflags)
cc_test(stringprintf_test SRCS printf_test.cc DEPS glog gflags)
cc_test(to_string_test SRCS to_string_test.cc)
if(NOT WITH_C_API AND WITH_FLUID)
file(GLOB STRING_HEADERS *.h)
install(FILES ${STRING_HEADERS} DESTINATION include/paddle/string)
install(FILES tinyformat/tinyformat.h DESTINATION include/paddle/string/tinyformat)
endif()
python/paddle/v2/dataset/wmt16.py
浏览文件 @ aece2905
......@@ -171,8 +171,9 @@ def train(src_dict_size, trg_dict_size, src_lang="en"):
callable: The train reader.
"""
assert (src_lang in ["en", "de"], ("An error language type. Only support: "
"en (for English); de(for Germany)"))
if src_lang not in ["en", "de"]:
raise ValueError("An error language type. Only support: "
"en (for English); de(for Germany).")
src_dict_size, trg_dict_size = __get_dict_size(src_dict_size, trg_dict_size,
src_lang)
......@@ -218,9 +219,9 @@ def test(src_dict_size, trg_dict_size, src_lang="en"):
callable: The test reader.
"""
assert (src_lang in ["en", "de"],
("An error language type. "
"Only support: en (for English); de(for Germany)"))
if src_lang not in ["en", "de"]:
raise ValueError("An error language type. "
"Only support: en (for English); de(for Germany).")
src_dict_size, trg_dict_size = __get_dict_size(src_dict_size, trg_dict_size,
src_lang)
......@@ -266,9 +267,9 @@ def validation(src_dict_size, trg_dict_size, src_lang="en"):
Returns:
callable: The validation reader.
"""
assert (src_lang in ["en", "de"],
("An error language type. "
"Only support: en (for English); de(for Germany)"))
if src_lang not in ["en", "de"]:
raise ValueError("An error language type. "
"Only support: en (for English); de(for Germany).")
src_dict_size, trg_dict_size = __get_dict_size(src_dict_size, trg_dict_size,
src_lang)
......@@ -304,9 +305,9 @@ def get_dict(lang, dict_size, reverse=False):
dict_path = os.path.join(paddle.v2.dataset.common.DATA_HOME,
"wmt16/%s_%d.dict" % (lang, dict_size))
assert (os.path.exists(dict_path), "Word dictionary does not exist. "
"Please invoke paddle.dataset.wmt16.train/test/validation "
"first to build the dictionary.")
assert os.path.exists(dict_path), "Word dictionary does not exist. "
"Please invoke paddle.dataset.wmt16.train/test/validation first "
"to build the dictionary."
tar_file = os.path.join(paddle.v2.dataset.common.DATA_HOME, "wmt16.tar.gz")
return __load_dict(tar_file, dict_size, lang, reverse)
......
python/paddle/v2/fluid/clip.py
浏览文件 @ aece2905
......@@ -12,14 +12,18 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import copy
import functools
import layers
import framework
from . import core
__all__ = [
'GradientClipByValue',
'ErrorClipByValue',
'GradientClipByValue',
'GradientClipByNorm',
'GradientClipByGlobalNorm',
'append_gradient_clip_ops',
'error_clip_callback',
]
......@@ -155,10 +159,11 @@ class GradientClipByGlobalNorm(BaseGradientClipAttr):
return param, new_grad
def gradient_clip_by_global_norm(clip_norm,
param_list=None,
group_name="default_group",
program=None):
def set_gradient_clip(clip, param_list=None, program=None):
if not isinstance(clip, BaseGradientClipAttr):
raise TypeError(
"'clip' should be an instance of BaseGradientClipAttr's derived class"
)
if program is None:
program = framework.default_main_program()
if param_list is None:
......@@ -171,8 +176,7 @@ def gradient_clip_by_global_norm(clip_norm,
)
for param in param_list:
param.gradient_clip_attr = GradientClipByGlobalNorm(clip_norm,
group_name)
param.gradient_clip_attr = copy.deepcopy(clip)
def append_gradient_clip_ops(param_grad):
......
python/paddle/v2/fluid/distribute_transpiler.py
浏览文件 @ aece2905
......@@ -38,14 +38,14 @@ def split_dense_variable(var_list,
min_block_size=1024,
max_block_size=1048576):
"""
We may need to split dense tensor to one or several blocks and put
We may need to split dense tensor to one or more blocks and put
them equally onto parameter server. One block is a sub-tensor
aligned by dim[0] of the tensor.
We need to have a minimal block size so that the calculations in
the parameter server side can gain better performance. By default
mininum block size is 1024. The max block size is used to prevent
too large block that may causing send error.
minimum block size is 1024. The max block size is used to prevent
very large blocks that may cause send error.
"""
blocks = []
for var in var_list:
......@@ -64,7 +64,7 @@ def split_dense_variable(var_list,
remains = block_size % dim1
if remains != 0:
block_size += dim1 - remains
# update split_count after align
# update split_count after aligning
split_count = int(math.ceil(var_numel / float(block_size)))
for block_id in xrange(split_count):
curr_block_size = min(block_size, var_numel - (
......@@ -83,18 +83,18 @@ class DistributeTranspiler:
trainers=1,
split_method=round_robin):
"""
Transpile the program to a distributed data-parallelism programs.
The main_program will be transform to use a remote parameter server
Transpile the program to distributed data-parallelism programs.
The main_program will be transformed to use a remote parameter server
to do parameter optimization. And the optimization graph will be put
in to a parameter server program.
into a parameter server program.
Use different methods to split trainable varialbles to different
Use different methods to split trainable variables to different
parameter servers.
:param optimize_ops: op list of optimization, should be the
return value of Optimizer.minimize
:type optimize_ops: list
:param program: program to optimize, default default_main_program
:param program: program to optimize, default is default_main_program
:param pservers: parameter server endpoints like "m1:6174,m2:6174"
:type pservers: string
:return: return a list of programs
......@@ -106,11 +106,11 @@ class DistributeTranspiler:
self.trainers = trainers
self.optimize_ops = optimize_ops
# steps to transpile:
# 1. split variable to multiple blocks, align by product(dim[1:]) (width).
# 1. split variable to multiple blocks, aligned by product(dim[1:]) (width).
# 2. modify trainer program add split_op to each Grad.
# 3. append send_op to trainer.
# 4. append concat_op to trainer to update local weights.
# 5. create new program as parameter server.
# 5. create new program for parameter server.
# 6. create parameter server program by split_method generated endpoint->VarBlock
pserver_endpoints = pservers.split(",")
......@@ -136,10 +136,10 @@ class DistributeTranspiler:
for b in param_blocks:
varname, block_id, _ = b.split(":")
send_outputs.append(param_var_mapping[varname][int(block_id)])
# let send_op know which endpoint to send which var, eplist is of the same
# order of send_inputs.
# let send_op know which endpoint to send which var to, eplist has the same
# order as send_inputs.
eplist = split_method(send_inputs, pserver_endpoints)
# create mapping of endpoint -> splited var to create pserver side program
# create mapping of endpoint -> split var to create pserver side program
self.param_grad_ep_mapping = dict()
for i, ep in enumerate(eplist):
param = send_outputs[i]
......@@ -149,6 +149,7 @@ class DistributeTranspiler:
self.param_grad_ep_mapping[ep]["params"].append(param)
self.param_grad_ep_mapping[ep]["grads"].append(grad)
# create send_op
send_op = program.global_block().append_op(
type="send",
inputs={"X": send_inputs},
......@@ -167,6 +168,7 @@ class DistributeTranspiler:
attrs={"axis": 0})
def _create_vars_from_blocklist(self, program, block_list):
# Create respective variables using the block_list
block_map = dict()
var_mapping = dict()
for block_str in block_list:
......@@ -207,11 +209,12 @@ class DistributeTranspiler:
dtype=var.dtype,
type=var.type,
lod_level=var.lod_level,
# HACK: let all param in pserver persistable so child
# HACK: let all param in pserver be persistable so the child
# program in recv can get them
persistable=True)
def _append_split_op(self, program, gradblocks):
# Split variables that need to be split and append respective ops
var_mapping = self._create_vars_from_blocklist(program, gradblocks)
for varname, splited_vars in var_mapping.iteritems():
# variable that don't need to split have empty splited_vars
......@@ -248,6 +251,7 @@ class DistributeTranspiler:
return self.program
def _create_var_for_trainers(self, block, var, trainers):
# For each trainer, create the necessary variables
var_list = []
for i in xrange(trainers):
var_each = block.create_var(
......@@ -262,7 +266,7 @@ class DistributeTranspiler:
param_shape):
"""
Returns the shape for optimizer inputs that need to be reshaped when
Param and Grad is splited to multiple servers.
Param and Grad is split to multiple servers.
"""
# HACK(typhoonzero): Should use functions of corresponding optimizer in
# optimizer.py to get the shape, do not bind this in the transpiler.
......@@ -396,7 +400,7 @@ class DistributeTranspiler:
dtype=var.dtype,
shape=new_shape)
# change outputs ParamOut variable
# change output's ParamOut variable
opt_op.outputs["ParamOut"] = new_inputs["Param"]
program.global_block().append_op(
type=opt_op.type,
......@@ -405,6 +409,7 @@ class DistributeTranspiler:
attrs=opt_op.attrs)
def _append_pserver_non_opt_ops(self, program, pserver_program, opt_op):
# Append the ops for parameters that do not need to be optimized/updated
for _, var in opt_op.inputs.iteritems():
program.global_block().create_var(
name=var.name,
......@@ -424,7 +429,7 @@ class DistributeTranspiler:
def get_pserver_program(self, endpoint):
"""
get pserver side program by endpoint
Get pserver side program using the endpoint
NOTE: assume blocks of the same variable is not distributed
on the same pserver, only change param/grad varnames for
......@@ -450,6 +455,7 @@ class DistributeTranspiler:
shape=v.shape)
# step6
optimize_sub_program = Program()
# Iterate through the ops and append ops as needed
for idx, opt_op in enumerate(self.optimize_ops):
is_op_on_pserver = self._is_op_on_pserver(endpoint,
self.optimize_ops, idx)
......@@ -461,6 +467,7 @@ class DistributeTranspiler:
else:
self._append_pserver_non_opt_ops(optimize_sub_program,
pserver_program, opt_op)
# Append the recv op
pserver_program.global_block().append_op(
type="recv",
inputs={"RX": self.param_grad_ep_mapping[endpoint]["grads"]
......@@ -486,7 +493,7 @@ class DistributeTranspiler:
"""
Get startup program for current parameter server.
Modify operator input variables if there are variables that
was splited to several blocks.
were split to several blocks.
"""
s_prog = Program()
orig_s_prog = framework.default_startup_program()
......
python/paddle/v2/fluid/evaluator.py
浏览文件 @ aece2905
......@@ -205,3 +205,63 @@ class ChunkEvaluator(Evaluator):
[precision], dtype='float32'), np.array(
[recall], dtype='float32'), np.array(
[f1_score], dtype='float32')
class EditDistance(Evaluator):
"""
Accumulate edit distance sum and sequence number from mini-batches and
compute the average edit_distance of all batches.
Args:
input: the sequences predicted by network.
label: the target sequences which must has same sequence count
with input.
ignored_tokens(list of int): Tokens that should be removed before
calculating edit distance.
Example:
exe = fluid.executor(place)
distance_evaluator = fluid.Evaluator.EditDistance(input, label)
for epoch in PASS_NUM:
distance_evaluator.reset(exe)
for data in batches:
loss, sum_distance = exe.run(fetch_list=[cost] + distance_evaluator.metrics)
avg_distance = distance_evaluator.eval(exe)
pass_distance = distance_evaluator.eval(exe)
In the above example:
'sum_distance' is the sum of the batch's edit distance.
'avg_distance' is the average of edit distance from the firt batch to the current batch.
'pass_distance' is the average of edit distance from all the pass.
"""
def __init__(self, input, label, ignored_tokens=None, **kwargs):
super(EditDistance, self).__init__("edit_distance", **kwargs)
main_program = self.helper.main_program
if main_program.current_block().idx != 0:
raise ValueError("You can only invoke Evaluator in root block")
self.total_error = self.create_state(
dtype='float32', shape=[1], suffix='total_error')
self.seq_num = self.create_state(
dtype='int64', shape=[1], suffix='seq_num')
error, seq_num = layers.edit_distance(
input=input, label=label, ignored_tokens=ignored_tokens)
#error = layers.cast(x=error, dtype='float32')
sum_error = layers.reduce_sum(error)
layers.sums(input=[self.total_error, sum_error], out=self.total_error)
layers.sums(input=[self.seq_num, seq_num], out=self.seq_num)
self.metrics.append(sum_error)
def eval(self, executor, eval_program=None):
if eval_program is None:
eval_program = Program()
block = eval_program.current_block()
with program_guard(main_program=eval_program):
total_error = _clone_var_(block, self.total_error)
seq_num = _clone_var_(block, self.seq_num)
seq_num = layers.cast(x=seq_num, dtype='float32')
out = layers.elementwise_div(x=total_error, y=seq_num)
return np.array(executor.run(eval_program, fetch_list=[out])[0])
python/paddle/v2/fluid/layers/math_op_patch.py
浏览文件 @ aece2905
......@@ -13,7 +13,7 @@
# limitations under the License.
from ..framework import Variable, unique_name
from ..registry import OpProtoHolder
from layer_function_generator import OpProtoHolder
__all__ = ['monkey_patch_variable']
......
python/paddle/v2/fluid/layers/nn.py
浏览文件 @ aece2905
......@@ -22,13 +22,41 @@ from ..param_attr import ParamAttr
from tensor import concat
__all__ = [
'fc', 'embedding', 'dynamic_lstm', 'gru_unit', 'linear_chain_crf',
'crf_decoding', 'cos_sim', 'cross_entropy', 'square_error_cost', 'accuracy',
'chunk_eval', 'sequence_conv', 'conv2d', 'sequence_pool', 'pool2d',
'batch_norm', 'beam_search_decode', 'conv2d_transpose', 'sequence_expand',
'lstm_unit', 'reduce_sum', 'reduce_mean', 'reduce_max', 'reduce_min',
'sequence_first_step', 'sequence_last_step', 'dropout', 'split',
'l2_normalize', 'matmul', 'warpctc', 'sequence_reshape'
'fc',
'embedding',
'dynamic_lstm',
'gru_unit',
'linear_chain_crf',
'crf_decoding',
'cos_sim',
'cross_entropy',
'square_error_cost',
'accuracy',
'chunk_eval',
'sequence_conv',
'conv2d',
'sequence_pool',
'pool2d',
'batch_norm',
'beam_search_decode',
'conv2d_transpose',
'sequence_expand',
'lstm_unit',
'reduce_sum',
'reduce_mean',
'reduce_max',
'reduce_min',
'sequence_first_step',
'sequence_last_step',
'dropout',
'split',
'ctc_greedy_decoder',
'edit_distance',
'l2_normalize',
'matmul',
'warpctc',
'sequence_reshape',
'transpose',
]
......@@ -43,14 +71,14 @@ def fc(input,
**Fully Connected Layer**
The fully connected layer can take multiple tensors as its inputs. It
creates a variable (one for each input tensor) called weights for each input
tensor, which represents a fully connected weight matrix from each input
unit to each output unit. The fully connected layer multiplies each input
tensor with its coresponding weight to produce an output Tensor. If
multiple input tensors are given, the results of multiple multiplications
will be sumed up. If bias_attr is not None, a biases variable will be
created and added to the output. Finally, if activation is not None,
it will be applied to the output as well.
creates a variable (one for each input tensor) called weights for each
input tensor, which represents a fully connected weight matrix from
each input unit to each output unit. The fully connected layer
multiplies each input tensor with its coresponding weight to produce
an output Tensor. If multiple input tensors are given, the results of
multiple multiplications will be sumed up. If bias_attr is not None,
a biases variable will be created and added to the output. Finally,
if activation is not None, it will be applied to the output as well.
This process can be formulated as follows:
......@@ -762,8 +790,8 @@ def conv2d(input,
<http://ufldl.stanford.edu/tutorial/supervised/FeatureExtractionUsingConvolution/>`_ .
If bias attribution and activation type are provided, bias is added to the output of the convolution,
and the corresponding activation function is applied to the final result.
For each input :math:`X`, the equation is:
For each input :math:`X`, the equation is:
.. math::
......@@ -780,14 +808,17 @@ def conv2d(input,
Example:
Input:
- Input:
Input shape: $(N, C_{in}, H_{in}, W_{in})$
Filter shape: $(C_{out}, C_{in}, H_f, W_f)$
Output:
- Output:
Output shape: $(N, C_{out}, H_{out}, W_{out})$
Where
.. math::
H_{out}&= \\frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]} + 1 \\\\
......@@ -830,13 +861,18 @@ def conv2d(input,
data = fluid.layers.data(name='data', shape=[3, 32, 32], dtype='float32')
conv2d = fluid.layers.conv2d(input=data, num_filters=2, filter_size=3, act="relu")
"""
if stride is None:
stride = [1, 1]
helper = LayerHelper('conv2d', **locals())
dtype = helper.input_dtype()
num_channels = input.shape[1]
l_type = 'conv2d'
if num_channels == groups and not use_cudnn:
l_type = 'depthwise_conv'
helper = LayerHelper(l_type, **locals())
dtype = helper.input_dtype()
if groups is None:
num_filter_channels = num_channels
else:
......@@ -869,7 +905,7 @@ def conv2d(input,
pre_bias = helper.create_tmp_variable(dtype)
helper.append_op(
type='conv2d',
type=l_type,
inputs={
'Input': input,
'Filter': filter_param,
......@@ -1184,13 +1220,51 @@ def conv2d_transpose(input,
use_cudnn=True,
name=None):
"""
The transpose of conv2d layer.
**Convlution2D transpose layer**
The convolution2D transpose layer calculates the output based on the input,
filter, and dilations, strides, paddings. Input(Input) and output(Output)
are in NCHW format. Where N is batch size, C is the number of channels,
H is the height of the feature, and W is the width of the feature.
Parameters(dilations, strides, paddings) are two elements. These two elements
represent height and width, respectively. The details of convolution transpose
layer, please refer to the following explanation and references `therein <http://www.matthewzeiler.com/wp-content/uploads/2017/07/cvpr2010.pdf>`_.
For each input :math:`X`, the equation is:
.. math::
Out = W \\ast X
In the above equation:
* :math:`X`: Input value, a tensor with NCHW format.
* :math:`W`: Filter value, a tensor with MCHW format.
* :math:`\\ast` : Convolution transpose operation.
* :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.
Example:
- Input:
Input shape: $(N, C_{in}, H_{in}, W_{in})$
Filter shape: $(C_{in}, C_{out}, H_f, W_f)$
- Output:
Output shape: $(N, C_{out}, H_{out}, W_{out})$
Where
.. math::
This layer is also known as deconvolution layer.
H_{out} &= (H_{in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (H_f - 1) + 1 \\\\
W_{out} &= (W_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (W_f - 1) + 1
Args:
input(Variable): The input image with [N, C, H, W] format.
num_filters(int): The number of filter. It is as same as the output
num_filters(int): The number of the filter. It is as same as the output
image channel.
output_size(int|tuple|None): The output image size. If output size is a
tuple, it must contain two integers, (image_H, image_W). This
......@@ -1198,24 +1272,33 @@ def conv2d_transpose(input,
filter_size(int|tuple|None): The filter size. If filter_size is a tuple,
it must contain two integers, (filter_size_H, filter_size_W).
Otherwise, the filter will be a square. None if use output size to
calculate filter_size
calculate filter_size.
padding(int|tuple): The padding size. If padding is a tuple, it must
contain two integers, (padding_H, padding_W). Otherwise, the
padding_H = padding_W = padding.
padding_H = padding_W = padding. Default: padding = 0.
stride(int|tuple): The stride size. If stride is a tuple, it must
contain two integers, (stride_H, stride_W). Otherwise, the
stride_H = stride_W = stride.
stride_H = stride_W = stride. Default: stride = 1.
dilation(int|tuple): The dilation size. If dilation is a tuple, it must
contain two integers, (dilation_H, dilation_W). Otherwise, the
dilation_H = dilation_W = dilation.
param_attr: Parameter Attribute.
dilation_H = dilation_W = dilation. Default: dilation = 1.
param_attr(ParamAttr): The parameters to the Conv2d_transpose Layer. Default: None
use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
library is installed. Default: True
name(str|None): A name for this layer(optional). If set None, the layer
will be named automatically.
Returns:
Variable: Output image.
Variable: The tensor variable storing the convolution transpose result.
Raises:
ValueError: If the shapes of input, filter_size, stride, padding and groups mismatch.
Examples:
.. code-block:: python
data = fluid.layers.data(name='data', shape=[3, 32, 32], dtype='float32')
conv2d_transpose = fluid.layers.conv2d_transpose(input=data, num_filters=2, filter_size=3)
"""
helper = LayerHelper("conv2d_transpose", **locals())
if not isinstance(input, Variable):
......@@ -1866,6 +1949,146 @@ def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
return out
def edit_distance(input,
label,
normalized=False,
ignored_tokens=None,
name=None):
"""
EditDistance operator computes the edit distances between a batch of hypothesis strings and their references. Edit distance, also called Levenshtein distance, measures how dissimilar two strings are by counting the minimum number of operations to transform one string into anthor. Here the operations include insertion, deletion, and substitution. For example, given hypothesis string A = "kitten" and reference B = "sitting", the edit distance is 3 for A will be transformed into B at least after two substitutions and one insertion:
"kitten" -> "sitten" -> "sittin" -> "sitting"
Input(Hyps) is a LoDTensor consisting of all the hypothesis strings with the total number denoted by `batch_size`, and the separation is specified by the LoD information. And the `batch_size` reference strings are arranged in order in the same way in the LoDTensor Input(Refs).
Output(Out) contains the `batch_size` results and each stands for the edit stance for a pair of strings respectively. If Attr(normalized) is true, the edit distance will be divided by the length of reference string.
Args:
input(Variable): The indices for hypothesis strings.
label(Variable): The indices for reference strings.
normalized(bool): Indicated whether to normalize the edit distance by the length of reference string.
ignored_tokens(list of int): Tokens that should be removed before calculating edit distance.
Returns:
Variable: sequence-to-sequence edit distance in shape [batch_size, 1].
Examples:
.. code-block:: python
x = fluid.layers.data(name='x', shape=[8], dtype='float32')
y = fluid.layers.data(name='y', shape=[7], dtype='float32')
cost = fluid.layers.edit_distance(input=x,label=y)
"""
helper = LayerHelper("edit_distance", **locals())
# remove some tokens from input and labels
if ignored_tokens is not None and len(ignored_tokens) > 0:
erased_input = helper.create_tmp_variable(dtype="int64")
erased_label = helper.create_tmp_variable(dtype="int64")
helper.append_op(
type="sequence_erase",
inputs={"X": [input]},
outputs={"Out": [erased_input]},
attrs={"tokens": ignored_tokens})
input = erased_input
helper.append_op(
type="sequence_erase",
inputs={"X": [label]},
outputs={"Out": [erase_label]},
attrs={"tokens": ignored_tokens})
label = erased_label
# edit distance op
edit_distance_out = helper.create_tmp_variable(dtype="int64")
sequence_num = helper.create_tmp_variable(dtype="int64")
helper.append_op(
type="edit_distance",
inputs={"Hyps": [input],
"Refs": [label]},
outputs={"Out": [edit_distance_out],
"SequenceNum": [sequence_num]},
attrs={"normalized": normalized})
return edit_distance_out, sequence_num
def ctc_greedy_decoder(input, blank, name=None):
"""
This op is used to decode sequences by greedy policy by below steps:
1. Get the indexes of max value for each row in input. a.k.a. numpy.argmax(input, axis=0).
2. For each sequence in result of step1, merge repeated tokens between two blanks and delete all blanks.
A simple example as below:
.. code-block:: text
Given:
input.data = [[0.6, 0.1, 0.3, 0.1],
[0.3, 0.2, 0.4, 0.1],
[0.1, 0.5, 0.1, 0.3],
[0.5, 0.1, 0.3, 0.1],
[0.5, 0.1, 0.3, 0.1],
[0.2, 0.2, 0.2, 0.4],
[0.2, 0.2, 0.1, 0.5],
[0.5, 0.1, 0.3, 0.1]]
input.lod = [[0, 4, 8]]
Then:
output.data = [[2],
[1],
[3]]
output.lod = [[0, 2, 3]]
Args:
input(Variable): (LoDTensor<float>), the probabilities of variable-length sequences, which is a 2-D Tensor with LoD information. It's shape is [Lp, num_classes + 1], where Lp is the sum of all input sequences' length and num_classes is the true number of classes. (not including the blank label).
blank(int): the blank label index of Connectionist Temporal Classification (CTC) loss, which is in thehalf-opened interval [0, num_classes + 1).
Returns:
Variable: CTC greedy decode result.
Examples:
.. code-block:: python
x = fluid.layers.data(name='x', shape=[8], dtype='float32')
cost = fluid.layers.ctc_greedy_decoder(input=x, blank=0)
"""
helper = LayerHelper("ctc_greedy_decoder", **locals())
# top 1 op
topk_out = helper.create_tmp_variable(dtype=input.dtype)
topk_indices = helper.create_tmp_variable(dtype="int64")
helper.append_op(
type="top_k",
inputs={"X": [input]},
outputs={"Out": [topk_out],
"Indices": [topk_indices]},
attrs={"k": 1})
# ctc align op
ctc_out = helper.create_tmp_variable(dtype="int64")
helper.append_op(
type="ctc_align",
inputs={"Input": [topk_indices]},
outputs={"Output": [ctc_out]},
attrs={"merge_repeated": True,
"blank": blank})
return ctc_out
def warpctc(input, label, blank=0, norm_by_times=False, **kwargs):
"""
An operator integrating the open source Warp-CTC library
......@@ -1890,7 +2113,7 @@ def warpctc(input, label, blank=0, norm_by_times=False, **kwargs):
Temporal Classification (CTC) loss, which is in the
half-opened interval [0, num_classes + 1).
norm_by_times: (bool, default: false), whether to normalize
the gradients by the number of time-step,which is also the
the gradients by the number of time-step, which is also the
sequence's length. There is no need to normalize the gradients
if warpctc layer was follewed by a mean_op.
......@@ -1971,3 +2194,41 @@ def sequence_reshape(input, new_dim):
outputs={'Out': [out]},
attrs={'new_dim': new_dim})
return out
def transpose(x, perm, name=None):
"""
**transpose Layer**
Permute the dimensions of `input` according to `perm`.
The `i`-th dimension of the returned tensor will correspond to the
perm[i]-th dimension of `input`.
Args:
input (Variable): (Tensor), A Tensor.
perm (list): A permutation of the dimensions of `input`.
Returns:
Variable: A transposed Tensor.
Examples:
.. code-block:: python
x = fluid.layers.data(name='x', shape=[5, 10, 15], dtype='float32')
x_transposed = layers.transpose(x, perm=[1, 0, 2])
"""
if len(perm) != len(x.shape):
raise ValueError(
"Input(perm) is the permutation of dimensions of Input(input). "
"It's length shoud be equal to Input(input)'s rank.")
helper = LayerHelper('transpose', **locals())
out = helper.create_tmp_variable(x.dtype)
helper.append_op(
type='transpose',
inputs={'X': [x]},
outputs={'Out': [out]},
attrs={'axis': perm})
return out
python/paddle/v2/fluid/layers/ops.py
浏览文件 @ aece2905
......@@ -45,10 +45,20 @@ __activations__ = [
]
__all__ = [
'mean', 'mul', 'reshape', 'scale', 'transpose',
'sigmoid_cross_entropy_with_logits', 'elementwise_add', 'elementwise_div',
'elementwise_sub', 'elementwise_mul', 'elementwise_max', 'elementwise_min',
'clip', 'clip_by_norm', 'sequence_softmax'
'mean',
'mul',
'reshape',
'scale',
'sigmoid_cross_entropy_with_logits',
'elementwise_add',
'elementwise_div',
'elementwise_sub',
'elementwise_mul',
'elementwise_max',
'elementwise_min',
'clip',
'clip_by_norm',
'sequence_softmax',
] + __activations__
for _OP in set(__all__):
......
python/paddle/v2/fluid/tests/book/test_understand_sentiment_lstm.py
浏览文件 @ aece2905
......@@ -65,13 +65,13 @@ def lstm_net(dict_dim, class_dim=2, emb_dim=32, seq_len=80, batch_size=50):
emb = fluid.layers.embedding(input=data, size=[dict_dim, emb_dim])
emb = fluid.layers.reshape(x=emb, shape=[batch_size, seq_len, emb_dim])
emb = fluid.layers.transpose(x=emb, axis=[1, 0, 2])
emb = fluid.layers.transpose(x=emb, perm=[1, 0, 2])
c_pre_init = fluid.layers.fill_constant(
dtype=emb.dtype, shape=[batch_size, emb_dim], value=0.0)
c_pre_init.stop_gradient = False
layer_1_out = lstm(emb, c_pre_init=c_pre_init, hidden_dim=emb_dim)
layer_1_out = fluid.layers.transpose(x=layer_1_out, axis=[1, 0, 2])
layer_1_out = fluid.layers.transpose(x=layer_1_out, perm=[1, 0, 2])
prediction = fluid.layers.fc(input=layer_1_out,
size=class_dim,
......
python/paddle/v2/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py
浏览文件 @ aece2905
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
import paddle.v2 as paddle
import paddle.v2.fluid as fluid
......
python/paddle/v2/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py
浏览文件 @ aece2905
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import print_function
import sys
......
python/paddle/v2/fluid/tests/test_bipartite_match_op.py 0 → 100644
浏览文件 @ aece2905
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
#Licensed under the Apache License, Version 2.0 (the "License");
#you may not use this file except in compliance with the License.
#You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
#Unless required by applicable law or agreed to in writing, software
#distributed under the License is distributed on an "AS IS" BASIS,
#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#See the License for the specific language governing permissions and
#limitations under the License.
import unittest
import numpy as np
from op_test import OpTest
def bipartite_match(distance, match_indices, match_dist):
"""Bipartite Matching algorithm.
Arg:
distance (numpy.array) : The distance of two entries with shape [M, N].
match_indices (numpy.array): the matched indices from column to row
with shape [1, N], it must be initialized to -1.
match_dist (numpy.array): The matched distance from column to row
with shape [1, N], it must be initialized to 0.
"""
match_pair = []
row, col = distance.shape
for i in range(row):
for j in range(col):
match_pair.append((i, j, distance[i][j]))
match_sorted = sorted(match_pair, key=lambda tup: tup[2], reverse=True)
row_indices = -1 * np.ones((row, ), dtype=np.int)
idx = 0
for i, j, dist in match_sorted:
if idx >= row:
break
if match_indices[j] == -1 and row_indices[i] == -1 and dist > 0:
match_indices[j] = i
row_indices[i] = j
match_dist[j] = dist
idx += 1
def batch_bipartite_match(distance, lod):
"""Bipartite Matching algorithm for batch input.
Arg:
distance (numpy.array) : The distance of two entries with shape [M, N].
lod (list of int): The offsets of each input in this batch.
"""
n = len(lod) - 1
m = distance.shape[1]
match_indices = -1 * np.ones((n, m), dtype=np.int)
match_dist = np.zeros((n, m), dtype=np.float32)
for i in range(len(lod) - 1):
bipartite_match(distance[lod[i]:lod[i + 1], :], match_indices[i, :],
match_dist[i, :])
return match_indices, match_dist
class TestBipartiteMatchOpForWithLoD(OpTest):
def setUp(self):
self.op_type = 'bipartite_match'
lod = [[0, 5, 11, 23]]
dist = np.random.random((23, 217)).astype('float32')
match_indices, match_dist = batch_bipartite_match(dist, lod[0])
self.inputs = {'DistMat': (dist, lod)}
self.outputs = {
'ColToRowMatchIndices': (match_indices),
'ColToRowMatchDis': (match_dist),
}
def test_check_output(self):
self.check_output()
class TestBipartiteMatchOpWithoutLoD(OpTest):
def setUp(self):
self.op_type = 'bipartite_match'
lod = [[0, 8]]
dist = np.random.random((8, 17)).astype('float32')
match_indices, match_dist = batch_bipartite_match(dist, lod[0])
self.inputs = {'DistMat': dist}
self.outputs = {
'ColToRowMatchIndices': match_indices,
'ColToRowMatchDis': match_dist,
}
def test_check_output(self):
self.check_output()
if __name__ == '__main__':
unittest.main()
python/paddle/v2/fluid/tests/test_conv2d_op.py
浏览文件 @ aece2905
......@@ -241,6 +241,30 @@ class TestCUDNNWith1x1(TestWith1x1):
self.op_type = "conv2d"
class TestDepthwiseConv(TestConv2dOp):
def init_test_case(self):
self.pad = [1, 1]
self.stride = [2, 2]
self.input_size = [2, 3, 5, 5] # NCHW
self.groups = 3
assert np.mod(self.input_size[1], self.groups) == 0
f_c = self.input_size[1] / self.groups
self.filter_size = [6, f_c, 3, 3]
self.op_type = "depthwise_conv"
class TestDepthwiseConv2(TestConv2dOp):
def init_test_case(self):
self.pad = [1, 1]
self.stride = [1, 1]
self.input_size = [2, 3, 5, 5] # NCHW
self.groups = 3
assert np.mod(self.input_size[1], self.groups) == 0
f_c = self.input_size[1] / self.groups
self.filter_size = [6, f_c, 3, 3]
self.op_type = "depthwise_conv"
# cudnn v5 does not support dilation conv.
# class TestCUDNNWithDilation(TestWithDilation):
# def init_op_type(self):
......
python/paddle/v2/fluid/tests/test_edit_distance_op.py
浏览文件 @ aece2905
......@@ -61,6 +61,7 @@ class TestEditDistanceOp(OpTest):
num_strs = len(x1_lod) - 1
distance = np.zeros((num_strs, 1)).astype("float32")
sequence_num = np.array(2).astype("int64")
for i in range(0, num_strs):
distance[i] = Levenshtein(
hyp=x1[x1_lod[i]:x1_lod[i + 1]],
......@@ -70,7 +71,7 @@ class TestEditDistanceOp(OpTest):
distance[i] = distance[i] / len_ref
self.attrs = {'normalized': normalized}
self.inputs = {'Hyps': (x1, [x1_lod]), 'Refs': (x2, [x2_lod])}
self.outputs = {'Out': distance}
self.outputs = {'Out': distance, 'SequenceNum': sequence_num}
def test_check_output(self):
self.check_output()
......@@ -89,6 +90,7 @@ class TestEditDistanceOpNormalized(OpTest):
num_strs = len(x1_lod) - 1
distance = np.zeros((num_strs, 1)).astype("float32")
sequence_num = np.array(3).astype("int64")
for i in range(0, num_strs):
distance[i] = Levenshtein(
hyp=x1[x1_lod[i]:x1_lod[i + 1]],
......@@ -98,7 +100,7 @@ class TestEditDistanceOpNormalized(OpTest):
distance[i] = distance[i] / len_ref
self.attrs = {'normalized': normalized}
self.inputs = {'Hyps': (x1, [x1_lod]), 'Refs': (x2, [x2_lod])}
self.outputs = {'Out': distance}
self.outputs = {'Out': distance, 'SequenceNum': sequence_num}
def test_check_output(self):
self.check_output()
......
python/paddle/v2/fluid/tests/test_gradient_clip.py
浏览文件 @ aece2905
......@@ -40,7 +40,8 @@ p_g = fluid.backward.append_backward(loss=avg_cost)
p_g_clip = fluid.backward.append_backward(loss=avg_cost_clip)
with fluid.program_guard(main_program=prog_clip):
fluid.clip.gradient_clip_by_global_norm(clip_norm=CLIP)
fluid.clip.set_gradient_clip(
fluid.clip.GradientClipByGlobalNorm(clip_norm=CLIP))
p_g_clip = fluid.clip.append_gradient_clip_ops(p_g_clip)
grad_list = [elem[1] for elem in p_g]
......
python/paddle/v2/fluid/tests/test_im2sequence_op.py 0 → 100644
浏览文件 @ aece2905
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
#Licensed under the Apache License, Version 2.0 (the "License");
#you may not use this file except in compliance with the License.
#You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
#Unless required by applicable law or agreed to in writing, software
#distributed under the License is distributed on an "AS IS" BASIS,
#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#See the License for the specific language governing permissions and
#limitations under the License.
import unittest
import numpy as np
from op_test import OpTest
def get_output_shape(attrs, in_shape):
img_height = in_shape[2]
img_width = in_shape[3]
paddings = attrs['paddings']
kernels = attrs['kernels']
strides = attrs['strides']
output_height = \
1 + \
(img_height + paddings[0] + paddings[2] - kernels[0] + strides[0] - 1) / \
strides[0]
output_width = \
1 + \
(img_width + paddings[1] + paddings[3] - kernels[1] + strides[1] - 1) / \
strides[1]
return output_height, output_width
def im2col(attrs, im, col):
"""
im: {CHW}
col:
{outputHeight, outputWidth, inputChannels, filterHeight, filterWidth}
"""
input_channels, input_height, input_width = im.shape
output_height, output_width, _, filter_height, filter_width = col.shape
stride_height, stride_width = attrs['strides']
padding_height, padding_width = attrs['paddings'][0:2]
for col_row_idx in range(0, output_height):
for col_col_idx in range(0, output_width):
for channel in range(0, input_channels):
for filter_row_idx in range(0, filter_height):
for filter_col_idx in range(0, filter_width):
im_row_offset = col_row_idx * stride_height \
+ filter_row_idx - padding_height
im_col_offset = col_col_idx * stride_width \
+ filter_col_idx - padding_width
if (im_row_offset < 0 or
im_row_offset >= input_height or
im_col_offset < 0 or
im_col_offset >= input_width):
col[col_row_idx][col_col_idx][channel][\
filter_row_idx][filter_col_idx] = 0.0
else:
im_offset = (channel * input_height + im_row_offset \
) * input_width + im_col_offset
col[col_row_idx][col_col_idx][channel][\
filter_row_idx][filter_col_idx] = im[channel][ \
im_row_offset][im_col_offset]
def Im2Sequence(inputs, attrs):
output_height, output_width = get_output_shape(attrs, inputs.shape)
img_channels = inputs.shape[1]
batch_size = inputs.shape[0]
out = np.zeros([
batch_size, output_height, output_width, img_channels,
attrs['kernels'][0], attrs['kernels'][1]
]).astype("float32")
for i in range(len(inputs)):
im2col(attrs, inputs[i], out[i])
out = out.reshape([
batch_size * output_height * output_width,
img_channels * attrs['kernels'][0] * attrs['kernels'][1]
])
return out
class TestBlockExpandOp(OpTest):
def config(self):
self.batch_size = 1
self.img_channels = 3
self.img_height = 4
self.img_width = 4
self.attrs = {
'kernels': [2, 2],
'strides': [1, 1],
'paddings': [1, 1, 1, 1]
}
def setUp(self):
self.config()
self.op_type = "im2sequence"
x = np.random.uniform(0.1, 1, [
self.batch_size, self.img_channels, self.img_height, self.img_width
]).astype("float32")
out = Im2Sequence(x, self.attrs)
self.inputs = {'X': x}
self.outputs = {'Out': out}
def test_check_output(self):
self.check_output()
def test_check_grad_normal(self):
self.check_grad(['X'], 'Out')
class TestBlockExpandOpCase2(TestBlockExpandOp):
def config(self):
self.batch_size = 2
self.img_channels = 3
self.img_height = 4
self.img_width = 5
self.attrs = {
'kernels': [2, 1],
'strides': [2, 1],
'paddings': [2, 1, 2, 1]
}
class TestBlockExpandOpCase3(TestBlockExpandOp):
def config(self):
self.batch_size = 3
self.img_channels = 1
self.img_height = 4
self.img_width = 5
self.attrs = {
'kernels': [2, 1],
'strides': [2, 1],
'paddings': [2, 0, 2, 0]
}
class TestBlockExpandOpCase4(TestBlockExpandOp):
def config(self):
self.batch_size = 2
self.img_channels = 2
self.img_height = 3
self.img_width = 3
self.attrs = {
'kernels': [2, 2],
'strides': [1, 1],
'paddings': [0, 0, 0, 0]
}
if __name__ == '__main__':
unittest.main()
python/paddle/v2/fluid/tests/test_reduce_op.py
浏览文件 @ aece2905
......@@ -20,7 +20,7 @@ from op_test import OpTest
class TestSumOp(OpTest):
def setUp(self):
self.op_type = "reduce_sum"
self.inputs = {'X': np.random.random((5, 6, 10)).astype("float32")}
self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")}
self.outputs = {'Out': self.inputs['X'].sum(axis=0)}
def test_check_output(self):
......@@ -33,7 +33,7 @@ class TestSumOp(OpTest):
class TestMeanOp(OpTest):
def setUp(self):
self.op_type = "reduce_mean"
self.inputs = {'X': np.random.random((5, 6, 2, 10)).astype("float32")}
self.inputs = {'X': np.random.random((5, 6, 2, 10)).astype("float64")}
self.attrs = {'dim': 1}
self.outputs = {'Out': self.inputs['X'].mean(axis=self.attrs['dim'])}
......@@ -49,7 +49,7 @@ class TestMaxOp(OpTest):
def setUp(self):
self.op_type = "reduce_max"
self.inputs = {'X': np.random.random((5, 6, 10)).astype("float32")}
self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")}
self.attrs = {'dim': -1}
self.outputs = {'Out': self.inputs['X'].max(axis=self.attrs['dim'])}
......@@ -62,7 +62,7 @@ class TestMinOp(OpTest):
def setUp(self):
self.op_type = "reduce_min"
self.inputs = {'X': np.random.random((5, 6, 10)).astype("float32")}
self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")}
self.attrs = {'dim': 2}
self.outputs = {'Out': self.inputs['X'].min(axis=self.attrs['dim'])}
......@@ -73,7 +73,7 @@ class TestMinOp(OpTest):
class TestKeepDimReduce(OpTest):
def setUp(self):
self.op_type = "reduce_sum"
self.inputs = {'X': np.random.random((5, 6, 10)).astype("float32")}
self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")}
self.attrs = {'dim': -2, 'keep_dim': True}
self.outputs = {
'Out': self.inputs['X'].sum(axis=self.attrs['dim'], keepdims=True)
......@@ -89,7 +89,7 @@ class TestKeepDimReduce(OpTest):
class Test1DReduce(OpTest):
def setUp(self):
self.op_type = "reduce_sum"
self.inputs = {'X': np.random.random(20).astype("float32")}
self.inputs = {'X': np.random.random(20).astype("float64")}
self.outputs = {'Out': self.inputs['X'].sum(axis=0)}
def test_check_output(self):
......@@ -102,7 +102,7 @@ class Test1DReduce(OpTest):
class TestReduceAll(OpTest):
def setUp(self):
self.op_type = "reduce_sum"
self.inputs = {'X': np.random.random((5, 6, 2, 10)).astype("float32")}
self.inputs = {'X': np.random.random((5, 6, 2, 10)).astype("float64")}
self.attrs = {'reduce_all': True}
self.outputs = {'Out': self.inputs['X'].sum()}
......
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