Skip to content

P
Paddle

PaddlePaddle / Paddle
大约 2 年 前同步成功

通知 2325
Star 20933
Fork 5424
P
Paddle
未验证 提交 e92fe9e6 编写于 作者: H helinwang 提交者： GitHub
浏览文件
操作

Merge branch 'develop' into remote_optimize

上级 151f579c 2ac46d53
隐藏空白更改
内联 并排
Showing with 3867 addition and 3346 deletion
.copyright.hook
浏览文件 @ e92fe9e6
...@@ -77,10 +77,13 @@ def lang_type(filename): ...@@ -77,10 +77,13 @@ def lang_type(filename):
elif filename.endswith(".proto"): elif filename.endswith(".proto"):
return "C" return "C"
else: else:
print("Unsupported filetype") print("Unsupported filetype %s", filename)
exit(0) exit(0)
PYTHON_ENCODE = re.compile("^[ \t\v]*#.*?coding[:=][ \t]*([-_.a-zA-Z0-9]+)")
def main(argv=None): def main(argv=None):
parser = argparse.ArgumentParser( parser = argparse.ArgumentParser(
description='Checker for copyright declaration.') description='Checker for copyright declaration.')
...@@ -89,8 +92,15 @@ def main(argv=None): ...@@ -89,8 +92,15 @@ def main(argv=None):
retv = 0 retv = 0
for filename in args.filenames: for filename in args.filenames:
first_line = io.open(filename).readline() fd = io.open(filename)
if "COPYRIGHT" in first_line.upper() : continue first_line = fd.readline()
if "COPYRIGHT" in first_line.upper(): continue
if filename.endswith(".py"):
second_line = fd.readline()
if first_line.startswith("#!") or PYTHON_ENCODE.match(
second_line) != None or PYTHON_ENCODE.match(
first_line) != None:
continue
original_contents = io.open(filename).read() original_contents = io.open(filename).read()
new_contents = generate_copyright( new_contents = generate_copyright(
COPYRIGHT, lang_type(filename)) + original_contents COPYRIGHT, lang_type(filename)) + original_contents
......
CODE_OF_CONDUCT_cn.md
浏览文件 @ e92fe9e6
# 貢獻者公約 # 参与者公约
## 我們的承諾 ## 我们的保证
為了促進一個開放透明且受歡迎的環境，我們作為貢獻者和維護者保證，無論年齡、種族、民族、性別認同和表達、體型、殘疾、經驗水平、國籍、個人表現、宗教或性別取向，在我們的專案以及社群的參與者都有不被騷擾的體驗。 为了促进一个开放透明且友好的环境，我们作为贡献者和维护者保证：无论年龄、种族、民族、性别认同和表达（方式）、体型、身体健全与否、经验水平、国籍、个人表现、宗教或性别取向，参与者在我们项目和社区中都免于骚扰。
## 我們的準則 ## 我们的标准
舉例來說有助於創造正面環境的行為包括： 有助于创造正面环境的行为包括但不限于：
* 使用歡迎和包容性語言 * 使用友好和包容性语言
* 尊重不同的觀點和經驗 * 尊重不同的观点和经历
* 優雅地接受建設性批評 * 耐心地接受建设性批评
* 關注在對於社群最好的事情上 * 关注对社区最有利的事情
* 對其他社群成員的表現友善 * 友善对待其他社区成员
舉例來說身為參與者不能接受的行為包括： 身为参与者不能接受的行为包括但不限于：
* 使用與性有關的言語或是圖像，以及不受歡迎的性騷擾 * 使用与性有关的言语或是图像，以及不受欢迎的性骚扰
* 酸民/反串/釣魚行為或進行侮辱/貶損的評論，人身攻擊及政治攻擊 * 捣乱/煽动/造谣的行为或进行侮辱/贬损的评论，人身攻击及政治攻击
* 公開或私下的騷擾 * 公开或私下的骚扰
* 未經許可地發布他人的個人資料，例如住址或是電子地址 * 未经许可地发布他人的个人资料，例如住址或是电子地址
* 其他可以被合理地認定為不恰當或者違反職業操守的行為 * 其他可以被合理地认定为不恰当或者违反职业操守的行为
## 我們的責任 ## 我们的责任
專案維護者有責任為"可接受的行為"準則做出詮釋，以及對已發生的不被接受的行為採取恰當且公平的糾正措施。 项目维护者有责任为「可接受的行为」标准做出诠释，以及对已发生的不被接受的行为采取恰当且公平的纠正措施。
專案維護者有權力及責任去刪除、編輯、拒絕與本行為準則有所違背的評論(comments)、提交(commits)、程式碼、wiki 編輯、問題(issues)和其他貢獻，以及專案維護者可暫時或永久性的禁止任何他們認為有不適當、威脅、冒犯、有害行為的貢獻者。 项目维护者有权利及责任去删除、编辑、拒绝与本行为标准有所违背的评论(comments)、提交(commits)、代码、wiki 编辑、问题(issues)和其他贡献，以及项目维护者可暂时或永久性的禁止任何他们认为有不适当、威胁、冒犯、有害行为的贡献者。
## 使用範圍 ## 使用范围
當一個人代表該專案或是其社群時，本行為準則適用於其專案平台和公共平台。 当一个人代表该项目或是其社区时，本行为标准适用于其项目平台和公共平台。
代表專案或是社群的情況，舉例來說包括使用官方專案的電子郵件地址、通過官方的社群媒體帳號發布或線上或線下事件中擔任指定代表。 代表项目或是社区的情况，举例来说包括使用官方项目的电子邮件地址、通过官方的社区媒体账号发布或线上或线下事件中担任指定代表。
該專案的呈現方式可由其專案維護者進行進一步的定義及解釋。 该项目的呈现方式可由其项目维护者进行进一步的定义及解释。
## 強制執行 ## 强制执行
可以透過paddle-dev@baidu.com，來聯繫專案團隊來報告濫用、騷擾或其他不被接受的行為。 可以通过paddle-dev@baidu.com，来联系项目团队来举报滥用、骚扰或其他不被接受的行为。
任何維護團隊認為有必要且適合的所有投訴都將進行審查及調查，並做出相對應的回應。專案小組有對事件回報者有保密的義務。具體執行的方針近一步細節可能會單獨公佈。 任何维护团队认为有必要且适合的所有投诉都将进行审查及调查，并做出相对应的回应。项目小组有对事件回报者有保密的义务。具体执行的方针近一步细节可能会单独公布。
沒有真誠的遵守或是執行本行為準則的專案維護人員，可能會因專案領導人或是其他成員的決定，暫時或是永久的取消其身份。 没有切实地遵守或是执行本行为标准的项目维护人员，可能会因项目领导人或是其他成员的决定，暂时或是永久地取消其参与资格。
## 來源 ## 来源
本行為準則改編自[貢獻者公約][首頁]，版本 1.4 本行为标准改编自[贡献者公约][主页]，版本 1.4
可在此觀看https://www.contributor-covenant.org/zh-tw/version/1/4/code-of-conduct.html 可在此观看https://www.contributor-covenant.org/zh-cn/version/1/4/code-of-conduct.html
[首頁]: https://www.contributor-covenant.org [主页]: https://www.contributor-covenant.org
adversarial/README.md 已删除 100644 → 0
浏览文件 @ 151f579c
# Advbox
Advbox is a Python toolbox to create adversarial examples that fool neural networks. It requires Python and paddle.
## How to use
1. train a model and save it's parameters. (like fluid_mnist.py)
2. load the parameters which is trained in step1, then reconstruct the model.(like mnist_tutorial_fgsm.py)
3. use advbox to generate the adversarial sample.
adversarial/advbox/__init__.py 已删除 100644 → 0
浏览文件 @ 151f579c
# Copyright (c) 2017 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.
"""
A set of tools for generating adversarial example on paddle platform
"""
adversarial/advbox/attacks/base.py 已删除 100644 → 0
浏览文件 @ 151f579c
# 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.
"""
The base model of the model.
"""
from abc import ABCMeta, abstractmethod
class Attack(object):
"""
Abstract base class for adversarial attacks. `Attack` represent an adversarial attack
which search an adversarial example. subclass should implement the _apply() method.
Args:
model(Model): an instance of the class advbox.base.Model.
"""
__metaclass__ = ABCMeta
def __init__(self, model):
self.model = model
def __call__(self, image_label):
"""
Generate the adversarial sample.
Args:
image_label(list): The image and label tuple list with one element.
"""
adv_img = self._apply(image_label)
return adv_img
@abstractmethod
def _apply(self, image_label):
"""
Search an adversarial example.
Args:
image_batch(list): The image and label tuple list with one element.
"""
raise NotImplementedError
adversarial/advbox/attacks/gradientsign.py 已删除 100644 → 0
浏览文件 @ 151f579c
# 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.
"""
This module provide the attack method for FGSM's implement.
"""
from __future__ import division
import numpy as np
from collections import Iterable
from .base import Attack
class GradientSignAttack(Attack):
"""
This attack was originally implemented by Goodfellow et al. (2015) with the
infinity norm (and is known as the "Fast Gradient Sign Method"). This is therefore called
the Fast Gradient Method.
Paper link: https://arxiv.org/abs/1412.6572
"""
def _apply(self, image_label, epsilons=1000):
assert len(image_label) == 1
pre_label = np.argmax(self.model.predict(image_label))
min_, max_ = self.model.bounds()
gradient = self.model.gradient(image_label)
gradient_sign = np.sign(gradient) * (max_ - min_)
if not isinstance(epsilons, Iterable):
epsilons = np.linspace(0, 1, num=epsilons + 1)
for epsilon in epsilons:
adv_img = image_label[0][0].reshape(
gradient_sign.shape) + epsilon * gradient_sign
adv_img = np.clip(adv_img, min_, max_)
adv_label = np.argmax(self.model.predict([(adv_img, 0)]))
if pre_label != adv_label:
return adv_img
FGSM = GradientSignAttack
class IteratorGradientSignAttack(Attack):
"""
This attack was originally implemented by Alexey Kurakin(Google Brain).
Paper link: https://arxiv.org/pdf/1607.02533.pdf
"""
def _apply(self, image_label, epsilons=100, steps=10):
"""
Apply the iterative gradient sign attack.
Args:
image_label(list): The image and label tuple list of one element.
epsilons(list|tuple|int): The epsilon (input variation parameter).
steps(int): The number of iterator steps.
Return:
numpy.ndarray: The adversarail sample generated by the algorithm.
"""
assert len(image_label) == 1
pre_label = np.argmax(self.model.predict(image_label))
gradient = self.model.gradient(image_label)
min_, max_ = self.model.bounds()
if not isinstance(epsilons, Iterable):
epsilons = np.linspace(0, 1, num=epsilons + 1)
for epsilon in epsilons:
adv_img = image_label[0][0].reshape(gradient.shape)
for _ in range(steps):
gradient = self.model.gradient([(adv_img, image_label[0][1])])
gradient_sign = np.sign(gradient) * (max_ - min_)
adv_img = adv_img + epsilon * gradient_sign
adv_img = np.clip(adv_img, min_, max_)
adv_label = np.argmax(self.model.predict([(adv_img, 0)]))
if pre_label != adv_label:
return adv_img
adversarial/advbox/models/__init__.py 已删除 100644 → 0
浏览文件 @ 151f579c
# Copyright (c) 2017 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.
"""
Paddle model for target of attack
"""
adversarial/advbox/models/base.py 已删除 100644 → 0
浏览文件 @ 151f579c
# 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.
"""
The base model of the model.
"""
from abc import ABCMeta
import abc
abstractmethod = abc.abstractmethod
class Model(object):
"""
Base class of model to provide attack.
Args:
bounds(tuple): The lower and upper bound for the image pixel.
channel_axis(int): The index of the axis that represents the color channel.
preprocess(tuple): Two element tuple used to preprocess the input. First
substract the first element, then divide the second element.
"""
__metaclass__ = ABCMeta
def __init__(self, bounds, channel_axis, preprocess=None):
assert len(bounds) == 2
assert channel_axis in [0, 1, 2, 3]
if preprocess is None:
preprocess = (0, 1)
self._bounds = bounds
self._channel_axis = channel_axis
self._preprocess = preprocess
def bounds(self):
"""
Return the upper and lower bounds of the model.
"""
return self._bounds
def channel_axis(self):
"""
Return the channel axis of the model.
"""
return self._channel_axis
def _process_input(self, input_):
res = input_
sub, div = self._preprocess
if sub != 0:
res = input_ - sub
assert div != 0
if div != 1:
res /= div
return res
@abstractmethod
def predict(self, image_batch):
"""
Calculate the prediction of the image batch.
Args:
image_batch(numpy.ndarray): image batch of shape (batch_size, height, width, channels).
Return:
numpy.ndarray: predictions of the images with shape (batch_size, num_of_classes).
"""
raise NotImplementedError
@abstractmethod
def num_classes(self):
"""
Determine the number of the classes
Return:
int: the number of the classes
"""
raise NotImplementedError
@abstractmethod
def gradient(self, image_batch):
"""
Calculate the gradient of the cross-entropy loss w.r.t the image.
Args:
image_batch(list): The image and label tuple list.
Return:
numpy.ndarray: gradient of the cross-entropy loss w.r.t the image with
the shape (height, width, channel).
"""
raise NotImplementedError
adversarial/advbox/models/paddle.py 已删除 100644 → 0
浏览文件 @ 151f579c
# 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 absolute_import
import numpy as np
import paddle.v2 as paddle
import paddle.v2.fluid as fluid
from paddle.v2.fluid.framework import program_guard
from .base import Model
class PaddleModel(Model):
"""
Create a PaddleModel instance.
When you need to generate a adversarial sample, you should construct an instance of PaddleModel.
Args:
program(paddle.v2.fluid.framework.Program): The program of the model which generate the adversarial sample.
input_name(string): The name of the input.
logits_name(string): The name of the logits.
predict_name(string): The name of the predict.
cost_name(string): The name of the loss in the program.
"""
def __init__(self,
program,
input_name,
logits_name,
predict_name,
cost_name,
bounds,
channel_axis=3,
preprocess=None):
super(PaddleModel, self).__init__(
bounds=bounds, channel_axis=channel_axis, preprocess=preprocess)
if preprocess is None:
preprocess = (0, 1)
self._program = program
self._place = fluid.CPUPlace()
self._exe = fluid.Executor(self._place)
self._input_name = input_name
self._logits_name = logits_name
self._predict_name = predict_name
self._cost_name = cost_name
# gradient
loss = self._program.block(0).var(self._cost_name)
param_grads = fluid.backward.append_backward(
loss, parameter_list=[self._input_name])
self._gradient = dict(param_grads)[self._input_name]
def predict(self, image_batch):
"""
Predict the label of the image_batch.
Args:
image_batch(list): The image and label tuple list.
Return:
numpy.ndarray: predictions of the images with shape (batch_size, num_of_classes).
"""
feeder = fluid.DataFeeder(
feed_list=[self._input_name, self._logits_name],
place=self._place,
program=self._program)
predict_var = self._program.block(0).var(self._predict_name)
predict = self._exe.run(self._program,
feed=feeder.feed(image_batch),
fetch_list=[predict_var])
return predict
def num_classes(self):
"""
Calculate the number of classes of the output label.
Return:
int: the number of classes
"""
predict_var = self._program.block(0).var(self._predict_name)
assert len(predict_var.shape) == 2
return predict_var.shape[1]
def gradient(self, image_batch):
"""
Calculate the gradient of the loss w.r.t the input.
Args:
image_batch(list): The image and label tuple list.
Return:
list: The list of the gradient of the image.
"""
feeder = fluid.DataFeeder(
feed_list=[self._input_name, self._logits_name],
place=self._place,
program=self._program)
grad, = self._exe.run(self._program,
feed=feeder.feed(image_batch),
fetch_list=[self._gradient])
return grad
adversarial/mnist_tutorial_fgsm.py 已删除 100644 → 0
浏览文件 @ 151f579c
# 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.
"""
FGSM demos on mnist using advbox tool.
"""
import paddle.v2 as paddle
import paddle.v2.fluid as fluid
import matplotlib.pyplot as plt
import numpy as np
from advbox.models.paddle import PaddleModel
from advbox.attacks.gradientsign import GradientSignAttack
def cnn_model(img):
"""
Mnist cnn model
Args:
img(Varaible): the input image to be recognized
Returns:
Variable: the label prediction
"""
#conv1 = fluid.nets.conv2d()
conv_pool_1 = fluid.nets.simple_img_conv_pool(
input=img,
num_filters=20,
filter_size=5,
pool_size=2,
pool_stride=2,
act='relu')
conv_pool_2 = fluid.nets.simple_img_conv_pool(
input=conv_pool_1,
num_filters=50,
filter_size=5,
pool_size=2,
pool_stride=2,
act='relu')
logits = fluid.layers.fc(input=conv_pool_2, size=10, act='softmax')
return logits
def main():
"""
Advbox demo which demonstrate how to use advbox.
"""
IMG_NAME = 'img'
LABEL_NAME = 'label'
img = fluid.layers.data(name=IMG_NAME, shape=[1, 28, 28], dtype='float32')
# gradient should flow
img.stop_gradient = False
label = fluid.layers.data(name=LABEL_NAME, shape=[1], dtype='int64')
logits = cnn_model(img)
cost = fluid.layers.cross_entropy(input=logits, label=label)
avg_cost = fluid.layers.mean(x=cost)
place = fluid.CPUPlace()
exe = fluid.Executor(place)
BATCH_SIZE = 1
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.mnist.train(), buf_size=500),
batch_size=BATCH_SIZE)
feeder = fluid.DataFeeder(
feed_list=[IMG_NAME, LABEL_NAME],
place=place,
program=fluid.default_main_program())
fluid.io.load_params(
exe, "./mnist/", main_program=fluid.default_main_program())
# advbox demo
m = PaddleModel(fluid.default_main_program(), IMG_NAME, LABEL_NAME,
logits.name, avg_cost.name, (-1, 1))
att = GradientSignAttack(m)
for data in train_reader():
# fgsm attack
adv_img = att(data)
plt.imshow(n[0][0], cmap='Greys_r')
plt.show()
#np.save('adv_img', adv_img)
break
if __name__ == '__main__':
main()
doc/api/v2/fluid/layers.rst
浏览文件 @ e92fe9e6
...@@ -499,3 +499,13 @@ swish ...@@ -499,3 +499,13 @@ swish
------ ------
.. autofunction:: paddle.v2.fluid.layers.swish .. autofunction:: paddle.v2.fluid.layers.swish
:noindex: :noindex:
l2_normalize
------------
.. autofunction:: paddle.v2.fluid.layers.l2_normalize
:noindex:
sequence_reshape
----------------
.. autofunction:: paddle.v2.fluid.layers.sequence_reshape
:noindex:
doc/design/error_clip.md
浏览文件 @ e92fe9e6
...@@ -46,12 +46,12 @@ class ErrorClipByValue(BaseErrorClipAttr): ...@@ -46,12 +46,12 @@ class ErrorClipByValue(BaseErrorClipAttr):
self.min = min self.min = min
def append_clip_op(self, block, grad_name): def append_clip_op(self, block, grad_name):
block.append_op( clip_op_desc = block.desc.append_op()
type="clip", clip_op_desc.set_type("clip")
inputs={"X": grad_name}, clip_op_desc.set_input("X", [grad_name])
outputs={"Out": grad_name}, clip_op_desc.set_output("Out", [grad_name])
attrs={"min": self.min, clip_op_desc.set_attr("min", self.min)
"max": self.max}) clip_op_desc.set_attr("max", self.max)
``` ```
The `BaseErrorClipAttr` have one main member functions: `append_clip_op(self, block, grad_name)`. The `BaseErrorClipAttr` have one main member functions: `append_clip_op(self, block, grad_name)`.
...@@ -80,6 +80,11 @@ def error_clip_callback(block, context): ...@@ -80,6 +80,11 @@ def error_clip_callback(block, context):
op_desc.output_arg_names()): op_desc.output_arg_names()):
fwd_var = block.var_recursive(grad_to_var[grad_n]) fwd_var = block.var_recursive(grad_to_var[grad_n])
error_clip = getattr(fwd_var, "error_clip", None) error_clip = getattr(fwd_var, "error_clip", None)
if not (error_clip is None or isinstance(error_clip,
BaseErrorClipAttr)):
raise TypeError(
"Variable's error_clip should be an instance of BaseErrorClipAttr or None."
)
if error_clip is not None: if error_clip is not None:
error_clip.append_clip_op(block, grad_n) error_clip.append_clip_op(block, grad_n)
``` ```
......
doc/design/switch_kernel.md
浏览文件 @ e92fe9e6
## Background ## Background
Every operator has many kernels because there are multiple data types, places, data layout that Fluid supports. We use the `KernelType` to describe kernel types that operators can hold. Every operator has many kernels because there are multiple data types, places, data layout, library type that Fluid supports. We use the `OpKernelType ` to describe kernel types that operators can hold.
The `KernelType` is as follows. The `OpKernelType ` is as follows:
``` ```cpp
struct KernelType { struct OpKernelType {
Place place_; Place place_;
DataType data_type_; DataType data_type_;
LayoutType layout_; DataLayout data_layout_;
LibraryType library_type_;
}; };
``` ```
The `place_` is a descriptor of the device and the computational library, e.g., `MKLDNNPlace`, `CUDAPlace`. - The `place_` is a descriptor of the device, e.g., CPUPlace, CUDAPlace.
The `data_type_` is the data type that this kernel performs on, e.g., `FP32`, `INT64`. Note that one kernel may have inputs with different data types. However, it will be a major `data_type`. For example, the `cross_entropy` takes `int64` as it label, and `double`/`float` as its input logit and output cost. The major `data_type` of `cross_entropy` is `float`/`double`. - The `data_type_` is the data type that this kernel performs on, e.g., `FP32`, `INT64`. Note that one kernel may have inputs with different data types. However, it will be a major `data_type`. For example, the `cross_entropy` takes `int64` as it label, and `double`/`float` as its input logit and output cost. The major `data_type` of `cross_entropy` is `float` or `double`.
The `layout` is useful for some computational library. One example is that MKLDNN uses many kinds of layout, such as `nChw8c`. Each kind of layout will invoke the different kernel. - The `data_layout_ ` is useful for some computational library. One example is that MKLDNN uses many kinds of layout, such as `nChw8c`. Each kind of layout will invoke the different kernel.
- The `library_type_` describes the computational library, e.g., `MKLDNN`, `CUDNN`.
## Problem ## Problem
...@@ -25,42 +28,72 @@ We register a kernel for every operator and every kernel type ideally. However, ...@@ -25,42 +28,72 @@ We register a kernel for every operator and every kernel type ideally. However,
2. Some operators will take too many memory. It is better to force them into CPU. However, the rest of operators in this neural network will be performed on GPU, i.e., model parallel problem. 2. Some operators will take too many memory. It is better to force them into CPU. However, the rest of operators in this neural network will be performed on GPU, i.e., model parallel problem.
3. Some layout and place are particular. One example is that MKLDNN uses `nChw8` and there is no other library uses `nChw8c`. 3. Some layout and place are particular. One example is that MKLDNN uses `nChw8` and there is no other library uses `nChw8c`.
Problems under these situations are similar. We can formalise this problem as follow. Take one situation to give a detailed explanation, if we have two Operators: OP1 and OP2, OP1 has one output `op1_to_op2`, and `op1_to_op2` is the input of OP2.
If OP1 and OP2 run on the same place(for example CPUPlace), then `op1_2_op2` can be used directly by OP2.
```
OP1(CPUPlace)
|
op1_2_op2
|
OP2(CPUPlace)
```
If OP1 and OP2 run one different place, then OP2 cannot `use op1_2_op2` directly.
Problems under these situations are similar. We can formalize this problem as follow.
We register kernels with types $KT = \{kt_1, kt_2, kt_3, ...\}$ for one operator. The inputs of this operator should be run on kernel type $kt_{?}$, which the $kt_{?} \notin KT$. How to cast the input of this operator from $kt_{?}$ to any of kernel type in $KT$. We register kernels with types $KT = \{kt_1, kt_2, kt_3, ...\}$ for one operator. The inputs of this operator should be run on kernel type $kt_{?}$, which the $kt_{?} \notin KT$. How to cast the input of this operator from $kt_{?}$ to any of kernel type in $KT$.
## Solution ## Solution: data transform
It is clearly that transforming inputs of an operator toadapt another kernel type is not related to the particular operator. So we should register these transformation methods as global methods. It is clear that transforming inputs of an operator to adapt another kernel type is not related to the particular operator. So we should register these transformation methods as global methods.
We can infer a kernel type from the inputs of an operators. We let this kernel type as `actual kernel type`, which means this kernel type is the actually kernel type that operator should be performed. We can infer kernel type for each input of an operator. We let this kernel type as `actual kernel type for var`, which means this kernel type is the kernel type that can process this input variable.
We can get a kernel type by 1) The configuration of operator description. (Users may want to force use `MKL` for `conv` operator). 2) The place of the current executor. (Executor is running on GPU). This kernel type is what we expect the operator will be performed on. We let this kernel type as `expect kernel type`. We can get a kernel type by 1) The configuration of operator description. (Users may want to force use `MKL` for `conv` operator). 2) The place of the current executor. (Executor is running on GPU). This kernel type is what we expect the operator will be performed on. We let this kernel type as `expect kernel type`.
We transform the input data from `actual` to `expect` if the expect kernel type is not as same as actual kernel type. We transform the input data from `actual` to `expect` if the actual kernel type is not as same as expect kernel type.
The algorithm is described as follow The algorithm is described as following
```cpp ```cpp
using DataTransformationFN = std::function<void(const Tensor& in, Tensor* out)>; void OperatorWithKernel::Run(
using KernelTypePair = std::pair<KernelType, KernelType>; const Scope& scope,
const platform::Place& place) const {
map<KernelTypePair, DataTransformationFN> g_data_transformation_; ExecutionContext ctx(...);
auto expected_kernel_key = this->GetExpectedKernelType(ctx);
void OpWithKernel::Run() {
vec<Tensor> inputs = ... Scope& new_scope = scope.NewScope();
auto actual_kernel_type = GetActualKernelType(inputs);
for (auto& var_name : this->Inputs()) {
// The expected kernel type is related to actual kernel type. auto* tensor_in = GetTensor(var_name);
// For the most operators, the expected kernel type is as same as auto kernel_type_for_var = this->GetKernelTypeForVar(...);
// actual kernel type. if (kernel_type_for_var.place_ != expected_kernel_key.place_) {
// auto* trans_var = new_scope.Var(var_name);
// So we pass `actual_kernel_type` as a parameter of auto* out = DataTransform(expected_kernel_key,
// GetExpectedKernelType kernel_type_for_var,
auto expect_kernel_type = GetExpectedKernelType(actual_kernel_type); *tensor_in);
CopyVariableWithTensor(...);
auto trans = g_data_transformation_[{actual_kernel_type, expect_kernel_type}]; }
}
kernel.run(trans(inputs));
auto kernel = kernels.find(expected_kernel_key);
kernel->Compute(ExecutionContext(...));
} }
``` ```
then the actual process for the multi-device above will be:
```
OP1(CPUPlace)
|
op1_2_op2(on CPU)
|
[transform](from CPU to GPU)
|
op1_2_op2(on GPU)
|
OP2(CUDAPlace)
```
doc/howto/dev/new_op_en.md
浏览文件 @ e92fe9e6
...@@ -4,7 +4,8 @@ ...@@ -4,7 +4,8 @@
- [Implementing C++ Types](#implementing-c-types) - [Implementing C++ Types](#implementing-c-types)
- [Defining ProtoMaker](#defining-protomaker) - [Defining ProtoMaker](#defining-protomaker)
- [Defining Operator](#defining-operator) - [Defining Operator](#defining-operator)
- [Registering Operator](#registering-operator) - [Defining OpKernel](#defining-opkernel)
- [Registering Operator and OpKernel](#registering-operator-and-opkernel)
- [Compilation](#compilation) - [Compilation](#compilation)
- [Python Binding](#python-binding) - [Python Binding](#python-binding)
- [Unit Tests](#unit-tests) - [Unit Tests](#unit-tests)
...@@ -16,12 +17,13 @@ ...@@ -16,12 +17,13 @@
Here are the base types needed. For details, please refer to the design docs. Here are the base types needed. For details, please refer to the design docs.
- `framework::OperatorBase`: Operator (Op)base class.
- `framework::OpKernel`: Base class for Op computation.
- `framework::OperatorWithKernel`: Inherited from OperatorBase, describing an operator with computation.
- `class OpProtoAndCheckerMaker`: Describes an Operator's input, output, attributes and description, mainly used to interface with Python API. - `class OpProtoAndCheckerMaker`: Describes an Operator's input, output, attributes and description, mainly used to interface with Python API.
- `framework::OperatorBase`: Operator (Op)base class.
- `framework::OpKernel`: Base class for Op computation kernel.
- `framework::OperatorWithKernel`: Inherited from OperatorBase, describing an operator with computation kernels.
An operator can be differentiated by whether in has kernel methods. An operator with kernel inherits from `OperatorWithKernel` while the ones without inherit from `OperatorBase`. This tutorial focuses on implementing operators with kernels. In short, an operator includes the following information: Operators can be categorized into two groups: operator with kernel(s) and operator without kernel(s). An operator with kernel(s) inherits from `OperatorWithKernel` while the one without kernel(s) inherits from `OperatorBase`. This tutorial focuses on implementing operators with kernels. In short, an operator includes the following information:
Information | Where is it defined Information | Where is it defined
...@@ -32,7 +34,7 @@ Kernel implementation | The kernel methods shared between CPU and CUDA are ...@@ -32,7 +34,7 @@ Kernel implementation | The kernel methods shared between CPU and CUDA are
Registering the Op | Ops are registered in `.cc` files; For Kernel registration, `.cc` files contain the CPU implementation, while `.cu` files contain the CUDA implementation. Registering the Op | Ops are registered in `.cc` files; For Kernel registration, `.cc` files contain the CPU implementation, while `.cu` files contain the CUDA implementation.
New Operator implementations are added to the list [paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators), with file names in the format `*_op.h` (if applicable), `*_op.cc`, `*_op.cu` (if applicable).** The system will use the naming scheme to automatically build operators and their corresponding Python extensions. ** New Operator implementations are added to the list [paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators), with file names in the format `*_op.h` (if applicable), `*_op.cc`, `*_op.cu` (if applicable).** The system will use the naming scheme to automatically build operators and their corresponding Python extensions.**
Let's take matrix multiplication operator, [MulOp](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc), as an example to introduce the writing of an Operator with Kernel. Let's take matrix multiplication operator, [MulOp](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc), as an example to introduce the writing of an Operator with Kernel.
...@@ -156,7 +158,8 @@ Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, w ...@@ -156,7 +158,8 @@ Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, w
- `typename T` denotes data type, such as `float` or `double`. - `typename T` denotes data type, such as `float` or `double`.
`MulKernel` types need to rewrite the interface for `Compute`. `MulKernel` types need to rewrite the interface for `Compute`.
- `Compute` takes one input variable `const framework::ExecutionContext& context`.
- `Compute` takes one input parameter: `const framework::ExecutionContext& context`.
- Compared with `InferShapeContext`, `ExecutionContext` includes device types, and can similarly extract input, output, and attribute variables. - Compared with `InferShapeContext`, `ExecutionContext` includes device types, and can similarly extract input, output, and attribute variables.
- `Compute` implements the computation logics of an `OpKernel`. - `Compute` implements the computation logics of an `OpKernel`.
...@@ -177,7 +180,7 @@ Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, w ...@@ -177,7 +180,7 @@ Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, w
}; };
``` ```
Note that **different devices (CPU, CUDA)share an Op definition; whether or not they share the same `OpKernel` depends on whether `Compute` calls functions that support both devices.** Note that **different devices (CPU, CUDA)share one Op definition; whether or not they share the same `OpKernel` depends on whether `Compute` calls functions can support both devices.**
`MulOp`'s CPU and CUDA share the same `Kernel`. A non-sharing `OpKernel` example can be seen in [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43). `MulOp`'s CPU and CUDA share the same `Kernel`. A non-sharing `OpKernel` example can be seen in [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43).
...@@ -188,13 +191,14 @@ This concludes the forward implementation of an operator. Next its operation and ...@@ -188,13 +191,14 @@ This concludes the forward implementation of an operator. Next its operation and
The definition of its corresponding backward operator, if applicable, is similar to that of an forward operator. **Note that a backward operator does not include a `ProtoMaker`**. The definition of its corresponding backward operator, if applicable, is similar to that of an forward operator. **Note that a backward operator does not include a `ProtoMaker`**.
### Registering Operator ### Registering Operator and OpKernel
- In `.cc` files, register forward and backward operator classes and the CPU kernel. - In `.cc` files, register forward and backward operator classes and the CPU kernel.
```cpp ```cpp
namespace ops = paddle::operators; namespace ops = paddle::operators;
REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad); REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUDeviceContext, float>); REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUDeviceContext, float>);
REGISTER_OP_CPU_KERNEL(mul_grad, REGISTER_OP_CPU_KERNEL(mul_grad,
ops::MulGradKernel<paddle::platform::CPUDeviceContext, float>); ops::MulGradKernel<paddle::platform::CPUDeviceContext, float>);
...@@ -204,6 +208,7 @@ The definition of its corresponding backward operator, if applicable, is similar ...@@ -204,6 +208,7 @@ The definition of its corresponding backward operator, if applicable, is similar
- `REGISTER_OP` registers the `ops::MulOp` class, type named `mul`, its type `ProtoMaker` is `ops::MulOpMaker`, registering `ops::MulOpGrad` as `mul_grad`. - `REGISTER_OP` registers the `ops::MulOp` class, type named `mul`, its type `ProtoMaker` is `ops::MulOpMaker`, registering `ops::MulOpGrad` as `mul_grad`.
- `REGISTER_OP_WITHOUT_GRADIENT` registers an operator without gradient. - `REGISTER_OP_WITHOUT_GRADIENT` registers an operator without gradient.
- `REGISTER_OP_CPU_KERNEL` registers `ops::MulKernel` class and specialized template types `paddle::platform::CPUPlace` and `float`, which also registers `ops::MulGradKernel`. - `REGISTER_OP_CPU_KERNEL` registers `ops::MulKernel` class and specialized template types `paddle::platform::CPUPlace` and `float`, which also registers `ops::MulGradKernel`.
...@@ -225,6 +230,7 @@ The definition of its corresponding backward operator, if applicable, is similar ...@@ -225,6 +230,7 @@ The definition of its corresponding backward operator, if applicable, is similar
Run the following commands to compile. Run the following commands to compile.
``` ```
# maybe you need to rerun cmake
make mul_op make mul_op
``` ```
......
doc/howto/dev/new_op_kernel_en.md 0 → 100644
浏览文件 @ e92fe9e6
## Add Kernels for a New Device
### Background
PaddlePaddle Fluid have hundreds of operators. Each operator could have one or more kernels. A kernel is an implementation of the operator for a certain device, which could be a hardware device, e.g., the CUDA GPU, or a library that utilizes a device, e.g., Intel MKL that makes full use of the Xeon CPU.
[This document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/new_op_en.md) explains how to add an operator, and its kernels. The kernels of an operator are indexed by a C++ type [`OpKernelType`](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/operator_kernel_type.md). An operator chooses the right kernel at runtime. This choosing mechanism is described [here](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md).
### Write Kernels for A New Device
#### Add A New Device
For some historical reaons, we misuse the word *library* for *device*. For example, we call the deivce type by *library type*. An example is the header file [`library_type.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/library_type.h#L24). We will correct this ASAP.
To register a new device, we need to add an enum value to `LibraryType`:
```
enum class LibraryType {
kPlain = 0,
kMKLDNN = 1,
kCUDNN = 2,
};
```
#### Add A New [Place](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L53)
If you have a new kind of Device, firstly you need to add a new kind of [`Place`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L53). For example `CUDAPlace`:
```cpp
struct CUDAPlace {
CUDAPlace() : CUDAPlace(0) {}
explicit CUDAPlace(int d) : device(d) {}
inline int GetDeviceId() const { return device; }
// needed for variant equality comparison
inline bool operator==(const CUDAPlace &o) const {
return device == o.device;
}
inline bool operator!=(const CUDAPlace &o) const { return !(*this == o); }
int device;
};
typedef boost::variant<CUDAPlace, CPUPlace> Place;
```
#### Add [device context]((https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h#L37))
After a new kind of Device is added, you should add a corresponding [DeviceContext](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h#L37) for it.
```cpp
class DeviceContext {
public:
virtual ~DeviceContext() {}
virtual Place GetPlace() const = 0;
virtual void Wait() const {}
};
```
#### Implement new [OpKernel](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L351) for your Device.
A detailed documentation can be found in [`new_op_and_kernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/new_op_en.md)
```cpp
class OpKernelBase {
public:
/**
* ExecutionContext is the only parameter of Kernel Run function.
* Run will get input/output variables, state such as momentum and
* device resource such as CUDA stream, cublas handle, etc. from
* ExecutionContext. User should construct it before run the Operator.
*/
virtual void Compute(const ExecutionContext& context) const = 0;
virtual ~OpKernelBase() = default;
};
template <typename T>
class OpKernel : public OpKernelBase {
public:
using ELEMENT_TYPE = T;
};
```
#### Register the OpKernel to framework
After writing the components described above, we should register the kernel to the framework.
We use `REGISTER_OP_KERNEL` to do the registration.
```cpp
REGISTER_OP_KERNEL(
op_type,
library_type,
place_type,
kernel0, kernel1, ...)
```
kernel0, kernel1 are kernels that have the same `op_type`, `library_type`, `place_type` but different `data_types`.
take [`conv2d`]((https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/conv_cudnn_op.cu.cc#L318)) as an example:
```cpp
REGISTER_OP_KERNEL(conv2d, CPU, paddle::platform::CPUPlace,
paddle::operators::GemmConvKernel<paddle::platform::CPUDeviceContext, float>,
paddle::operators::GemmConvKernel<paddle::platform::CPUDeviceContext, double>);
REGISTER_OP_KERNEL(conv2d, CUDNN, ::paddle::platform::CUDAPlace,
paddle::operators::CUDNNConvOpKernel<float>,
paddle::operators::CUDNNConvOpKernel<double>);
```
In the code above:
- `conv2d` is the type/name of the operator
- `CUDNN/CPU` is `library`
- `paddle::platform::CUDAPlace/CPUPlace` is `place`
- template parameter `float/double` on `CUDNNConvOpKernel<T>` is `data_type`.
doc/howto/usage/cluster/fluid_cluster_train_en.md 0 → 100644
浏览文件 @ e92fe9e6
# Fluid Distributed Training
## Introduction
In this article, we'll explain how to config and run distributed training jobs with PaddlePaddle Fluid in a bare metal cluster.
## Preparations
### Get your 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.
### 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).
### Update 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:
``` python
import paddle.v2 as paddle
import paddle.v2.fluid as fluid
x = fluid.layers.data(name='x', shape=[13], dtype='float32')
y_predict = fluid.layers.fc(input=x, size=1, act=None)
y = fluid.layers.data(name='y', shape=[1], dtype='float32')
cost = fluid.layers.square_error_cost(input=y_predict, label=y)
avg_cost = fluid.layers.mean(x=cost)
sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
sgd_optimizer.minimize(avg_cost)
BATCH_SIZE = 20
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.uci_housing.train(), buf_size=500),
batch_size=BATCH_SIZE)
place = fluid.CPUPlace()
feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
PASS_NUM = 100
for pass_id in range(PASS_NUM):
fluid.io.save_persistables(exe, "./fit_a_line.model/")
fluid.io.load_persistables(exe, "./fit_a_line.model/")
for data in train_reader():
avg_loss_value, = exe.run(fluid.default_main_program(),
feed=feeder.feed(data),
fetch_list=[avg_cost])
if avg_loss_value[0] < 10.0:
exit(0) # if avg cost less than 10.0, we think our code is good.
exit(1)
```
We created a simple fully connected neural networks 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.
#### 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.
![parameter server architect](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).
Now we need to create program 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.
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.
Optimize OPs and gradient parameters can be found from the return values of optimizer's minimize function.
To put them together:
``` python
... #define the program, cost, and create sgd optimizer
optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) #get optimize OPs and gradient parameters
t = fluid.DistributeTranspiler() # create 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)
... #create executor
# in pserver, run this
#current_endpoint here means current pserver IP:PORT you wish to run on
pserver_prog = t.get_pserver_program(current_endpoint)
pserver_startup = t.get_startup_program(current_endpoint, pserver_prog)
exe.run(pserver_startup)
exe.run(pserver_prog)
# in trainer, run this
... # define data reader
exe.run(fluid.default_startup_program())
for pass_id in range(100):
for data in train_reader():
exe.run(t.get_trainer_program())
```
### 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:
``` bash
PSERVERS=192.168.1.2:6174 SERVER_ENDPOINT=192.168.1.2:6174 TRAINING_ROLE=PSERVER python notest_dist_fit_a_line.py
```
*please note we assume that your parameter server runs at 192.168.1.2:6174*
Wait until the prompt `Server listening on 192.168.1.2:6174`
Then in 2 of your trainer node 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*
Now you have 2 trainers and 1 parameter server up and running.
doc/v1_api_tutorials/README.md 已删除 100644 → 0
浏览文件 @ 151f579c
The tutorials in v1_api_tutorials are using v1_api currently, and will be upgraded to v2_api later.
Thus, v1_api_tutorials is a temporary directory. We decide not to maintain it and will delete it in future.
Please go to [PaddlePaddle/book](https://github.com/PaddlePaddle/book) and
[PaddlePaddle/models](https://github.com/PaddlePaddle/models) to learn PaddlePaddle.
doc/v1_api_tutorials/embedding_model/index_cn.md 已删除 100644 → 0
浏览文件 @ 151f579c
# 中文词向量模型的使用 #
----------
本文档介绍如何在PaddlePaddle平台上,使用预训练的标准格式词向量模型。
在此感谢 @lipeng 提出的代码需求，并给出的相关模型格式的定义。
## 介绍 ###
### 中文字典 ###
我们的字典使用内部的分词工具对百度知道和百度百科的语料进行分词后产生。分词风格如下： "《红楼梦》"将被分为 "《"，"红楼梦"，"》"，和 "《红楼梦》"。字典采用UTF8编码，输出有2列：词本身和词频。字典共包含 3206326个词和4个特殊标记：
- `<s>`: 分词序列的开始
- `<e>`: 分词序列的结束
- `PALCEHOLDER_JUST_IGNORE_THE_EMBEDDING`: 占位符，没有实际意义
- `<unk>`: 未知词
### 中文词向量的预训练模型 ###
遵循文章 [A Neural Probabilistic Language Model](http://www.jmlr.org/papers/volume3/bengio03a/bengio03a.pdf)中介绍的方法，模型采用 n-gram 语言模型，结构如下图：6元上下文作为输入层->全连接层->softmax层 。对应于字典，我们预训练得到4种不同维度的词向量，分别为：32维、64维、128维和256维。
<center>![](./neural-n-gram-model.png)</center>
<center>Figure 1. neural-n-gram-model</center>
### 下载和数据抽取 ###
运行以下的命令下载和获取我们的字典和预训练模型：
cd $PADDLE_ROOT/demo/model_zoo/embedding
./pre_DictAndModel.sh
## 中文短语改写的例子 ##
以下示范如何使用预训练的中文字典和词向量进行短语改写。
### 数据的准备和预处理 ###
首先，运行以下的命令下载数据集。该数据集（utf8编码）包含20个训练样例，5个测试样例和2个生成式样例。
cd $PADDLE_ROOT/demo/seqToseq/data
./paraphrase_data.sh
第二步，将数据处理成规范格式，在训练数集上训练生成词向量字典（数据将保存在 `$PADDLE_SOURCE_ROOT/demo/seqToseq/data/pre-paraphrase`）:
cd $PADDLE_ROOT/demo/seqToseq/
python preprocess.py -i data/paraphrase [--mergeDict]
- 其中，如果使用`--mergeDict`选项，源语言短语和目标语言短语的字典将被合并（源语言和目标语言共享相同的编码字典）。本实例中，源语言和目标语言都是相同的语言，因此可以使用该选项。
### 使用用户指定的词向量字典 ###
使用如下命令，从预训练模型中，根据用户指定的字典，抽取对应的词向量构成新的词表:
cd $PADDLE_ROOT/demo/model_zoo/embedding
python extract_para.py --preModel PREMODEL --preDict PREDICT --usrModel USRMODEL--usrDict USRDICT -d DIM
- `--preModel PREMODEL`: 预训练词向量字典模型的路径
- `--preDict PREDICT`: 预训练模型使用的字典的路径
- `--usrModel USRMODEL`: 抽取出的新词表的保存路径
- `--usrDict USRDICT`: 用户指定新的字典的路径，用于构成新的词表
- `-d DIM`: 参数（词向量）的维度
此处，你也可以简单的运行以下的命令：
cd $PADDLE_ROOT/demo/seqToseq/data/
./paraphrase_model.sh
运行成功以后，你将会看到以下的模型结构：
paraphrase_model
|--- _source_language_embedding
|--- _target_language_embedding
### 在PaddlePaddle平台训练模型 ###
首先，配置模型文件，配置如下（可以参考保存在 `demo/seqToseq/paraphrase/train.conf`的配置）:
from seqToseq_net import *
is_generating = False
################## Data Definition #####################
train_conf = seq_to_seq_data(data_dir = "./data/pre-paraphrase",
job_mode = job_mode)
############## Algorithm Configuration ##################
settings(
learning_method = AdamOptimizer(),
batch_size = 50,
learning_rate = 5e-4)
################# Network configure #####################
gru_encoder_decoder(train_conf, is_generating, word_vector_dim = 32)
这个配置与`demo/seqToseq/translation/train.conf` 基本相同
然后，使用以下命令进行模型训练:
cd $PADDLE_SOURCE_ROOT/demo/seqToseq/paraphrase
./train.sh
其中，`train.sh` 与`demo/seqToseq/translation/train.sh` 基本相同，只有2个配置不一样:
- `--init_model_path`: 初始化模型的路径配置为`data/paraphrase_modeldata/paraphrase_model`
- `--load_missing_parameter_strategy`：如果参数模型文件缺失，除词向量模型外的参数将使用正态分布随机初始化
如果用户想要了解详细的数据集的格式、模型的结构和训练过程，请查看 [Text generation Tutorial](../text_generation/index_cn.md).
## 可选功能 ##
### 观测词向量
PaddlePaddle 平台为想观测词向量的用户提供了将二进制词向量模型转换为文本模型的功能:
cd $PADDLE_ROOT/demo/model_zoo/embedding
python paraconvert.py --b2t -i INPUT -o OUTPUT -d DIM
- `-i INPUT`: 输入的（二进制）词向量模型名称
- `-o OUTPUT`: 输出的文本模型名称
- `-d DIM`: （词向量）参数维度
运行完以上命令，用户可以在输出的文本模型中看到:
0,4,32156096
-0.7845433,1.1937413,-0.1704215,0.4154715,0.9566584,-0.5558153,-0.2503305, ......
0.0000909,0.0009465,-0.0008813,-0.0008428,0.0007879,0.0000183,0.0001984, ......
......
- 其中，第一行是`PaddlePaddle` 输出文件的格式说明，包含3个属性：:
- `PaddlePaddle`的版本号，本例中为0
- 浮点数占用的字节数，本例中为4
- 总计的参数个数，本例中为32,156,096
- 其余行是（词向量）参数行（假设词向量维度为32）
- 每行打印32个参数以','分隔
- 共有32,156,096/32 = 1,004,877行，也就是说，模型共包含1,004,877个被向量化的词
### 词向量模型的修正
`PaddlePaddle` 为想修正词向量模型的用户提供了将文本词向量模型转换为二进制模型的命令:
cd $PADDLE_ROOT/demo/model_zoo/embedding
python paraconvert.py --t2b -i INPUT -o OUTPUT
- `-i INPUT`: 输入的文本词向量模型名称
- `-o OUTPUT`: 输出的二进制词向量模型名称
请注意，输入的文本格式如下:
-0.7845433,1.1937413,-0.1704215,0.4154715,0.9566584,-0.5558153,-0.2503305, ......
0.0000909,0.0009465,-0.0008813,-0.0008428,0.0007879,0.0000183,0.0001984, ......
......
- 输入文本中没有头部（格式说明）行
- （输入文本）每行存储一个词，以逗号','分隔
doc/v1_api_tutorials/embedding_model/index_en.md 已删除 100644 → 0
浏览文件 @ 151f579c
# Chinese Word Embedding Model Tutorial #
----------
This tutorial is to guide you through the process of using a Pretrained Chinese Word Embedding Model in the PaddlePaddle standard format.
We thank @lipeng for the pull request that defined the model schemas and pretrained the models.
## Introduction ###
### Chinese Word Dictionary ###
Our Chinese-word dictionary is created on Baidu ZhiDao and Baidu Baike by using in-house word segmentor. For example, the participle of "《红楼梦》" is "《"，"红楼梦"，"》"，and "《红楼梦》". Our dictionary (using UTF-8 format) has has two columns: word and its frequency. The total word count is 3206326, including 4 special token:
- `<s>`: the start of a sequence
- `<e>`: the end of a sequence
- `PALCEHOLDER_JUST_IGNORE_THE_EMBEDDING`: a placeholder, just ignore it and its embedding
- `<unk>`: a word not included in dictionary
### Pretrained Chinese Word Embedding Model ###
Inspired by paper [A Neural Probabilistic Language Model](http://www.jmlr.org/papers/volume3/bengio03a/bengio03a.pdf), our model architecture (**Embedding joint of six words->FullyConnect->SoftMax**) is as following graph. And for our dictionary, we pretrain four models with different word vector dimenstions, i.e 32, 64, 128, 256.
<center>![](./neural-n-gram-model.png)</center>
<center>Figure 1. neural-n-gram-model</center>
### Download and Extract ###
To download and extract our dictionary and pretrained model, run the following commands.
cd $PADDLE_ROOT/demo/model_zoo/embedding
./pre_DictAndModel.sh
## Chinese Paraphrasing Example ##
We provide a paraphrasing task to show the usage of pretrained Chinese Word Dictionary and Embedding Model.
### Data Preparation and Preprocess ###
First, run the following commands to download and extract the in-house dataset. The dataset (using UTF-8 format) has 20 training samples, 5 testing samples and 2 generating samples.
cd $PADDLE_ROOT/demo/seqToseq/data
./paraphrase_data.sh
Second, preprocess data and build dictionary on train data by running the following commands, and the preprocessed dataset is stored in `$PADDLE_SOURCE_ROOT/demo/seqToseq/data/pre-paraphrase`:
cd $PADDLE_ROOT/demo/seqToseq/
python preprocess.py -i data/paraphrase [--mergeDict]
- `--mergeDict`: if using this option, the source and target dictionary are merged, i.e, two dictionaries have the same context. Here, as source and target data are all chinese words, this option can be used.
### User Specified Embedding Model ###
The general command of extracting desired parameters from the pretrained embedding model based on user dictionary is:
cd $PADDLE_ROOT/demo/model_zoo/embedding
python extract_para.py --preModel PREMODEL --preDict PREDICT --usrModel USRMODEL--usrDict USRDICT -d DIM
- `--preModel PREMODEL`: the name of pretrained embedding model
- `--preDict PREDICT`: the name of pretrained dictionary
- `--usrModel USRMODEL`: the name of extracted embedding model
- `--usrDict USRDICT`: the name of user specified dictionary
- `-d DIM`: dimension of parameter
Here, you can simply run the command:
cd $PADDLE_ROOT/demo/seqToseq/data/
./paraphrase_model.sh
And you will see following embedding model structure:
paraphrase_model
|--- _source_language_embedding
|--- _target_language_embedding
### Training Model in PaddlePaddle ###
First, create a model config file, see example `demo/seqToseq/paraphrase/train.conf`:
from seqToseq_net import *
is_generating = False
################## Data Definition #####################
train_conf = seq_to_seq_data(data_dir = "./data/pre-paraphrase",
job_mode = job_mode)
############## Algorithm Configuration ##################
settings(
learning_method = AdamOptimizer(),
batch_size = 50,
learning_rate = 5e-4)
################# Network configure #####################
gru_encoder_decoder(train_conf, is_generating, word_vector_dim = 32)
This config is almost the same as `demo/seqToseq/translation/train.conf`.
Then, train the model by running the command:
cd $PADDLE_SOURCE_ROOT/demo/seqToseq/paraphrase
./train.sh
where `train.sh` is almost the same as `demo/seqToseq/translation/train.sh`, the only difference is following two command arguments:
- `--init_model_path`: path of the initialization model, here is `data/paraphrase_model`
- `--load_missing_parameter_strategy`: operations when model file is missing, here use a normal distibution to initialize the other parameters except for the embedding layer
For users who want to understand the dataset format, model architecture and training procedure in detail, please refer to [Text generation Tutorial](../text_generation/index_en.md).
## Optional Function ##
### Embedding Parameters Observation
For users who want to observe the embedding parameters, this function can convert a PaddlePaddle binary embedding model to a text model by running the command:
cd $PADDLE_ROOT/demo/model_zoo/embedding
python paraconvert.py --b2t -i INPUT -o OUTPUT -d DIM
- `-i INPUT`: the name of input binary embedding model
- `-o OUTPUT`: the name of output text embedding model
- `-d DIM`: the dimension of parameter
You will see parameters like this in output text model:
0,4,32156096
-0.7845433,1.1937413,-0.1704215,0.4154715,0.9566584,-0.5558153,-0.2503305, ......
0.0000909,0.0009465,-0.0008813,-0.0008428,0.0007879,0.0000183,0.0001984, ......
......
- 1st line is **PaddlePaddle format file head**, it has 3 attributes:
- version of PaddlePaddle, here is 0
- sizeof(float), here is 4
- total number of parameter, here is 32156096
- Other lines print the paramters (assume `<dim>` = 32)
- each line print 32 paramters splitted by ','
- there is 32156096/32 = 1004877 lines, meaning there is 1004877 embedding words
### Embedding Parameters Revision
For users who want to revise the embedding parameters, this function can convert a revised text embedding model to a PaddlePaddle binary model by running the command:
cd $PADDLE_ROOT/demo/model_zoo/embedding
python paraconvert.py --t2b -i INPUT -o OUTPUT
- `-i INPUT`: the name of input text embedding model.
- `-o OUTPUT`: the name of output binary embedding model
Note that the format of input text model is as follows:
-0.7845433,1.1937413,-0.1704215,0.4154715,0.9566584,-0.5558153,-0.2503305, ......
0.0000909,0.0009465,-0.0008813,-0.0008428,0.0007879,0.0000183,0.0001984, ......
......
- there is no file header in 1st line
- each line stores parameters for one word, the separator is commas ','
doc/v1_api_tutorials/gan/index_en.md 已删除 100644 → 0
浏览文件 @ 151f579c
# Generative Adversarial Networks (GAN)
This demo implements GAN training described in the original [GAN paper](https://arxiv.org/abs/1406.2661) and deep convolutional generative adversarial networks [DCGAN paper](https://arxiv.org/abs/1511.06434).
The high-level structure of GAN is shown in Figure. 1 below. It is composed of two major parts: a generator and a discriminator, both of which are based on neural networks. The generator takes in some kind of noise with a known distribution and transforms it into an image. The discriminator takes in an image and determines whether it is artificially generated by the generator or a real image. So the generator and the discriminator are in a competitive game in which generator is trying to generate image to look as real as possible to fool the discriminator, while the discriminator is trying to distinguish between real and fake images.
<center>![](./gan.png)</center>
<p align="center">
Figure 1. GAN-Model-Structure
<a href="https://ishmaelbelghazi.github.io/ALI/">figure credit</a>
</p>
The generator and discriminator take turn to be trained using SGD. The objective function of the generator is for its generated images being classified as real by the discriminator, and the objective function of the discriminator is to correctly classify real and fake images. When the GAN model is trained to converge to the equilibrium state, the generator will transform the given noise distribution to the distribution of real images, and the discriminator will not be able to distinguish between real and fake images at all.
## Implementation of GAN Model Structure
Since GAN model involves multiple neural networks, it requires to use paddle python API. So the code walk-through below can also partially serve as an introduction to the usage of Paddle Python API.
There are three networks defined in gan_conf.py, namely **generator_training**, **discriminator_training** and **generator**. The relationship to the model structure we defined above is that **discriminator_training** is the discriminator, **generator** is the generator, and the **generator_training** combined the generator and discriminator since training generator would require the discriminator to provide loss function. This relationship is described in the following code:
```python
if is_generator_training:
noise = data_layer(name="noise", size=noise_dim)
sample = generator(noise)
if is_discriminator_training:
sample = data_layer(name="sample", size=sample_dim)
if is_generator_training or is_discriminator_training:
label = data_layer(name="label", size=1)
prob = discriminator(sample)
cost = cross_entropy(input=prob, label=label)
classification_error_evaluator(
input=prob, label=label, name=mode + '_error')
outputs(cost)
if is_generator:
noise = data_layer(name="noise", size=noise_dim)
outputs(generator(noise))
```
In order to train the networks defined in gan_conf.py, one first needs to initialize a Paddle environment, parse the config, create GradientMachine from the config and create trainer from GradientMachine as done in the code chunk below:
```python
import py_paddle.swig_paddle as api
# init paddle environment
api.initPaddle('--use_gpu=' + use_gpu, '--dot_period=10',
'--log_period=100', '--gpu_id=' + args.gpu_id,
'--save_dir=' + "./%s_params/" % data_source)
# Parse config
gen_conf = parse_config(conf, "mode=generator_training,data=" + data_source)
dis_conf = parse_config(conf, "mode=discriminator_training,data=" + data_source)
generator_conf = parse_config(conf, "mode=generator,data=" + data_source)
# Create GradientMachine
dis_training_machine = api.GradientMachine.createFromConfigProto(
dis_conf.model_config)
gen_training_machine = api.GradientMachine.createFromConfigProto(
gen_conf.model_config)
generator_machine = api.GradientMachine.createFromConfigProto(
generator_conf.model_config)
# Create trainer
dis_trainer = api.Trainer.create(dis_conf, dis_training_machine)
gen_trainer = api.Trainer.create(gen_conf, gen_training_machine)
```
In order to balance the strength between generator and discriminator, we schedule to train whichever one is performing worse by comparing their loss function value. The loss function value can be calculated by a forward pass through the GradientMachine.
```python
def get_training_loss(training_machine, inputs):
outputs = api.Arguments.createArguments(0)
training_machine.forward(inputs, outputs, api.PASS_TEST)
loss = outputs.getSlotValue(0).copyToNumpyMat()
return numpy.mean(loss)
```
After training one network, one needs to sync the new parameters to the other networks. The code below demonstrates one example of such use case:
```python
# Train the gen_training
gen_trainer.trainOneDataBatch(batch_size, data_batch_gen)
# Copy the parameters from gen_training to dis_training and generator
copy_shared_parameters(gen_training_machine,
dis_training_machine)
copy_shared_parameters(gen_training_machine, generator_machine)
```
## A Toy Example
With the infrastructure explained above, we can now walk you through a toy example of generating two dimensional uniform distribution using 10 dimensional Gaussian noise.
The Gaussian noises are generated using the code below:
```python
def get_noise(batch_size, noise_dim):
return numpy.random.normal(size=(batch_size, noise_dim)).astype('float32')
```
The real samples (2-D uniform) are generated using the code below:
```python
# synthesize 2-D uniform data in gan_trainer.py:114
def load_uniform_data():
data = numpy.random.rand(1000000, 2).astype('float32')
return data
```
The generator and discriminator network are built using fully-connected layer and batch_norm layer, and are defined in gan_conf.py.
To train the GAN model, one can use the command below. The flag -d specifies the training data (cifar, mnist or uniform) and flag --useGpu specifies whether to use gpu for training (0 is cpu, 1 is gpu).
```bash
$python gan_trainer.py -d uniform --useGpu 1
```
The generated samples can be found in ./uniform_samples/ and one example is shown below as Figure 2. One can see that it roughly recovers the 2D uniform distribution.
<center>![](./uniform_sample.png)</center>
<p align="center">
Figure 2. Uniform Sample
</p>
## MNIST Example
### Data preparation
To download the MNIST data, one can use the following commands:
```bash
$cd data/
$./get_mnist_data.sh
```
### Model description
Following the DC-Gan paper (https://arxiv.org/abs/1511.06434), we use convolution/convolution-transpose layer in the discriminator/generator network to better deal with images. The details of the network structures are defined in gan_conf_image.py.
### Training the model
To train the GAN model on mnist data, one can use the following command:
```bash
$python gan_trainer.py -d mnist --useGpu 1
```
The generated sample images can be found at ./mnist_samples/ and one example is shown below as Figure 3.
<center>![](./mnist_sample.png)</center>
<p align="center">
Figure 3. MNIST Sample
</p>
doc/v1_api_tutorials/imagenet_model/resnet_model_cn.md 已删除 100644 → 0
浏览文件 @ 151f579c
# Model Zoo - ImageNet #
[ImageNet](http://www.image-net.org/) 是通用物体分类领域一个众所周知的数据库。本教程提供了一个用于ImageNet上的卷积分类网络模型。
## ResNet 介绍
论文 [Deep Residual Learning for Image Recognition](http://arxiv.org/abs/1512.03385) 中提出的ResNet网络结构在2015年ImageNet大规模视觉识别竞赛(ILSVRC 2015)的分类任务中赢得了第一名。他们提出残差学习的框架来简化网络的训练，所构建网络结构的的深度比之前使用的网络有大幅度的提高。下图展示的是基于残差的连接方式。左图构造网络模块的方式被用于34层的网络中，而右图的瓶颈连接模块用于50层，101层和152层的网络结构中。
<center>![resnet_block](./resnet_block.jpg)</center>
<center>图 1. ResNet 网络模块</center>
本教程中我们给出了三个ResNet模型，这些模型都是由原作者提供的模型<https://github.com/KaimingHe/deep-residual-networks>转换过来的。我们使用PaddlePaddle在ILSVRC的验证集共50,000幅图像上测试了模型的分类错误率，其中输入图像的颜色通道顺序为**BGR**，保持宽高比缩放到短边为256，只截取中心方形的图像区域。分类错误率和模型大小由下表给出。
<center>
<table border="2" cellspacing="0" cellpadding="6" rules="all" frame="border">
<colgroup>
<col class="left" />
<col class="left" />
<col class="left" />
</colgroup>
<thead>
<tr>
<th scope="col" class="left">ResNet</th>
<th scope="col" class="left">Top-1</th>
<th scope="col" class="left">Model Size</th>
</tr>
</thead>
<tbody>
<tr>
<td class="left">ResNet-50</td>
<td class="left">24.9%</td>
<td class="left">99M</td>
</tr>
<tr>
<td class="left">ResNet-101</td>
<td class="left">23.7%</td>
<td class="left">173M</td>
</tr>
<tr>
<td class="left">ResNet-152</td>
<td class="left">23.2%</td>
<td class="left">234M</td>
</tr>
</tbody>
</table></center>
<br>
## ResNet 模型
50层，101层和152层的网络配置文件可参照```demo/model_zoo/resnet/resnet.py```。你也可以通过在命令行参数中增加一个参数如```--config_args=layer_num=50```来指定网络层的数目。
### 网络可视化
你可以通过执行下面的命令来得到ResNet网络的结构可视化图。该脚本会生成一个dot文件，然后可以转换为图片。需要安装graphviz来转换dot文件为图片。
```
cd demo/model_zoo/resnet
./net_diagram.sh
```
### 模型下载
```
cd demo/model_zoo/resnet
./get_model.sh
```
你可以执行上述命令来下载所有的模型和均值文件，如果下载成功，这些文件将会被保存在```demo/model_zoo/resnet/model```路径下。
```
mean_meta_224 resnet_101 resnet_152 resnet_50
```
* resnet_50: 50层网络模型。
* resnet_101: 101层网络模型。
* resnet_152: 152层网络模型。
* mean\_meta\_224: 均值图像文件，图像大小为3 x 224 x 224，颜色通道顺序为**BGR**。你也可以使用这三个值: 103.939, 116.779, 123.68。
### 参数信息
* **卷积层权重**
由于每个卷积层后面连接的是batch normalization层，因此该层中没有偏置(bias)参数，并且只有一个权重。
形状: `(Co, ky, kx, Ci)`
* Co: 输出特征图的通道数目
* ky: 滤波器核在垂直方向上的尺寸
* kx: 滤波器核在水平方向上的尺寸
* Ci: 输入特征图的通道数目
二维矩阵: (Co * ky * kx, Ci), 行优先次序存储。
* **全连接层权重**
二维矩阵: (输入层尺寸, 本层尺寸), 行优先次序存储。
* **[Batch Normalization](<http://arxiv.org/abs/1502.03167>) 层权重**
本层有四个参数，实际上只有.w0和.wbias是需要学习的参数，另外两个分别是滑动均值和方差。在测试阶段它们将会被加载到模型中。下表展示了batch normalization层的参数。
<center>
<table border="2" cellspacing="0" cellpadding="6" rules="all" frame="border">
<colgroup>
<col class="left" />
<col class="left" />
<col class="left" />
</colgroup>
<thead>
<tr>
<th scope="col" class="left">参数名</th>
<th scope="col" class="left">尺寸</th>
<th scope="col" class="left">含义</th>
</tr>
</thead>
<tbody>
<tr>
<td class="left">_res2_1_branch1_bn.w0</td>
<td class="left">256</td>
<td class="left">gamma, 缩放参数</td>
</tr>
<tr>
<td class="left">_res2_1_branch1_bn.w1</td>
<td class="left">256</td>
<td class="left">特征图均值</td>
</tr>
<tr>
<td class="left">_res2_1_branch1_bn.w2</td>
<td class="left">256</td>
<td class="left">特征图方差</td>
</tr>
<tr>
<td class="left">_res2_1_branch1_bn.wbias</td>
<td class="left">256</td>
<td class="left">beta, 偏置参数</td>
</tr>
</tbody>
</table></center>
<br>
### 参数读取
使用者可以使用下面的Python脚本来读取参数值:
```
import sys
import numpy as np
def load(file_name):
with open(file_name, 'rb') as f:
f.read(16) # skip header for float type.
return np.fromfile(f, dtype=np.float32)
if __name__=='__main__':
weight = load(sys.argv[1])
```
或者直接使用下面的shell命令:
```
od -j 16 -f _res2_1_branch1_bn.w0
```
## 特征提取
我们提供了C++和Python接口来提取特征。下面的例子使用了`demo/model_zoo/resnet/example`中的数据，详细地展示了整个特征提取的过程。
### C++接口
首先，在配置文件中的`define_py_data_sources2`里指定图像数据列表，具体请参照示例`demo/model_zoo/resnet/resnet.py`。
```
train_list = 'train.list' if not is_test else None
# mean.meta is mean file of ImageNet dataset.
# mean.meta size : 3 x 224 x 224.
# If you use three mean value, set like:
# "mean_value:103.939,116.779,123.68;"
args={
'mean_meta': "model/mean_meta_224/mean.meta",
'image_size': 224, 'crop_size': 224,
'color': True,'swap_channel:': [2, 1, 0]}
define_py_data_sources2(train_list,
'example/test.list',
module="example.image_list_provider",
obj="processData",
args=args)
```
第二步，在`resnet.py`文件中指定要提取特征的网络层的名字。例如，
```
Outputs("res5_3_branch2c_conv", "res5_3_branch2c_bn")
```
第三步，在`extract_fea_c++.sh`文件中指定模型路径和输出的目录，然后执行下面的命令。
```
cd demo/model_zoo/resnet
./extract_fea_c++.sh
```
如果执行成功，特征将会存到`fea_output/rank-00000`文件中，如下所示。同时你可以使用`load_feature.py`文件中的`load_feature_c`接口来加载该文件。
```
-0.115318 -0.108358 ... -0.087884;-1.27664 ... -1.11516 -2.59123;
-0.126383 -0.116248 ... -0.00534909;-1.42593 ... -1.04501 -1.40769;
```
* 每行存储的是一个样本的特征。其中，第一行存的是图像`example/dog.jpg`的特征，第二行存的是图像`example/cat.jpg`的特征。
* 不同层的特征由分号`;`隔开，并且它们的顺序与`Outputs()`中指定的层顺序一致。这里，左边是`res5_3_branch2c_conv`层的特征，右边是`res5_3_branch2c_bn`层特征。
### Python接口
示例`demo/model_zoo/resnet/classify.py`中展示了如何使用Python来提取特征。下面的例子同样使用了`./example/test.list`中的数据。执行的命令如下：
```
cd demo/model_zoo/resnet
./extract_fea_py.sh
```
extract_fea_py.sh:
```
python classify.py \
--job=extract \
--conf=resnet.py\
--use_gpu=1 \
--mean=model/mean_meta_224/mean.meta \
--model=model/resnet_50 \
--data=./example/test.list \
--output_layer="res5_3_branch2c_conv,res5_3_branch2c_bn" \
--output_dir=features
```
* \--job=extract: 指定工作模式来提取特征。
* \--conf=resnet.py: 网络配置文件。
* \--use_gpu=1: 指定是否使用GPU。
* \--model=model/resnet_50: 模型路径。
* \--data=./example/test.list: 数据列表。
* \--output_layer="xxx,xxx": 指定提取特征的层。
* \--output_dir=features: 输出目录。
如果运行成功，你将会看到特征存储在`features/batch_0`文件中，该文件是由cPickle产生的。你可以使用`load_feature.py`中的`load_feature_py`接口来打开该文件，它将返回如下的字典：
```
{
'cat.jpg': {'res5_3_branch2c_conv': array([[-0.12638293, -0.116248 , -0.11883899, ..., -0.00895038, 0.01994277, -0.00534909]], dtype=float32), 'res5_3_branch2c_bn': array([[-1.42593431, -1.28918779, -1.32414699, ..., -1.45933616, -1.04501402, -1.40769434]], dtype=float32)},
'dog.jpg': {'res5_3_branch2c_conv': array([[-0.11531784, -0.10835785, -0.08809858, ...,0.0055237, 0.01505112, -0.08788397]], dtype=float32), 'res5_3_branch2c_bn': array([[-1.27663755, -1.18272924, -0.90937918, ..., -1.25178063, -1.11515927, -2.59122872]], dtype=float32)}
}
```
仔细观察，这些特征值与上述使用C++接口提取的结果是一致的。
## 预测
`classify.py`文件也可以用于对样本进行预测。我们提供了一个示例脚本`predict.sh`，它使用50层的ResNet模型来对`example/test.list`中的数据进行预测。
```
cd demo/model_zoo/resnet
./predict.sh
```
predict.sh调用了`classify.py`:
```
python classify.py \
--job=predict \
--conf=resnet.py\
--multi_crop \
--model=model/resnet_50 \
--use_gpu=1 \
--data=./example/test.list
```
* \--job=extract: 指定工作模型进行预测。
* \--conf=resnet.py: 网络配置文件。network configure.
* \--multi_crop: 使用10个裁剪图像块，预测概率取平均。
* \--use_gpu=1: 指定是否使用GPU。
* \--model=model/resnet_50: 模型路径。
* \--data=./example/test.list: 数据列表。
如果运行成功，你将会看到如下结果，其中156和285是这些图像的分类标签。
```
Label of example/dog.jpg is: 156
Label of example/cat.jpg is: 282
```
doc/v1_api_tutorials/imagenet_model/resnet_model_en.md 已删除 100644 → 0
浏览文件 @ 151f579c
# Model Zoo - ImageNet #
[ImageNet](http://www.image-net.org/) is a popular dataset for generic object classification. This tutorial provides convolutional neural network(CNN) models for ImageNet.
## ResNet Introduction
ResNets from paper [Deep Residual Learning for Image Recognition](http://arxiv.org/abs/1512.03385) won the 1st place on the ILSVRC 2015 classification task. They present residual learning framework to ease the training of networks that are substantially deeper than those used previously. The residual connections are shown in following figure. The left building block is used in network of 34 layers and the right bottleneck building block is used in network of 50, 101, 152 layers .
<center>![resnet_block](./resnet_block.jpg)</center>
<center>Figure 1. ResNet Block</center>
We present three ResNet models, which are converted from the models provided by the authors <https://github.com/KaimingHe/deep-residual-networks>. The classfication errors tested in PaddlePaddle on 50,000 ILSVRC validation set with input images channel order of **BGR** by single scale with the shorter side of 256 and single crop as following table.
<center>
<table border="2" cellspacing="0" cellpadding="6" rules="all" frame="border">
<colgroup>
<col class="left" />
<col class="left" />
<col class="left" />
</colgroup>
<thead>
<tr>
<th scope="col" class="left">ResNet</th>
<th scope="col" class="left">Top-1</th>
<th scope="col" class="left">Model Size</th>
</tr>
</thead>
<tbody>
<tr>
<td class="left">ResNet-50</td>
<td class="left">24.9%</td>
<td class="left">99M</td>
</tr>
<tr>
<td class="left">ResNet-101</td>
<td class="left">23.7%</td>
<td class="left">173M</td>
</tr>
<tr>
<td class="left">ResNet-152</td>
<td class="left">23.2%</td>
<td class="left">234M</td>
</tr>
</tbody>
</table></center>
<br>
## ResNet Model
See ```demo/model_zoo/resnet/resnet.py```. This config contains network of 50, 101 and 152 layers. You can specify layer number by adding argument like ```--config_args=layer_num=50``` in command line arguments.
### Network Visualization
You can get a diagram of ResNet network by running the following commands. The script generates dot file and then converts dot file to PNG file, which needs to install graphviz to convert.
```
cd demo/model_zoo/resnet
./net_diagram.sh
```
### Model Download
```
cd demo/model_zoo/resnet
./get_model.sh
```
You can run above command to download all models and mean file and save them in ```demo/model_zoo/resnet/model``` if downloading successfully.
```
mean_meta_224 resnet_101 resnet_152 resnet_50
```
* resnet_50: model of 50 layers.
* resnet_101: model of 101 layers.
* resnet_152: model of 152 layers.
* mean\_meta\_224: mean file with 3 x 224 x 224 size in **BGR** order. You also can use three mean values: 103.939, 116.779, 123.68.
### Parameter Info
* **Convolution Layer Weight**
As batch normalization layer is connected after each convolution layer, there is no parameter of bias and only one weight in this layer.
shape: `(Co, ky, kx, Ci)`
* Co: channle number of output feature map.
* ky: filter size in vertical direction.
* kx: filter size in horizontal direction.
* Ci: channle number of input feature map.
2-Dim matrix: (Co * ky * kx, Ci), saved in row-major order.
* **Fully connected Layer Weight**
2-Dim matrix: (input layer size, this layer size), saved in row-major order.
* **[Batch Normalization](<http://arxiv.org/abs/1502.03167>) Layer Weight**
There are four parameters in this layer. In fact, only .w0 and .wbias are the learned parameters. The other two are therunning mean and variance respectively. They will be loaded in testing. Following table shows parameters of a batch normzalization layer.
<center>
<table border="2" cellspacing="0" cellpadding="6" rules="all" frame="border">
<colgroup>
<col class="left" />
<col class="left" />
<col class="left" />
</colgroup>
<thead>
<tr>
<th scope="col" class="left">Parameter Name</th>
<th scope="col" class="left">Number</th>
<th scope="col" class="left">Meaning</th>
</tr>
</thead>
<tbody>
<tr>
<td class="left">_res2_1_branch1_bn.w0</td>
<td class="left">256</td>
<td class="left">gamma, scale parameter</td>
</tr>
<tr>
<td class="left">_res2_1_branch1_bn.w1</td>
<td class="left">256</td>
<td class="left">mean value of feature map</td>
</tr>
<tr>
<td class="left">_res2_1_branch1_bn.w2</td>
<td class="left">256</td>
<td class="left">variance of feature map</td>
</tr>
<tr>
<td class="left">_res2_1_branch1_bn.wbias</td>
<td class="left">256</td>
<td class="left">beta, shift parameter</td>
</tr>
</tbody>
</table></center>
<br>
### Parameter Observation
Users who want to observe the parameters can use Python to read:
```
import sys
import numpy as np
def load(file_name):
with open(file_name, 'rb') as f:
f.read(16) # skip header for float type.
return np.fromfile(f, dtype=np.float32)
if __name__=='__main__':
weight = load(sys.argv[1])
```
or simply use following shell command:
```
od -j 16 -f _res2_1_branch1_bn.w0
```
## Feature Extraction
We provide both C++ and Python interfaces to extract features. The following examples use data in `demo/model_zoo/resnet/example` to show the extracting process in detail.
### C++ Interface
First, specify image data list in `define_py_data_sources2` in the config, see example `demo/model_zoo/resnet/resnet.py`.
```
train_list = 'train.list' if not is_test else None
# mean.meta is mean file of ImageNet dataset.
# mean.meta size : 3 x 224 x 224.
# If you use three mean value, set like:
# "mean_value:103.939,116.779,123.68;"
args={
'mean_meta': "model/mean_meta_224/mean.meta",
'image_size': 224, 'crop_size': 224,
'color': True,'swap_channel:': [2, 1, 0]}
define_py_data_sources2(train_list,
'example/test.list',
module="example.image_list_provider",
obj="processData",
args=args)
```
Second, specify layers to extract features in `Outputs()` of `resnet.py`. For example,
```
Outputs("res5_3_branch2c_conv", "res5_3_branch2c_bn")
```
Third, specify model path and output directory in `extract_fea_c++.sh`, and then run the following commands.
```
cd demo/model_zoo/resnet
./extract_fea_c++.sh
```
If successful, features are saved in `fea_output/rank-00000` as follows. And you can use `load_feature_c` interface in `load_feature.py ` to load such a file.
```
-0.115318 -0.108358 ... -0.087884;-1.27664 ... -1.11516 -2.59123;
-0.126383 -0.116248 ... -0.00534909;-1.42593 ... -1.04501 -1.40769;
```
* Each line stores features of a sample. Here, the first line stores features of `example/dog.jpg` and second line stores features of `example/cat.jpg`.
* Features of different layers are splitted by `;`, and their order is consistent with the layer order in `Outputs()`. Here, the left features are `res5_3_branch2c_conv` layer and right features are `res5_3_branch2c_bn` layer.
### Python Interface
`demo/model_zoo/resnet/classify.py` is an example to show how to use Python to extract features. Following example still uses data of `./example/test.list`. Command is as follows:
```
cd demo/model_zoo/resnet
./extract_fea_py.sh
```
extract_fea_py.sh:
```
python classify.py \
--job=extract \
--conf=resnet.py\
--use_gpu=1 \
--mean=model/mean_meta_224/mean.meta \
--model=model/resnet_50 \
--data=./example/test.list \
--output_layer="res5_3_branch2c_conv,res5_3_branch2c_bn" \
--output_dir=features
```
* \--job=extract: specify job mode to extract feature.
* \--conf=resnet.py: network configure.
* \--use_gpu=1: speficy GPU mode.
* \--model=model/resnet_5: model path.
* \--data=./example/test.list: data list.
* \--output_layer="xxx,xxx": specify layers to extract features.
* \--output_dir=features: output diretcoty.
If run successfully, you will see features saved in `features/batch_0`, this file is produced with cPickle. You can use `load_feature_py` interface in `load_feature.py` to open the file, and it returns a dictionary as follows:
```
{
'cat.jpg': {'res5_3_branch2c_conv': array([[-0.12638293, -0.116248 , -0.11883899, ..., -0.00895038, 0.01994277, -0.00534909]], dtype=float32), 'res5_3_branch2c_bn': array([[-1.42593431, -1.28918779, -1.32414699, ..., -1.45933616, -1.04501402, -1.40769434]], dtype=float32)},
'dog.jpg': {'res5_3_branch2c_conv': array([[-0.11531784, -0.10835785, -0.08809858, ...,0.0055237, 0.01505112, -0.08788397]], dtype=float32), 'res5_3_branch2c_bn': array([[-1.27663755, -1.18272924, -0.90937918, ..., -1.25178063, -1.11515927, -2.59122872]], dtype=float32)}
}
```
Observed carefully, these feature values are consistent with the above results extracted by C++ interface.
## Prediction
`classify.py` also can be used to predict. We provide an example script `predict.sh` to predict data in `example/test.list` using a ResNet model with 50 layers.
```
cd demo/model_zoo/resnet
./predict.sh
```
predict.sh calls the `classify.py`:
```
python classify.py \
--job=predict \
--conf=resnet.py\
--multi_crop \
--model=model/resnet_50 \
--use_gpu=1 \
--data=./example/test.list
```
* \--job=extract: speficy job mode to predict.
* \--conf=resnet.py: network configure.
* \--multi_crop: use 10 crops and average predicting probability.
* \--use_gpu=1: speficy GPU mode.
* \--model=model/resnet_50: model path.
* \--data=./example/test.list: data list.
If run successfully, you will see following results, where 156 and 285 are labels of the images.
```
Label of example/dog.jpg is: 156
Label of example/cat.jpg is: 282
```
doc/v1_api_tutorials/quick_start/index_cn.rst 已删除 100644 → 0
浏览文件 @ 151f579c
=============
快速入门教程
=============
我们将以 `文本分类问题 <https://en.wikipedia.org/wiki/Document_classification>`_ 为例,
介绍PaddlePaddle的基本使用方法。
安装
====
请参考 :ref:`install_steps` 安装PaddlePaddle。
使用概述
========
**文本分类问题**：对于给定的一条文本，我们从提前给定的类别集合中选择其所属类别。
比如, 在购物网站上，通过查看买家对某个产品的评价反馈, 评估该产品的质量。
- 这个显示器很棒！ （好评）
- 用了两个月之后这个显示器屏幕碎了。（差评）
使用PaddlePaddle, 每一个任务流程都可以被划分为如下五个步骤。
.. image:: src/Pipeline_cn.jpg
:align: center
:scale: 80%
1. 数据格式准备
- 本例每行保存一条样本，类别Id和文本信息用 ``Tab`` 间隔，文本中的单词用空格分隔（如果不切词，则字与字之间用空格分隔），例如：``类别Id '\t' 这 个 显 示 器 很 棒 ！``
2. 向系统传送数据
- PaddlePaddle可以执行用户的python脚本程序来读取各种格式的数据文件。
- 本例的所有字符都将转换为连续整数表示的Id传给模型。
3. 描述网络结构和优化算法
- 本例由易到难展示4种不同的文本分类网络配置：逻辑回归模型，词向量模型，卷积模型，时序模型。
- 常用优化算法包括Momentum, RMSProp，AdaDelta，AdaGrad，Adam，Adamax等，本例采用Adam优化方法，加了L2正则和梯度截断。
4. 训练模型
5. 应用模型
数据格式准备
------------
接下来我们将展示如何用PaddlePaddle训练一个文本分类模型，将 `Amazon电子产品评论数据 <http://jmcauley.ucsd.edu/data/amazon/>`_ 分为好评(正样本)和差评(负样本)两种类别。
`源代码 <https://github.com/PaddlePaddle/Paddle>`_ 的 ``demo/quick_start`` 目录里提供了该数据的下载脚本和预处理脚本，你只需要在命令行输入以下命令，就能够很方便的完成数据下载和相应的预处理工作。
.. code-block:: bash
cd demo/quick_start
./data/get_data.sh
./preprocess.sh
数据预处理完成之后，通过配置类似于 ``dataprovider_*.py`` 的数据读取脚本和类似于 ``trainer_config.*.py`` 的训练模型脚本，PaddlePaddle将以设置参数的方式来设置
相应的数据读取脚本和训练模型脚本。接下来，我们将对这两个步骤给出了详细的解释，你也可以先跳过本文的解释环节，直接进入训练模型章节, 使用 ``sh train.sh`` 开始训练模型，
查看`train.sh`内容，通过 **自底向上法** (bottom-up approach)来帮助你理解PaddlePaddle的内部运行机制。
向系统传送数据
==============
Python脚本读取数据
------------------
`DataProvider` 是PaddlePaddle负责提供数据的模块，主要职责在于将训练数据传入内存或者显存，让模型能够得到训练更新，其包括两个函数：
* initializer：PaddlePaddle会在调用读取数据的Python脚本之前，先调用initializer函数。在下面例子里，我们在initialzier函数里初始化词表，并且在随后的读取数据过程中填充词表。
* process：PaddlePaddle调用process函数来读取数据。每次读取一条数据后，process函数会用yield语句输出这条数据，从而能够被PaddlePaddle 捕获 (harvest)。
``dataprovider_bow.py`` 文件给出了完整例子：
.. literalinclude:: ../../../demo/quick_start/dataprovider_bow.py
:language: python
:lines: 21-70
:linenos:
:emphasize-lines: 8,33
详细内容请参见 :ref:`api_dataprovider` 。
配置中的数据加载定义
--------------------
在模型配置中通过 ``define_py_data_sources2`` 接口来加载数据：
.. literalinclude:: ../../../demo/quick_start/trainer_config.emb.py
:language: python
:lines: 19-35
:linenos:
:emphasize-lines: 12
以下是对上述数据加载的解释：
- data/train.list,data/test.list: 指定训练数据和测试数据
- module="dataprovider_bow": 处理数据的Python脚本文件
- obj="process": 指定生成数据的函数
- args={"dictionary": word_dict}: 额外的参数，这里指定词典
更详细数据格式和用例请参考 :ref:`api_pydataprovider2` 。
模型网络结构
============
本小节我们将介绍模型网络结构。
.. image:: src/PipelineNetwork_cn.jpg
:align: center
:scale: 80%
我们将以最基本的逻辑回归网络作为起点，并逐渐展示更加深入的功能。更详细的网络配置连接请参考 :ref:`api_trainer_config_helpers_layers` 。
所有配置都能在 `源代码 <https://github.com/PaddlePaddle/Paddle>`_ 的 ``demo/quick_start`` 目录下找到。
逻辑回归模型
------------
具体流程如下:
.. image:: src/NetLR_cn.jpg
:align: center
:scale: 80%
- 获取利用 `one-hot vector <https://en.wikipedia.org/wiki/One-hot>`_ 表示的每个单词，维度是词典大小
.. code-block:: python
word = data_layer(name="word", size=word_dim)
- 获取该条样本类别Id，维度是类别个数。
.. code-block:: python
label = data_layer(name="label", size=label_dim)
- 利用逻辑回归模型对该向量进行分类，同时会计算分类准确率
.. code-block:: python
# Define a fully connected layer with logistic activation (also called softmax activation).
output = fc_layer(input=word,
size=label_dim,
act_type=SoftmaxActivation())
# Define cross-entropy classification loss and error.
classification_cost(input=output, label=label)
- input: 除去data层，每个层都有一个或多个input,多个input以list方式输入
- size: 该层神经元个数
- act_type: 激活函数类型
**效果总结**：我们将在后面介绍训练和预测流程的脚本。在此为方便对比不同网络结构，我们总结了各个网络的复杂度和效果。
===================== =============================== =================
网络名称 参数数量 错误率
===================== =============================== =================
逻辑回归 252 KB 8.652 %
===================== =============================== =================
词向量模型
----------
embedding模型需要稍微改变提供数据的Python脚本，即 ``dataprovider_emb.py``，词向量模型、
卷积模型、时序模型均使用该脚本。其中文本输入类型定义为整数时序类型integer_value_sequence。
.. code-block:: python
def initializer(settings, dictionary, **kwargs):
settings.word_dict = dictionary
settings.input_types = [
# Define the type of the first input as sequence of integer.
# The value of the integers range from 0 to len(dictrionary)-1
integer_value_sequence(len(dictionary)),
# Define the second input for label id
integer_value(2)]
@provider(init_hook=initializer)
def process(settings, file_name):
...
# omitted, it is same as the data provider for LR model
该模型依然使用逻辑回归分类网络的框架， 只是将句子用连续向量表示替换为用稀疏向量表示， 即对第三步进行替换。句子表示的计算更新为两步：
.. image:: src/NetContinuous_cn.jpg
:align: center
:scale: 80%
- 利用单词Id查找该单词对应的连续向量(维度为word_dim)， 输入N个单词，输出为N个word_dim维度向量
.. code-block:: python
emb = embedding_layer(input=word, size=word_dim)
- 将该句话包含的所有单词向量求平均, 得到句子的表示
.. code-block:: python
avg = pooling_layer(input=emb, pooling_type=AvgPooling())
其它部分和逻辑回归网络结构一致。
**效果总结：**
===================== =============================== ==================
网络名称 参数数量 错误率
===================== =============================== ==================
词向量模型 15 MB 8.484 %
===================== =============================== ==================
卷积模型
-----------
卷积网络是一种特殊的从词向量表示到句子表示的方法， 也就是将词向量模型进一步演化为三个新步骤。
.. image:: src/NetConv_cn.jpg
:align: center
:scale: 80%
文本卷积分可为三个步骤:
1. 首先，从每个单词左右两端分别获取k个相邻的单词, 拼接成一个新的向量；
2. 其次，对该向量进行非线性变换(例如Sigmoid变换), 使其转变为维度为hidden_dim的新向量；
3. 最后，对整个新向量集合的每一个维度取最大值来表示最后的句子。
这三个步骤可配置为:
.. code-block:: python
text_conv = sequence_conv_pool(input=emb,
context_start=k,
context_len=2 * k + 1)
**效果总结：**
===================== =============================== ========================
网络名称 参数数量 错误率
===================== =============================== ========================
卷积模型 16 MB 5.628 %
===================== =============================== ========================
时序模型
----------
.. image:: src/NetRNN_cn.jpg
:align: center
:scale: 80%
时序模型，也称为RNN模型, 包括简单的 `RNN模型 <https://en.wikipedia.org/wiki/Recurrent_neural_network>`_, `GRU模型 <https://en.wikipedia.org/wiki/Gated_recurrent_unit>`_ 和 `LSTM模型 <https://en.wikipedia.org/wiki/Long_short-term_memory>`_ 等等。
- GRU模型配置：
.. code-block:: python
gru = simple_gru(input=emb, size=gru_size)
- LSTM模型配置：
.. code-block:: python
lstm = simple_lstm(input=emb, size=lstm_size)
本次试验，我们采用单层LSTM模型，并使用了Dropout，**效果总结：**
===================== =============================== =========================
网络名称 参数数量 错误率
===================== =============================== =========================
时序模型 16 MB 4.812 %
===================== =============================== =========================
优化算法
=========
`优化算法 <http://www.paddlepaddle.org/doc/ui/api/trainer_config_helpers/optimizers_index.html>`_ 包括
Momentum, RMSProp，AdaDelta，AdaGrad，ADAM，Adamax等，这里采用Adam优化方法，同时使用了L2正则(L2 Regularization)和梯度截断(Gradient Clipping)。
.. code-block:: python
settings(batch_size=128,
learning_rate=2e-3,
learning_method=AdamOptimizer(),
regularization=L2Regularization(8e-4),
gradient_clipping_threshold=25)
训练模型
=========
在数据加载和网络配置完成之后， 我们就可以训练模型了。
.. image:: src/PipelineTrain_cn.jpg
:align: center
:scale: 80%
训练模型，我们只需要运行 ``train.sh`` 训练脚本：
.. code-block:: bash
./train.sh
``train.sh`` 中包含了训练模型的基本命令。训练时所需设置的主要参数如下：
.. code-block:: bash
paddle train \
--config=trainer_config.py \
--log_period=20 \
--save_dir=./output \
--num_passes=15 \
--use_gpu=false
这里只简单介绍了单机训练，如何进行分布式训练，请参考 :ref:`cluster_train` 。
预测
=====
当模型训练好了之后，我们就可以进行预测了。
.. image:: src/PipelineTest_cn.jpg
:align: center
:scale: 80%
之前配置文件中 ``test.list`` 指定的数据将会被测试，这里直接通过预测脚本 ``predict.sh`` 进行预测,
更详细的说明，请参考 :ref:`api_swig_py_paddle` 。
.. code-block:: bash
model="output/pass-00003"
paddle train \
--config=trainer_config.lstm.py \
--use_gpu=false \
--job=test \
--init_model_path=$model \
--config_args=is_predict=1 \
--predict_output_dir=. \
mv rank-00000 result.txt
这里以 ``output/pass-00003`` 为例进行预测，用户可以根据训练日志，选择测试结果最好的模型来预测。
预测结果以文本的形式保存在 ``result.txt`` 中，一行为一个样本，格式如下：
.. code-block:: bash
预测ID;ID为0的概率 ID为1的概率
预测ID;ID为0的概率 ID为1的概率
总体效果总结
==============
在 ``/demo/quick_start`` 目录下，能够找到这里使用的所有数据, 网络配置, 训练脚本等等。
对于Amazon-Elec测试集(25k), 如下表格，展示了上述网络模型的训练效果:
===================== =============================== ============= ==================================
网络名称 参数数量 错误率 配置文件
===================== =============================== ============= ==================================
逻辑回归模型 252 KB 8.652% trainer_config.lr.py
词向量模型 15 MB 8.484% trainer_config.emb.py
卷积模型 16 MB 5.628% trainer_config.cnn.py
时序模型 16 MB 4.812% trainer_config.lstm.py
===================== =============================== ============= ==================================
附录
=====
命令行参数
----------
* \--config：网络配置
* \--save_dir：模型存储路径
* \--log_period：每隔多少batch打印一次日志
* \--num_passes：训练轮次，一个pass表示过一遍所有训练样本
* \--config_args：命令指定的参数会传入网络配置中。
* \--init_model_path：指定初始化模型路径，可用在测试或训练时指定初始化模型。
默认一个pass保存一次模型，也可以通过saving_period_by_batches设置每隔多少batch保存一次模型。
可以通过show_parameter_stats_period设置打印参数信息等。
其他参数请参考 命令行参数文档（链接待补充）。
输出日志
---------
.. code-block:: bash
TrainerInternal.cpp:160] Batch=20 samples=2560 AvgCost=0.628761 CurrentCost=0.628761 Eval: classification_error_evaluator=0.304297 CurrentEval: classification_error_evaluator=0.304297
模型训练会看到类似上面这样的日志信息，详细的参数解释，请参考如下表格：
=========================================== ==============================================================
名称 解释
=========================================== ==============================================================
Batch=20 表示过了20个batch
samples=2560 表示过了2560个样本
AvgCost 每个pass的第0个batch到当前batch所有样本的平均cost
CurrentCost 当前log_period个batch所有样本的平均cost
Eval: classification_error_evaluator 每个pass的第0个batch到当前batch所有样本的平均分类错误率
CurrentEval: classification_error_evaluator 当前log_period个batch所有样本的平均分类错误率
=========================================== ==============================================================
doc/v1_api_tutorials/quick_start/index_en.md 已删除 100644 → 0
浏览文件 @ 151f579c
# Quick Start
This tutorial will teach the basics of deep learning (DL), including how to implement many different models in PaddlePaddle. You will learn how to:
- Prepare data into the standardized format that PaddlePaddle accepts.
- Write data providers that read data into PaddlePaddle.
- Configure neural networks in PaddlePaddle layer by layer.
- Train models.
- Perform inference with trained models.
## Install
To get started, please install PaddlePaddle on your computer. Throughout this tutorial, you will learn by implementing different DL models for text classification.
To install PaddlePaddle, please follow the instructions here: <a href = "../../getstarted/build_and_install/index_en.html" >Build and Install</a>.
## Overview
For the first step, you will use PaddlePaddle to build a **text classification** system. For example, suppose you run an e-commence website, and you want to analyze the sentiment of user reviews to evaluate product quality.
For example, given the input
```
This monitor is fantastic.
```
Your classifier should output “positive”, since this text snippet shows that the user is satisfied with the product. Given this input:
```
The monitor breaks down two months after purchase.
```
the classifier should output “negative“.
To build your text classification system, your code will need to perform five steps:
<center> ![](./src/Pipeline_en.jpg) </center>
- Preprocess data into a standardized format.
- Provide data to the learning model.
- Specify the neural network structure.
- Train the model.
- Inference (make prediction on test examples).
1. Preprocess data into standardized format
- In the text classification example, you will start with a text file with one training example per line. Each line contains category id (in machine learning, often denoted the target y), followed by the input text (often denoted x); these two elements are separated by a Tab. For example: ```positive [tab] This monitor is fantastic```. You will preprocess this raw data into a format that Paddle can use.
2. Provide data to the learning model.
- You can write data providers in Python. For any required data preprocessing step, you can add the preprocessing code to the PyDataProvider Python file.
- In our text classification example, every word or character will be converted into an integer id, specified in a dictionary file. It perform a dictionary lookup in PyDataProvider to get the id.
3. Specify neural network structure. (From easy to hard, we provide 4 kinds of network configurations)
- A logistic regression model.
- A word embedding model.
- A convolutional neural network model.
- A sequential recurrent neural network model.
- You will also learn different learning algorithms.
4. Training model.
5. Inference.
## Preprocess data into standardized format
In this example, you are going to use [Amazon electronic product review dataset](http://jmcauley.ucsd.edu/data/amazon/) to build a bunch of deep neural network models for text classification. Each text in this dataset is a product review. This dataset has two categories: “positive” and “negative”. Positive means the reviewer likes the product, while negative means the reviewer does not like the product.
`demo/quick_start` in the [source code](https://github.com/PaddlePaddle/Paddle) provides script for downloading the preprocessed data as shown below. (If you want to process the raw data, you can use the script `demo/quick_start/data/proc_from_raw_data/get_data.sh`).
```bash
cd demo/quick_start
./data/get_data.sh
```
## Transfer Data to Model
### Write Data Provider with Python
The following `dataprovider_bow.py` gives a complete example of writing data provider with Python. It includes the following parts:
* initalizer： define the additional meta-data of the data provider and the types of the input data.
* process： Each `yield` returns a data sample. In this case, it return the text representation and category id. The order of features in the returned result needs to be consistent with the definition of the input types in `initalizer`.
```python
from paddle.trainer.PyDataProvider2 import *
# id of the word not in dictionary
UNK_IDX = 0
# initializer is called by the framework during initialization.
# It allows the user to describe the data types and setup the
# necessary data structure for later use.
# `settings` is an object. initializer need to properly fill settings.input_types.
# initializer can also store other data structures needed to be used at process().
# In this example, dictionary is stored in settings.
# `dictionay` and `kwargs` are arguments passed from trainer_config.lr.py
def initializer(settings, dictionary, **kwargs):
# Put the word dictionary into settings
settings.word_dict = dictionary
# setting.input_types specifies what the data types the data provider
# generates.
settings.input_types = [
# The first input is a sparse_binary_vector,
# which means each dimension of the vector is either 0 or 1. It is the
# bag-of-words (BOW) representation of the texts.
sparse_binary_vector(len(dictionary)),
# The second input is an integer. It represents the category id of the
# sample. 2 means there are two labels in the dataset.
# (1 for positive and 0 for negative)
integer_value(2)]
# Delaring a data provider. It has an initializer 'data_initialzer'.
# It will cache the generated data of the first pass in memory, so that
# during later pass, no on-the-fly data generation will be needed.
# `setting` is the same object used by initializer()
# `file_name` is the name of a file listed train_list or test_list file given
# to define_py_data_sources2(). See trainer_config.lr.py.
@provider(init_hook=initializer, cache=CacheType.CACHE_PASS_IN_MEM)
def process(settings, file_name):
# Open the input data file.
with open(file_name, 'r') as f:
# Read each line.
for line in f:
# Each line contains the label and text of the comment, separated by \t.
label, comment = line.strip().split('\t')
# Split the words into a list.
words = comment.split()
# convert the words into a list of ids by looking them up in word_dict.
word_vector = [settings.word_dict.get(w, UNK_IDX) for w in words]
# Return the features for the current comment. The first is a list
# of ids representing a 0-1 binary sparse vector of the text,
# the second is the integer id of the label.
yield word_vector, int(label)
```
### Define Python Data Provider in Configuration files.
You need to add a data provider definition `define_py_data_sources2` in our network configuration. This definition specifies:
- The path of the training and testing data (`data/train.list`, `data/test.list`).
- The location of the data provider file (`dataprovider_bow`).
- The function to call to get data. (`process`).
- Additional arguments or data. Here it passes the path of word dictionary.
```python
from paddle.trainer_config_helpers import *
file = "data/dict.txt"
word_dict = dict()
with open(dict_file, 'r') as f:
for i, line in enumerate(f):
w = line.strip().split()[0]
word_dict[w] = i
# define the data sources for the model.
# We need to use different process for training and prediction.
# For training, the input data includes both word IDs and labels.
# For prediction, the input data only includs word Ids.
define_py_data_sources2(train_list='data/train.list',
test_list='data/test.list',
module="dataprovider_bow",
obj="process",
args={"dictionary": word_dict})
```
You can refer to the following link for more detailed examples and data formats: <a href = "../../api/v1/data_provider/pydataprovider2_en.html">PyDataProvider2</a>.
## Network Architecture
We will describe four kinds of network architectures in this section.
<center> ![](./src/PipelineNetwork_en.jpg) </center>
First, you will build a logistic regression model. Later, you will also get chance to build other more powerful network architectures.
For more detailed documentation, you could refer to: <a href = "../../api/v1/trainer_config_helpers/layers.html">layer documentation</a>. All configuration files are in `demo/quick_start` directory.
### Logistic Regression
The architecture is illustrated in the following picture:
<center> ![](./src/NetLR_en.png) </center>
- You need define the data for text features. The size of the data layer is the number of words in the dictionary.
```python
word = data_layer(name="word", size=voc_dim)
```
- You also need to define the category id for each example. The size of the data layer is the number of labels.
```python
label = data_layer(name="label", size=label_dim)
```
- It uses logistic regression model to classify the vector, and it will output the classification error during training.
- Each layer has an *input* argument that specifies its input layer. Some layers can have multiple input layers. You can use a list of the input layers as input in that case.
- *size* for each layer means the number of neurons of the layer.
- *act_type* means activation function applied to the output of each neuron independently.
- Some layers can have additional special inputs. For example, `classification_cost` needs ground truth label as input to compute classification loss and error.
```python
# Define a fully connected layer with logistic activation (also called softmax activation).
output = fc_layer(input=word,
size=label_dim,
act_type=SoftmaxActivation())
# Define cross-entropy classification loss and error.
classification_cost(input=output, label=label)
```
Performance summary: You can refer to the training and testing scripts later. In order to compare different network architectures, the model complexity and test classification error are listed in the following table:
<html>
<center>
<table border="2" cellspacing="0" cellpadding="6" rules="all" frame="border">
<thead>
<th scope="col" class="left">Network name</th>
<th scope="col" class="left">Number of parameters</th>
<th scope="col" class="left">Test error</th>
</tr>
</thead>
<tbody>
<tr>
<td class="left">Logistic regression</td>
<td class="left">252 KB</td>
<td class="left">8.652%</td>
</tr>
</tbody>
</table></center>
</html>
<br>
### Word Embedding Model
In order to use the word embedding model, you need to change the data provider a little bit to make the input words as a sequence of word IDs. The revised data provider `dataprovider_emb.py` is listed below. You only need to change initializer() for the type of the first input. It is changed from sparse_binary_vector to sequence of intergers. process() remains the same. This data provider can also be used for later sequence models.
```python
def initializer(settings, dictionary, **kwargs):
# Put the word dictionary into settings
settings.word_dict = dictionary
settings.input_types = [
# Define the type of the first input as a sequence of integers.
integer_value_sequence(len(dictionary)),
# Define the second input for label id
integer_value(2)]
@provider(init_hook=initializer)
def process(settings, file_name):
...
# omitted, it is same as the data provider for LR model
```
This model is very similar to the framework of logistic regression, but it uses word embedding vectors instead of a sparse vectors to represent words.
<center> ![](./src/NetContinuous_en.png) </center>
- It can look up the dense word embedding vector in the dictionary (its words embedding vector is `word_dim`). The input is a sequence of N words, the output is N word_dim dimensional vectors.
```python
emb = embedding_layer(input=word, dim=word_dim)
```
- It averages all the word embedding in a sentence to get its sentence representation.
```python
avg = pooling_layer(input=emb, pooling_type=AvgPooling())
```
The other parts of the model are the same as logistic regression network.
The performance is summarized in the following table:
<html>
<center>
<table border="2" cellspacing="0" cellpadding="6" rules="all" frame="border">
<thead>
<th scope="col" class="left">Network name</th>
<th scope="col" class="left">Number of parameters</th>
<th scope="col" class="left">Test error</th>
</tr>
</thead>
<tbody>
<tr>
<td class="left">Word embedding model</td>
<td class="left">15 MB</td>
<td class="left">8.484%</td>
</tr>
</tbody>
</table>
</html></center>
<br>
### Convolutional Neural Network Model
Convolutional neural network converts a sequence of word embeddings into a sentence representation using temporal convolutions. You will transform the fully connected layer of the word embedding model to 3 new sub-steps.
<center> ![](./src/NetConv_en.png) </center>
Text convolution has 3 steps:
1. Get K nearest neighbor context of each word in a sentence, stack them into a 2D vector representation.
2. Apply temporal convolution to this representation to produce a new hidden_dim dimensional vector.
3. Apply max-pooling to the new vectors at all the time steps in a sentence to get a sentence representation.
```python
# context_len means convolution kernel size.
# context_start means the start of the convolution. It can be negative. In that case, zero padding is applied.
text_conv = sequence_conv_pool(input=emb,
context_start=k,
context_len=2 * k + 1)
```
The performance is summarized in the following table：
<html>
<center>
<table border="2" cellspacing="0" cellpadding="6" rules="all" frame="border">
<thead>
<th scope="col" class="left">Network name</th>
<th scope="col" class="left">Number of parameters</th>
<th scope="col" class="left">Test error</th>
</tr>
</thead>
<tbody>
<tr>
<td class="left">Convolutional model</td>
<td class="left">16 MB</td>
<td class="left">5.628%</td>
</tr>
</tbody>
</table></center>
<br>
### Recurrent Model
<center> ![](./src/NetRNN_en.png) </center>
You can use Recurrent neural network as our time sequence model, including simple RNN model, GRU model, and LSTM model。
- GRU model can be specified via：
```python
gru = simple_gru(input=emb, size=gru_size)
```
- LSTM model can be specified via：
```python
lstm = simple_lstm(input=emb, size=lstm_size)
```
You can use single layer LSTM model with Dropout for our text classification problem. The performance is summarized in the following table:
<html>
<center>
<table border="2" cellspacing="0" cellpadding="6" rules="all" frame="border">
<thead>
<th scope="col" class="left">Network name</th>
<th scope="col" class="left">Number of parameters</th>
<th scope="col" class="left">Test error</th>
</tr>
</thead>
<tbody>
<tr>
<td class="left">Recurrent model</td>
<td class="left">16 MB</td>
<td class="left">4.812%</td>
</tr>
</tbody>
</table></center>
</html>
<br>
## Optimization Algorithm
<a href = "../../api/v1/trainer_config_helpers/optimizers.html">Optimization algorithms</a> include Momentum, RMSProp, AdaDelta, AdaGrad, Adam, and Adamax. You can use Adam optimization method here, with L2 regularization and gradient clipping, because Adam has been proved to work very well for training recurrent neural network.
```python
settings(batch_size=128,
learning_rate=2e-3,
learning_method=AdamOptimizer(),
regularization=L2Regularization(8e-4),
gradient_clipping_threshold=25)
```
## Training Model
After completing data preparation and network architecture specification, you will run the training script.
<center> ![](./src/PipelineTrain_en.png) </center>
Training script: our training script is in `train.sh` file. The training arguments are listed below:
```bash
paddle train \
--config=trainer_config.py \
--log_period=20 \
--save_dir=./output \
--num_passes=15 \
--use_gpu=false
```
We do not provide examples on how to train on clusters here. If you want to train on clusters, please follow the <a href = "../../howto/usage/cluster/cluster_train_en.html">distributed training</a> documentation or other demos for more details.
## Inference
You can use the trained model to perform prediction on the dataset with no labels. You can also evaluate the model on dataset with labels to obtain its test accuracy.
<center> ![](./src/PipelineTest_en.png) </center>
The test script is listed below. PaddlePaddle can evaluate a model on the data with labels specified in `test.list`.
```bash
paddle train \
--config=trainer_config.lstm.py \
--use_gpu=false \
--job=test \
--init_model_path=./output/pass-0000x
```
We will give an example of performing prediction using Recurrent model on a dataset with no labels. You can refer to <a href = "../../api/v1/predict/swig_py_paddle_en.html">Python Prediction API</a> tutorial，or other <a href = "../../tutorials/index_en.html">demo</a> for the prediction process using Python. You can also use the following script for inference or evaluation.
inference script (predict.sh)：
```bash
model="output/pass-00003"
paddle train \
--config=trainer_config.lstm.py \
--use_gpu=false \
--job=test \
--init_model_path=$model \
--config_args=is_predict=1 \
--predict_output_dir=. \
mv rank-00000 result.txt
```
User can choose the best model base on the training log instead of model `output/pass-00003`. There are several differences between training and inference network configurations.
- You do not need labels during inference.
- Outputs need to be specified to the classification probability layer (the output of softmax layer), or the id of maximum probability (`max_id` layer). An example to output the id and probability is given in the code snippet.
- batch_size = 1.
- You need to specify the location of `test_list` in the test data.
The results in `result.txt` is as follows, each line is one sample.
```
predicted_label_id;probability_of_label_0 probability_of_label_1 # the first sample
predicted_label_id;probability_of_label_0 probability_of_label_1 # the second sample
```
```python
is_predict = get_config_arg('is_predict', bool, False)
trn = 'data/train.list' if not is_predict else None
tst = 'data/test.list' if not is_predict else 'data/pred.list'
obj = 'process' if not is_predict else 'process_pre'
batch_size = 128 if not is_predict else 1
if is_predict:
maxid = maxid_layer(output)
outputs([maxid,output])
else:
label = data_layer(name="label", size=2)
cls = classification_cost(input=output, label=label) outputs(cls)
```
## Summary
The scripts of data downloading, network configurations, and training scrips are in `/demo/quick_start`. The following table summarizes the performance of our network architecture on Amazon-Elec dataset(25k):
<center>
<table border="2" cellspacing="0" cellpadding="6" rules="all" frame="border">
<thead>
<th scope="col" class="left">Network name</th>
<th scope="col" class="left">Number of parameters</th>
<th scope="col" class="left">Error rate</th>
<th scope="col" class="left">Configuration file name</th>
</tr>
</thead>
<tbody>
<tr>
<td class="left">Logistic regression model(BOW)</td>
<td class="left"> 252KB </td>
<td class="left">8.652%</td>
<td class="left">trainer_config.lr.py</td>
</tr>
<tr>
<td class="left">Word embedding</td>
<td class="left"> 15MB </td>
<td class="left"> 8.484%</td>
<td class="left">trainer_config.emb.py</td>
</tr>
<tr>
<td class="left">Convolution model</td>
<td class="left"> 16MB </td>
<td class="left"> 5.628%</td>
<td class="left">trainer_config.cnn.py</td>
</tr>
<tr>
<td class="left">Time sequence model</td>
<td class="left"> 16MB </td>
<td class="left"> 4.812%</td>
<td class="left">trainer_config.lstm.py</td>
</tr>
</tbody>
</table>
</center>
<br>
## Appendix
### Command Line Argument
* \--config：network architecture path.
* \--save_dir：model save directory.
* \--log_period：the logging period per batch.
* \--num_passes：number of training passes. One pass means the training would go over the whole training dataset once.
* \--config_args：Other configuration arguments.
* \--init_model_path：The path of the initial model parameter.
By default, the trainer will save model every pass. You can also specify `saving_period_by_batches` to set the frequency of batch saving. You can use `show_parameter_stats_period` to print the statistics of the parameters, which are very useful for tuning parameters. Other command line arguments can be found in <a href = "../../howto/usage/cmd_parameter/index_en.html">command line argument documentation</a>。
### Log
```
TrainerInternal.cpp:160] Batch=20 samples=2560 AvgCost=0.628761 CurrentCost=0.628761 Eval: classification_error_evaluator=0.304297 CurrentEval: classification_error_evaluator=0.304297
```
During model training, you will see the log like the examples above:
<center>
<table border="2" cellspacing="0" cellpadding="6" rules="all" frame="border">
<thead>
<th scope="col" class="left">Name</th>
<th scope="col" class="left">Explanation</th>
</tr>
</thead>
<tr>
<td class="left">Batch=20</td>
<td class="left"> You have trained 20 batches. </td>
</tr>
<tr>
<td class="left">samples=2560</td>
<td class="left"> You have trained 2560 examples. </td>
</tr>
<tr>
<td class="left">AvgCost</td>
<td class="left"> The average cost from the first batch to the current batch. </td>
</tr>
<tr>
<td class="left">CurrentCost</td>
<td class="left"> the average cost of the last log_period batches </td>
</tr>
<tr>
<td class="left">Eval: classification_error_evaluator</td>
<td class="left"> The average classification error from the first batch to the current batch.</td>
</tr>
<tr>
<td class="left">CurrentEval: classification_error_evaluator</td>
<td class="left"> The average error rate of the last log_period batches </td>
</tr>
</tbody>
</table>
</center>
<br>
paddle/framework/CMakeLists.txt
浏览文件 @ e92fe9e6
...@@ -33,8 +33,13 @@ cc_library(scope SRCS scope.cc DEPS glog threadpool) ...@@ -33,8 +33,13 @@ cc_library(scope SRCS scope.cc DEPS glog threadpool)
cc_test(scope_test SRCS scope_test.cc DEPS scope) cc_test(scope_test SRCS scope_test.cc DEPS scope)
cc_library(data_device_transform SRCS data_device_transform.cc DEPS tensor) cc_library(data_device_transform SRCS data_device_transform.cc DEPS tensor)
nv_test(data_device_transform_test SRCS data_device_transform_test.cu
DEPS operator op_registry init math_function)
cc_library(data_type_transform SRCS data_type_transform.cc DEPS tensor) cc_library(data_type_transform SRCS data_type_transform.cc DEPS tensor)
cc_library(data_layout_transform SRCS data_layout_transform.cc DEPS tensor math_function) cc_library(data_layout_transform SRCS data_layout_transform.cc DEPS tensor math_function)
cc_test(data_layout_transform_test SRCS data_layout_transform_test.cc DEPS data_layout_transform)
cc_library(data_transform SRCS data_transform.cc DEPS math_function tensor cc_library(data_transform SRCS data_transform.cc DEPS math_function tensor
framework_proto selected_rows data_device_transform data_type_transform data_layout_transform) framework_proto selected_rows data_device_transform data_type_transform data_layout_transform)
...@@ -82,5 +87,3 @@ cc_test(init_test SRCS init_test.cc DEPS init) ...@@ -82,5 +87,3 @@ 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(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto)
cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc) cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc)
nv_test(data_device_transform_test SRCS data_device_transform_test.cu
DEPS operator op_registry init math_function)
paddle/framework/data_device_transform_test.cu
浏览文件 @ e92fe9e6
...@@ -150,6 +150,7 @@ TEST(Operator, CPUtoGPU) { ...@@ -150,6 +150,7 @@ TEST(Operator, CPUtoGPU) {
// get output // get output
auto* output2 = scope.Var("OUT2"); auto* output2 = scope.Var("OUT2");
gpu_op->Run(scope, cuda_place); gpu_op->Run(scope, cuda_place);
VLOG(3) << "after gpu_op run";
// auto* output2_ptr = output2->Get<LoDTensor>().data<float>(); // auto* output2_ptr = output2->Get<LoDTensor>().data<float>();
DeviceContextPool& pool = DeviceContextPool::Instance(); DeviceContextPool& pool = DeviceContextPool::Instance();
......
paddle/framework/data_layout_transform.cc
浏览文件 @ e92fe9e6
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. /* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License"); Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License. you may not use this file except in compliance with the License.
...@@ -14,12 +14,23 @@ limitations under the License. */ ...@@ -14,12 +14,23 @@ limitations under the License. */
#include "paddle/framework/data_layout_transform.h" #include "paddle/framework/data_layout_transform.h"
#include "paddle/framework/tensor.h"
#include "paddle/operators/math/math_function.h" #include "paddle/operators/math/math_function.h"
namespace paddle { namespace paddle {
namespace framework { namespace framework {
std::vector<int> GetAxis(const DataLayout& from, const DataLayout& to) {
PADDLE_ENFORCE_NE(from, to,
"layout transform should transform different layout");
if (from == DataLayout::kNCHW && to == DataLayout::kNHWC) {
return {0, 2, 3, 1};
} else if (from == DataLayout::kNHWC && to == DataLayout::kNCHW) {
return {0, 3, 1, 2};
} else {
PADDLE_THROW("unsupported transform");
}
}
struct CastDataLayout { struct CastDataLayout {
CastDataLayout(const platform::DeviceContext* ctx, CastDataLayout(const platform::DeviceContext* ctx,
const std::vector<int>& axis, const framework::Tensor& in, const std::vector<int>& axis, const framework::Tensor& in,
...@@ -44,38 +55,36 @@ struct CastDataLayout { ...@@ -44,38 +55,36 @@ struct CastDataLayout {
} }
}; };
void TransDataLayout(const std::vector<int>& axis, void TransDataLayout(const OpKernelType& kernel_type_for_var,
const platform::DeviceContext* ctx, const OpKernelType& expected_kernel_type, const Tensor& in,
const KernelTypePair& kernel_pair, const Variable& in, Tensor* out) {
Variable* out) {
PADDLE_ENFORCE(in.IsType<Tensor>(), "Only support Tensor transform!.");
PADDLE_ENFORCE( PADDLE_ENFORCE(
platform::places_are_same_class(kernel_pair.first.place_, platform::places_are_same_class(kernel_type_for_var.place_,
kernel_pair.second.place_), expected_kernel_type.place_),
"TransDataLayout only support DataLayout transform on same place!"); "TransDataLayout only support DataLayout transform on same place!");
PADDLE_ENFORCE(kernel_pair.first.data_type_ == kernel_pair.second.data_type_,
"TransDataLayout only support Datatype are same!");
auto src = in.Get<Tensor>(); PADDLE_ENFORCE(arity(in.dims()) == 4, "Input Arity only support 4!");
auto* dst = out->GetMutable<Tensor>();
PADDLE_ENFORCE(arity(src.dims()) == 4, "Input Arity Only Suppport 4!"); auto& pool = platform::DeviceContextPool::Instance();
auto src_dim = src.dims(); auto src_dim = in.dims();
std::vector<int64_t> dst_dim; std::vector<int64_t> dst_dim;
auto axis = GetAxis(kernel_type_for_var.data_layout_,
expected_kernel_type.data_layout_);
dst_dim.resize(axis.size()); dst_dim.resize(axis.size());
for (size_t i = 0; i < axis.size(); i++) { for (size_t i = 0; i < axis.size(); i++) {
dst_dim[i] = src_dim[axis[i]]; dst_dim[i] = src_dim[axis[i]];
} }
dst->Resize(make_ddim(dst_dim)); out->Resize(make_ddim(dst_dim));
auto place = kernel_pair.second.place_; out->mutable_data(expected_kernel_type.place_, in.type());
dst->mutable_data(place, src.type());
auto src_type = kernel_pair.first.data_type_; framework::VisitDataType(
framework::VisitDataType(src_type, CastDataLayout(ctx, axis, src, dst)); framework::ToDataType(in.type()),
CastDataLayout(pool.Get(expected_kernel_type.place_), axis, in, out));
dst->set_layout(kernel_pair.second.data_layout_); out->set_layout(expected_kernel_type.data_layout_);
} }
} // namespace framework } // namespace framework
......
paddle/framework/data_layout_transform.h
浏览文件 @ e92fe9e6
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. /* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License"); Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License. you may not use this file except in compliance with the License.
...@@ -15,17 +15,17 @@ limitations under the License. */ ...@@ -15,17 +15,17 @@ limitations under the License. */
#pragma once #pragma once
#include "paddle/framework/op_kernel_type.h" #include "paddle/framework/op_kernel_type.h"
#include "paddle/framework/tensor.h"
#include "paddle/framework/variable.h" #include "paddle/framework/variable.h"
namespace paddle { namespace paddle {
namespace framework { namespace framework {
using KernelTypePair = std::pair<OpKernelType, OpKernelType>; std::vector<int> GetAxis(const DataLayout& from, const DataLayout& to);
void TransDataLayout(const std::vector<int>& axis, void TransDataLayout(const OpKernelType& kernel_type_for_var,
const platform::DeviceContext* ctx, const OpKernelType& expected_kernel_type, const Tensor& in,
const KernelTypePair& kernel_pair, const Variable& in, Tensor* out);
Variable* out);
} // namespace framework } // namespace framework
} // namespace paddle } // namespace paddle
paddle/framework/data_layout_transform_test.cc 0 → 100644
浏览文件 @ e92fe9e6
/* 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. */
#include "paddle/framework/data_layout_transform.h"
#include "gtest/gtest.h"
#include "paddle/platform/device_context.h"
TEST(DataTransform, DataLayoutFunction) {
using namespace paddle::framework;
using namespace paddle::platform;
auto place = CPUPlace();
Tensor in = Tensor();
Tensor out = Tensor();
in.mutable_data<double>(make_ddim({2, 3, 1, 2}), place);
in.set_layout(DataLayout::kNHWC);
auto kernel_nhwc = OpKernelType(proto::DataType::FP32, place,
DataLayout::kNHWC, LibraryType::kPlain);
auto kernel_ncwh = OpKernelType(proto::DataType::FP32, place,
DataLayout::kNCHW, LibraryType::kPlain);
TransDataLayout(kernel_nhwc, kernel_ncwh, in, &out);
EXPECT_TRUE(out.layout() == DataLayout::kNCHW);
EXPECT_TRUE(out.dims() == make_ddim({2, 2, 3, 1}));
TransDataLayout(kernel_ncwh, kernel_nhwc, in, &out);
EXPECT_TRUE(in.layout() == DataLayout::kNHWC);
EXPECT_TRUE(in.dims() == make_ddim({2, 3, 1, 2}));
}
\ No newline at end of file
paddle/framework/data_transform.cc
浏览文件 @ e92fe9e6
...@@ -15,18 +15,43 @@ limitations under the License. */ ...@@ -15,18 +15,43 @@ limitations under the License. */
#include "paddle/framework/data_transform.h" #include "paddle/framework/data_transform.h"
#include "paddle/framework/data_device_transform.h" #include "paddle/framework/data_device_transform.h"
#include "paddle/framework/data_layout_transform.h"
namespace paddle { namespace paddle {
namespace framework { namespace framework {
static void PassTensorData(Tensor* from, Tensor* to) {
to->ShareDataWith(*from);
*from = Tensor();
}
void DataTransform(const OpKernelType& expected_kernel_type, void DataTransform(const OpKernelType& expected_kernel_type,
const OpKernelType& kernel_type_for_var, const OpKernelType& kernel_type_for_var,
const Tensor& input_tensor, Tensor* out) { const Tensor& input_tensor, Tensor* output_tensor) {
bool transformed = false;
Tensor in;
in.ShareDataWith(input_tensor);
Tensor out;
// do layout transform
if (NeedTransformLayout(expected_kernel_type.data_layout_,
kernel_type_for_var.data_layout_)) {
TransDataLayout(kernel_type_for_var, expected_kernel_type, in, &out);
transformed = true;
PassTensorData(&out, &in);
}
// do device transform
if (!platform::is_same_place(kernel_type_for_var.place_, if (!platform::is_same_place(kernel_type_for_var.place_,
expected_kernel_type.place_)) { expected_kernel_type.place_)) {
DeviceTransform(input_tensor, expected_kernel_type.place_, out); DeviceTransform(in, expected_kernel_type.place_, &out);
transformed = true;
PassTensorData(&out, &in);
} }
PADDLE_ENFORCE_NOT_NULL(out, "out should not be null");
PADDLE_ENFORCE(transformed, "no transform is done, please check!");
// get output data
output_tensor->ShareDataWith(in);
} }
void CopyVariableWithTensor(const Variable& in_var, const Tensor& tensor, void CopyVariableWithTensor(const Variable& in_var, const Tensor& tensor,
......
paddle/framework/lod_tensor.cc
浏览文件 @ e92fe9e6
...@@ -286,48 +286,86 @@ void DeserializeFromStream(std::istream &is, LoDTensor *tensor, ...@@ -286,48 +286,86 @@ void DeserializeFromStream(std::istream &is, LoDTensor *tensor,
DeserializeFromStream(is, static_cast<Tensor *>(tensor), dev_ctx); DeserializeFromStream(is, static_cast<Tensor *>(tensor), dev_ctx);
} }
// TODO(tonyyang-svail): make this function support LoD
std::vector<LoDTensor> LoDTensor::SplitLoDTensor( std::vector<LoDTensor> LoDTensor::SplitLoDTensor(
const std::vector<platform::Place> places) const { const std::vector<platform::Place> places) const {
check_memory_size(); check_memory_size();
PADDLE_ENFORCE(lod().empty(), "Disable parallel lod for now"); int batch_size =
PADDLE_ENFORCE(dims()[0] % places.size() == 0, lod().empty() ? dims()[0] : static_cast<int>(lod()[0].size()) - 1;
"Batch size should be divided by places size"); size_t result_size = std::min(static_cast<size_t>(batch_size), places.size());
size_t remainder = batch_size % places.size();
std::vector<LoDTensor> lods;
for (size_t place_idx = 0; place_idx < places.size(); ++place_idx) { std::vector<LoDTensor> results;
int begin = place_idx * dims()[0] / places.size(); results.reserve(result_size);
int end = (place_idx + 1) * dims()[0] / places.size();
int step_width = static_cast<int>(batch_size / result_size);
for (size_t i = 0; i < result_size; ++i) {
int begin = static_cast<int>(i * step_width);
int end = static_cast<int>((i + 1) * step_width);
if (i + 1 == places.size()) { // last
end += remainder;
}
auto src = Slice(begin, end);
auto &dst_place = places[place_idx];
LoDTensor dst; LoDTensor dst;
framework::Copy(src, dst_place, &dst); if (lod().empty()) {
auto src = Slice(begin, end);
lods.emplace_back(dst); auto &dst_place = places[i];
framework::Copy(src, dst_place, &dst);
} else {
auto lod_and_offset = GetSubLoDAndAbsoluteOffset(lod(), begin, end, 0);
auto &offset = lod_and_offset.second;
auto src = Slice(offset.first, offset.second);
auto &dst_place = places[i];
framework::Copy(src, dst_place, &dst);
LoD my_lod;
for (auto &l : lod_and_offset.first) {
std::vector<size_t> v{0};
for (auto &ll : l) {
v.push_back(ll + v.back());
}
my_lod.emplace_back(v);
}
dst.set_lod(my_lod);
}
results.emplace_back(dst);
} }
return lods; return results;
} }
// TODO(tonyyang-svail): make this function support LoD
void LoDTensor::MergeLoDTensor( void LoDTensor::MergeLoDTensor(
const std::vector<const LoDTensor *> &lod_tensors, const std::vector<const LoDTensor *> &lod_tensors,
platform::Place dst_place) { platform::Place dst_place) {
PADDLE_ENFORCE(!lod_tensors.empty()); PADDLE_ENFORCE(!lod_tensors.empty());
framework::DDim new_dim = lod_tensors[0]->dims(); framework::DDim new_dim = lod_tensors[0]->dims();
std::type_index new_type = lod_tensors[0]->type(); std::type_index new_type = lod_tensors[0]->type();
auto new_layout = lod_tensors[0]->layout(); framework::DataLayout new_layout = lod_tensors[0]->layout();
for (auto *lod : lod_tensors) { LoD new_lod = lod_tensors[0]->lod();
PADDLE_ENFORCE(new_dim == lod->dims()); for (size_t i = 1; i < lod_tensors.size(); ++i) {
PADDLE_ENFORCE(new_type == lod->type()); auto *t = lod_tensors[i];
PADDLE_ENFORCE(new_layout == lod->layout()); PADDLE_ENFORCE_EQ(new_type.hash_code(), t->type().hash_code());
PADDLE_ENFORCE_EQ(new_layout, t->layout());
PADDLE_ENFORCE_EQ(framework::product(new_dim) / new_dim[0],
framework::product(t->dims()) / t->dims()[0]);
new_dim[0] += t->dims()[0];
auto &lod = t->lod();
for (size_t j = 0; j < lod.size(); ++j) {
auto &sub_lod = new_lod[j];
auto &offset = sub_lod.back();
for (size_t k = 1; k < lod[j].size(); ++k) {
sub_lod.push_back(lod[j][k] + offset);
}
}
} }
new_dim[0] *= lod_tensors.size();
Resize(new_dim); Resize(new_dim);
set_layout(new_layout); set_layout(new_layout);
set_lod(new_lod);
mutable_data(dst_place, new_type); mutable_data(dst_place, new_type);
int begin = 0; int begin = 0;
for (auto *src : lod_tensors) { for (auto *src : lod_tensors) {
int end = begin + src->dims()[0]; int end = begin + src->dims()[0];
......
paddle/framework/lod_tensor_test.cc
浏览文件 @ e92fe9e6
...@@ -100,6 +100,71 @@ TEST(LoD, ToAbsOffset) { ...@@ -100,6 +100,71 @@ TEST(LoD, ToAbsOffset) {
EXPECT_EQ(abs_lod, expected); EXPECT_EQ(abs_lod, expected);
} }
TEST(LoD, SplitLoDTensor) {
LoD lod;
lod.push_back(std::vector<size_t>({0, 2, 4, 5, 6}));
lod.push_back(std::vector<size_t>({0, 1, 6, 8, 13, 15, 20}));
platform::CPUPlace place;
LoDTensor lod_tensor;
lod_tensor.Resize({20, 1});
float* dst_ptr = lod_tensor.mutable_data<float>(place);
for (int i = 0; i < lod_tensor.numel(); ++i) {
dst_ptr[i] = i;
}
lod_tensor.set_lod(lod);
std::vector<platform::Place> places{platform::CPUPlace(),
platform::CPUPlace()};
LoD lod0;
lod0.push_back(std::vector<size_t>({0, 2, 4}));
lod0.push_back(std::vector<size_t>({0, 1, 6, 8, 13}));
LoD lod1;
lod1.push_back(std::vector<size_t>({0, 1, 2}));
lod1.push_back(std::vector<size_t>({0, 2, 7}));
auto lods = lod_tensor.SplitLoDTensor(places);
EXPECT_EQ(lods[0].lod(), lod0);
EXPECT_EQ(lods[1].lod(), lod1);
}
TEST(LoD, MergeLoDTensor) {
LoD lod;
lod.push_back(std::vector<size_t>({0, 2, 4, 5, 6}));
lod.push_back(std::vector<size_t>({0, 1, 6, 8, 13, 15, 20}));
platform::CPUPlace place;
LoDTensor lod_tensor0;
LoD lod0;
lod0.push_back(std::vector<size_t>({0, 2, 4}));
lod0.push_back(std::vector<size_t>({0, 1, 6, 8, 13}));
lod_tensor0.set_lod(lod0);
lod_tensor0.Resize({13, 1});
float* dst_ptr = lod_tensor0.mutable_data<float>(place);
for (int i = 0; i < lod_tensor0.numel(); ++i) {
dst_ptr[i] = i;
}
LoDTensor lod_tensor1;
LoD lod1;
lod1.push_back(std::vector<size_t>({0, 1, 2}));
lod1.push_back(std::vector<size_t>({0, 2, 7}));
lod_tensor1.set_lod(lod1);
lod_tensor1.Resize({7, 1});
dst_ptr = lod_tensor1.mutable_data<float>(place);
for (int i = 0; i < lod_tensor1.numel(); ++i) {
dst_ptr[i] = i;
}
std::vector<const LoDTensor*> lods{&lod_tensor0, &lod_tensor1};
LoDTensor lod_tensor;
lod_tensor.MergeLoDTensor(lods, place);
EXPECT_EQ(lod_tensor.lod(), lod);
}
TEST(LoD, CheckLoD) { TEST(LoD, CheckLoD) {
LoD relative_lod; LoD relative_lod;
relative_lod.push_back(std::vector<size_t>({0, 2})); relative_lod.push_back(std::vector<size_t>({0, 2}));
......
paddle/framework/op_kernel_type.h
浏览文件 @ e92fe9e6
...@@ -85,9 +85,14 @@ inline std::string KernelTypeToString(const OpKernelType& kernel_key) { ...@@ -85,9 +85,14 @@ inline std::string KernelTypeToString(const OpKernelType& kernel_key) {
return stream.str(); return stream.str();
} }
inline bool NeedTransformLayout(const DataLayout& l, const DataLayout& r) {
return l != DataLayout::kAnyLayout && r != DataLayout::kAnyLayout && l != r;
}
inline bool TransFromNeeded(const OpKernelType& l, const OpKernelType& r) { inline bool TransFromNeeded(const OpKernelType& l, const OpKernelType& r) {
return (!platform::places_are_same_class(l.place_, r.place_)) || return (!platform::places_are_same_class(l.place_, r.place_)) ||
(l.data_type_ != r.data_type_) || (l.data_layout_ != r.data_layout_); (l.data_type_ != r.data_type_) ||
NeedTransformLayout(l.data_layout_, r.data_layout_);
} }
} // namespace framework } // namespace framework
......
paddle/framework/operator.cc
浏览文件 @ e92fe9e6
...@@ -485,9 +485,15 @@ void OperatorWithKernel::Run(const Scope& scope, ...@@ -485,9 +485,15 @@ void OperatorWithKernel::Run(const Scope& scope,
// } // }
auto expected_kernel_key = this->GetExpectedKernelType(ctx); auto expected_kernel_key = this->GetExpectedKernelType(ctx);
VLOG(3) << "expected_kernel_key:" << expected_kernel_key; VLOG(3) << "expected_kernel_key:" << expected_kernel_key;
auto kernel_iter = kernels.find(expected_kernel_key);
if (kernel_iter == kernels.end()) {
PADDLE_THROW("op %s does not have kernel for %s", type_,
KernelTypeToString(expected_kernel_key));
}
// do data transform
Scope& new_scope = scope.NewScope(); Scope& new_scope = scope.NewScope();
for (auto& var_name_item : this->Inputs()) { for (auto& var_name_item : this->Inputs()) {
...@@ -520,8 +526,6 @@ void OperatorWithKernel::Run(const Scope& scope, ...@@ -520,8 +526,6 @@ void OperatorWithKernel::Run(const Scope& scope,
} }
} }
auto kernel_iter = kernels.find(expected_kernel_key);
auto* new_dev_ctx = pool.Get(expected_kernel_key.place_); auto* new_dev_ctx = pool.Get(expected_kernel_key.place_);
kernel_iter->second->Compute( kernel_iter->second->Compute(
ExecutionContext(*this, new_scope, *new_dev_ctx)); ExecutionContext(*this, new_scope, *new_dev_ctx));
......
paddle/gserver/tests/sequence_recurrent_group.py
浏览文件 @ e92fe9e6
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. # Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
# #
#Licensed under the Apache License, Version 2.0 (the "License"); # Licensed under the Apache License, Version 2.0 (the "License");
#you may not use this file except in compliance with the License. # you may not use this file except in compliance with the License.
#You may obtain a copy of the License at # You may obtain a copy of the License at
# #
# http://www.apache.org/licenses/LICENSE-2.0 # http://www.apache.org/licenses/LICENSE-2.0
# #
#Unless required by applicable law or agreed to in writing, software # Unless required by applicable law or agreed to in writing, software
#distributed under the License is distributed on an "AS IS" BASIS, # distributed under the License is distributed on an "AS IS" BASIS,
#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
#limitations under the License. # limitations under the License.
from paddle.trainer_config_helpers import * from paddle.trainer_config_helpers import *
######################## data source ################################ ######################## data source ################################
......
paddle/inference/CMakeLists.txt
浏览文件 @ e92fe9e6
...@@ -8,27 +8,6 @@ cc_library(paddle_fluid_api ...@@ -8,27 +8,6 @@ cc_library(paddle_fluid_api
# Merge all modules into a simgle static library # Merge all modules into a simgle static library
cc_library(paddle_fluid DEPS paddle_fluid_api ${FLUID_CORE_MODULES}) cc_library(paddle_fluid DEPS paddle_fluid_api ${FLUID_CORE_MODULES})
# ptools
# just for testing, we may need to change the storing format for inference_model
# and move the dependent of pickle.
# download from http://www.picklingtools.com/
# build in the C++ sub-directory, using command
# make -f Makefile.Linux libptools.so
set(PTOOLS_LIB)
set(PTOOLS_ROOT $ENV{PTOOLS_ROOT} CACHE PATH "Folder contains PicklingTools")
find_path(PTOOLS_INC_DIR chooseser.h PATHS ${PTOOLS_ROOT}/C++)
find_library(PTOOLS_SHARED_LIB NAMES ptools PATHS ${PTOOLS_ROOT}/C++)
if(PTOOLS_INC_DIR AND PTOOLS_SHARED_LIB)
add_definitions(-DPADDLE_USE_PTOOLS)
set(PTOOLS_LIB ptools)
message(STATUS "Found PicklingTools: ${PTOOLS_SHARED_LIB}")
add_library(${PTOOLS_LIB} SHARED IMPORTED GLOBAL)
set_property(TARGET ${PTOOLS_LIB} PROPERTY IMPORTED_LOCATION ${PTOOLS_SHARED_LIB})
include_directories(${PTOOLS_ROOT}/C++)
include_directories(${PTOOLS_ROOT}/C++/opencontainers_1_8_5/include)
add_definitions(-DOC_NEW_STYLE_INCLUDES) # used in ptools
endif()
add_executable(example example.cc) add_executable(example example.cc)
if(APPLE) if(APPLE)
set(OPTIONAL_LINK_FLAGS) set(OPTIONAL_LINK_FLAGS)
......
paddle/inference/example.cc
浏览文件 @ e92fe9e6
...@@ -18,33 +18,21 @@ limitations under the License. */ ...@@ -18,33 +18,21 @@ limitations under the License. */
#include "paddle/inference/inference.h" #include "paddle/inference/inference.h"
DEFINE_string(dirname, "", "Directory of the inference model."); DEFINE_string(dirname, "", "Directory of the inference model.");
DEFINE_string(feed_var_names, "", "Names of feeding variables");
DEFINE_string(fetch_var_names, "", "Names of fetching variables");
int main(int argc, char** argv) { int main(int argc, char** argv) {
google::ParseCommandLineFlags(&argc, &argv, true); google::ParseCommandLineFlags(&argc, &argv, true);
if (FLAGS_dirname.empty() || FLAGS_feed_var_names.empty() || if (FLAGS_dirname.empty()) {
FLAGS_fetch_var_names.empty()) {
// Example: // Example:
// ./example --dirname=recognize_digits_mlp.inference.model // ./example --dirname=recognize_digits_mlp.inference.model
// --feed_var_names="x" std::cout << "Usage: ./example --dirname=path/to/your/model" << std::endl;
// --fetch_var_names="fc_2.tmp_2"
std::cout << "Usage: ./example --dirname=path/to/your/model "
"--feed_var_names=x --fetch_var_names=y"
<< std::endl;
exit(1); exit(1);
} }
std::cout << "FLAGS_dirname: " << FLAGS_dirname << std::endl; std::cout << "FLAGS_dirname: " << FLAGS_dirname << std::endl;
std::cout << "FLAGS_feed_var_names: " << FLAGS_feed_var_names << std::endl;
std::cout << "FLAGS_fetch_var_names: " << FLAGS_fetch_var_names << std::endl;
std::string dirname = FLAGS_dirname; std::string dirname = FLAGS_dirname;
std::vector<std::string> feed_var_names = {FLAGS_feed_var_names};
std::vector<std::string> fetch_var_names = {FLAGS_fetch_var_names};
paddle::InferenceEngine* engine = new paddle::InferenceEngine(); paddle::InferenceEngine* engine = new paddle::InferenceEngine();
engine->LoadInferenceModel(dirname, feed_var_names, fetch_var_names); engine->LoadInferenceModel(dirname);
paddle::framework::LoDTensor input; paddle::framework::LoDTensor input;
srand(time(0)); srand(time(0));
......
paddle/inference/inference.cc
浏览文件 @ e92fe9e6
...@@ -25,19 +25,37 @@ limitations under the License. */ ...@@ -25,19 +25,37 @@ limitations under the License. */
namespace paddle { namespace paddle {
void InferenceEngine::LoadInferenceModel(const std::string& dirname) {
std::string model_filename = dirname + "/__model__.dat";
LOG(INFO) << "loading model from " << model_filename;
std::ifstream inputfs(model_filename, std::ios::in | std::ios::binary);
std::string program_desc_str;
inputfs.seekg(0, std::ios::end);
program_desc_str.resize(inputfs.tellg());
inputfs.seekg(0, std::ios::beg);
LOG(INFO) << "program_desc_str's size: " << program_desc_str.size();
inputfs.read(&program_desc_str[0], program_desc_str.size());
inputfs.close();
program_ = new framework::ProgramDesc(program_desc_str);
GenerateLoadProgram(dirname);
framework::BlockDesc* global_block = program_->MutableBlock(0);
feed_var_names_.clear();
fetch_var_names_.clear();
for (auto* op : global_block->AllOps()) {
if (op->Type() == "feed") {
feed_var_names_.insert(feed_var_names_.begin(), op->Output("Out")[0]);
} else if (op->Type() == "fetch") {
fetch_var_names_.push_back(op->Input("X")[0]);
}
}
}
void InferenceEngine::LoadInferenceModel( void InferenceEngine::LoadInferenceModel(
const std::string& dirname, const std::string& dirname,
const std::vector<std::string>& feed_var_names, const std::vector<std::string>& feed_var_names,
const std::vector<std::string>& fetch_var_names) { const std::vector<std::string>& fetch_var_names) {
#ifdef PADDLE_USE_PTOOLS
std::string model_filename = dirname + "/__model__";
LOG(INFO) << "Using PicklingTools, loading model from " << model_filename;
Val v;
LoadValFromFile(model_filename.c_str(), v, SERIALIZE_P0);
std::string program_desc_str = v["program_desc_str"];
LOG(INFO) << "program_desc_str's size: " << program_desc_str.size();
// PicklingTools cannot parse the vector of strings correctly.
#else
std::string model_filename = dirname + "/__model__.dat"; std::string model_filename = dirname + "/__model__.dat";
LOG(INFO) << "loading model from " << model_filename; LOG(INFO) << "loading model from " << model_filename;
std::ifstream inputfs(model_filename, std::ios::in | std::ios::binary); std::ifstream inputfs(model_filename, std::ios::in | std::ios::binary);
...@@ -48,7 +66,7 @@ void InferenceEngine::LoadInferenceModel( ...@@ -48,7 +66,7 @@ void InferenceEngine::LoadInferenceModel(
LOG(INFO) << "program_desc_str's size: " << program_desc_str.size(); LOG(INFO) << "program_desc_str's size: " << program_desc_str.size();
inputfs.read(&program_desc_str[0], program_desc_str.size()); inputfs.read(&program_desc_str[0], program_desc_str.size());
inputfs.close(); inputfs.close();
#endif
program_ = new framework::ProgramDesc(program_desc_str); program_ = new framework::ProgramDesc(program_desc_str);
GenerateLoadProgram(dirname); GenerateLoadProgram(dirname);
...@@ -62,7 +80,7 @@ void InferenceEngine::LoadInferenceModel( ...@@ -62,7 +80,7 @@ void InferenceEngine::LoadInferenceModel(
} }
bool InferenceEngine::IsParameter(const framework::VarDesc* var) { bool InferenceEngine::IsParameter(const framework::VarDesc* var) {
if (var->Persistable()) { if (var->Persistable() && var->Name() != "feed" && var->Name() != "fetch") {
// There are many unreachable variables in the program // There are many unreachable variables in the program
for (size_t i = 0; i < program_->Size(); ++i) { for (size_t i = 0; i < program_->Size(); ++i) {
const framework::BlockDesc& block = program_->Block(i); const framework::BlockDesc& block = program_->Block(i);
......
paddle/inference/inference.h
浏览文件 @ e92fe9e6
...@@ -28,6 +28,7 @@ public: ...@@ -28,6 +28,7 @@ public:
delete load_program_; delete load_program_;
} }
void LoadInferenceModel(const std::string& dirname);
void LoadInferenceModel(const std::string& dirname, void LoadInferenceModel(const std::string& dirname,
const std::vector<std::string>& feed_var_names, const std::vector<std::string>& feed_var_names,
const std::vector<std::string>& fetch_var_names); const std::vector<std::string>& fetch_var_names);
......
paddle/operators/CMakeLists.txt
浏览文件 @ e92fe9e6
...@@ -178,14 +178,13 @@ foreach(src ${GENERAL_OPS}) ...@@ -178,14 +178,13 @@ foreach(src ${GENERAL_OPS})
endforeach() endforeach()
file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(logical_and);\nUSE_NO_KERNEL_OP(read_from_array);\n") file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(logical_and);\nUSE_NO_KERNEL_OP(read_from_array);\n")
set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library") set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")
cc_test(gather_test SRCS gather_test.cc DEPS tensor) cc_test(gather_test SRCS gather_test.cc DEPS tensor)
cc_test(net_op_test SRCS net_op_test.cc DEPS net_op) cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
cc_test(scatter_test SRCS scatter_test.cc DEPS tensor) cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)
cc_test(beam_search_decode_op_test SRCS beam_search_decode_op_test.cc DEPS lod_tensor) cc_test(beam_search_decode_op_test SRCS beam_search_decode_op_test.cc DEPS lod_tensor)
cc_test(beam_search_op_test SRCS beam_search_op_test.cc DEPS lod_tensor beam_search_op)
cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor paddle_memory) cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor paddle_memory)
if(WITH_GPU) if(WITH_GPU)
cc_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context) cc_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
......
paddle/operators/beam_search_op.cc
浏览文件 @ e92fe9e6
...@@ -29,7 +29,7 @@ void BeamSearch::operator()(const framework::LoDTensor &pre_ids, ...@@ -29,7 +29,7 @@ void BeamSearch::operator()(const framework::LoDTensor &pre_ids,
PruneEndidCandidates(pre_ids, &selected_items); PruneEndidCandidates(pre_ids, &selected_items);
// calculate the output tensor's height // calculate the output tensor's height
size_t num_instances = std::accumulate( size_t num_instances = std::accumulate(
std::begin(items), std::end(items), 0, std::begin(selected_items), std::end(selected_items), 0,
[](size_t a, std::vector<Item> &b) { return a + b.size(); }); [](size_t a, std::vector<Item> &b) { return a + b.size(); });
// the output tensor shape should be [num_instances, 1] // the output tensor shape should be [num_instances, 1]
auto dims = framework::make_ddim( auto dims = framework::make_ddim(
...@@ -48,12 +48,20 @@ void BeamSearch::operator()(const framework::LoDTensor &pre_ids, ...@@ -48,12 +48,20 @@ void BeamSearch::operator()(const framework::LoDTensor &pre_ids,
size_t low_offset = 0; size_t low_offset = 0;
for (auto &items : selected_items) { for (auto &items : selected_items) {
low_level.push_back(low_offset); low_level.push_back(low_offset);
sort(items.begin(), items.end(), [](const Item &a, const Item &b) {
if (a.offset < b.offset) {
return true;
}
return a.id < b.id;
});
for (auto &item : items) { for (auto &item : items) {
ids_data[low_offset] = item.id; ids_data[low_offset] = item.id;
scores_data[low_offset] = item.score; scores_data[low_offset] = item.score;
low_offset++; low_offset++;
} }
} }
low_level.push_back(low_offset);
// fill lod // fill lod
auto abs_lod = framework::ToAbsOffset(ids_->lod()); auto abs_lod = framework::ToAbsOffset(ids_->lod());
auto &high_level = abs_lod[lod_level_]; auto &high_level = abs_lod[lod_level_];
...@@ -64,16 +72,21 @@ void BeamSearch::operator()(const framework::LoDTensor &pre_ids, ...@@ -64,16 +72,21 @@ void BeamSearch::operator()(const framework::LoDTensor &pre_ids,
selected_scores->set_lod(lod); selected_scores->set_lod(lod);
} }
void BeamSearch::PruneEndidCandidates(const framework::LoDTensor &pre_ids, int BeamSearch::PruneEndidCandidates(const framework::LoDTensor &pre_ids,
std::vector<std::vector<Item>> *items) { std::vector<std::vector<Item>> *items) {
auto *pre_ids_data = pre_ids.data<int64_t>(); auto *pre_ids_data = pre_ids.data<int64_t>();
int res = 0;
for (size_t offset = 0; offset < items->size(); offset++) { for (size_t offset = 0; offset < items->size(); offset++) {
auto prefix_id = pre_ids_data[offset]; auto prefix_id = pre_ids_data[offset];
if (prefix_id == end_id_) { if (prefix_id == end_id_) {
items->at(offset).clear(); items->at(offset).clear();
} else {
res++;
} }
} }
return res;
} }
std::vector<std::vector<BeamSearch::Item>> BeamSearch::ToMap( std::vector<std::vector<BeamSearch::Item>> BeamSearch::ToMap(
...@@ -121,11 +134,7 @@ bool BeamSearch::NextItemSet(std::vector<BeamSearch::Item> *items) { ...@@ -121,11 +134,7 @@ bool BeamSearch::NextItemSet(std::vector<BeamSearch::Item> *items) {
auto ids = *ids_; auto ids = *ids_;
auto scores = *scores_; auto scores = *scores_;
auto source_abs_two_level_lod = framework::SliceInLevel(
ids.lod(), lod_level_, sent_offset_, sent_offset_ + 1);
source_abs_two_level_lod = framework::ToAbsOffset(source_abs_two_level_lod);
auto abs_lod = framework::ToAbsOffset(ids.lod()); auto abs_lod = framework::ToAbsOffset(ids.lod());
PADDLE_ENFORCE_GE(source_abs_two_level_lod.size(), 2UL);
auto *ids_data = ids.data<int64_t>(); auto *ids_data = ids.data<int64_t>();
auto *scores_data = scores.data<float>(); auto *scores_data = scores.data<float>();
......
paddle/operators/beam_search_op.h
浏览文件 @ e92fe9e6
...@@ -73,7 +73,15 @@ namespace operators { ...@@ -73,7 +73,15 @@ namespace operators {
* second level: * second level:
* [0, 2, 4] * [0, 2, 4]
* *
* tensor's data * id tensor's data
* [[
* 4,
* 1,
* 3,
* 8,
* ]]
*
* score tensor's data
* [[ * [[
* 0.5, * 0.5,
* 0.3, * 0.3,
...@@ -137,16 +145,21 @@ class BeamSearch { ...@@ -137,16 +145,21 @@ class BeamSearch {
Item() {} Item() {}
Item(size_t offset, size_t id, float score) Item(size_t offset, size_t id, float score)
: offset(offset), id(id), score(score) {} : offset(offset), id(id), score(score) {}
// offset in the lod_level_+1 // offset in the higher lod level.
size_t offset; size_t offset;
// // prefix id in the lower lod level.
// size_t prefix;
// the candidate id // the candidate id
id_t id; id_t id;
// the corresponding score // the corresponding score
score_t score; score_t score;
}; };
void PruneEndidCandidates(const framework::LoDTensor& pre_ids, /*
std::vector<std::vector<Item>>* items); * Delete all the records that follows the end token.
*/
int PruneEndidCandidates(const framework::LoDTensor& pre_ids,
std::vector<std::vector<Item>>* items);
/* /*
* Transform the items into a map whose key is offset, value is the items. * Transform the items into a map whose key is offset, value is the items.
......
paddle/operators/beam_search_op_test.cc 0 → 100644
浏览文件 @ e92fe9e6
/* 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/beam_search_op.h"
#include <gtest/gtest.h>
#include <vector>
namespace paddle {
namespace test {
using std::vector;
using framework::LoDTensor;
using framework::LoD;
using operators::BeamSearch;
using paddle::platform::CPUPlace;
using std::cout;
using std::endl;
void CreateInput(LoDTensor* ids, LoDTensor* scores) {
LoD lod;
vector<size_t> level0({0, 1, 4});
vector<size_t> level1({0, 1, 2, 3, 4});
lod.push_back(level0);
lod.push_back(level1);
ids->set_lod(lod);
scores->set_lod(lod);
auto dims = framework::make_ddim(vector<int64_t>({4, 3}));
ids->Resize(dims);
scores->Resize(dims);
CPUPlace place;
auto* ids_data = ids->mutable_data<int64_t>(place);
auto* scores_data = scores->mutable_data<float>(place);
vector<int64_t> _ids({4, 2, 5, 2, 1, 3, 3, 5, 2, 8, 2, 1});
vector<float> _scores(
{0.5, 0.3, 0.2, 0.6, 0.3, 0.1, 0.9, 0.5, 0.1, 0.7, 0.5, 0.1});
for (int i = 0; i < 12; i++) {
ids_data[i] = _ids[i];
scores_data[i] = _scores[i];
}
}
TEST(beam_search_op, run) {
CPUPlace place;
LoDTensor ids, scores;
CreateInput(&ids, &scores);
LoDTensor pre_ids;
pre_ids.Resize(framework::make_ddim(vector<int64_t>(4, 1)));
for (int i = 0; i < 4; i++) {
pre_ids.mutable_data<int64_t>(place)[i] = i + 1;
}
BeamSearch beamsearch(ids, scores, (int64_t)0, (int64_t)2, 0);
LoDTensor sids, sscores;
beamsearch(pre_ids, &sids, &sscores);
LOG(INFO) << "score: " << sscores << endl;
ASSERT_EQ(sids.lod(), sscores.lod());
vector<int> tids({2, 4, 3, 8});
vector<float> tscores({0.3, 0.5, 0.9, 0.7});
for (int i = 0; i < 4; i++) {
ASSERT_EQ(tids[i], sids.data<int64_t>()[i]);
ASSERT_EQ(tscores[i], sscores.data<float>()[i]);
}
}
} // namespace test
} // namespace paddle
paddle/operators/clip_op.cc
浏览文件 @ e92fe9e6
...@@ -51,8 +51,8 @@ class ClipOpMaker : public framework::OpProtoAndCheckerMaker { ...@@ -51,8 +51,8 @@ class ClipOpMaker : public framework::OpProtoAndCheckerMaker {
AddComment(R"DOC( AddComment(R"DOC(
Clip Operator. Clip Operator.
The clip operator limits the value of given input within an interval. The interval is The clip operator limits the value of given input within an interval. The
specified with arguments 'min' and 'max': interval is specified with arguments 'min' and 'max':
$$ $$
Out = \min(\max(X, min), max) Out = \min(\max(X, min), max)
......
paddle/operators/conv_op.cc
浏览文件 @ e92fe9e6
...@@ -70,6 +70,13 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const { ...@@ -70,6 +70,13 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType ConvOp::GetExpectedKernelType( framework::OpKernelType ConvOp::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const { const framework::ExecutionContext& ctx) const {
bool use_cudnn = ctx.Attr<bool>("use_cudnn"); bool use_cudnn = ctx.Attr<bool>("use_cudnn");
use_cudnn &= platform::is_gpu_place(ctx.GetPlace());
#ifdef PADDLE_WITH_CUDA
if (platform::is_gpu_place(ctx.GetPlace())) {
auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
use_cudnn &= dev_ctx.cudnn_handle() != nullptr;
}
#endif
framework::LibraryType library_; framework::LibraryType library_;
if (use_cudnn) { if (use_cudnn) {
library_ = framework::LibraryType::kCUDNN; library_ = framework::LibraryType::kCUDNN;
...@@ -283,6 +290,14 @@ void ConvOpGrad::InferShape(framework::InferShapeContext* ctx) const { ...@@ -283,6 +290,14 @@ void ConvOpGrad::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType ConvOpGrad::GetExpectedKernelType( framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const { const framework::ExecutionContext& ctx) const {
bool use_cudnn = ctx.Attr<bool>("use_cudnn"); bool use_cudnn = ctx.Attr<bool>("use_cudnn");
use_cudnn &= platform::is_gpu_place(ctx.GetPlace());
#ifdef PADDLE_WITH_CUDA
if (platform::is_gpu_place(ctx.GetPlace())) {
auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
use_cudnn &= dev_ctx.cudnn_handle() != nullptr;
}
#endif
framework::LibraryType library_; framework::LibraryType library_;
if (use_cudnn) { if (use_cudnn) {
library_ = framework::LibraryType::kCUDNN; library_ = framework::LibraryType::kCUDNN;
......
paddle/operators/conv_transpose_op.cc
浏览文件 @ e92fe9e6
...@@ -61,6 +61,13 @@ void ConvTransposeOp::InferShape(framework::InferShapeContext* ctx) const { ...@@ -61,6 +61,13 @@ void ConvTransposeOp::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType ConvTransposeOp::GetExpectedKernelType( framework::OpKernelType ConvTransposeOp::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const { const framework::ExecutionContext& ctx) const {
bool use_cudnn = ctx.Attr<bool>("use_cudnn"); bool use_cudnn = ctx.Attr<bool>("use_cudnn");
use_cudnn &= platform::is_gpu_place(ctx.GetPlace());
#ifdef PADDLE_WITH_CUDA
if (platform::is_gpu_place(ctx.GetPlace())) {
auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
use_cudnn &= dev_ctx.cudnn_handle() != nullptr;
}
#endif
framework::LibraryType library_; framework::LibraryType library_;
if (use_cudnn) { if (use_cudnn) {
library_ = framework::LibraryType::kCUDNN; library_ = framework::LibraryType::kCUDNN;
...@@ -263,6 +270,13 @@ void ConvTransposeOpGrad::InferShape(framework::InferShapeContext* ctx) const { ...@@ -263,6 +270,13 @@ void ConvTransposeOpGrad::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType ConvTransposeOpGrad::GetExpectedKernelType( framework::OpKernelType ConvTransposeOpGrad::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const { const framework::ExecutionContext& ctx) const {
bool use_cudnn = ctx.Attr<bool>("use_cudnn"); bool use_cudnn = ctx.Attr<bool>("use_cudnn");
use_cudnn &= platform::is_gpu_place(ctx.GetPlace());
#ifdef PADDLE_WITH_CUDA
if (platform::is_gpu_place(ctx.GetPlace())) {
auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
use_cudnn &= dev_ctx.cudnn_handle() != nullptr;
}
#endif
framework::LibraryType library_; framework::LibraryType library_;
if (use_cudnn) { if (use_cudnn) {
library_ = framework::LibraryType::kCUDNN; library_ = framework::LibraryType::kCUDNN;
......
paddle/operators/ctc_align_op.cc 0 → 100644
浏览文件 @ e92fe9e6
/* 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/ctc_align_op.h"
namespace paddle {
namespace operators {
class CTCAlignOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Input"),
"Input of CTCAlignOp should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("Output"),
"Output of CTCAlignOp should not be null.");
auto input_dims = ctx->GetInputDim("Input");
// TODO(wanghaoshuang): it is tricky to set the wrong dimension here.
ctx->SetOutputDim("Output", input_dims);
}
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("Input")->type()),
ctx.device_context());
}
};
class CTCAlignOpMaker : public framework::OpProtoAndCheckerMaker {
public:
CTCAlignOpMaker(OpProto* proto, OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("Input",
"(LodTensor, default: LoDTensor<int>), Its shape is "
"[Lp, 1], where Lp is the sum of all input sequences' length.");
AddOutput("Output", "(Tensor, default: Tensor<int>), The align result.");
AddAttr<int>("blank",
"(int, default: 0), the blank label setted in Connectionist "
"Temporal Classification (CTC) op.")
.SetDefault(0);
AddAttr<bool>("merge_repeated",
"(bool, default: true), whether to "
"merge repeated elements between two blanks. ")
.SetDefault(true);
AddComment(R"DOC(
CTCAlign op is used to merge repeated elements between two blanks
and then delete all blanks in sequence.
Given:
Input.data = [0, 1, 2, 2, 0, 4, 0, 4, 5, 0, 6,
6, 0, 0, 7, 7, 7, 0]
Input.dims = {18, 1}
Input.LoD = [[0, 11, 18]]
And:
blank = 0
merge_repeated = True
Then:
Output.data = [1, 2, 4, 4, 5, 6,
6, 7]
Output.dims = {8, 1}
Output.LoD = [[0, 6, 8]]
)DOC");
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(ctc_align, ops::CTCAlignOp, ops::CTCAlignOpMaker,
paddle::framework::EmptyGradOpMaker);
REGISTER_OP_CPU_KERNEL(
ctc_align, ops::CTCAlignKernel<paddle::platform::CPUDeviceContext, int>,
ops::CTCAlignKernel<paddle::platform::CPUDeviceContext, int64_t>);
paddle/operators/ctc_align_op.cu 0 → 100644
浏览文件 @ e92fe9e6
/* 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 <stdio.h>
#include <thrust/device_vector.h>
#include <thrust/host_vector.h>
#include "paddle/operators/ctc_align_op.h"
namespace paddle {
namespace operators {
template <typename T>
__global__ void MergeAndDelCudaKernel(const int64_t num_token, const T* tokens,
const size_t num_seq, size_t* lod0,
const int blank, const int merge_repeated,
size_t* out_lod0, T* output) {
int ouput_idx = 0;
out_lod0[0] = 0;
for (int i = 0; i < num_seq; ++i) {
T pre_token = -1;
for (int j = lod0[i]; j < lod0[i + 1]; ++j) {
if (tokens[j] != blank && !(merge_repeated && tokens[j] == pre_token)) {
output[ouput_idx] = tokens[j];
++ouput_idx;
}
pre_token = tokens[j];
}
out_lod0[i + 1] = ouput_idx;
}
}
template <typename T>
class CTCAlignOpCUDAKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
"It must use CUDAPlace.");
const size_t level = 0;
auto* input = ctx.Input<LoDTensor>("Input");
auto* output = ctx.Output<LoDTensor>("Output");
auto input_lod = framework::ToAbsOffset(input->lod());
const T* tokens = input->data<T>();
const int64_t num_tokens = input->dims()[0];
const size_t num_seq = input_lod[level].size() - 1;
const int blank = ctx.Attr<int>("blank");
const int merge_repeated =
static_cast<int>(ctx.Attr<bool>("merge_repeated"));
// prepare a lod to record lod information while merging elements
thrust::device_vector<size_t> dev_out_lod0(input_lod[level].size());
size_t* dev_out_lod0_ptr = thrust::raw_pointer_cast(dev_out_lod0.data());
// merge elements and delete blank
T* output_data = output->mutable_data<T>({num_tokens, 1}, ctx.GetPlace());
auto stream = ctx.cuda_device_context().stream();
MergeAndDelCudaKernel<T><<<1, 1, 0, stream>>>(
num_tokens, tokens, num_seq, input_lod[level].data(), blank,
merge_repeated, dev_out_lod0_ptr, output_data);
// set output lod
thrust::host_vector<size_t> host_out_lod0(dev_out_lod0.begin(),
dev_out_lod0.end());
framework::LoD out_lod;
out_lod.push_back(host_out_lod0);
output->set_lod(out_lod);
// resize output dims
output->Resize({static_cast<int64_t>(host_out_lod0.back()), 1});
}
};
} // namespace operators
} // namespace paddle
REGISTER_OP_CUDA_KERNEL(ctc_align, paddle::operators::CTCAlignOpCUDAKernel<int>,
paddle::operators::CTCAlignOpCUDAKernel<int64_t>);
paddle/operators/ctc_align_op.h 0 → 100644
浏览文件 @ e92fe9e6
/* 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 <string.h>
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
using LoDTensor = framework::LoDTensor;
template <typename DeviceContext, typename T>
class CTCAlignKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* input = ctx.Input<LoDTensor>("Input");
auto* output = ctx.Output<LoDTensor>("Output");
const size_t level = 0;
auto input_lod = framework::ToAbsOffset(input->lod());
// check input dims and lod
auto input_dims = input->dims();
PADDLE_ENFORCE_EQ(input_dims[0],
static_cast<int64_t>(input_lod[level].back()),
"The first dimension of Input(Input) should be equal to "
"the sum of all sequences' lengths.");
const size_t num_sequences = input_lod[level].size() - 1;
size_t blank = static_cast<size_t>(ctx.Attr<int>("blank"));
bool merge_repeated = ctx.Attr<bool>("merge_repeated");
// merge repeated tokens and delete blank
T* output_data = output->mutable_data<T>(ctx.GetPlace());
size_t output_idx = 0;
std::vector<size_t> output_lod0(1, 0);
const T* input_data = input->data<T>();
for (size_t seq_idx = 0; seq_idx < num_sequences; ++seq_idx) {
T prev_token = -1;
for (size_t i = input_lod[level][seq_idx];
i < input_lod[level][seq_idx + 1]; ++i) {
if (input_data[i] != blank &&
!(merge_repeated && input_data[i] == prev_token)) {
output_data[output_idx] = input_data[i];
++output_idx;
}
prev_token = input_data[i];
}
output_lod0.push_back(output_idx);
}
// set output lod
framework::LoD output_lod;
output_lod.push_back(output_lod0);
output->set_lod(output_lod);
// resize output dims
output->Resize({static_cast<int64_t>(output_lod0.back()), 1});
}
};
} // namespace operators
} // namespace paddle
paddle/operators/detail/grpc_client.cc
浏览文件 @ e92fe9e6
...@@ -63,9 +63,6 @@ bool RPCClient::AsyncGetVariable(const std::string& ep, ...@@ -63,9 +63,6 @@ bool RPCClient::AsyncGetVariable(const std::string& ep,
sendrecv::VariableMessage req; sendrecv::VariableMessage req;
req.set_varname(var_name); req.set_varname(var_name);
auto* var = scope.FindVar(var_name);
SerializeToMessage(var_name, var, ctx, &req);
// varhandle // varhandle
VarHandle var_h; VarHandle var_h;
var_h.ep = ep; var_h.ep = ep;
......
paddle/operators/detail/grpc_server.cc
浏览文件 @ e92fe9e6
...@@ -83,11 +83,13 @@ class RequestGet final : public RequestBase { ...@@ -83,11 +83,13 @@ class RequestGet final : public RequestBase {
public: public:
explicit RequestGet(sendrecv::SendRecvService::AsyncService* service, explicit RequestGet(sendrecv::SendRecvService::AsyncService* service,
grpc::ServerCompletionQueue* cq, framework::Scope* scope, grpc::ServerCompletionQueue* cq, framework::Scope* scope,
const platform::DeviceContext* dev_ctx) const platform::DeviceContext* dev_ctx,
SimpleBlockQueue<char>* queue)
: RequestBase(service, cq), : RequestBase(service, cq),
responder_(&ctx_), responder_(&ctx_),
scope_(scope), scope_(scope),
dev_ctx_(dev_ctx) { dev_ctx_(dev_ctx),
queue_(queue) {
service_->RequestGetVariable(&ctx_, &request_, &responder_, cq_, cq_, this); service_->RequestGetVariable(&ctx_, &request_, &responder_, cq_, cq_, this);
} }
...@@ -103,6 +105,7 @@ class RequestGet final : public RequestBase { ...@@ -103,6 +105,7 @@ class RequestGet final : public RequestBase {
// TODO(gongwb): check var's info. // TODO(gongwb): check var's info.
responder_.Finish(reply_, grpc::Status::OK, this); responder_.Finish(reply_, grpc::Status::OK, this);
status_ = FINISH; status_ = FINISH;
queue_->Push('c');
} }
protected: protected:
...@@ -111,8 +114,15 @@ class RequestGet final : public RequestBase { ...@@ -111,8 +114,15 @@ class RequestGet final : public RequestBase {
ServerAsyncResponseWriter<sendrecv::VariableMessage> responder_; ServerAsyncResponseWriter<sendrecv::VariableMessage> responder_;
framework::Scope* scope_; framework::Scope* scope_;
const platform::DeviceContext* dev_ctx_; const platform::DeviceContext* dev_ctx_;
SimpleBlockQueue<char>* queue_;
}; };
void AsyncGRPCServer::WaitClientGet(int count) {
for (int i = 0; i < count; ++i) {
var_get_queue_.Pop();
}
}
void AsyncGRPCServer::RunSyncUpdate() { void AsyncGRPCServer::RunSyncUpdate() {
grpc::ServerBuilder builder; grpc::ServerBuilder builder;
builder.AddListeningPort(address_, grpc::InsecureServerCredentials()); builder.AddListeningPort(address_, grpc::InsecureServerCredentials());
...@@ -152,7 +162,6 @@ void AsyncGRPCServer::ShutdownQueue() { ...@@ -152,7 +162,6 @@ void AsyncGRPCServer::ShutdownQueue() {
} }
// This URL explains why shutdown is complicate: // This URL explains why shutdown is complicate:
// https://stackoverflow.com/questions/35708348/grpc-what-is-the-recommended-way-to-shut-down-an-asynchronous-server-in-c
void AsyncGRPCServer::ShutDown() { void AsyncGRPCServer::ShutDown() {
server_->Shutdown(); server_->Shutdown();
ShutdownQueue(); ShutdownQueue();
...@@ -173,10 +182,12 @@ void AsyncGRPCServer::TryToRegisterNewGetOne() { ...@@ -173,10 +182,12 @@ void AsyncGRPCServer::TryToRegisterNewGetOne() {
if (is_shut_down_) { if (is_shut_down_) {
return; return;
} }
RequestGet* get = new RequestGet(&service_, cq_get_.get(), scope_, dev_ctx_); RequestGet* get = new RequestGet(&service_, cq_get_.get(), scope_, dev_ctx_,
&var_get_queue_);
VLOG(4) << "create Requestget status:" << get->Status(); VLOG(4) << "create Requestget status:" << get->Status();
} }
// FIXME(typhoonzero): remove wait argument and change cq_name to enum.
void AsyncGRPCServer::HandleRequest(bool wait, grpc::ServerCompletionQueue* cq, void AsyncGRPCServer::HandleRequest(bool wait, grpc::ServerCompletionQueue* cq,
std::string cq_name, std::string cq_name,
std::function<void()> TryToRegisterNewOne) { std::function<void()> TryToRegisterNewOne) {
...@@ -191,9 +202,9 @@ void AsyncGRPCServer::HandleRequest(bool wait, grpc::ServerCompletionQueue* cq, ...@@ -191,9 +202,9 @@ void AsyncGRPCServer::HandleRequest(bool wait, grpc::ServerCompletionQueue* cq,
} }
PADDLE_ENFORCE(tag); PADDLE_ENFORCE(tag);
if (wait && !done_) { // FIXME(typhoonzero): de-couple the barriers with recv_op
Wait(); if (cq_name == "cq_get") WaitCond(1);
} if (cq_name == "cq_send") WaitCond(0);
RequestBase* base = (RequestBase*)tag; RequestBase* base = (RequestBase*)tag;
// reference: // reference:
...@@ -225,22 +236,18 @@ void AsyncGRPCServer::HandleRequest(bool wait, grpc::ServerCompletionQueue* cq, ...@@ -225,22 +236,18 @@ void AsyncGRPCServer::HandleRequest(bool wait, grpc::ServerCompletionQueue* cq,
} }
} }
void AsyncGRPCServer::Wait() { void AsyncGRPCServer::WaitCond(int cond) {
std::unique_lock<std::mutex> lock(this->mutex_); std::unique_lock<std::mutex> lock(this->barrier_mutex_);
condition_.wait(lock, [=] { return this->done_ == true; }); barrier_condition_.wait(lock,
} [=] { return this->barrier_cond_step_ == cond; });
void AsyncGRPCServer::Reset() {
std::lock_guard<std::mutex> lock(this->mutex_);
done_ = false;
} }
void AsyncGRPCServer::Done() { void AsyncGRPCServer::SetCond(int cond) {
{ {
std::lock_guard<std::mutex> lock(this->mutex_); std::lock_guard<std::mutex> lock(this->barrier_mutex_);
done_ = true; barrier_cond_step_ = cond;
} }
condition_.notify_all(); barrier_condition_.notify_all();
} }
} // namespace detail } // namespace detail
......
paddle/operators/detail/grpc_server.h
浏览文件 @ e92fe9e6
...@@ -41,9 +41,10 @@ class AsyncGRPCServer final : public sendrecv::SendRecvService::Service { ...@@ -41,9 +41,10 @@ class AsyncGRPCServer final : public sendrecv::SendRecvService::Service {
void RunSyncUpdate(); void RunSyncUpdate();
void Reset(); // functions to sync server barrier status.
void WaitCond(int cond);
void Done(); void SetCond(int cond);
void WaitClientGet(int count);
void SetScope(framework::Scope *scope) { scope_ = scope; } void SetScope(framework::Scope *scope) { scope_ = scope; }
...@@ -56,7 +57,6 @@ class AsyncGRPCServer final : public sendrecv::SendRecvService::Service { ...@@ -56,7 +57,6 @@ class AsyncGRPCServer final : public sendrecv::SendRecvService::Service {
void ShutDown(); void ShutDown();
protected: protected:
void Wait();
void HandleRequest(bool wait, grpc::ServerCompletionQueue *cq, void HandleRequest(bool wait, grpc::ServerCompletionQueue *cq,
std::string cq_name, std::string cq_name,
std::function<void()> TryToRegisterNewOne); std::function<void()> TryToRegisterNewOne);
...@@ -78,11 +78,12 @@ class AsyncGRPCServer final : public sendrecv::SendRecvService::Service { ...@@ -78,11 +78,12 @@ class AsyncGRPCServer final : public sendrecv::SendRecvService::Service {
const platform::DeviceContext *dev_ctx_; const platform::DeviceContext *dev_ctx_;
// received variable from RPC, operators fetch variable from this queue. // received variable from RPC, operators fetch variable from this queue.
SimpleBlockQueue<MessageWithName> var_recv_queue_; SimpleBlockQueue<MessageWithName> var_recv_queue_;
SimpleBlockQueue<char> var_get_queue_;
// condition of the sub program // condition of the sub program
std::mutex mutex_; std::mutex barrier_mutex_;
volatile mutable bool done_; mutable int barrier_cond_step_;
std::condition_variable condition_; std::condition_variable barrier_condition_;
std::unique_ptr<std::thread> t_send_; std::unique_ptr<std::thread> t_send_;
std::unique_ptr<std::thread> t_get_; std::unique_ptr<std::thread> t_get_;
......
paddle/operators/edit_distance_op.cc
浏览文件 @ e92fe9e6
...@@ -49,10 +49,10 @@ class EditDistanceOpMaker : public framework::OpProtoAndCheckerMaker { ...@@ -49,10 +49,10 @@ class EditDistanceOpMaker : public framework::OpProtoAndCheckerMaker {
EditDistanceOpMaker(OpProto *proto, OpAttrChecker *op_checker) EditDistanceOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) { : OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("Hyps", AddInput("Hyps",
"(2-D LoDTensor<int>, 2nd dim. equal to 1) " "(2-D LoDTensor<int64_t>, 2nd dim. equal to 1) "
"The indices for hypothesis strings."); "The indices for hypothesis strings.");
AddInput("Refs", AddInput("Refs",
"(2-D LoDTensor<int>, 2nd dim. equal to 1) " "(2-D LoDTensor<int64_t>, 2nd dim. equal to 1) "
"The indices for reference strings."); "The indices for reference strings.");
AddAttr<bool>("normalized", AddAttr<bool>("normalized",
"(bool, default false) Indicated whether to normalize " "(bool, default false) Indicated whether to normalize "
...@@ -66,22 +66,22 @@ class EditDistanceOpMaker : public framework::OpProtoAndCheckerMaker { ...@@ -66,22 +66,22 @@ class EditDistanceOpMaker : public framework::OpProtoAndCheckerMaker {
EditDistance operator computes the edit distances between a batch of hypothesis EditDistance operator computes the edit distances between a batch of hypothesis
strings and their references. strings and their references.
Edit distance, also called Levenshtein distance, measures how dissimilar two strings 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. are by counting the minimum number of operations to transform one string into anthor.
Here the operations include insertion, deletion, and substitution. For example, Here the operations include insertion, deletion, and substitution. For example,
given hypothesis string A = "kitten" and reference B = "sitting", the edit distance 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 is 3 for A will be transformed into B at least after two substitutions and one
insertion: insertion:
"kitten" -> "sitten" -> "sittin" -> "sitting" "kitten" -> "sitten" -> "sittin" -> "sitting"
Input(Hyps) is a LoDTensor consisting of all the hypothesis strings with the total 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. 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 And the `batch_size` reference strings are arranged in order in the same way in the
LoDTensor Input(Refs). LoDTensor Input(Refs).
Output(Out) contains the `batch_size` results and each stands for the edit stance 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 for a pair of strings respectively. If Attr(normalized) is true, the edit distance
will be divided by the length of reference string. will be divided by the length of reference string.
)DOC"); )DOC");
} }
......
paddle/operators/edit_distance_op.cu
浏览文件 @ e92fe9e6
...@@ -39,8 +39,8 @@ __global__ void FillFirstColumn(T* dist, const int M, const int N) { ...@@ -39,8 +39,8 @@ __global__ void FillFirstColumn(T* dist, const int M, const int N) {
} }
template <typename T> template <typename T>
__global__ void Levenshtein(T* dist, const int* x1, const int* x2, const int M, __global__ void Levenshtein(T* dist, const int64_t* x1, const int64_t* x2,
const int N, const int start) { const int M, const int N, const int start) {
int idx = blockDim.x * blockIdx.x + threadIdx.x; int idx = blockDim.x * blockIdx.x + threadIdx.x;
int offset = N; int offset = N;
int index = start + idx * offset; int index = start + idx * offset;
...@@ -113,8 +113,8 @@ class EditDistanceGPUKernel : public framework::OpKernel<T> { ...@@ -113,8 +113,8 @@ class EditDistanceGPUKernel : public framework::OpKernel<T> {
dist_t.Resize({m + 1, n + 1}); dist_t.Resize({m + 1, n + 1});
dist_t.mutable_data<T>(ctx.GetPlace()); dist_t.mutable_data<T>(ctx.GetPlace());
auto dist = dist_t.data<T>(); auto dist = dist_t.data<T>();
auto x1 = x1_t->data<int>() + hyp_lod[num]; auto x1 = x1_t->data<int64_t>() + hyp_lod[num];
auto x2 = x2_t->data<int>() + ref_lod[num]; auto x2 = x2_t->data<int64_t>() + ref_lod[num];
FillFirstColumn<T><<<1 + m / PADDLE_CUDA_NUM_THREADS, FillFirstColumn<T><<<1 + m / PADDLE_CUDA_NUM_THREADS,
PADDLE_CUDA_NUM_THREADS, 0, stream>>>(dist, m, n); PADDLE_CUDA_NUM_THREADS, 0, stream>>>(dist, m, n);
......
paddle/operators/edit_distance_op.h
浏览文件 @ e92fe9e6
...@@ -60,8 +60,8 @@ class EditDistanceKernel : public framework::OpKernel<T> { ...@@ -60,8 +60,8 @@ class EditDistanceKernel : public framework::OpKernel<T> {
dist_t.Resize({m + 1, n + 1}); dist_t.Resize({m + 1, n + 1});
dist_t.mutable_data<T>(ctx.GetPlace()); dist_t.mutable_data<T>(ctx.GetPlace());
auto dist = dist_t.data<T>(); auto dist = dist_t.data<T>();
auto x1 = x1_t->data<int>() + hyp_lod[num]; auto x1 = x1_t->data<int64_t>() + hyp_lod[num];
auto x2 = x2_t->data<int>() + ref_lod[num]; auto x2 = x2_t->data<int64_t>() + ref_lod[num];
for (int64_t i = 0; i < m + 1; ++i) { for (int64_t i = 0; i < m + 1; ++i) {
dist[i * (n + 1)] = i; dist[i * (n + 1)] = i;
} }
......
paddle/operators/elementwise_add_op.h
浏览文件 @ e92fe9e6
...@@ -28,39 +28,7 @@ template <typename DeviceContext, typename T> ...@@ -28,39 +28,7 @@ template <typename DeviceContext, typename T>
class ElementwiseAddKernel : public framework::OpKernel<T> { class ElementwiseAddKernel : public framework::OpKernel<T> {
public: public:
void Compute(const framework::ExecutionContext& ctx) const override { void Compute(const framework::ExecutionContext& ctx) const override {
using Tensor = framework::Tensor; ElementwiseComputeEx<AddFunctor<T>, DeviceContext, T>(ctx);
auto* x = ctx.Input<Tensor>("X");
auto* y = ctx.Input<Tensor>("Y");
auto* z = ctx.Output<Tensor>("Out");
z->mutable_data<T>(ctx.GetPlace());
TransformFunctor<AddFunctor<T>, T, DeviceContext> functor(
x, y, z, ctx.template device_context<DeviceContext>(), AddFunctor<T>());
auto x_dims = x->dims();
auto y_dims = y->dims();
PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
"Rank of first input must >= rank of second input.");
if (x_dims == y_dims) {
functor.Run();
return;
}
int axis = ctx.Attr<int>("axis");
axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(),
"Axis should be in range [0, x_dims)");
int pre, n, post;
get_mid_dims(x_dims, y_dims, axis, pre, n, post);
if (post == 1) {
functor.RunRowWise(n, pre);
return;
} else {
functor.RunMidWise(n, pre, post);
return;
}
} }
}; };
......
paddle/operators/elementwise_div_op.h
浏览文件 @ e92fe9e6
...@@ -19,11 +19,16 @@ limitations under the License. */ ...@@ -19,11 +19,16 @@ limitations under the License. */
namespace paddle { namespace paddle {
namespace operators { namespace operators {
template <typename T>
struct DivFunctor {
inline HOSTDEVICE T operator()(T a, T b) const { return a / b; }
};
template <typename DeviceContext, typename T> template <typename DeviceContext, typename T>
class ElementwiseDivKernel : public framework::OpKernel<T> { class ElementwiseDivKernel : public framework::OpKernel<T> {
public: public:
void Compute(const framework::ExecutionContext& ctx) const override { void Compute(const framework::ExecutionContext& ctx) const override {
ElementwiseCompute<EigenDivFunctor, DeviceContext, T>(ctx); ElementwiseComputeEx<DivFunctor<T>, DeviceContext, T>(ctx);
} }
}; };
......
paddle/operators/elementwise_max_op.cc 0 → 100644
浏览文件 @ e92fe9e6
/* 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/elementwise_max_op.h"
#include "paddle/operators/elementwise_op.h"
namespace paddle {
namespace operators {
class ElementwiseMaxOpMaker : public ElementwiseOpMaker {
public:
ElementwiseMaxOpMaker(OpProto* proto, OpAttrChecker* op_checker)
: ElementwiseOpMaker(proto, op_checker) {
SetComment("Max", "Out = max(X, Y)");
AddComment(comment_);
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(elementwise_max, ops::ElementwiseOp, ops::ElementwiseMaxOpMaker,
elementwise_max_grad, ops::ElementwiseOpGrad);
REGISTER_OP_CPU_KERNEL(
elementwise_max,
ops::ElementwiseMaxKernel<paddle::platform::CPUDeviceContext, float>,
ops::ElementwiseMaxKernel<paddle::platform::CPUDeviceContext, double>,
ops::ElementwiseMaxKernel<paddle::platform::CPUDeviceContext, int>,
ops::ElementwiseMaxKernel<paddle::platform::CPUDeviceContext, int64_t>);
REGISTER_OP_CPU_KERNEL(
elementwise_max_grad,
ops::ElementwiseMaxGradKernel<paddle::platform::CPUDeviceContext, float>,
ops::ElementwiseMaxGradKernel<paddle::platform::CPUDeviceContext, double>,
ops::ElementwiseMaxGradKernel<paddle::platform::CPUDeviceContext, int>,
ops::ElementwiseMaxGradKernel<paddle::platform::CPUDeviceContext, int64_t>);
paddle/operators/elementwise_max_op.cu 0 → 100644
浏览文件 @ e92fe9e6
/* 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/elementwise_max_op.h"
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(
elementwise_max,
ops::ElementwiseMaxKernel<paddle::platform::CUDADeviceContext, float>,
ops::ElementwiseMaxKernel<paddle::platform::CUDADeviceContext, double>,
ops::ElementwiseMaxKernel<paddle::platform::CUDADeviceContext, int>,
ops::ElementwiseMaxKernel<paddle::platform::CUDADeviceContext, int64_t>);
REGISTER_OP_CUDA_KERNEL(
elementwise_max_grad,
ops::ElementwiseMaxGradKernel<paddle::platform::CUDADeviceContext, float>,
ops::ElementwiseMaxGradKernel<paddle::platform::CUDADeviceContext, double>,
ops::ElementwiseMaxGradKernel<paddle::platform::CUDADeviceContext, int>,
ops::ElementwiseMaxGradKernel<paddle::platform::CUDADeviceContext,
int64_t>);
paddle/operators/elementwise_max_op.h 0 → 100644
浏览文件 @ e92fe9e6
/* 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/operators/elementwise_op_function.h"
namespace paddle {
namespace operators {
template <typename T>
struct MaxFunctor {
inline HOSTDEVICE T operator()(T a, T b) const { return a > b ? a : b; }
};
template <typename DeviceContext, typename T>
class ElementwiseMaxKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
ElementwiseComputeEx<MaxFunctor<T>, DeviceContext, T>(ctx);
}
};
template <typename T>
struct ElementwiseMaxGradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = (x_e > y_e).template cast<T>() * dz_e;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = (x_e <= y_e).template cast<T>() * dz_e;
}
}
};
template <typename T>
struct ElementwiseMaxBroadCastGradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ, typename Pre, typename N>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
.broadcast(Eigen::DSizes<int, 2>(pre, 1))
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = (x_e > y_e_bcast).template cast<T>() * dz_e;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = ((x_e <= y_e_bcast).template cast<T>() * dz_e)
.reshape(Eigen::DSizes<int, 2>(pre, n))
.sum(Eigen::array<int, 1>{{0}});
}
}
};
template <typename T>
struct ElementwiseMaxBroadCast2GradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ, typename Pre, typename N, typename Post>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n,
Post post) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
.broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = (x_e > y_e_bcast).template cast<T>() * dz_e;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = ((x_e <= y_e_bcast).template cast<T>() * dz_e)
.reshape(Eigen::DSizes<int, 3>(pre, n, post))
.sum(Eigen::array<int, 2>{{0, 2}});
}
}
};
template <typename DeviceContext, typename T>
class ElementwiseMaxGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
ElementwiseGradCompute<DeviceContext, T, ElementwiseMaxGradFunctor<T>,
ElementwiseMaxBroadCastGradFunctor<T>,
ElementwiseMaxBroadCast2GradFunctor<T>>(ctx);
}
};
} // namespace operators
} // namespace paddle
paddle/operators/elementwise_min_op.cc 0 → 100644
浏览文件 @ e92fe9e6
/* 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/elementwise_min_op.h"
#include "paddle/operators/elementwise_op.h"
namespace paddle {
namespace operators {
class ElementwiseMinOpMaker : public ElementwiseOpMaker {
public:
ElementwiseMinOpMaker(OpProto* proto, OpAttrChecker* op_checker)
: ElementwiseOpMaker(proto, op_checker) {
SetComment("Max", "Out = min(X, Y)");
AddComment(comment_);
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(elementwise_min, ops::ElementwiseOp, ops::ElementwiseMinOpMaker,
elementwise_min_grad, ops::ElementwiseOpGrad);
REGISTER_OP_CPU_KERNEL(
elementwise_min,
ops::ElementwiseMinKernel<paddle::platform::CPUDeviceContext, float>,
ops::ElementwiseMinKernel<paddle::platform::CPUDeviceContext, double>,
ops::ElementwiseMinKernel<paddle::platform::CPUDeviceContext, int>,
ops::ElementwiseMinKernel<paddle::platform::CPUDeviceContext, int64_t>);
REGISTER_OP_CPU_KERNEL(
elementwise_min_grad,
ops::ElementwiseMinGradKernel<paddle::platform::CPUDeviceContext, float>,
ops::ElementwiseMinGradKernel<paddle::platform::CPUDeviceContext, double>,
ops::ElementwiseMinGradKernel<paddle::platform::CPUDeviceContext, int>,
ops::ElementwiseMinGradKernel<paddle::platform::CPUDeviceContext, int64_t>);
paddle/operators/elementwise_min_op.cu 0 → 100644
浏览文件 @ e92fe9e6
/* 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/elementwise_min_op.h"
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(
elementwise_min,
ops::ElementwiseMinKernel<paddle::platform::CUDADeviceContext, float>,
ops::ElementwiseMinKernel<paddle::platform::CUDADeviceContext, double>,
ops::ElementwiseMinKernel<paddle::platform::CUDADeviceContext, int>,
ops::ElementwiseMinKernel<paddle::platform::CUDADeviceContext, int64_t>);
REGISTER_OP_CUDA_KERNEL(
elementwise_min_grad,
ops::ElementwiseMinGradKernel<paddle::platform::CUDADeviceContext, float>,
ops::ElementwiseMinGradKernel<paddle::platform::CUDADeviceContext, double>,
ops::ElementwiseMinGradKernel<paddle::platform::CUDADeviceContext, int>,
ops::ElementwiseMinGradKernel<paddle::platform::CUDADeviceContext,
int64_t>);
paddle/operators/elementwise_min_op.h 0 → 100644
浏览文件 @ e92fe9e6
/* 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/operators/elementwise_op_function.h"
namespace paddle {
namespace operators {
template <typename T>
struct MinFunctor {
inline HOSTDEVICE T operator()(T a, T b) const { return a < b ? a : b; }
};
template <typename DeviceContext, typename T>
class ElementwiseMinKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
ElementwiseComputeEx<MinFunctor<T>, DeviceContext, T>(ctx);
}
};
template <typename T>
struct ElementwiseMinGradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = (x_e < y_e).template cast<T>() * dz_e;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = (x_e >= y_e).template cast<T>() * dz_e;
}
}
};
template <typename T>
struct ElementwiseMinBroadCastGradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ, typename Pre, typename N>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
.broadcast(Eigen::DSizes<int, 2>(pre, 1))
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = (x_e < y_e_bcast).template cast<T>() * dz_e;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = ((x_e >= y_e_bcast).template cast<T>() * dz_e)
.reshape(Eigen::DSizes<int, 2>(pre, n))
.sum(Eigen::array<int, 1>{{0}});
}
}
};
template <typename T>
struct ElementwiseMinBroadCast2GradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ, typename Pre, typename N, typename Post>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n,
Post post) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
.broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = (x_e < y_e_bcast).template cast<T>() * dz_e;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = ((x_e >= y_e_bcast).template cast<T>() * dz_e)
.reshape(Eigen::DSizes<int, 3>(pre, n, post))
.sum(Eigen::array<int, 2>{{0, 2}});
}
}
};
template <typename DeviceContext, typename T>
class ElementwiseMinGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
ElementwiseGradCompute<DeviceContext, T, ElementwiseMinGradFunctor<T>,
ElementwiseMinBroadCastGradFunctor<T>,
ElementwiseMinBroadCast2GradFunctor<T>>(ctx);
}
};
} // namespace operators
} // namespace paddle
paddle/operators/elementwise_mul_op.h
浏览文件 @ e92fe9e6
...@@ -18,11 +18,16 @@ limitations under the License. */ ...@@ -18,11 +18,16 @@ limitations under the License. */
namespace paddle { namespace paddle {
namespace operators { namespace operators {
template <typename T>
struct MulFunctor {
inline HOSTDEVICE T operator()(T a, T b) const { return a * b; }
};
template <typename DeviceContext, typename T> template <typename DeviceContext, typename T>
class ElementwiseMulKernel : public framework::OpKernel<T> { class ElementwiseMulKernel : public framework::OpKernel<T> {
public: public:
void Compute(const framework::ExecutionContext& ctx) const override { void Compute(const framework::ExecutionContext& ctx) const override {
ElementwiseCompute<EigenMulFunctor, DeviceContext, T>(ctx); ElementwiseComputeEx<MulFunctor<T>, DeviceContext, T>(ctx);
} }
}; };
......
paddle/operators/elementwise_op.h
浏览文件 @ e92fe9e6
...@@ -26,9 +26,9 @@ class ElementwiseOp : public framework::OperatorWithKernel { ...@@ -26,9 +26,9 @@ class ElementwiseOp : public framework::OperatorWithKernel {
using Tensor = framework::Tensor; using Tensor = framework::Tensor;
void InferShape(framework::InferShapeContext* ctx) const override { void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"), PADDLE_ENFORCE(ctx->HasInput("X"),
"Input(X) of elementwise op should not be null"); "Input(X) of elementwise op should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Y"), PADDLE_ENFORCE(ctx->HasInput("Y"),
"Input(Y) of elementwise op should not be null"); "Input(Y) of elementwise op should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("Out"), PADDLE_ENFORCE(ctx->HasOutput("Out"),
"Output(Out) of elementwise op should not be null."); "Output(Out) of elementwise op should not be null.");
...@@ -45,12 +45,12 @@ class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker { ...@@ -45,12 +45,12 @@ class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker {
public: public:
ElementwiseOpMaker(OpProto* proto, OpAttrChecker* op_checker) ElementwiseOpMaker(OpProto* proto, OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) { : OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "(Tensor) The first input tensor of elementwise op"); AddInput("X", "(Tensor), The first input tensor of elementwise op.");
AddInput("Y", "(Tensor) The second input tensor of elementwise op"); AddInput("Y", "(Tensor), The second input tensor of elementwise op.");
AddOutput("Out", "The output of elementwise op"); AddOutput("Out", "The output of elementwise op.");
AddAttr<int>("axis", AddAttr<int>("axis",
"(int, default -1) The starting dimension index " "(int, default -1). The start dimension index "
"for broadcasting Y onto X") "for broadcasting Y onto X.")
.SetDefault(-1) .SetDefault(-1)
.EqualGreaterThan(-1); .EqualGreaterThan(-1);
comment_ = R"DOC( comment_ = R"DOC(
...@@ -58,19 +58,18 @@ Limited Elementwise {name} Operator. ...@@ -58,19 +58,18 @@ Limited Elementwise {name} Operator.
The equation is: The equation is:
.. math:: $${equation}$$
{equation}
X is a tensor of any dimension and the dimensions of tensor Y must be smaller than $X$ is a tensor of any dimension and the dimensions of tensor $Y$ must be
or equal to the dimensions of X. smaller than or equal to the dimensions of $X$.
There are two cases for this operator: There are two cases for this operator:
1. The shape of Y is same with X; 1. The shape of $Y$ is same with $X$;
2. The shape of Y is a subset of X. 2. The shape of $Y$ is a subset of $X$.
For case 2: For case 2:
Y will be broadcasted to match the shape of X and axis should be $Y$ will be broadcasted to match the shape of $X$ and axis should be
the starting dimension index for broadcasting Y onto X. set to index of the start dimension to broadcast $Y$ onto $X$.
For example For example
.. code-block:: python .. code-block:: python
...@@ -81,7 +80,8 @@ For example ...@@ -81,7 +80,8 @@ For example
shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1 shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0 shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0
Either of the inputs X and Y or none can carry the LoD (Level of Details) information. However, the output only shares the LoD information with input X. Either of the inputs $X$ and $Y$ or none can carry the LoD (Level of Details)
information. However, the output only shares the LoD information with input $X$.
)DOC"; )DOC";
AddComment(comment_); AddComment(comment_);
......
paddle/operators/elementwise_op_function.h
浏览文件 @ e92fe9e6
...@@ -340,6 +340,13 @@ void ElementwiseGradCompute(const framework::ExecutionContext& ctx) { ...@@ -340,6 +340,13 @@ void ElementwiseGradCompute(const framework::ExecutionContext& ctx) {
return; return;
} }
if (y_dims.size() == 1 && y_dims[0] == 1) {
// y is a scalar
auto extended_dims = framework::vectorize(x_dims);
extended_dims.push_back(1);
x_dims = framework::make_ddim(extended_dims);
}
int axis = ctx.Attr<int>("axis"); int axis = ctx.Attr<int>("axis");
axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis); axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
...@@ -356,5 +363,50 @@ void ElementwiseGradCompute(const framework::ExecutionContext& ctx) { ...@@ -356,5 +363,50 @@ void ElementwiseGradCompute(const framework::ExecutionContext& ctx) {
return; return;
} }
} }
template <typename Functor, typename DeviceContext, typename T>
void ElementwiseComputeEx(const framework::ExecutionContext& ctx) {
using Tensor = framework::Tensor;
auto* x = ctx.Input<Tensor>("X");
auto* y = ctx.Input<Tensor>("Y");
auto* z = ctx.Output<Tensor>("Out");
z->mutable_data<T>(ctx.GetPlace());
TransformFunctor<Functor, T, DeviceContext> functor(
x, y, z, ctx.template device_context<DeviceContext>(), Functor());
auto x_dims = x->dims();
auto y_dims = y->dims();
PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
"Rank of first input must >= rank of second input.");
if (x_dims == y_dims) {
functor.Run();
return;
}
if (y_dims.size() == 1 && y_dims[0] == 1) {
// y is a scalar
auto extended_dims = framework::vectorize(x_dims);
extended_dims.push_back(1);
x_dims = framework::make_ddim(extended_dims);
}
int axis = ctx.Attr<int>("axis");
axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(),
"Axis should be in range [0, x_dims)");
int pre, n, post;
get_mid_dims(x_dims, y_dims, axis, pre, n, post);
if (post == 1) {
functor.RunRowWise(n, pre);
return;
} else {
functor.RunMidWise(n, pre, post);
return;
}
}
} // namespace operators } // namespace operators
} // namespace paddle } // namespace paddle
paddle/operators/elementwise_sub_op.h
浏览文件 @ e92fe9e6
...@@ -18,11 +18,16 @@ limitations under the License. */ ...@@ -18,11 +18,16 @@ limitations under the License. */
namespace paddle { namespace paddle {
namespace operators { namespace operators {
template <typename T>
struct SubFunctor {
inline HOSTDEVICE T operator()(T a, T b) const { return a - b; }
};
template <typename DeviceContext, typename T> template <typename DeviceContext, typename T>
class ElementwiseSubKernel : public framework::OpKernel<T> { class ElementwiseSubKernel : public framework::OpKernel<T> {
public: public:
void Compute(const framework::ExecutionContext& ctx) const override { void Compute(const framework::ExecutionContext& ctx) const override {
ElementwiseCompute<EigenSubFunctor, DeviceContext, T>(ctx); ElementwiseComputeEx<SubFunctor<T>, DeviceContext, T>(ctx);
} }
}; };
......
paddle/operators/expand_op.cc
浏览文件 @ e92fe9e6
...@@ -58,21 +58,21 @@ class ExpandOpMaker : public framework::OpProtoAndCheckerMaker { ...@@ -58,21 +58,21 @@ class ExpandOpMaker : public framework::OpProtoAndCheckerMaker {
ExpandOpMaker(OpProto* proto, OpAttrChecker* op_checker) ExpandOpMaker(OpProto* proto, OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) { : OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", AddInput("X",
"(Tensor, default Tensor<float>) A tensor with rank in [1, 6]." "(Tensor, default Tensor<float>). A tensor with rank in [1, 6]."
"X is the input tensor to be expanded."); "X is the input to be expanded.");
AddOutput("Out", AddOutput("Out",
"(Tensor, default Tensor<float>) A tensor with rank in [1, 6]." "(Tensor, default Tensor<float>). A tensor with rank in [1, 6]."
"The rank of Output(Out) is same as Input(X) except that each " "The rank of Output(Out) have the same with Input(X). "
"dimension size of Output(Out) is equal to corresponding " "After expanding, size of each dimension of Output(Out) is equal "
"dimension size of Input(X) multiplying corresponding value of " "to size of the corresponding dimension of Input(X) multiplying "
"Attr(expand_times)."); "the corresponding value given by Attr(expand_times).");
AddAttr<std::vector<int>>("expand_times", AddAttr<std::vector<int>>("expand_times",
"Expand times number for each dimension."); "Expand times number for each dimension.");
AddComment(R"DOC( AddComment(R"DOC(
Expand operator tiles the input by given times number. You should set times Expand operator tiles the input by given times number. You should set times
number for each dimension by providing attribute 'expand_times'. The rank of X number for each dimension by providing attribute 'expand_times'. The rank of X
should be in [1, 6]. Please notice that size of 'expand_times' must be same with should be in [1, 6]. Please note that size of 'expand_times' must be the same
X's rank. Following is a using case: with X's rank. Following is a using case:
Input(X) is a 3-D tensor with shape [2, 3, 1]: Input(X) is a 3-D tensor with shape [2, 3, 1]:
......
paddle/operators/linear_chain_crf_op.cc
浏览文件 @ e92fe9e6
...@@ -187,7 +187,7 @@ class LinearChainCRFOp : public framework::OperatorWithKernel { ...@@ -187,7 +187,7 @@ class LinearChainCRFOp : public framework::OperatorWithKernel {
const framework::ExecutionContext& ctx) const override { const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType( return framework::OpKernelType(
framework::ToDataType(ctx.Input<LoDTensor>("Emission")->type()), framework::ToDataType(ctx.Input<LoDTensor>("Emission")->type()),
ctx.device_context()); platform::CPUPlace());
} }
}; };
...@@ -248,7 +248,7 @@ class LinearChainCRFGradOp : public framework::OperatorWithKernel { ...@@ -248,7 +248,7 @@ class LinearChainCRFGradOp : public framework::OperatorWithKernel {
framework::ToDataType( framework::ToDataType(
ctx.Input<LoDTensor>(framework::GradVarName("LogLikelihood")) ctx.Input<LoDTensor>(framework::GradVarName("LogLikelihood"))
->type()), ->type()),
ctx.device_context()); platform::CPUPlace());
} }
}; };
......
paddle/operators/linear_chain_crf_op.h
浏览文件 @ e92fe9e6
...@@ -65,57 +65,14 @@ class LinearChainCRFOpKernel : public framework::OpKernel<T> { ...@@ -65,57 +65,14 @@ class LinearChainCRFOpKernel : public framework::OpKernel<T> {
const size_t level = 0; const size_t level = 0;
const size_t seq_num = in_lod[level].size() - 1; const size_t seq_num = in_lod[level].size() - 1;
// These local variables hold the inputs and outputs, garanteeing them on const LoDTensor* emission_weights = ctx.Input<LoDTensor>("Emission");
// CPU memory, to provide a consistent reference. const Tensor* transition_weights = ctx.Input<Tensor>("Transition");
// TODO(caoying) Fix this by moving all these local variables into the const LoDTensor* label = ctx.Input<LoDTensor>("Label");
// class's data members once we can profile the whole training process.
LoDTensor* emission_weights = nullptr; Tensor* emission_exps = ctx.Output<Tensor>("EmissionExps");
LoDTensor emission_weight_tensor; Tensor* transition_exps = ctx.Output<Tensor>("TransitionExps");
Tensor* transition_weights = nullptr; Tensor* alpha = ctx.Output<Tensor>("Alpha");
Tensor transition_weight_tensor; Tensor* ll = ctx.Output<Tensor>("LogLikelihood");
LoDTensor* label = nullptr;
LoDTensor label_tensor;
Tensor* emission_exps = nullptr;
Tensor emission_exps_tensor;
Tensor* transition_exps = nullptr;
Tensor transition_exps_tensor;
Tensor* alpha = nullptr;
Tensor alpha_tensor;
Tensor* ll = nullptr;
Tensor ll_tensor;
if (platform::is_gpu_place(ctx.GetPlace())) {
emission_weights = &emission_weight_tensor;
transition_weights = &transition_weight_tensor;
label = &label_tensor;
CopyInputsToCpuMemory(
ctx.device_context(), *ctx.Input<LoDTensor>("Emission"),
*ctx.Input<Tensor>("Transition"), *ctx.Input<LoDTensor>("Label"),
emission_weights, transition_weights, label);
emission_exps = &emission_exps_tensor;
emission_exps->Resize(emission_weights->dims());
transition_exps = &transition_exps_tensor;
transition_exps->Resize(transition_weights->dims());
alpha = &alpha_tensor;
alpha->Resize(ctx.Output<Tensor>("Alpha")->dims());
ll = &ll_tensor;
} else {
emission_weights =
const_cast<LoDTensor*>(ctx.Input<LoDTensor>("Emission"));
transition_weights = const_cast<Tensor*>(ctx.Input<Tensor>("Transition"));
label = const_cast<LoDTensor*>(ctx.Input<LoDTensor>("Label"));
emission_exps = ctx.Output<Tensor>("EmissionExps");
transition_exps = ctx.Output<Tensor>("TransitionExps");
alpha = ctx.Output<Tensor>("Alpha");
ll = ctx.Output<Tensor>("LogLikelihood");
}
// Because the computation codes only runs on CPU, here the memory for all // Because the computation codes only runs on CPU, here the memory for all
// the outputs is FIXED to be allocated on the CPU memory. // the outputs is FIXED to be allocated on the CPU memory.
...@@ -173,61 +130,9 @@ class LinearChainCRFOpKernel : public framework::OpKernel<T> { ...@@ -173,61 +130,9 @@ class LinearChainCRFOpKernel : public framework::OpKernel<T> {
one_seq, one_seq_row_max, one_seq_exps, *transition_weights, one_seq, one_seq_row_max, one_seq_exps, *transition_weights,
*transition_exps, one_seq_label, &one_seq_alpha); *transition_exps, one_seq_label, &one_seq_alpha);
} }
if (platform::is_gpu_place(ctx.GetPlace())) {
CopyOutputsToGpuMemory(
ctx.device_context(), *emission_exps, *transition_exps, *alpha, *ll,
ctx.Output<Tensor>("EmissionExps"),
ctx.Output<Tensor>("TransitionExps"), ctx.Output<Tensor>("Alpha"),
ctx.Output<Tensor>("LogLikelihood"));
}
}; };
private: private:
void CopyInputsToCpuMemory(const platform::DeviceContext& ctx,
const LoDTensor& emission_weights_src,
const Tensor& transition_weights_src,
const LoDTensor& label_src,
LoDTensor* emission_weights_dst,
Tensor* transition_weights_dst,
LoDTensor* label_dst) const {
// Copy the inputs from GPU memory to CPU memory if this operators runs on
// GPU device.
auto copyLoDTensor = [](const platform::DeviceContext& ctx,
const LoDTensor& src, LoDTensor* dst) {
dst->mutable_data<T>(src.dims(), platform::CPUPlace());
framework::Copy(src, platform::CPUPlace(), ctx, dst);
};
copyLoDTensor(ctx, emission_weights_src, emission_weights_dst);
copyLoDTensor(ctx, label_src, label_dst);
transition_weights_dst->mutable_data<T>(transition_weights_src.dims(),
platform::CPUPlace());
framework::Copy(transition_weights_src, platform::CPUPlace(), ctx,
transition_weights_dst);
}
void CopyOutputsToGpuMemory(const platform::DeviceContext& ctx,
const Tensor& emission_exps_src,
const Tensor& transition_exps_src,
const Tensor& alpha_src, const Tensor& ll_src,
Tensor* emission_exps_dst,
Tensor* transition_exps_dst, Tensor* alpha_dst,
Tensor* ll_dst) const {
// Copy the forward results from CPU memory to GPU memory if this
// operators runs on GPU device.
auto copyTensor = [](const platform::DeviceContext& ctx, const Tensor& src,
Tensor* dst) {
dst->mutable_data<T>(platform::CUDAPlace());
framework::Copy(src, platform::CUDAPlace(), ctx, dst);
};
copyTensor(ctx, emission_exps_src, emission_exps_dst);
copyTensor(ctx, transition_exps_src, transition_exps_dst);
copyTensor(ctx, alpha_src, alpha_dst);
copyTensor(ctx, ll_src, ll_dst);
}
T ForwardOneSequence(const Tensor& emission, const Tensor& emission_row_max, T ForwardOneSequence(const Tensor& emission, const Tensor& emission_row_max,
const Tensor& emission_exps, const Tensor& trans_weights, const Tensor& emission_exps, const Tensor& trans_weights,
const Tensor& trans_weight_exps, const Tensor& label, const Tensor& trans_weight_exps, const Tensor& label,
...@@ -296,63 +201,17 @@ class LinearChainCRFGradOpKernel : public framework::OpKernel<T> { ...@@ -296,63 +201,17 @@ class LinearChainCRFGradOpKernel : public framework::OpKernel<T> {
auto lod = ctx.Input<LoDTensor>("Label")->lod(); auto lod = ctx.Input<LoDTensor>("Label")->lod();
PADDLE_ENFORCE(lod.size(), "Input(Label) must be a sequence."); PADDLE_ENFORCE(lod.size(), "Input(Label) must be a sequence.");
// These local variables hold the inputs and outputs, garanteeing them on const Tensor* label = ctx.Input<LoDTensor>("Label");
// CPU memory, to provide a consistent reference. const Tensor* emission_exps = ctx.Input<Tensor>("EmissionExps");
// TODO(caoying) Fix this by moving all these local variables into the const Tensor* transition_exps = ctx.Input<Tensor>("TransitionExps");
// class's data members once we can profile the training process, or const Tensor* alpha = ctx.Input<Tensor>("Alpha");
// implementing a real GPU kernel for CRF. const T* ll_grad =
Tensor* label = nullptr; ctx.Input<Tensor>(framework::GradVarName("LogLikelihood"))->data<T>();
Tensor label_tensor;
Tensor* emission_exps = nullptr;
Tensor emission_exps_tensor;
Tensor* transition_exps = nullptr;
Tensor transition_exps_tensor;
Tensor* alpha = nullptr;
Tensor alpha_tensor;
Tensor ll_grad_tensor;
T* ll_grad = nullptr;
Tensor* emission_grad = nullptr;
Tensor emission_grad_tensor;
Tensor* transition_grad = nullptr;
Tensor transition_grad_tensor;
if (platform::is_gpu_place(ctx.GetPlace())) {
label = &label_tensor;
emission_exps = &emission_exps_tensor;
transition_exps = &transition_exps_tensor;
alpha = &alpha_tensor;
CopyInputsToCpuMemory(
ctx.device_context(), *ctx.Input<LoDTensor>("Label"),
*ctx.Input<Tensor>("EmissionExps"),
*ctx.Input<Tensor>("TransitionExps"), *ctx.Input<Tensor>("Alpha"),
*ctx.Input<Tensor>(framework::GradVarName("LogLikelihood")), label,
emission_exps, transition_exps, alpha, &ll_grad_tensor);
ll_grad = ll_grad_tensor.data<T>();
if (ctx.Output<Tensor>(framework::GradVarName("Emission"))) {
emission_grad = &emission_grad_tensor;
emission_grad->Resize(emission_exps->dims());
}
if (ctx.Output<Tensor>(framework::GradVarName("Transition"))) { Tensor* emission_grad =
transition_grad = &transition_grad_tensor; ctx.Output<Tensor>(framework::GradVarName("Emission"));
transition_grad->Resize(transition_exps->dims()); Tensor* transition_grad =
} ctx.Output<Tensor>(framework::GradVarName("Transition"));
} else {
label = const_cast<LoDTensor*>(ctx.Input<LoDTensor>("Label"));
emission_exps = const_cast<Tensor*>(ctx.Input<Tensor>("EmissionExps"));
transition_exps =
const_cast<Tensor*>(ctx.Input<Tensor>("TransitionExps"));
alpha = const_cast<Tensor*>(ctx.Input<Tensor>("Alpha"));
ll_grad = const_cast<Tensor*>(
ctx.Input<Tensor>(framework::GradVarName("LogLikelihood")))
->data<T>();
emission_grad = ctx.Output<Tensor>(framework::GradVarName("Emission"));
transition_grad =
ctx.Output<Tensor>(framework::GradVarName("Transition"));
}
// TODO(caoying) Fix this constraint. When the Input(Emission) is from the // TODO(caoying) Fix this constraint. When the Input(Emission) is from the
// data reader operator, it can have no gradients. // data reader operator, it can have no gradients.
...@@ -389,58 +248,9 @@ class LinearChainCRFGradOpKernel : public framework::OpKernel<T> { ...@@ -389,58 +248,9 @@ class LinearChainCRFGradOpKernel : public framework::OpKernel<T> {
one_seq_emission_exps, *transition_exps, one_seq_alpha, one_seq_label, one_seq_emission_exps, *transition_exps, one_seq_alpha, one_seq_label,
&one_seq_beta, transition_grad, &one_seq_emission_grad); &one_seq_beta, transition_grad, &one_seq_emission_grad);
} }
if (platform::is_gpu_place(ctx.GetPlace())) {
CopyOutputsToGpuMemory(
ctx.device_context(), emission_grad, transition_grad,
ctx.Output<Tensor>(framework::GradVarName("Emission")),
ctx.Output<Tensor>(framework::GradVarName("Transition")));
}
}; };
private: private:
void CopyInputsToCpuMemory(const platform::DeviceContext& ctx,
const LoDTensor& label_src,
const Tensor& emission_exps_src,
const Tensor& transition_exps_src,
const Tensor& alpha_src, const Tensor& ll_grad_src,
Tensor* label_dst, Tensor* emission_exps_dst,
Tensor* transition_exps_dst, Tensor* alpha_dst,
Tensor* ll_grad_dst) const {
// Copy the inputs from GPU memory to CPU memory when this operators runs on
// GPU device.
label_dst->mutable_data<T>(label_src.dims(), platform::CPUPlace());
framework::Copy(label_src, platform::CPUPlace(), ctx, label_dst);
auto copyTensor = [](const platform::DeviceContext& ctx, const Tensor& src,
Tensor* dst) {
dst->mutable_data<T>(src.dims(), platform::CPUPlace());
framework::Copy(src, platform::CPUPlace(), ctx, dst);
};
copyTensor(ctx, emission_exps_src, emission_exps_dst);
copyTensor(ctx, transition_exps_src, transition_exps_dst);
copyTensor(ctx, alpha_src, alpha_dst);
copyTensor(ctx, ll_grad_src, ll_grad_dst);
}
void CopyOutputsToGpuMemory(const platform::DeviceContext& ctx,
const Tensor* emission_grad_src,
const Tensor* transition_grad_src,
Tensor* emission_grad_dst,
Tensor* transition_grad_dst) const {
// Copy the backward results from CPU memory to GPU
// memory if this operators runs on GPU device.
auto copyTensor = [](const platform::DeviceContext& ctx, const Tensor* src,
Tensor* dst) {
if (src && dst) {
dst->mutable_data<T>(platform::CUDAPlace());
framework::Copy(*src, platform::CUDAPlace(), ctx, dst);
}
};
copyTensor(ctx, emission_grad_src, emission_grad_dst);
copyTensor(ctx, transition_grad_src, transition_grad_dst);
}
void BackwardOneSequence(const platform::CPUDeviceContext& ctx, void BackwardOneSequence(const platform::CPUDeviceContext& ctx,
const T ll_grad, const Tensor& emission_exps, const T ll_grad, const Tensor& emission_exps,
const Tensor& transition_exps, const Tensor& alpha, const Tensor& transition_exps, const Tensor& alpha,
......
paddle/operators/lstm_op.cc
浏览文件 @ e92fe9e6
...@@ -117,7 +117,7 @@ class LSTMOpMaker : public framework::OpProtoAndCheckerMaker { ...@@ -117,7 +117,7 @@ class LSTMOpMaker : public framework::OpProtoAndCheckerMaker {
AddInput("C0", AddInput("C0",
"(Tensor, optional) the initial cell state is an optional " "(Tensor, optional) the initial cell state is an optional "
"input. This is a tensor with shape (N x D), where N is the " "input. This is a tensor with shape (N x D), where N is the "
"batch size. `H0` and `C0` can be NULL but only at the same time") "batch size. `H0` and `C0` can be NULL but only at the same time.")
.AsDispensable(); .AsDispensable();
AddInput("Weight", AddInput("Weight",
"(Tensor) the learnable hidden-hidden weights." "(Tensor) the learnable hidden-hidden weights."
......
paddle/operators/math/sequence_padding.cc
浏览文件 @ e92fe9e6
...@@ -32,7 +32,8 @@ class PaddingLoDTensorFunctor<platform::CPUDeviceContext, T> { ...@@ -32,7 +32,8 @@ class PaddingLoDTensorFunctor<platform::CPUDeviceContext, T> {
framework::LoD abs_offset_lod = framework::ToAbsOffset(lod); framework::LoD abs_offset_lod = framework::ToAbsOffset(lod);
auto seq_dims = seq.dims(); auto seq_dims = seq.dims();
PADDLE_ENFORCE_EQ(seq_dims[0], abs_offset_lod[level].back(), PADDLE_ENFORCE_EQ(seq_dims[0],
static_cast<int64_t>(abs_offset_lod[level].back()),
"The first dimension of LoDTensor seq should be " "The first dimension of LoDTensor seq should be "
"equal to the sum of all sequences's length."); "equal to the sum of all sequences's length.");
...@@ -41,32 +42,32 @@ class PaddingLoDTensorFunctor<platform::CPUDeviceContext, T> { ...@@ -41,32 +42,32 @@ class PaddingLoDTensorFunctor<platform::CPUDeviceContext, T> {
"The input padding should be a 3-D Tensor of shape " "The input padding should be a 3-D Tensor of shape "
"[max_sequence_length, num_sequences, sequence_width]."); "[max_sequence_length, num_sequences, sequence_width].");
const size_t max_sequence_length = MaximumSequenceLength(lod, level); const int64_t max_sequence_length = MaximumSequenceLength(lod, level);
PADDLE_ENFORCE_EQ(padding_dims[0], max_sequence_length, PADDLE_ENFORCE_EQ(padding_dims[0], max_sequence_length,
"The first dimension of Tensor padding should be the " "The first dimension of Tensor padding should be the "
"maximum length of all sequences in LoDTensor seq."); "maximum length of all sequences in LoDTensor seq.");
const size_t num_sequences = abs_offset_lod[level].size() - 1; const int64_t num_sequences = abs_offset_lod[level].size() - 1;
PADDLE_ENFORCE_EQ(padding_dims[1], num_sequences, PADDLE_ENFORCE_EQ(padding_dims[1], num_sequences,
"The second dimension of Tensor padding should be the " "The second dimension of Tensor padding should be the "
"number of sequences in LoDTensor seq."); "number of sequences in LoDTensor seq.");
const size_t sequence_width = seq.numel() / seq_dims[0]; const int64_t sequence_width = seq.numel() / seq_dims[0];
PADDLE_ENFORCE_EQ(padding_dims[2], sequence_width, PADDLE_ENFORCE_EQ(padding_dims[2], sequence_width,
"The third dimension of Tensor padding should be the " "The third dimension of Tensor padding should be the "
"width of sequence in LoDTensor seq."); "width of sequence in LoDTensor seq.");
const T* seq_data = seq.data<T>(); const T* seq_data = seq.data<T>();
T* padding_data = padding.data<T>(); T* padding_data = padding.data<T>();
for (size_t i = 0; i < max_sequence_length; ++i) { for (int64_t i = 0; i < max_sequence_length; ++i) {
for (size_t j = 0; j < num_sequences; ++j) { for (int64_t j = 0; j < num_sequences; ++j) {
size_t start_pos = abs_offset_lod[level][j]; int64_t start_pos = abs_offset_lod[level][j];
size_t sequence_length = abs_offset_lod[level][j + 1] - start_pos; int64_t sequence_length = abs_offset_lod[level][j + 1] - start_pos;
if (i < sequence_length) { if (i < sequence_length) {
// i > 0 => sequence_length > 0 // i > 0 => sequence_length > 0
T scale = T scale =
norm_by_times ? (1.0f / static_cast<T>(sequence_length)) : 1.0f; norm_by_times ? (1.0f / static_cast<T>(sequence_length)) : 1.0f;
for (size_t k = 0; k < sequence_width; ++k) { for (int64_t k = 0; k < sequence_width; ++k) {
padding_data[(i * num_sequences + j) * sequence_width + k] = padding_data[(i * num_sequences + j) * sequence_width + k] =
seq_data[(start_pos + i) * sequence_width + k] * scale; seq_data[(start_pos + i) * sequence_width + k] * scale;
} }
...@@ -93,7 +94,8 @@ class UnpaddingLoDTensorFunctor<platform::CPUDeviceContext, T> { ...@@ -93,7 +94,8 @@ class UnpaddingLoDTensorFunctor<platform::CPUDeviceContext, T> {
framework::LoD abs_offset_lod = framework::ToAbsOffset(lod); framework::LoD abs_offset_lod = framework::ToAbsOffset(lod);
auto seq_dims = seq.dims(); auto seq_dims = seq.dims();
PADDLE_ENFORCE_EQ(seq_dims[0], abs_offset_lod[level].back(), PADDLE_ENFORCE_EQ(seq_dims[0],
static_cast<int64_t>(abs_offset_lod[level].back()),
"The first dimension of LoDTensor seq should be " "The first dimension of LoDTensor seq should be "
"equal to the sum of all sequences's length."); "equal to the sum of all sequences's length.");
...@@ -102,31 +104,31 @@ class UnpaddingLoDTensorFunctor<platform::CPUDeviceContext, T> { ...@@ -102,31 +104,31 @@ class UnpaddingLoDTensorFunctor<platform::CPUDeviceContext, T> {
"The input padding should be a 3-D Tensor of shape " "The input padding should be a 3-D Tensor of shape "
"[max_sequnece_length, num_sequences, sequence_width]."); "[max_sequnece_length, num_sequences, sequence_width].");
const size_t max_sequence_length = MaximumSequenceLength(lod, level); const int64_t max_sequence_length = MaximumSequenceLength(lod, level);
PADDLE_ENFORCE_EQ(padding_dims[0], max_sequence_length, PADDLE_ENFORCE_EQ(padding_dims[0], max_sequence_length,
"The first dimension of Tensor padding should be " "The first dimension of Tensor padding should be "
"the maximum length of all sequences in LoDTensor seq."); "the maximum length of all sequences in LoDTensor seq.");
const size_t num_sequences = abs_offset_lod[level].size() - 1; const int64_t num_sequences = abs_offset_lod[level].size() - 1;
PADDLE_ENFORCE_EQ(padding_dims[1], num_sequences, PADDLE_ENFORCE_EQ(padding_dims[1], num_sequences,
"The second dimension of Tensor padding should be " "The second dimension of Tensor padding should be "
"the number of sequences in LoDTensor seq."); "the number of sequences in LoDTensor seq.");
const size_t sequence_width = seq.numel() / seq_dims[0]; const int64_t sequence_width = seq.numel() / seq_dims[0];
PADDLE_ENFORCE_EQ(padding_dims[2], sequence_width, PADDLE_ENFORCE_EQ(padding_dims[2], sequence_width,
"The third dimension of Tensor padding should be the " "The third dimension of Tensor padding should be the "
"width of sequence in LoDTensor seq."); "width of sequence in LoDTensor seq.");
const T* padding_data = padding.data<T>(); const T* padding_data = padding.data<T>();
T* seq_data = seq.data<T>(); T* seq_data = seq.data<T>();
for (size_t i = 0; i < num_sequences; ++i) { for (int64_t i = 0; i < num_sequences; ++i) {
size_t start_pos = abs_offset_lod[level][i]; int64_t start_pos = abs_offset_lod[level][i];
size_t sequence_length = abs_offset_lod[level][i + 1] - start_pos; int64_t sequence_length = abs_offset_lod[level][i + 1] - start_pos;
for (size_t j = 0; j < sequence_length; ++j) { for (int64_t j = 0; j < sequence_length; ++j) {
// sequence_width > j > 0 // sequence_width > j > 0
T scale = T scale =
norm_by_times ? (1.0f / static_cast<T>(sequence_length)) : 1.0f; norm_by_times ? (1.0f / static_cast<T>(sequence_length)) : 1.0f;
for (size_t k = 0; k < sequence_width; ++k) { for (int64_t k = 0; k < sequence_width; ++k) {
seq_data[(start_pos + j) * sequence_width + k] = seq_data[(start_pos + j) * sequence_width + k] =
padding_data[(j * num_sequences + i) * sequence_width + k] * padding_data[(j * num_sequences + i) * sequence_width + k] *
scale; scale;
......
paddle/operators/math/sequence_padding.cu
浏览文件 @ e92fe9e6
...@@ -71,7 +71,8 @@ class PaddingLoDTensorFunctor<platform::CUDADeviceContext, T> { ...@@ -71,7 +71,8 @@ class PaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
framework::LoD abs_offset_lod = framework::ToAbsOffset(lod); framework::LoD abs_offset_lod = framework::ToAbsOffset(lod);
auto seq_dims = seq.dims(); auto seq_dims = seq.dims();
PADDLE_ENFORCE_EQ(seq_dims[0], abs_offset_lod[level].back(), PADDLE_ENFORCE_EQ(seq_dims[0],
static_cast<int64_t>(abs_offset_lod[level].back()),
"The first dimension of LoDTensor seq should be " "The first dimension of LoDTensor seq should be "
"equal to the sum of all sequences's length."); "equal to the sum of all sequences's length.");
...@@ -80,17 +81,17 @@ class PaddingLoDTensorFunctor<platform::CUDADeviceContext, T> { ...@@ -80,17 +81,17 @@ class PaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
"The input padding should be a 3-D Tensor of shape " "The input padding should be a 3-D Tensor of shape "
"[max_sequence_length, num_sequences, sequence_width]."); "[max_sequence_length, num_sequences, sequence_width].");
size_t max_sequence_length = MaximumSequenceLength(lod, level); int64_t max_sequence_length = MaximumSequenceLength(lod, level);
PADDLE_ENFORCE_EQ(padding_dims[0], max_sequence_length, PADDLE_ENFORCE_EQ(padding_dims[0], max_sequence_length,
"The first dimension of Tensor padding should be the " "The first dimension of Tensor padding should be the "
"maximum length of all sequences in LoDTensor seq."); "maximum length of all sequences in LoDTensor seq.");
const size_t num_sequences = abs_offset_lod[level].size() - 1; const int64_t num_sequences = abs_offset_lod[level].size() - 1;
PADDLE_ENFORCE_EQ(padding_dims[1], num_sequences, PADDLE_ENFORCE_EQ(padding_dims[1], num_sequences,
"The second dimension of Tensor padding should be the " "The second dimension of Tensor padding should be the "
"number of sequences in LoDTensor seq."); "number of sequences in LoDTensor seq.");
const size_t sequence_width = seq.numel() / seq_dims[0]; const int64_t sequence_width = seq.numel() / seq_dims[0];
PADDLE_ENFORCE_EQ(padding_dims[2], sequence_width, PADDLE_ENFORCE_EQ(padding_dims[2], sequence_width,
"The third dimension of Tensor padding should be the " "The third dimension of Tensor padding should be the "
"width of sequence in LoDTensor seq."); "width of sequence in LoDTensor seq.");
...@@ -101,7 +102,7 @@ class PaddingLoDTensorFunctor<platform::CUDADeviceContext, T> { ...@@ -101,7 +102,7 @@ class PaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
return; return;
} }
const size_t kBlockSize = 512; const int64_t kBlockSize = 512;
/* At least use 32 threads to copy sequence_width elements, /* At least use 32 threads to copy sequence_width elements,
* and at least 8 elements for each thread. * and at least 8 elements for each thread.
...@@ -143,7 +144,8 @@ class UnpaddingLoDTensorFunctor<platform::CUDADeviceContext, T> { ...@@ -143,7 +144,8 @@ class UnpaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
framework::LoD abs_offset_lod = framework::ToAbsOffset(lod); framework::LoD abs_offset_lod = framework::ToAbsOffset(lod);
auto seq_dims = seq.dims(); auto seq_dims = seq.dims();
PADDLE_ENFORCE_EQ(seq_dims[0], abs_offset_lod[level].back(), PADDLE_ENFORCE_EQ(seq_dims[0],
static_cast<int64_t>(abs_offset_lod[level].back()),
"The first dimension of LoDTensor seq should be " "The first dimension of LoDTensor seq should be "
"equal to the sum of all sequences's length."); "equal to the sum of all sequences's length.");
...@@ -152,17 +154,17 @@ class UnpaddingLoDTensorFunctor<platform::CUDADeviceContext, T> { ...@@ -152,17 +154,17 @@ class UnpaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
"The input padding should be a 3-D Tensor of shape " "The input padding should be a 3-D Tensor of shape "
"[max_sequnece_length, num_sequences, sequence_width]."); "[max_sequnece_length, num_sequences, sequence_width].");
size_t max_sequence_length = MaximumSequenceLength(lod, level); int64_t max_sequence_length = MaximumSequenceLength(lod, level);
PADDLE_ENFORCE_EQ(padding_dims[0], max_sequence_length, PADDLE_ENFORCE_EQ(padding_dims[0], max_sequence_length,
"The first dimension of Tensor padding should be " "The first dimension of Tensor padding should be "
"the maximum length of all sequences in LoDTensor seq."); "the maximum length of all sequences in LoDTensor seq.");
const size_t num_sequences = abs_offset_lod[level].size() - 1; const int64_t num_sequences = abs_offset_lod[level].size() - 1;
PADDLE_ENFORCE_EQ(padding_dims[1], num_sequences, PADDLE_ENFORCE_EQ(padding_dims[1], num_sequences,
"The second dimension of Tensor padding should be " "The second dimension of Tensor padding should be "
"the number of sequences in LoDTensor seq."); "the number of sequences in LoDTensor seq.");
const size_t sequence_width = seq.numel() / seq_dims[0]; const int64_t sequence_width = seq.numel() / seq_dims[0];
PADDLE_ENFORCE_EQ(padding_dims[2], sequence_width, PADDLE_ENFORCE_EQ(padding_dims[2], sequence_width,
"The third dimension of Tensor padding should be the " "The third dimension of Tensor padding should be the "
"width of sequence in LoDTensor seq."); "width of sequence in LoDTensor seq.");
...@@ -173,7 +175,7 @@ class UnpaddingLoDTensorFunctor<platform::CUDADeviceContext, T> { ...@@ -173,7 +175,7 @@ class UnpaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
return; return;
} }
const size_t kBlockSize = 512; const int64_t kBlockSize = 512;
/* At least use 32 threads to copy sequence_width elements, /* At least use 32 threads to copy sequence_width elements,
* and at least 8 elements for each thread. * and at least 8 elements for each thread.
......
paddle/operators/math/sequence_padding_test.cc
浏览文件 @ e92fe9e6
...@@ -31,7 +31,7 @@ void TestSequencePadding(const paddle::framework::LoD& lod, ...@@ -31,7 +31,7 @@ void TestSequencePadding(const paddle::framework::LoD& lod,
cpu_seq.set_lod(lod); cpu_seq.set_lod(lod);
cpu_seq.mutable_data<T>(seq_dims, paddle::platform::CPUPlace()); cpu_seq.mutable_data<T>(seq_dims, paddle::platform::CPUPlace());
for (size_t i = 0; i < cpu_seq.numel(); ++i) { for (int64_t i = 0; i < cpu_seq.numel(); ++i) {
cpu_seq.data<T>()[i] = static_cast<T>(i); cpu_seq.data<T>()[i] = static_cast<T>(i);
} }
...@@ -69,7 +69,7 @@ void TestSequencePadding(const paddle::framework::LoD& lod, ...@@ -69,7 +69,7 @@ void TestSequencePadding(const paddle::framework::LoD& lod,
EXPECT_EQ(cpu_seq.numel(), cpu_seq_back.numel()); EXPECT_EQ(cpu_seq.numel(), cpu_seq_back.numel());
EXPECT_EQ(cpu_seq.dims(), cpu_seq_back.dims()); EXPECT_EQ(cpu_seq.dims(), cpu_seq_back.dims());
for (size_t i = 0; i < cpu_seq.numel(); ++i) { for (int64_t i = 0; i < cpu_seq.numel(); ++i) {
EXPECT_EQ(cpu_seq.data<T>()[i], cpu_seq_back.data<T>()[i]); EXPECT_EQ(cpu_seq.data<T>()[i], cpu_seq_back.data<T>()[i]);
} }
......
paddle/operators/parallel_do_op.cc
浏览文件 @ e92fe9e6
...@@ -30,16 +30,13 @@ static constexpr char kParallelScopes[] = "parallel_scopes"; ...@@ -30,16 +30,13 @@ static constexpr char kParallelScopes[] = "parallel_scopes";
static constexpr char kParallelBlock[] = "sub_block"; static constexpr char kParallelBlock[] = "sub_block";
// using ParallelScopeVar = std::vector<framework::Scope *>;
using LoDTensor = framework::LoDTensor; using LoDTensor = framework::LoDTensor;
using OperatorBase = framework::OperatorBase;
void SplitTensorAndMoveTensorToScopes( static void SplitTensorAndMoveTensorToScopes(
const framework::Scope &scope, const framework::Scope &scope, std::vector<framework::Scope *> *sub_scopes,
const std::vector<framework::Scope *> &sub_scopes,
const std::vector<platform::Place> &places, const std::vector<platform::Place> &places,
const std::vector<std::string> &names) { const std::vector<std::string> &names) {
PADDLE_ENFORCE_EQ(sub_scopes.size(), places.size()); size_t num_sub_scopes = 0;
for (auto &argu : names) { for (auto &argu : names) {
auto *var = scope.FindVar(argu); auto *var = scope.FindVar(argu);
const auto &tensor = var->Get<LoDTensor>(); const auto &tensor = var->Get<LoDTensor>();
...@@ -48,13 +45,31 @@ void SplitTensorAndMoveTensorToScopes( ...@@ -48,13 +45,31 @@ void SplitTensorAndMoveTensorToScopes(
for (auto &lod : lod_tensors) { for (auto &lod : lod_tensors) {
VLOG(3) << lod.dims(); VLOG(3) << lod.dims();
} }
if (num_sub_scopes == 0) {
num_sub_scopes = lod_tensors.size();
} else {
PADDLE_ENFORCE_EQ(num_sub_scopes, lod_tensors.size());
}
PADDLE_ENFORCE_NE(num_sub_scopes, 0);
if (sub_scopes->size() == 0) {
sub_scopes->reserve(num_sub_scopes);
for (size_t i = 0; i < num_sub_scopes; ++i) {
sub_scopes->emplace_back(&scope.NewScope());
}
}
for (size_t i = 0; i < sub_scopes.size(); ++i) { for (size_t i = 0; i < lod_tensors.size(); ++i) {
*sub_scopes[i]->Var(argu)->GetMutable<LoDTensor>() = lod_tensors[i]; *(*sub_scopes)[i]->Var(argu)->GetMutable<LoDTensor>() = lod_tensors[i];
} }
} }
} }
void WaitOnPlace(const platform::Place place) {
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
auto &dev_ctx = *pool.Get(place);
dev_ctx.Wait();
}
void WaitOnPlaces(const std::vector<platform::Place> places) { void WaitOnPlaces(const std::vector<platform::Place> places) {
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
...@@ -70,7 +85,7 @@ class ParallelDoOp : public framework::OperatorBase { ...@@ -70,7 +85,7 @@ class ParallelDoOp : public framework::OperatorBase {
const framework::VariableNameMap &inputs, const framework::VariableNameMap &inputs,
const framework::VariableNameMap &outputs, const framework::VariableNameMap &outputs,
const framework::AttributeMap &attrs) const framework::AttributeMap &attrs)
: OperatorBase(type, inputs, outputs, attrs) {} : framework::OperatorBase(type, inputs, outputs, attrs) {}
void Run(const framework::Scope &scope, void Run(const framework::Scope &scope,
const platform::Place &place) const override { const platform::Place &place) const override {
...@@ -85,19 +100,17 @@ class ParallelDoOp : public framework::OperatorBase { ...@@ -85,19 +100,17 @@ class ParallelDoOp : public framework::OperatorBase {
auto &sub_scopes = *scope.FindVar(Output(kParallelScopes)) auto &sub_scopes = *scope.FindVar(Output(kParallelScopes))
->GetMutable<std::vector<framework::Scope *>>(); ->GetMutable<std::vector<framework::Scope *>>();
for (size_t place_idx = 0; place_idx < places.size(); ++place_idx) {
sub_scopes.push_back(&scope.NewScope());
}
// split input // split input
SplitTensorAndMoveTensorToScopes(scope, sub_scopes, places, SplitTensorAndMoveTensorToScopes(scope, &sub_scopes, places,
Inputs(kInputs)); Inputs(kInputs));
// copy parameter // copy parameter
for (auto &param : Inputs(kParameters)) { for (auto &param : Inputs(kParameters)) {
PADDLE_ENFORCE(scope.FindVar(param)->IsType<LoDTensor>(), PADDLE_ENFORCE(scope.FindVar(param)->IsType<LoDTensor>(),
"Only support parameter type as LoDTensor"); "Only support parameter type as LoDTensor");
auto &src = scope.FindVar(param)->Get<LoDTensor>(); auto &src = scope.FindVar(param)->Get<LoDTensor>();
for (size_t i = 0; i < places.size(); ++i) { for (size_t i = 0; i < sub_scopes.size(); ++i) {
auto &place = places[i]; auto &place = places[i];
auto *sub_scope = sub_scopes[i]; auto *sub_scope = sub_scopes[i];
auto *dst = sub_scope->Var(param)->GetMutable<LoDTensor>(); auto *dst = sub_scope->Var(param)->GetMutable<LoDTensor>();
...@@ -108,9 +121,7 @@ class ParallelDoOp : public framework::OperatorBase { ...@@ -108,9 +121,7 @@ class ParallelDoOp : public framework::OperatorBase {
std::vector<std::future<void>> workers; std::vector<std::future<void>> workers;
workers.reserve(places.size()); workers.reserve(places.size());
for (size_t place_idx = 0; place_idx < places.size(); ++place_idx) { for (size_t place_idx = 0; place_idx < sub_scopes.size(); ++place_idx) {
VLOG(3) << "Run " << place_idx;
auto &place = places[place_idx]; auto &place = places[place_idx];
auto *cur_scope = sub_scopes[place_idx]; auto *cur_scope = sub_scopes[place_idx];
...@@ -157,21 +168,16 @@ ParallelDo Operator. ...@@ -157,21 +168,16 @@ ParallelDo Operator.
} }
}; };
class ParallelDoGradOp : public OperatorBase { class ParallelDoGradOp : public framework::OperatorBase {
public: public:
ParallelDoGradOp(const std::string &type, ParallelDoGradOp(const std::string &type,
const framework::VariableNameMap &inputs, const framework::VariableNameMap &inputs,
const framework::VariableNameMap &outputs, const framework::VariableNameMap &outputs,
const framework::AttributeMap &attrs) const framework::AttributeMap &attrs)
: OperatorBase(type, inputs, outputs, attrs) {} : framework::OperatorBase(type, inputs, outputs, attrs) {}
void Run(const framework::Scope &scope, void Run(const framework::Scope &scope,
const platform::Place &place) const override { const platform::Place &place) const override {
// // get device context from pool
// platform::DeviceContextPool &pool =
// platform::DeviceContextPool::Instance();
// auto &dev_ctx = *pool.Get(place);
auto *block = Attr<framework::BlockDesc *>(kParallelBlock); auto *block = Attr<framework::BlockDesc *>(kParallelBlock);
auto *program = block->Program(); auto *program = block->Program();
...@@ -181,26 +187,16 @@ class ParallelDoGradOp : public OperatorBase { ...@@ -181,26 +187,16 @@ class ParallelDoGradOp : public OperatorBase {
auto &places = scope.FindVar(Input(kPlaces))->Get<platform::PlaceList>(); auto &places = scope.FindVar(Input(kPlaces))->Get<platform::PlaceList>();
// feed output@grad // feed output@grad
SplitTensorAndMoveTensorToScopes(scope, sub_scopes, places, SplitTensorAndMoveTensorToScopes(
Inputs(framework::GradVarName(kOutputs))); scope, const_cast<std::vector<framework::Scope *> *>(&sub_scopes),
places, Inputs(framework::GradVarName(kOutputs)));
WaitOnPlaces(places); WaitOnPlaces(places);
// for debugging
for (auto &s : Inputs(framework::GradVarName(kOutputs))) {
VLOG(3) << s;
VLOG(3) << scope.FindVar(s)->Get<LoDTensor>();
for (auto *sub_scope : sub_scopes) {
VLOG(3) << sub_scope->FindVar(s)->Get<LoDTensor>();
}
}
// exe run // exe run
std::vector<std::future<void>> workers; std::vector<std::future<void>> workers;
for (size_t place_idx = 0; place_idx < places.size(); ++place_idx) { for (size_t i = 0; i < sub_scopes.size(); ++i) {
VLOG(3) << "Run " << place_idx; auto &place = places[i];
auto *cur_scope = sub_scopes[i];
auto &place = places[place_idx];
auto *cur_scope = sub_scopes[place_idx];
// execute // execute
workers.emplace_back(framework::Async([program, cur_scope, place, block] { workers.emplace_back(framework::Async([program, cur_scope, place, block] {
...@@ -216,33 +212,40 @@ class ParallelDoGradOp : public OperatorBase { ...@@ -216,33 +212,40 @@ class ParallelDoGradOp : public OperatorBase {
// merge grad // merge grad
for (auto &s : Outputs(framework::GradVarName(kParameters))) { for (auto &s : Outputs(framework::GradVarName(kParameters))) {
VLOG(3) << "merge grad " << s; auto &result = sub_scopes[0]->FindVar(s)->Get<LoDTensor>();
std::string tmp_name;
auto &t = sub_scopes[0]->FindVar(s)->Get<LoDTensor>(); auto *tmp = sub_scopes[0]->Var(&tmp_name)->GetMutable<LoDTensor>();
VLOG(3) << t;
for (size_t i = 1; i < sub_scopes.size(); ++i) {
std::string s_buf = s + "@BUF"; auto &tensor_to_merge = sub_scopes[i]->FindVar(s)->Get<LoDTensor>();
auto *t_buf = sub_scopes[0]->Var(s_buf)->GetMutable<LoDTensor>(); if (!(places[i] == places[0])) {
framework::Copy(tensor_to_merge, places[0], tmp);
for (size_t place_idx = 1; place_idx < places.size(); ++place_idx) { WaitOnPlace(places[0]);
auto &tt = sub_scopes[place_idx]->FindVar(s)->Get<LoDTensor>(); } else {
VLOG(3) << place_idx; tmp->ShareDataWith(tensor_to_merge);
VLOG(3) << tt; }
framework::Copy(tt, places[0], t_buf);
auto sum_op = framework::OpRegistry::CreateOp( auto sum_op = framework::OpRegistry::CreateOp(
"sum", {{"X", {s, s_buf}}}, {{"Out", {s}}}, "sum", {{"X", {s, tmp_name}}}, {{"Out", {s}}},
framework::AttributeMap{}); framework::AttributeMap{});
sum_op->Run(*sub_scopes[0], places[0]); sum_op->Run(*sub_scopes[0], places[0]);
WaitOnPlaces(places); WaitOnPlace(places[0]);
} }
VLOG(3) << t; VLOG(3) << result;
framework::Copy(t, place, scope.FindVar(s)->GetMutable<LoDTensor>()); framework::Copy(result, place, scope.FindVar(s)->GetMutable<LoDTensor>());
} }
WaitOnPlaces(places);
} }
}; };
std::ostream &operator<<(std::ostream &sout,
const std::vector<std::string> &strs) {
std::copy(strs.begin(), strs.end(),
std::ostream_iterator<std::string>(sout, ","));
return sout;
}
class ParallelDoGradOpDescMaker : public framework::SingleGradOpDescMaker { class ParallelDoGradOpDescMaker : public framework::SingleGradOpDescMaker {
public: public:
using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
...@@ -283,18 +286,30 @@ class ParallelDoGradOpShapeInference : public framework::InferShapeBase { ...@@ -283,18 +286,30 @@ class ParallelDoGradOpShapeInference : public framework::InferShapeBase {
void operator()(framework::InferShapeContext *ctx) const override { void operator()(framework::InferShapeContext *ctx) const override {
std::vector<std::string> input{kParameters, kInputs}; std::vector<std::string> input{kParameters, kInputs};
std::vector<std::string> output{kOutputs}; std::vector<std::string> output{kOutputs};
for (auto &s : input) {
PADDLE_ENFORCE(ctx->HasInputs(s)); PADDLE_ENFORCE(ctx->HasInputs(kParameters));
PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName(s)), PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName(kParameters)));
"Cannot find the gradient variable %s", PADDLE_ENFORCE(ctx->HasInput(kInputs));
framework::GradVarName(s));
}
for (auto &s : output) { for (auto &s : output) {
PADDLE_ENFORCE(ctx->HasInputs(s)); PADDLE_ENFORCE(ctx->HasInputs(s));
} }
for (auto &s : input) {
ctx->SetOutputsDim(framework::GradVarName(s), ctx->GetInputsDim(s)); ctx->SetOutputsDim(framework::GradVarName(kParameters),
ctx->GetInputsDim(kParameters));
auto i_dims = ctx->GetInputsDim(kInputs);
auto ig_names = ctx->Outputs(framework::GradVarName(kInputs));
for (size_t i = 0; i < ig_names.size(); ++i) {
auto &ig_name = ig_names[i];
if (ig_name == framework::kEmptyVarName) {
continue;
}
ctx->SetDims({ig_name}, {i_dims[i]});
} }
if (ctx->HasInputs(kParameters)) { if (ctx->HasInputs(kParameters)) {
PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName(kParameters))); PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName(kParameters)));
ctx->SetOutputsDim(framework::GradVarName(kParameters), ctx->SetOutputsDim(framework::GradVarName(kParameters),
......
paddle/operators/pool_op.cc
浏览文件 @ e92fe9e6
...@@ -64,6 +64,13 @@ void PoolOp::InferShape(framework::InferShapeContext *ctx) const { ...@@ -64,6 +64,13 @@ void PoolOp::InferShape(framework::InferShapeContext *ctx) const {
framework::OpKernelType PoolOp::GetExpectedKernelType( framework::OpKernelType PoolOp::GetExpectedKernelType(
const framework::ExecutionContext &ctx) const { const framework::ExecutionContext &ctx) const {
bool use_cudnn = ctx.Attr<bool>("use_cudnn"); bool use_cudnn = ctx.Attr<bool>("use_cudnn");
use_cudnn &= platform::is_gpu_place(ctx.GetPlace());
#ifdef PADDLE_WITH_CUDA
if (platform::is_gpu_place(ctx.GetPlace())) {
auto &dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
use_cudnn &= dev_ctx.cudnn_handle() != nullptr;
}
#endif
framework::LibraryType library_; framework::LibraryType library_;
if (use_cudnn) { if (use_cudnn) {
library_ = framework::LibraryType::kCUDNN; library_ = framework::LibraryType::kCUDNN;
...@@ -88,6 +95,13 @@ void PoolOpGrad::InferShape(framework::InferShapeContext *ctx) const { ...@@ -88,6 +95,13 @@ void PoolOpGrad::InferShape(framework::InferShapeContext *ctx) const {
framework::OpKernelType PoolOpGrad::GetExpectedKernelType( framework::OpKernelType PoolOpGrad::GetExpectedKernelType(
const framework::ExecutionContext &ctx) const { const framework::ExecutionContext &ctx) const {
bool use_cudnn = ctx.Attr<bool>("use_cudnn"); bool use_cudnn = ctx.Attr<bool>("use_cudnn");
use_cudnn &= platform::is_gpu_place(ctx.GetPlace());
#ifdef PADDLE_WITH_CUDA
if (platform::is_gpu_place(ctx.GetPlace())) {
auto &dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
use_cudnn &= dev_ctx.cudnn_handle() != nullptr;
}
#endif
framework::LibraryType library_; framework::LibraryType library_;
if (use_cudnn) { if (use_cudnn) {
library_ = framework::LibraryType::kCUDNN; library_ = framework::LibraryType::kCUDNN;
......
paddle/operators/recv_op.cc
浏览文件 @ e92fe9e6
...@@ -27,6 +27,7 @@ limitations under the License. */ ...@@ -27,6 +27,7 @@ limitations under the License. */
#include "paddle/operators/detail/grpc_server.h" #include "paddle/operators/detail/grpc_server.h"
#include "paddle/operators/detail/sendrecvop_utils.h" #include "paddle/operators/detail/sendrecvop_utils.h"
#include "paddle/operators/detail/simple_block_queue.h" #include "paddle/operators/detail/simple_block_queue.h"
#include "paddle/string/printf.h"
#define LISTEN_TERMINATE_MESSAGE "TERMINATE@RECV" #define LISTEN_TERMINATE_MESSAGE "TERMINATE@RECV"
...@@ -34,6 +35,9 @@ namespace paddle { ...@@ -34,6 +35,9 @@ namespace paddle {
namespace operators { namespace operators {
constexpr char kOptimizeBlock[] = "OptimizeBlock"; constexpr char kOptimizeBlock[] = "OptimizeBlock";
constexpr int kCondStart = 0;
constexpr int kCondRunning = 1;
constexpr int kCondDone = 2;
void RunServer(std::shared_ptr<detail::AsyncGRPCServer> service) { void RunServer(std::shared_ptr<detail::AsyncGRPCServer> service) {
service->RunSyncUpdate(); service->RunSyncUpdate();
...@@ -79,42 +83,39 @@ class RecvOp : public framework::OperatorBase { ...@@ -79,42 +83,39 @@ class RecvOp : public framework::OperatorBase {
if (grads_counter_.find(varname) == grads_counter_.end()) { if (grads_counter_.find(varname) == grads_counter_.end()) {
grads_counter_[varname] = 0; grads_counter_[varname] = 0;
} }
char ret[256]; return string::Sprintf("%s.trainer_%d", varname, grads_counter_[varname]++);
snprintf(ret, sizeof(ret), "%s.trainer_%d", varname.c_str(),
grads_counter_[varname]++);
return std::string(ret);
} }
void Run(const framework::Scope &scope, void Run(const framework::Scope &scope,
const platform::Place &dev_place) const override { const platform::Place &dev_place) const override {
// FIXME(typhoonzero): no new scopes for every run.
framework::Scope &recv_scope = scope.NewScope();
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
auto &dev_ctx = *pool.Get(dev_place); auto &dev_ctx = *pool.Get(dev_place);
framework::Scope &recv_scope = scope.NewScope();
// FIXME(Yancey1989): initialize rpc server with laze mode. // FIXME(Yancey1989): initialize rpc server with laze mode.
rpc_service_->SetScope(&recv_scope); rpc_service_->SetScope(&recv_scope);
rpc_service_->SetDevCtx(&dev_ctx); rpc_service_->SetDevCtx(&dev_ctx);
auto param_list = Attr<std::vector<std::string>>("ParamList"); auto param_list = Attr<std::vector<std::string>>("ParamList");
auto grad_list = Attr<std::vector<std::string>>("GradList"); auto grad_list = Attr<std::vector<std::string>>("GradList");
auto trainer_count = Attr<int>("Trainers"); auto fan_in = Attr<int>("Fanin");
size_t param_count = param_list.size(); size_t param_count = param_list.size();
rpc_service_->Reset(); auto *block = Attr<framework::BlockDesc *>(kOptimizeBlock);
auto *program = block->Program();
framework::Executor executor(dev_place);
// TODO(typhoonzero): change this to a while_op for every cluster-batch. // TODO(typhoonzero): change this to a while_op for every cluster-batch.
bool exit_flag = false; bool exit_flag = false;
VLOG(4) << "param_count:" << param_count int64_t barrier_size = param_count * fan_in;
<< " trainer_count:" << trainer_count;
while (!exit_flag) { while (!exit_flag) {
// TODO(gognwb): simply this loop. // Get from multiple trainers, we don't care about the order in which
// Get from multiple trainers, we don't care about order in which // the gradients arrives, just add suffix 0~n and merge the gradient.
// the gradient arrives, just add suffix 0~n then average the gradient. rpc_service_->SetCond(0);
for (size_t i = 0; i < param_count * trainer_count; ++i) { for (size_t i = 0; i < barrier_size; ++i) {
// blocking get one var from client.
const detail::MessageWithName &v = rpc_service_->Get(); const detail::MessageWithName &v = rpc_service_->Get();
auto grad_var_name = v.first; auto grad_var_name = v.first;
if (grad_var_name == LISTEN_TERMINATE_MESSAGE) { if (grad_var_name == LISTEN_TERMINATE_MESSAGE) {
VLOG(4) << "received LISTEN_TERMINATE_MESSAGE and RunOp.Run() exit"; LOG(INFO) << "received terminate message and exit";
exit_flag = true; exit_flag = true;
break; break;
} }
...@@ -123,47 +124,32 @@ class RecvOp : public framework::OperatorBase { ...@@ -123,47 +124,32 @@ class RecvOp : public framework::OperatorBase {
if (it != grad_list.end()) { if (it != grad_list.end()) {
param_var_name = param_list[it - grad_list.begin()]; param_var_name = param_list[it - grad_list.begin()];
} else { } else {
LOG(ERROR) << "grad have no paired param found!\"" << grad_var_name LOG(ERROR) << "grad have no paired param:" << grad_var_name;
<< "\"";
} }
VLOG(3) << "recved grad: " << grad_var_name VLOG(3) << "recved grad: " << grad_var_name
<< " updating param: " << param_var_name; << " updating param: " << param_var_name;
if (fan_in > 1) {
auto *merged_grad = recv_scope.FindVar(grad_var_name);
if (merged_grad == nullptr) {
auto *ptr = recv_scope.Var(grad_var_name);
CreateTensorFromMessageType(ptr, v.second.type());
VLOG(3) << "Create Variable " << grad_var_name
<< " on recv scope, which pointer is " << ptr << " type is "
<< v.second.type();
}
if (trainer_count > 1) {
grad_var_name = this->GetGradVarNameForTrainer(grad_var_name); grad_var_name = this->GetGradVarNameForTrainer(grad_var_name);
} }
auto *var = recv_scope.FindVar(grad_var_name);
auto *var = recv_scope.Var(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");
}
detail::DeserializeFromMessage(v.second, dev_ctx, var); detail::DeserializeFromMessage(v.second, dev_ctx, var);
} }
if (exit_flag) { if (exit_flag) {
break; break;
} }
rpc_service_->Reset();
auto *block = Attr<framework::BlockDesc *>(kOptimizeBlock);
auto *program = block->Program();
framework::Executor executor(dev_place);
// Run sub graph to get optimized tensor
try { try {
executor.Run(*program, &recv_scope, block->ID(), /*global_block*/ executor.Run(*program, &recv_scope, block->ID(), /*global_block*/
false /*create_local_scope*/, false /*create_vars*/); false /*create_local_scope*/, false /*create_vars*/);
} catch (std::exception &e) { } catch (std::exception &e) {
LOG(ERROR) << "run sub program error " << e.what(); LOG(ERROR) << "run sub program error " << e.what();
} }
rpc_service_->SetCond(1);
rpc_service_->Done(); rpc_service_->WaitClientGet(barrier_size);
grads_counter_.clear(); grads_counter_.clear();
} // while(true) } // while(true)
} }
...@@ -199,7 +185,7 @@ This operator will recv tensor from send_op ...@@ -199,7 +185,7 @@ This operator will recv tensor from send_op
"GradList", "type list of string", "GradList", "type list of string",
"grad->param name mapping to find which param to optimize.") "grad->param name mapping to find which param to optimize.")
.SetDefault({}); .SetDefault({});
AddAttr<int>("Trainers", "type int", AddAttr<int>("Fanin", "type int",
"Number of trainers in the current cluster job") "Number of trainers in the current cluster job")
.SetDefault(1); .SetDefault(1);
} }
......
paddle/operators/reduce_op.cc
浏览文件 @ e92fe9e6
...@@ -129,7 +129,7 @@ If reduce_all is true, just reduce along all dimensions and output a scalar. ...@@ -129,7 +129,7 @@ If reduce_all is true, just reduce along all dimensions and output a scalar.
} }
void SetComment(std::string name, std::string op) { void SetComment(std::string name, std::string op) {
Replace(comment_, "{ReduceOP}", name); Replace(comment_, "{ReduceOp}", name);
Replace(comment_, "{reduce}", op); Replace(comment_, "{reduce}", op);
} }
}; };
......
paddle/operators/send_op.cc
浏览文件 @ e92fe9e6
...@@ -41,10 +41,13 @@ class SendOp : public framework::OperatorBase { ...@@ -41,10 +41,13 @@ class SendOp : public framework::OperatorBase {
platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
auto& ctx = *pool.Get(place); auto& ctx = *pool.Get(place);
for (size_t i = 0; i < ins.size(); i++) { for (size_t i = 0; i < ins.size(); i++) {
VLOG(3) << "sending " << ins[i];
client_.AsyncSendVariable(epmap[i], ctx, scope, ins[i]); client_.AsyncSendVariable(epmap[i], ctx, scope, ins[i]);
} }
PADDLE_ENFORCE(client_.Wait());
for (size_t i = 0; i < outs.size(); i++) { for (size_t i = 0; i < outs.size(); i++) {
VLOG(3) << "getting " << outs[i];
client_.AsyncGetVariable(epmap[i], ctx, scope, outs[i]); client_.AsyncGetVariable(epmap[i], ctx, scope, outs[i]);
} }
......
paddle/operators/sequence_erase_op.cc
浏览文件 @ e92fe9e6
...@@ -86,4 +86,5 @@ REGISTER_OP_WITHOUT_GRADIENT(sequence_erase, ops::SequenceEraseOp, ...@@ -86,4 +86,5 @@ REGISTER_OP_WITHOUT_GRADIENT(sequence_erase, ops::SequenceEraseOp,
ops::SequenceEraseOpMaker); ops::SequenceEraseOpMaker);
REGISTER_OP_CPU_KERNEL( REGISTER_OP_CPU_KERNEL(
sequence_erase, sequence_erase,
ops::SequenceEraseKernel<paddle::platform::CPUDeviceContext, int32_t>); ops::SequenceEraseKernel<paddle::platform::CPUDeviceContext, int32_t>,
ops::SequenceEraseKernel<paddle::platform::CPUDeviceContext, int64_t>);
paddle/operators/sequence_erase_op.cu
浏览文件 @ e92fe9e6
...@@ -23,27 +23,22 @@ using platform::PADDLE_CUDA_NUM_THREADS; ...@@ -23,27 +23,22 @@ using platform::PADDLE_CUDA_NUM_THREADS;
using LoDTensor = framework::LoDTensor; using LoDTensor = framework::LoDTensor;
template <typename T> template <typename T>
__global__ void LabelErasedIdx(const T* in_dat, const int in_len, __global__ void LabelErasedIdx(const T* in_dat, const int64_t in_len,
const T* tokens, const int tokens_len, const int* tokens, const size_t tokens_len,
int* num_erased) { size_t* num_erased) {
int index = blockIdx.x * blockDim.x + threadIdx.x; int index = blockIdx.x * blockDim.x + threadIdx.x;
if (index < in_len) { if (index < in_len) {
int erased = 0; for (size_t i = 0; i < tokens_len; ++i) {
for (int i = 0; i < tokens_len; ++i) {
if (in_dat[index] == tokens[i]) { if (in_dat[index] == tokens[i]) {
erased = 1; num_erased[index + 1] = 1;
break;
} }
} }
num_erased[index + 1] = erased;
if (index == 0) {
num_erased[0] = 0;
}
} }
} }
template <typename T> __global__ void GetOutLod(const size_t* num_erased, const size_t* in_lod,
__global__ void GetOutLod(const T* num_erased, const int* in_lod, const size_t lod_len, size_t* out_lod0) {
const int lod_len, int* out_lod0) {
int index = blockIdx.x * blockDim.x + threadIdx.x; int index = blockIdx.x * blockDim.x + threadIdx.x;
if (index < lod_len) { if (index < lod_len) {
out_lod0[index] = in_lod[index] - num_erased[in_lod[index]]; out_lod0[index] = in_lod[index] - num_erased[in_lod[index]];
...@@ -51,11 +46,11 @@ __global__ void GetOutLod(const T* num_erased, const int* in_lod, ...@@ -51,11 +46,11 @@ __global__ void GetOutLod(const T* num_erased, const int* in_lod,
} }
template <typename T> template <typename T>
__global__ void SetOutput(const T* in_dat, const int in_len, __global__ void SetOutput(const T* in_dat, const int64_t in_len,
const int* num_erased, T* out_dat) { const size_t* num_erased, T* out_dat) {
int index = blockIdx.x * blockDim.x + threadIdx.x; int index = blockIdx.x * blockDim.x + threadIdx.x;
if (index < in_len) { if (index < in_len) {
if (in_dat[index] != in_dat[index + 1]) { if (num_erased[index] == num_erased[index + 1]) {
out_dat[index - num_erased[index]] = in_dat[index]; out_dat[index - num_erased[index]] = in_dat[index];
} }
} }
...@@ -72,53 +67,44 @@ class SequenceEraseOpCUDAKernel : public framework::OpKernel<T> { ...@@ -72,53 +67,44 @@ class SequenceEraseOpCUDAKernel : public framework::OpKernel<T> {
PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now."); PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now.");
PADDLE_ENFORCE_EQ(lod[0].back(), (size_t)in->numel(), PADDLE_ENFORCE_EQ(lod[0].back(), (size_t)in->numel(),
"The actual size mismatches with the LoD information."); "The actual size mismatches with the LoD information.");
auto tokens = ctx.Attr<std::vector<T>>("tokens"); auto tokens = ctx.Attr<std::vector<int>>("tokens");
auto tokens_len = tokens.size();
auto in_len = in->numel(); auto in_len = in->numel();
auto in_dat = in->data<T>(); auto in_dat = in->data<T>();
auto lod0 = lod[0]; // Copy tokens to GPU
thrust::device_vector<int> dev_tokens(tokens.begin(), tokens.end());
thrust::host_vector<T> host_tokens(tokens_len); int* dev_tokens_ptr = thrust::raw_pointer_cast(dev_tokens.data());
for (size_t i = 0; i < tokens.size(); ++i) {
host_tokens[i] = tokens[i];
}
thrust::device_vector<T> dev_tokens = host_tokens;
thrust::device_vector<int> num_erased(in_len + 1);
T* dev_tokens_ptr = thrust::raw_pointer_cast(dev_tokens.data());
int* num_erased_ptr = thrust::raw_pointer_cast(num_erased.data());
// Count number of elements to be erased
thrust::device_vector<size_t> num_erased(in_len + 1, 0);
size_t* num_erased_ptr = thrust::raw_pointer_cast(num_erased.data());
auto stream = ctx.cuda_device_context().stream(); auto stream = ctx.cuda_device_context().stream();
LabelErasedIdx<<<(in_len - 1) / PADDLE_CUDA_NUM_THREADS + 1, LabelErasedIdx<<<(in_len - 1) / PADDLE_CUDA_NUM_THREADS + 1,
PADDLE_CUDA_NUM_THREADS, 0, stream>>>( PADDLE_CUDA_NUM_THREADS, 0, stream>>>(
in_dat, in_len, dev_tokens_ptr, tokens_len, num_erased_ptr); in_dat, in_len, dev_tokens_ptr, tokens.size(), num_erased_ptr);
thrust::inclusive_scan(num_erased.begin() + 1, num_erased.end(), thrust::inclusive_scan(num_erased.begin() + 1, num_erased.end(),
num_erased.begin() + 1); num_erased.begin() + 1);
// Calc LoD // Copy LoD to GPU
auto lod0 = lod[0];
auto lod_len = lod0.size(); auto lod_len = lod0.size();
thrust::host_vector<int> host_lod(lod_len); thrust::device_vector<size_t> dev_in_lod = lod0;
for (size_t i = 0; i < lod_len; ++i) { size_t* dev_in_lod_ptr = thrust::raw_pointer_cast(dev_in_lod.data());
host_lod[i] = lod0[i];
} // Calc output LoD
thrust::device_vector<int> dev_in_lod = host_lod; thrust::device_vector<size_t> dev_out_lod(lod_len);
thrust::device_vector<int> dev_out_lod(lod_len); size_t* dev_out_lod_ptr = thrust::raw_pointer_cast(dev_out_lod.data());
int* dev_in_lod_ptr = thrust::raw_pointer_cast(dev_in_lod.data());
int* dev_out_lod_ptr = thrust::raw_pointer_cast(dev_out_lod.data());
GetOutLod<<<(lod_len - 1) / PADDLE_CUDA_NUM_THREADS + 1, GetOutLod<<<(lod_len - 1) / PADDLE_CUDA_NUM_THREADS + 1,
PADDLE_CUDA_NUM_THREADS, 0, stream>>>( PADDLE_CUDA_NUM_THREADS, 0, stream>>>(
num_erased_ptr, dev_in_lod_ptr, lod_len, dev_out_lod_ptr); num_erased_ptr, dev_in_lod_ptr, lod_len, dev_out_lod_ptr);
thrust::host_vector<int> host_out_lod = dev_out_lod;
std::vector<int> out_lod0(lod_len, 0); // Set LoD for output
for (size_t i = 0; i < lod_len; i++) { thrust::host_vector<size_t> out_lod0 = dev_out_lod;
out_lod0[i] = host_out_lod[i];
}
framework::LoD out_lod; framework::LoD out_lod;
out_lod.push_back(out_lod0); out_lod.push_back(out_lod0);
out->set_lod(out_lod); out->set_lod(out_lod);
// Set output // Set output
out->Resize({out_lod0.back(), 1}); out->Resize({static_cast<int64_t>(out_lod0.back()), 1});
auto out_dat = out->mutable_data<T>(ctx.GetPlace()); auto out_dat = out->mutable_data<T>(ctx.GetPlace());
SetOutput<<<(in_len - 1) / PADDLE_CUDA_NUM_THREADS + 1, SetOutput<<<(in_len - 1) / PADDLE_CUDA_NUM_THREADS + 1,
PADDLE_CUDA_NUM_THREADS, 0, stream>>>(in_dat, in_len, PADDLE_CUDA_NUM_THREADS, 0, stream>>>(in_dat, in_len,
...@@ -130,4 +116,5 @@ class SequenceEraseOpCUDAKernel : public framework::OpKernel<T> { ...@@ -130,4 +116,5 @@ class SequenceEraseOpCUDAKernel : public framework::OpKernel<T> {
} // namespace paddle } // namespace paddle
REGISTER_OP_CUDA_KERNEL(sequence_erase, REGISTER_OP_CUDA_KERNEL(sequence_erase,
paddle::operators::SequenceEraseOpCUDAKernel<int32_t>); paddle::operators::SequenceEraseOpCUDAKernel<int32_t>,
paddle::operators::SequenceEraseOpCUDAKernel<int64_t>);
paddle/operators/sequence_expand_op.cc
浏览文件 @ e92fe9e6
...@@ -58,7 +58,7 @@ This operator expands input(X) according to LOD of input(Y). ...@@ -58,7 +58,7 @@ This operator expands input(X) according to LOD of input(Y).
Following are cases to better explain how this works: Following are cases to better explain how this works:
Case 1: Case 1:
Given 2-level a LoDTensor input(X) Given a 2-level LoDTensor input(X)
X.lod = [[0, 2, 3], X.lod = [[0, 2, 3],
[0, 1, 3, 4]] [0, 1, 3, 4]]
X.data = [a, b, c, d] X.data = [a, b, c, d]
...@@ -75,9 +75,8 @@ then we get 2-level LoDTensor ...@@ -75,9 +75,8 @@ then we get 2-level LoDTensor
Case 2: Case 2:
Given a 0-level LoDTensor input(X) Given a common Tensor input(X)
X.data = [a, b, c] X.data = [a, b, c]
X.lod = NULL
X.dims = [3, 1] X.dims = [3, 1]
and input(Y) and input(Y)
Y.lod = [[0, 2, 3, 6]] Y.lod = [[0, 2, 3, 6]]
...@@ -89,9 +88,8 @@ then we get 1-level LoDTensor ...@@ -89,9 +88,8 @@ then we get 1-level LoDTensor
Case 3: Case 3:
Given a 0-level LoDTensor input(X) Given a common Tensor input(X)
X.data = [[a, b], [c, d], [e, f]] X.data = [[a, b], [c, d], [e, f]]
X.lod = NULL
X.dims = [3, 2] X.dims = [3, 2]
and input(Y) and input(Y)
Y.lod = [[0, 2, 3, 6]] Y.lod = [[0, 2, 3, 6]]
......
paddle/operators/sequence_reshape_op.cc 0 → 100644
浏览文件 @ e92fe9e6
/* 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. */
#include "paddle/operators/sequence_reshape_op.h"
#include "paddle/framework/ddim.h"
namespace paddle {
namespace operators {
class SequenceReshapeOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"),
"Input(X) of SequenceReshapeOp should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("Out"),
"Output(Out) of SequenceReshapeOp should not be null.");
auto x_dims = ctx->GetInputDim("X");
auto x_numel = product(x_dims);
PADDLE_ENFORCE_EQ(x_dims.size(), 2U, "Rank of Input(X) should be 2.");
int new_dim = ctx->Attrs().Get<int>("new_dim");
ctx->SetOutputDim("Out",
{x_numel / new_dim, static_cast<int64_t>(new_dim)});
}
};
class SequenceReshapeOpMaker : public framework::OpProtoAndCheckerMaker {
public:
SequenceReshapeOpMaker(OpProto* proto, OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X",
"(LoDTensor, default LoDTensor<float>) A 2-D LoDTensor with shape "
"being [N, M].");
AddOutput("Out",
"(LoDTensor, default LoDTensor<float>) A 2-D LoDTensor with "
"shape [T, new_dim] where T is calculated based on X.lod, M and "
"new_dim.");
AddAttr<int>("new_dim", "Sequence dimension of the output LoDTensor.");
AddComment(R"DOC(
Sequence Reshape Operator.
This operator will rearrange the input sequences. The new dimension is set by
attribute and length of each sequence may change longer or shorter which is
decided by original length, original dimension and new dimension. The following
example will help to illustrate the function of this operator:
x is a LoDTensor:
x.lod = [[0, 2, 6]]
x.data = [[1, 2], [3, 4],
[5, 6], [7, 8], [9, 10], [11, 12]]
x.dims = [6, 2]
set new_dim = 4
then out is a LoDTensor:
out.lod = [[0, 1, 3]]
out.data = [[1, 2, 3, 4],
[5, 6, 7, 8], [9, 10, 11, 12]]
out.dims = [3, 4]
Currently, only 1-level LoDTensor is supported and please make sure (original
length * original dimension) can be divided by new_dim with no remainder for
each sequence.
)DOC");
}
};
class SequenceReshapeGradOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(
ctx->HasInput(framework::GradVarName("Out")),
"Input(Out@GRAD) of SequenceReshapeGradOp should not be null.");
PADDLE_ENFORCE(ctx->HasInput("X"),
"Input(X) of SequenceReshapeGradOp should not be null.");
ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
ctx->ShareLoD("X", /*->*/ framework::GradVarName("X"));
}
};
class SequenceReshapeGradOpMaker : public framework::SingleGradOpDescMaker {
public:
using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
protected:
std::unique_ptr<framework::OpDesc> Apply() const override {
auto* op_desc_ptr = new framework::OpDesc();
op_desc_ptr->SetType("sequence_reshape_grad");
op_desc_ptr->SetInput("X", Input("X"));
op_desc_ptr->SetInput(framework::GradVarName("Out"), OutputGrad("Out"));
op_desc_ptr->SetOutput(framework::GradVarName("X"), InputGrad("X"));
op_desc_ptr->SetAttrMap(Attrs());
return std::unique_ptr<framework::OpDesc>(op_desc_ptr);
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(sequence_reshape, ops::SequenceReshapeOp,
ops::SequenceReshapeOpMaker, ops::SequenceReshapeGradOpMaker);
REGISTER_OPERATOR(sequence_reshape_grad, ops::SequenceReshapeGradOp);
REGISTER_OP_CPU_KERNEL(
sequence_reshape,
ops::SequenceReshapeKernel<paddle::platform::CPUDeviceContext, float>,
ops::SequenceReshapeKernel<paddle::platform::CPUDeviceContext, double>,
ops::SequenceReshapeKernel<paddle::platform::CPUDeviceContext, int>,
ops::SequenceReshapeKernel<paddle::platform::CPUDeviceContext, int64_t>);
REGISTER_OP_CPU_KERNEL(
sequence_reshape_grad,
ops::SequenceReshapeGradKernel<paddle::platform::CPUDeviceContext, float>,
ops::SequenceReshapeGradKernel<paddle::platform::CPUDeviceContext, double>,
ops::SequenceReshapeGradKernel<paddle::platform::CPUDeviceContext, int64_t>,
ops::SequenceReshapeGradKernel<paddle::platform::CPUDeviceContext, int>);
paddle/operators/sequence_reshape_op.cu 0 → 100644
浏览文件 @ e92fe9e6
/* 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. */
#include "paddle/operators/sequence_reshape_op.h"
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(
sequence_reshape,
ops::SequenceReshapeKernel<paddle::platform::CUDADeviceContext, float>,
ops::SequenceReshapeKernel<paddle::platform::CUDADeviceContext, double>,
ops::SequenceReshapeKernel<paddle::platform::CUDADeviceContext, int>,
ops::SequenceReshapeKernel<paddle::platform::CUDADeviceContext, int64_t>);
REGISTER_OP_CUDA_KERNEL(
sequence_reshape_grad,
ops::SequenceReshapeGradKernel<paddle::platform::CUDADeviceContext, float>,
ops::SequenceReshapeGradKernel<paddle::platform::CUDADeviceContext, double>,
ops::SequenceReshapeGradKernel<paddle::platform::CUDADeviceContext,
int64_t>,
ops::SequenceReshapeGradKernel<paddle::platform::CUDADeviceContext, int>);
paddle/operators/sequence_reshape_op.h 0 → 100644
浏览文件 @ e92fe9e6
/* 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. */
#pragma once
#include "paddle/framework/op_registry.h"
#include "paddle/operators/math/math_function.h"
namespace paddle {
namespace operators {
using LoDTensor = framework::LoDTensor;
template <typename DeviceContext, typename T>
class SequenceReshapeKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
auto* in = context.Input<LoDTensor>("X");
auto* out = context.Output<LoDTensor>("Out");
int out_width = context.Attr<int>("new_dim");
auto in_dims = in->dims();
int64_t in_width = in_dims[1];
auto& in_lod = in->lod();
PADDLE_ENFORCE_EQ(in_lod.size(), 1UL,
"Only support one level sequence now.");
PADDLE_ENFORCE_EQ(
in_dims[0], in_lod[0].back(),
"Inconsistent size between X.shape[0] and X.lod()[0].back().");
auto in_lod_l0 = in_lod[0];
int seq_num = in_lod_l0.size() - 1;
if (in_width == out_width) {
out->set_lod(in->lod());
} else {
auto& out_lod = *out->mutable_lod();
out_lod.resize(1);
out_lod[0].resize(seq_num + 1);
out_lod[0][0] = 0;
for (int i = 0; i < seq_num; ++i) {
size_t seq_len = in_lod_l0[i + 1] - in_lod_l0[i];
size_t offset = 0;
offset = (seq_len * in_width) / out_width;
PADDLE_ENFORCE_EQ(offset * out_width, seq_len * in_width,
"Please make sure (sequence_length * dimension) can "
"be divided by new_dim with no remainder for each "
"sequence. The %dth sequence is invalid.",
i + 1);
out_lod[0][i + 1] = out_lod[0][i] + offset;
}
}
framework::Copy(*in, context.GetPlace(), out);
out->Resize({static_cast<int64_t>(out->lod()[0].back()), out_width});
}
};
template <typename DeviceContext, typename T>
class SequenceReshapeGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
auto* x_tensor_ptr = context.Input<LoDTensor>("X");
auto* outg_tensor_ptr =
context.Input<LoDTensor>(framework::GradVarName("Out"));
auto* xg_tensor_ptr =
context.Output<LoDTensor>(framework::GradVarName("X"));
xg_tensor_ptr->mutable_data<T>(context.GetPlace());
framework::Copy(*outg_tensor_ptr, context.GetPlace(), xg_tensor_ptr);
xg_tensor_ptr->Resize(x_tensor_ptr->dims());
}
};
} // namespace operators
} // namespace paddle
paddle/operators/split_selected_rows_op.cc 0 → 100644
浏览文件 @ e92fe9e6
/* 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/split_selected_rows_op.h"
namespace paddle {
namespace operators {
class SplitSelectedRowsOpMaker : public framework::OpProtoAndCheckerMaker {
public:
SplitSelectedRowsOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "The input SelectedRows.");
AddOutput("Out", "The outputs of input SelectedRows.").AsDuplicable();
AddAttr<std::vector<int>>("rows_sections", "Rows section for output.")
.SetDefault(std::vector<int>({}));
AddAttr<std::vector<int>>("height_sections",
"Height for each output SelectedRows.")
.SetDefault(std::vector<int>({}));
AddComment(R"DOC(
Split a SelectedRows with a specified rows section.
height_sections is only needed when need to split the dims of the original tensor.
Example:
Input:
X.rows = {0, 7, 5}
X.height = 12
Attr:
rows_sections = {1, 2}
height_sections = {}
Out:
out0.rows = {0}
out0.height = 12
out1.rows = {7, 5}
out2.height = 12
)DOC");
}
};
class SplitSelectedRowsOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext *ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"), "SplitSelectedRowsOp must has input X.");
PADDLE_ENFORCE(ctx->HasOutputs("Out"),
"SplitSelectedRowsOp must has output Out.");
std::vector<int> height_sections =
ctx->Attrs().Get<std::vector<int>>("height_sections");
std::vector<int> rows_sections =
ctx->Attrs().Get<std::vector<int>>("rows_sections");
PADDLE_ENFORCE_EQ(
rows_sections.size(), ctx->Outputs("Out").size(),
"The size of rows section should be the same with Outputs size.");
int64_t n = ctx->Outputs("Out").size();
std::vector<framework::DDim> outs_dims;
outs_dims.reserve(n);
// make output dims
for (int64_t i = 0; i < n; ++i) {
auto dims = ctx->GetInputDim("X");
if (height_sections.size()) {
PADDLE_ENFORCE_EQ(
height_sections.size(), static_cast<size_t>(n),
"The size of height section should be the same with height"
" section size.");
dims[0] = height_sections[i];
}
outs_dims.push_back(dims);
}
ctx->SetOutputsDim("Out", outs_dims);
}
};
class SplitSelectedRowsGradMaker : public framework::SingleGradOpDescMaker {
public:
using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
protected:
std::unique_ptr<framework::OpDesc> Apply() const override {
auto *grad_op = new framework::OpDesc();
grad_op->SetType("sum");
grad_op->SetInput("X", OutputGrad("Out"));
grad_op->SetOutput("Out", InputGrad("X"));
grad_op->SetAttrMap(Attrs());
return std::unique_ptr<framework::OpDesc>(grad_op);
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(split_selected_rows, ops::SplitSelectedRowsOp,
ops::SplitSelectedRowsOpMaker,
ops::SplitSelectedRowsGradMaker);
REGISTER_OP_CPU_KERNEL(
split_selected_rows,
ops::SplitSelectedRowsOpKernel<paddle::platform::CPUPlace, float>);
paddle/operators/split_selected_rows_op.cu 0 → 100644
浏览文件 @ e92fe9e6
/* 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/split_selected_rows_op.h"
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(
split_selected_rows,
ops::SplitSelectedRowsOpKernel<paddle::platform::CUDADeviceContext, float>);
paddle/operators/split_selected_rows_op.h 0 → 100644
浏览文件 @ e92fe9e6
/* 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 <vector>
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
template <typename DeviceContext, typename T>
class SplitSelectedRowsOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* x = ctx.Input<framework::SelectedRows>("X");
auto outs = ctx.MultiOutput<framework::SelectedRows>("Out");
auto rows_sections = ctx.Attr<std::vector<int>>("rows_sections");
auto height_sections = ctx.Attr<std::vector<int>>("height_sections");
int64_t n = outs.size();
int offset = 0;
for (int64_t i = 0; i < n; ++i) {
framework::Vector<int64_t> out_rows;
for (int64_t j = 0; j < rows_sections[i]; ++j) {
out_rows.push_back(x->rows()[offset + j]);
}
auto& out = outs[i];
auto x_dims = x->GetCompleteDims();
x_dims[0] = rows_sections[i];
out->mutable_value()->mutable_data<T>(x_dims, ctx.GetPlace());
framework::Copy(x->value().Slice(offset, rows_sections[i] + offset),
x->place(), ctx.device_context(), out->mutable_value());
outs[i]->set_rows(out_rows);
if (height_sections.size()) {
outs[i]->set_height(height_sections[i]);
}
offset += rows_sections[i];
}
}
};
} // namespace operators
} // namespace paddle
paddle/string/to_string.h
浏览文件 @ e92fe9e6
...@@ -15,9 +15,15 @@ limitations under the License. */ ...@@ -15,9 +15,15 @@ limitations under the License. */
#pragma once #pragma once
#include <sstream> #include <sstream>
#include <string> #include <string>
#include <typeindex>
namespace paddle { namespace paddle {
namespace string { namespace string {
inline std::ostream& operator<<(std::ostream& s, const std::type_index& t) {
s << t.name();
return s;
}
template <typename T> template <typename T>
inline std::string to_string(T v) { inline std::string to_string(T v) {
std::ostringstream sout; std::ostringstream sout;
...@@ -25,6 +31,11 @@ inline std::string to_string(T v) { ...@@ -25,6 +31,11 @@ inline std::string to_string(T v) {
return sout.str(); return sout.str();
} }
template <>
inline std::string to_string(std::type_index t) {
return t.name();
}
// Faster std::string/const char* type // Faster std::string/const char* type
template <> template <>
inline std::string to_string(std::string v) { inline std::string to_string(std::string v) {
......
python/paddle/trainer_config_helpers/evaluators.py
浏览文件 @ e92fe9e6
...@@ -16,13 +16,22 @@ from paddle.trainer.config_parser import * ...@@ -16,13 +16,22 @@ from paddle.trainer.config_parser import *
from default_decorators import * from default_decorators import *
__all__ = [ __all__ = [
"evaluator_base", "classification_error_evaluator", "auc_evaluator", "evaluator_base",
"pnpair_evaluator", "precision_recall_evaluator", "ctc_error_evaluator", "classification_error_evaluator",
"chunk_evaluator", "sum_evaluator", "column_sum_evaluator", "auc_evaluator",
"value_printer_evaluator", "gradient_printer_evaluator", "pnpair_evaluator",
"maxid_printer_evaluator", "maxframe_printer_evaluator", "precision_recall_evaluator",
"seqtext_printer_evaluator", "classification_error_printer_evaluator", "ctc_error_evaluator",
"detection_map_evaluator" "chunk_evaluator",
"sum_evaluator",
"column_sum_evaluator",
"value_printer_evaluator",
"gradient_printer_evaluator",
"maxid_printer_evaluator",
"maxframe_printer_evaluator",
"seqtext_printer_evaluator",
"classification_error_printer_evaluator",
"detection_map_evaluator",
] ]
......
python/paddle/v2/fluid/distribute_transpiler.py
浏览文件 @ e92fe9e6
...@@ -407,7 +407,7 @@ class DistributeTranspiler: ...@@ -407,7 +407,7 @@ class DistributeTranspiler:
outputs=opt_op.outputs, outputs=opt_op.outputs,
attrs=opt_op.attrs) attrs=opt_op.attrs)
def get_pserver_program(self, endpoint, optimize_ops): def get_pserver_program(self, endpoint):
""" """
get pserver side program by endpoint get pserver side program by endpoint
...@@ -420,11 +420,24 @@ class DistributeTranspiler: ...@@ -420,11 +420,24 @@ class DistributeTranspiler:
pserver_program = Program() pserver_program = Program()
for v in self.param_grad_ep_mapping[endpoint]["params"]: for v in self.param_grad_ep_mapping[endpoint]["params"]:
self._clone_var(pserver_program.global_block(), v) self._clone_var(pserver_program.global_block(), v)
for v in self.param_grad_ep_mapping[endpoint]["grads"]:
# create vars for each trainer in global scope, so
# we don't need to create them when grad arrives.
pserver_program.global_block().create_var(
name=v.name, persistable=True, dtype=v.dtype, shape=v.shape)
for trainer_id in xrange(self.trainers):
print("create variable for program: %s.trainer_%d" %
(v.name, trainer_id))
pserver_program.global_block().create_var(
name="%s.trainer_%d" % (v.name, trainer_id),
persistable=True,
dtype=v.dtype,
shape=v.shape)
# step6 # step6
optimize_sub_program = Program() optimize_sub_program = Program()
for idx, opt_op in enumerate(optimize_ops): for idx, opt_op in enumerate(self.optimize_ops):
is_op_on_pserver = self._is_op_on_pserver(endpoint, optimize_ops, is_op_on_pserver = self._is_op_on_pserver(endpoint,
idx) self.optimize_ops, idx)
if not is_op_on_pserver: if not is_op_on_pserver:
continue continue
if opt_op.inputs.has_key("Grad"): if opt_op.inputs.has_key("Grad"):
...@@ -449,7 +462,7 @@ class DistributeTranspiler: ...@@ -449,7 +462,7 @@ class DistributeTranspiler:
p.name p.name
for p in self.param_grad_ep_mapping[endpoint]["grads"] for p in self.param_grad_ep_mapping[endpoint]["grads"]
], ],
"Trainers": self.trainers "Fanin": self.trainers
}) })
pserver_program.sync_with_cpp() pserver_program.sync_with_cpp()
return pserver_program return pserver_program
......
python/paddle/v2/fluid/framework.py
浏览文件 @ e92fe9e6
...@@ -116,8 +116,8 @@ def _debug_string_(proto, throw_on_error=True): ...@@ -116,8 +116,8 @@ def _debug_string_(proto, throw_on_error=True):
""" """
error_fields = list() error_fields = list()
if not proto.IsInitialized(error_fields) and throw_on_error: if not proto.IsInitialized(error_fields) and throw_on_error:
raise ValueError("{0} are not initialized\nThe message is {1}".format( raise ValueError("{0} are not initialized.\nThe message is {1}:\n".
error_fields, proto)) format(error_fields, proto))
return proto.__str__() return proto.__str__()
...@@ -374,12 +374,13 @@ class Operator(object): ...@@ -374,12 +374,13 @@ class Operator(object):
>>> outputs={"Out": [var1]}) >>> outputs={"Out": [var1]})
Args: Args:
block(Block): The block has the current operator block(Block): The block has the current operator.
desc(core.OpDesc): The protobuf description desc(core.OpDesc): The protobuf description.
type(str): The type of operator. type(str): The type of operator.
inputs(dict): The input dictionary. Key is the input parameter name. inputs(dict): The input dictionary. Key is the input parameter name.
Value is a list of variables. Value is a list of variables.
outputs(dict): The output dictionary. Has same format with inputs outputs(dict): The output dictionary which has the same format with
inputs.
attrs(dict): The attributes dictionary. Key is attribute name. Value attrs(dict): The attributes dictionary. Key is attribute name. Value
is the attribute value. The attribute type should be as same as is the attribute value. The attribute type should be as same as
the type registered in C++ the type registered in C++
...@@ -436,10 +437,11 @@ class Operator(object): ...@@ -436,10 +437,11 @@ class Operator(object):
for m in proto.outputs: for m in proto.outputs:
need.add(m.name) need.add(m.name)
if not given == need: if not given == need:
raise ValueError( raise ValueError(("Incorrect setting for output(s) of "
"Incorrect setting for output(s) of operator \"%s\". Need: [%s] Given: [%s]" "operator \"%s\". Need: [%s] Given: [%s]") %
% (type, ", ".join(str(e) for e in need), ", ".join( (type, ", ".join(str(e)
str(e) for e in given))) for e in need), ", ".join(
str(e) for e in given)))
for out_proto in proto.outputs: for out_proto in proto.outputs:
out_args = outputs[out_proto.name] out_args = outputs[out_proto.name]
...@@ -818,9 +820,8 @@ class Program(object): ...@@ -818,9 +820,8 @@ class Program(object):
if isinstance(t, Variable): if isinstance(t, Variable):
t = t.op t = t.op
else: else:
raise ValueError( raise ValueError(("All targets of prune() can only be "
"All targets of prune() can only be Variable or Operator." "Variable or Operator."))
)
targets_idx.append([t.block.idx, t.idx]) targets_idx.append([t.block.idx, t.idx])
res = Program() res = Program()
......
python/paddle/v2/fluid/io.py
浏览文件 @ e92fe9e6
...@@ -15,6 +15,7 @@ import os ...@@ -15,6 +15,7 @@ import os
import cPickle as pickle import cPickle as pickle
from paddle.v2.fluid.framework import Program, Parameter, default_main_program, Variable from paddle.v2.fluid.framework import Program, Parameter, default_main_program, Variable
from . import core
__all__ = [ __all__ = [
'save_vars', 'save_vars',
...@@ -191,6 +192,33 @@ def get_inference_program(target_vars, main_program=None): ...@@ -191,6 +192,33 @@ def get_inference_program(target_vars, main_program=None):
return inference_program return inference_program
def prepend_feed_ops(inference_program, feeded_var_names):
global_block = inference_program.global_block()
feed_var = global_block.create_var(
name='feed', type=core.VarDesc.VarType.FEED_MINIBATCH, persistable=True)
for i, name in enumerate(feeded_var_names):
out = global_block.var(name)
global_block.prepend_op(
type='feed',
inputs={'X': [feed_var]},
outputs={'Out': [out]},
attrs={'col': i})
def append_fetch_ops(inference_program, fetch_var_names):
global_block = inference_program.global_block()
fetch_var = global_block.create_var(
name='fetch', type=core.VarDesc.VarType.FETCH_LIST, persistable=True)
for i, name in enumerate(fetch_var_names):
global_block.append_op(
type='fetch',
inputs={'X': [name]},
outputs={'Out': [fetch_var]},
attrs={'col': i})
def save_inference_model(dirname, def save_inference_model(dirname,
feeded_var_names, feeded_var_names,
target_vars, target_vars,
...@@ -241,6 +269,9 @@ def save_inference_model(dirname, ...@@ -241,6 +269,9 @@ def save_inference_model(dirname,
"fetch_var_names": fetch_var_names "fetch_var_names": fetch_var_names
}, f, -1) }, f, -1)
prepend_feed_ops(inference_program, feeded_var_names)
append_fetch_ops(inference_program, fetch_var_names)
# Save only programDesc of inference_program in binary format # Save only programDesc of inference_program in binary format
# in another file: __model__.dat # in another file: __model__.dat
with open(model_file_name + ".dat", "wb") as fp: with open(model_file_name + ".dat", "wb") as fp:
......
python/paddle/v2/fluid/layers/io.py
浏览文件 @ e92fe9e6
...@@ -28,9 +28,9 @@ def data(name, ...@@ -28,9 +28,9 @@ def data(name,
**Data Layer** **Data Layer**
This function takes in the input and based on whether data has This function takes in the input and based on whether data has
to be returned back as a minibatch, it creates the global variable using to be returned back as a minibatch, it creates the global variable by using
the helper functions. The global variables can be accessed by all the the helper functions. The global variables can be accessed by all the
following operations and layers in the graph. following operators in the graph.
All the input variables of this function are passed in as local variables All the input variables of this function are passed in as local variables
to the LayerHelper constructor. to the LayerHelper constructor.
......
python/paddle/v2/fluid/layers/nn.py
浏览文件 @ e92fe9e6
...@@ -22,35 +22,13 @@ from ..param_attr import ParamAttr ...@@ -22,35 +22,13 @@ from ..param_attr import ParamAttr
from tensor import concat from tensor import concat
__all__ = [ __all__ = [
'fc', 'fc', 'embedding', 'dynamic_lstm', 'gru_unit', 'linear_chain_crf',
'embedding', 'crf_decoding', 'cos_sim', 'cross_entropy', 'square_error_cost', 'accuracy',
'dynamic_lstm', 'chunk_eval', 'sequence_conv', 'conv2d', 'sequence_pool', 'pool2d',
'gru_unit', 'batch_norm', 'beam_search_decode', 'conv2d_transpose', 'sequence_expand',
'linear_chain_crf', 'lstm_unit', 'reduce_sum', 'reduce_mean', 'reduce_max', 'reduce_min',
'crf_decoding', 'sequence_first_step', 'sequence_last_step', 'dropout', 'split',
'cos_sim', 'l2_normalize', 'matmul', 'warpctc', 'sequence_reshape'
'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',
'matmul',
] ]
...@@ -228,6 +206,102 @@ def dynamic_lstm(input, ...@@ -228,6 +206,102 @@ def dynamic_lstm(input,
cell_activation='tanh', cell_activation='tanh',
candidate_activation='tanh', candidate_activation='tanh',
dtype='float32'): dtype='float32'):
"""
**Dynamic LSTM Layer**
The defalut implementation is diagonal/peephole connection
(https://arxiv.org/pdf/1402.1128.pdf), the formula is as follows:
.. math::
i_t & = \sigma(W_{ix}x_{t} + W_{ih}h_{t-1} + W_{ic}c_{t-1} + b_i)
f_t & = \sigma(W_{fx}x_{t} + W_{fh}h_{t-1} + W_{fc}c_{t-1} + b_f)
\\tilde{c_t} & = act_g(W_{cx}x_t + W_{ch}h_{t-1} + b_c)
o_t & = \sigma(W_{ox}x_{t} + W_{oh}h_{t-1} + W_{oc}c_t + b_o)
c_t & = f_t \odot c_{t-1} + i_t \odot \\tilde{c_t}
h_t & = o_t \odot act_h(c_t)
where the :math:`W` terms denote weight matrices (e.g. :math:`W_{xi}` is
the matrix of weights from the input gate to the input), :math:`W_{ic}, \
W_{fc}, W_{oc}` are diagonal weight matrices for peephole connections. In
our implementation, we use vectors to reprenset these diagonal weight
matrices. The :math:`b` terms denote bias vectors (:math:`b_i` is the input
gate bias vector), :math:`\sigma` is the non-line activations, such as
logistic sigmoid function, and :math:`i, f, o` and :math:`c` are the input
gate, forget gate, output gate, and cell activation vectors, respectively,
all of which have the same size as the cell output activation vector :math:`h`.
The :math:`\odot` is the element-wise product of the vectors. :math:`act_g`
and :math:`act_h` are the cell input and cell output activation functions
and `tanh` is usually used for them. :math:`\\tilde{c_t}` is also called
candidate hidden state, which is computed based on the current input and
the previous hidden state.
Set `use_peepholes` to `False` to disable peephole connection. The formula
is omitted here, please refer to the paper
http://www.bioinf.jku.at/publications/older/2604.pdf for details.
Note that these :math:`W_{xi}x_{t}, W_{xf}x_{t}, W_{xc}x_{t}, W_{xo}x_{t}`
operations on the input :math:`x_{t}` are NOT included in this operator.
Users can choose to use fully-connect layer before LSTM layer.
Args:
input(Variable): The input of dynamic_lstm layer, which supports
variable-time length input sequence. The underlying
tensor in this Variable is a matrix with shape
(T X 4D), where T is the total time steps in this
mini-batch, D is the hidden size.
size(int): 4 * hidden size.
param_attr(ParamAttr): The parameter attribute for the learnable
hidden-hidden weights.
- The shape is (D x 4D), where D is the hidden
size.
- Weights = {:math:`W_{ch}, W_{ih}, \
W_{fh}, W_{oh}`}
bias_attr(ParamAttr): The bias attribute for the learnable bias
weights, which contains two parts, input-hidden
bias weights and peephole connections weights if
setting `use_peepholes` to `True`.
1. `use_peepholes = False`
- The shape is (1 x 4D).
- Biases = {:math:`b_c, b_i, b_f, b_o`}.
2. `use_peepholes = True`
- The shape is (1 x 7D).
- Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
W_{fc}, W_{oc}`}.
use_peepholes(bool): Whether to enable diagonal/peephole connections,
default `True`.
is_reverse(bool): Whether to compute reversed LSTM, default `False`.
gate_activation(str): The activation for input gate, forget gate and
output gate. Choices = ["sigmoid", "tanh", "relu",
"identity"], default "sigmoid".
cell_activation(str): The activation for cell output. Choices = ["sigmoid",
"tanh", "relu", "identity"], default "tanh".
candidate_activation(str): The activation for candidate hidden state.
Choices = ["sigmoid", "tanh", "relu", "identity"],
default "tanh".
dtype(str): Data type. Choices = ["float32", "float64"], default "float32".
Returns:
tuple: The hidden state, and cell state of LSTM. The shape of both \
is (T x D), and lod is the same with the `input`.
Examples:
.. code-block:: python
hidden_dim = 512
forward_proj = fluid.layers.fc(input=input_seq, size=hidden_dim * 4,
act=None, bias_attr=None)
forward, _ = fluid.layers.dynamic_lstm(
input=forward_proj, size=hidden_dim * 4, use_peepholes=False)
"""
helper = LayerHelper('lstm', **locals()) helper = LayerHelper('lstm', **locals())
size = size / 4 size = size / 4
weight = helper.create_parameter( weight = helper.create_parameter(
...@@ -675,6 +749,7 @@ def conv2d(input, ...@@ -675,6 +749,7 @@ def conv2d(input,
groups=None, groups=None,
param_attr=None, param_attr=None,
bias_attr=None, bias_attr=None,
use_cudnn=True,
act=None): act=None):
""" """
**Convlution2D Layer** **Convlution2D Layer**
...@@ -738,6 +813,8 @@ def conv2d(input, ...@@ -738,6 +813,8 @@ def conv2d(input,
connected to the second half of the input channels. Default: groups=1 connected to the second half of the input channels. Default: groups=1
param_attr(ParamAttr): The parameters to the Conv2d Layer. Default: None param_attr(ParamAttr): The parameters to the Conv2d Layer. Default: None
bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None
use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
library is installed. Default: True
act(str): Activation type. Default: None act(str): Activation type. Default: None
Returns: Returns:
...@@ -773,6 +850,8 @@ def conv2d(input, ...@@ -773,6 +850,8 @@ def conv2d(input,
stride = [stride, stride] stride = [stride, stride]
if isinstance(padding, int): if isinstance(padding, int):
padding = [padding, padding] padding = [padding, padding]
if not isinstance(use_cudnn, bool):
raise ValueError("use_cudnn should be True or False")
input_shape = input.shape input_shape = input.shape
filter_shape = [num_filters, num_filter_channels] + filter_size filter_shape = [num_filters, num_filter_channels] + filter_size
...@@ -796,9 +875,12 @@ def conv2d(input, ...@@ -796,9 +875,12 @@ def conv2d(input,
'Filter': filter_param, 'Filter': filter_param,
}, },
outputs={"Output": pre_bias}, outputs={"Output": pre_bias},
attrs={'strides': stride, attrs={
'paddings': padding, 'strides': stride,
'groups': groups}) 'paddings': padding,
'groups': groups,
'use_cudnn': use_cudnn
})
pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2) pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
...@@ -946,7 +1028,9 @@ def pool2d(input, ...@@ -946,7 +1028,9 @@ def pool2d(input,
pool_type, pool_type,
pool_stride=None, pool_stride=None,
pool_padding=None, pool_padding=None,
global_pooling=False): global_pooling=False,
use_cudnn=True,
name=None):
""" """
This function adds the operator for pooling in 2 dimensions, using the This function adds the operator for pooling in 2 dimensions, using the
pooling configurations mentioned in input parameters. pooling configurations mentioned in input parameters.
...@@ -965,6 +1049,8 @@ def pool2d(input, ...@@ -965,6 +1049,8 @@ def pool2d(input,
pool_stride = [pool_stride, pool_stride] pool_stride = [pool_stride, pool_stride]
if isinstance(pool_padding, int): if isinstance(pool_padding, int):
pool_padding = [pool_padding, pool_padding] pool_padding = [pool_padding, pool_padding]
if not isinstance(use_cudnn, bool):
raise ValueError("use_cudnn should be True or False")
helper = LayerHelper('pool2d', **locals()) helper = LayerHelper('pool2d', **locals())
dtype = helper.input_dtype() dtype = helper.input_dtype()
...@@ -979,7 +1065,8 @@ def pool2d(input, ...@@ -979,7 +1065,8 @@ def pool2d(input,
"ksize": pool_size, "ksize": pool_size,
"global_pooling": global_pooling, "global_pooling": global_pooling,
"strides": pool_stride, "strides": pool_stride,
"paddings": pool_padding "paddings": pool_padding,
"use_cudnn": use_cudnn
}) })
return pool_out return pool_out
...@@ -992,7 +1079,8 @@ def batch_norm(input, ...@@ -992,7 +1079,8 @@ def batch_norm(input,
epsilon=1e-05, epsilon=1e-05,
param_attr=None, param_attr=None,
bias_attr=None, bias_attr=None,
data_layout='NCHW'): data_layout='NCHW',
name=None):
""" """
This function helps create an operator to implement This function helps create an operator to implement
the BatchNorm layer using the configurations from the input parameters. the BatchNorm layer using the configurations from the input parameters.
...@@ -1068,7 +1156,7 @@ def batch_norm(input, ...@@ -1068,7 +1156,7 @@ def batch_norm(input,
return helper.append_activation(batch_norm_out) return helper.append_activation(batch_norm_out)
def beam_search_decode(ids, scores): def beam_search_decode(ids, scores, name=None):
helper = LayerHelper('beam_search_decode', **locals()) helper = LayerHelper('beam_search_decode', **locals())
sentence_ids = helper.create_tmp_variable(dtype=ids.dtype) sentence_ids = helper.create_tmp_variable(dtype=ids.dtype)
sentence_scores = helper.create_tmp_variable(dtype=ids.dtype) sentence_scores = helper.create_tmp_variable(dtype=ids.dtype)
...@@ -1092,7 +1180,9 @@ def conv2d_transpose(input, ...@@ -1092,7 +1180,9 @@ def conv2d_transpose(input,
padding=None, padding=None,
stride=None, stride=None,
dilation=None, dilation=None,
param_attr=None): param_attr=None,
use_cudnn=True,
name=None):
""" """
The transpose of conv2d layer. The transpose of conv2d layer.
...@@ -1119,8 +1209,10 @@ def conv2d_transpose(input, ...@@ -1119,8 +1209,10 @@ def conv2d_transpose(input,
contain two integers, (dilation_H, dilation_W). Otherwise, the contain two integers, (dilation_H, dilation_W). Otherwise, the
dilation_H = dilation_W = dilation. dilation_H = dilation_W = dilation.
param_attr: Parameter Attribute. param_attr: Parameter Attribute.
main_program(Program): the main program use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
startup_program(Program): the startup program library is installed. Default: True
name(str|None): A name for this layer(optional). If set None, the layer
will be named automatically.
Returns: Returns:
Variable: Output image. Variable: Output image.
...@@ -1147,6 +1239,10 @@ def conv2d_transpose(input, ...@@ -1147,6 +1239,10 @@ def conv2d_transpose(input,
elif dilation is not None: elif dilation is not None:
op_attr['dilations'] = dilation op_attr['dilations'] = dilation
if not isinstance(use_cudnn, bool):
raise ValueError("use_cudnn should be True or False")
op_attr['use_cudnn'] = use_cudnn
if filter_size is None: if filter_size is None:
if output_size is None: if output_size is None:
raise ValueError("output_size must be set when filter_size is None") raise ValueError("output_size must be set when filter_size is None")
...@@ -1184,7 +1280,7 @@ def conv2d_transpose(input, ...@@ -1184,7 +1280,7 @@ def conv2d_transpose(input,
return out return out
def sequence_expand(x, y): def sequence_expand(x, y, name=None):
"""Sequence Expand Layer. This layer will expand the input variable **x** """Sequence Expand Layer. This layer will expand the input variable **x**
according to LoD information of **y**. And the following examples will according to LoD information of **y**. And the following examples will
explain how sequence_expand works: explain how sequence_expand works:
...@@ -1228,6 +1324,8 @@ def sequence_expand(x, y): ...@@ -1228,6 +1324,8 @@ def sequence_expand(x, y):
Args: Args:
x (Variable): The input variable which is a Tensor or LoDTensor. x (Variable): The input variable which is a Tensor or LoDTensor.
y (Variable): The input variable which is a LoDTensor. y (Variable): The input variable which is a LoDTensor.
name(str|None): A name for this layer(optional). If set None, the layer
will be named automatically.
Returns: Returns:
Variable: The expanded variable which is a LoDTensor. Variable: The expanded variable which is a LoDTensor.
...@@ -1254,7 +1352,8 @@ def lstm_unit(x_t, ...@@ -1254,7 +1352,8 @@ def lstm_unit(x_t,
cell_t_prev, cell_t_prev,
forget_bias=0.0, forget_bias=0.0,
param_attr=None, param_attr=None,
bias_attr=None): bias_attr=None,
name=None):
"""Lstm unit layer. The equation of a lstm step is: """Lstm unit layer. The equation of a lstm step is:
.. math:: .. math::
...@@ -1301,6 +1400,8 @@ def lstm_unit(x_t, ...@@ -1301,6 +1400,8 @@ def lstm_unit(x_t,
initializer, name etc. initializer, name etc.
bias_attr (ParamAttr): The attributes of bias weights, if not False, bias_attr (ParamAttr): The attributes of bias weights, if not False,
bias weights will be created and be set to default value. bias weights will be created and be set to default value.
name(str|None): A name for this layer(optional). If set None, the layer
will be named automatically.
Returns: Returns:
tuple: The hidden value and cell value of lstm unit. tuple: The hidden value and cell value of lstm unit.
...@@ -1366,7 +1467,7 @@ def lstm_unit(x_t, ...@@ -1366,7 +1467,7 @@ def lstm_unit(x_t,
return h, c return h, c
def reduce_sum(input, dim=None, keep_dim=False): def reduce_sum(input, dim=None, keep_dim=False, name=None):
""" """
Computes the sum of tensor elements over the given dimension. Computes the sum of tensor elements over the given dimension.
...@@ -1380,6 +1481,8 @@ def reduce_sum(input, dim=None, keep_dim=False): ...@@ -1380,6 +1481,8 @@ def reduce_sum(input, dim=None, keep_dim=False):
keep_dim (bool): Whether to reserve the reduced dimension in the keep_dim (bool): Whether to reserve the reduced dimension in the
output Tensor. The result tensor will have one fewer dimension output Tensor. The result tensor will have one fewer dimension
than the :attr:`input` unless :attr:`keep_dim` is true. than the :attr:`input` unless :attr:`keep_dim` is true.
name(str|None): A name for this layer(optional). If set None, the layer
will be named automatically.
Returns: Returns:
Variable: The reduced Tensor variable. Variable: The reduced Tensor variable.
...@@ -1410,7 +1513,7 @@ def reduce_sum(input, dim=None, keep_dim=False): ...@@ -1410,7 +1513,7 @@ def reduce_sum(input, dim=None, keep_dim=False):
return out return out
def reduce_mean(input, dim=None, keep_dim=False): def reduce_mean(input, dim=None, keep_dim=False, name=None):
""" """
Computes the mean of tensor elements over the given dimension. Computes the mean of tensor elements over the given dimension.
...@@ -1424,6 +1527,8 @@ def reduce_mean(input, dim=None, keep_dim=False): ...@@ -1424,6 +1527,8 @@ def reduce_mean(input, dim=None, keep_dim=False):
keep_dim (bool): Whether to reserve the reduced dimension in the keep_dim (bool): Whether to reserve the reduced dimension in the
output Tensor. The result tensor will have one fewer dimension output Tensor. The result tensor will have one fewer dimension
than the :attr:`input` unless :attr:`keep_dim` is true. than the :attr:`input` unless :attr:`keep_dim` is true.
name(str|None): A name for this layer(optional). If set None, the layer
will be named automatically.
Returns: Returns:
Variable: The reduced Tensor variable. Variable: The reduced Tensor variable.
...@@ -1454,7 +1559,7 @@ def reduce_mean(input, dim=None, keep_dim=False): ...@@ -1454,7 +1559,7 @@ def reduce_mean(input, dim=None, keep_dim=False):
return out return out
def reduce_max(input, dim=None, keep_dim=False): def reduce_max(input, dim=None, keep_dim=False, name=None):
""" """
Computes the maximum of tensor elements over the given dimension. Computes the maximum of tensor elements over the given dimension.
...@@ -1468,6 +1573,8 @@ def reduce_max(input, dim=None, keep_dim=False): ...@@ -1468,6 +1573,8 @@ def reduce_max(input, dim=None, keep_dim=False):
keep_dim (bool): Whether to reserve the reduced dimension in the keep_dim (bool): Whether to reserve the reduced dimension in the
output Tensor. The result tensor will have one fewer dimension output Tensor. The result tensor will have one fewer dimension
than the :attr:`input` unless :attr:`keep_dim` is true. than the :attr:`input` unless :attr:`keep_dim` is true.
name(str|None): A name for this layer(optional). If set None, the layer
will be named automatically.
Returns: Returns:
Variable: The reduced Tensor variable. Variable: The reduced Tensor variable.
...@@ -1498,7 +1605,7 @@ def reduce_max(input, dim=None, keep_dim=False): ...@@ -1498,7 +1605,7 @@ def reduce_max(input, dim=None, keep_dim=False):
return out return out
def reduce_min(input, dim=None, keep_dim=False): def reduce_min(input, dim=None, keep_dim=False, name=None):
""" """
Computes the minimum of tensor elements over the given dimension. Computes the minimum of tensor elements over the given dimension.
...@@ -1512,6 +1619,8 @@ def reduce_min(input, dim=None, keep_dim=False): ...@@ -1512,6 +1619,8 @@ def reduce_min(input, dim=None, keep_dim=False):
keep_dim (bool): Whether to reserve the reduced dimension in the keep_dim (bool): Whether to reserve the reduced dimension in the
output Tensor. The result tensor will have one fewer dimension output Tensor. The result tensor will have one fewer dimension
than the :attr:`input` unless :attr:`keep_dim` is true. than the :attr:`input` unless :attr:`keep_dim` is true.
name(str|None): A name for this layer(optional). If set None, the layer
will be named automatically.
Returns: Returns:
Variable: The reduced Tensor variable. Variable: The reduced Tensor variable.
...@@ -1542,20 +1651,22 @@ def reduce_min(input, dim=None, keep_dim=False): ...@@ -1542,20 +1651,22 @@ def reduce_min(input, dim=None, keep_dim=False):
return out return out
def split(input, num_or_sections, dim=-1): def split(input, num_or_sections, dim=-1, name=None):
""" """
Splits the tensor into multiple sub-tensors. Split the input tensor into multiple sub-tensors.
Args: Args:
input (Variable): The input variable which is a Tensor or LoDTensor. input (Variable): The input variable which is a Tensor or LoDTensor.
num_or_sections (int|list): If :attr:`num_or_sections` is an integer, num_or_sections (int|list): If :attr:`num_or_sections` is an integer,
then the integer indicates the number of equal sized sub-tensors then the integer indicates the number of equal sized sub-tensors
that the tensor will be divided into. If :attr:`num_or_sections` that the tensor will be divided into. If :attr:`num_or_sections`
is a list of integers, the length of list indicates the number of is a list of integers, the length of list indicates the number of
sub-tensors and the integers indicate the sizes of sub-tensors' sub-tensors and the integers indicate the sizes of sub-tensors'
:attr:`dim` dimension orderly. :attr:`dim` dimension orderly.
dim (int): The dimension along which to split. If :math:`dim < 0`, the dim (int): The dimension along which to split. If :math:`dim < 0`, the
dimension to split along is :math:`rank(input) + dim`. dimension to split along is :math:`rank(input) + dim`.
name(str|None): A name for this layer(optional). If set None, the layer
will be named automatically.
Returns: Returns:
List: The list of segmented tensor variables. List: The list of segmented tensor variables.
...@@ -1600,31 +1711,112 @@ def split(input, num_or_sections, dim=-1): ...@@ -1600,31 +1711,112 @@ def split(input, num_or_sections, dim=-1):
return outs return outs
def l2_normalize(x, axis, epsilon=1e-12, name=None):
"""
**L2 normalize Layer**
The l2 normalize layer normalizes `x` along dimension `axis` using an L2
norm. For a 1-D tensor (`dim` is fixed to 0), this layer computes
output = x / sqrt(max(sum(x**2), epsilon))
For `x` with more dimensions, this layer independently normalizes each 1-D
slice along dimension `axis`.
Args:
x(Variable|list): The input tensor to l2_normalize layer.
axis(int): Dimension along which to normalize the input.
epsilon(float): A lower bound value for `x`'s l2 norm. sqrt(epsilon) will
be used as the divisor if the l2 norm of `x` is less than
sqrt(epsilon).
name(str|None): A name for this layer(optional). If set None, the layer
will be named automatically.
Returns:
Variable: The output tensor variable.
Examples:
.. code-block:: python
data = fluid.layers.data(name="data",
shape=(3, 17, 13),
dtype="float32")
fc = fluid.layers.l2_normalize(x=data, axis=1)
"""
if len(x.shape) == 1: axis = 0
helper = LayerHelper("l2_normalize", **locals())
square = helper.create_tmp_variable(dtype=x.dtype)
helper.append_op(type="square", inputs={"X": x}, outputs={"Out": square})
reduced_sum = helper.create_tmp_variable(dtype=x.dtype)
helper.append_op(
type="reduce_sum",
inputs={"X": square},
outputs={"Out": reduced_sum},
attrs={
"dim": 1 if axis is None else axis,
"keep_dim": True,
"reduce_all": False
})
# TODO(caoying) A lower bound value epsilon for the norm is needed to
# imporve the numeric stability of reciprocal. This requires a maximum_op.
rsquare = helper.create_tmp_variable(dtype=x.dtype)
helper.append_op(
type="reciprocal", inputs={"X": reduced_sum}, outputs={"Out": rsquare})
# TODO(caoying) the current elementwise_mul operator does not support a
# general broadcast rule which broadcasts input(Y) to have the same
# dimension with Input(X) starting from a specified dimension. So this
# exanpsion is requred. Once a general broadcast rule is spported, this
# expanding canbe removed.
rsquare_expanded = helper.create_tmp_variable(dtype=x.dtype)
expand_times = [1] * len(x.shape)
expand_times[axis] = int(x.shape[axis])
helper.append_op(
type="expand",
inputs={"X": rsquare},
outputs={"Out": rsquare_expanded},
attrs={"expand_times": expand_times})
out = helper.create_tmp_variable(dtype=x.dtype)
helper.append_op(
type="elementwise_mul",
inputs={"X": x,
"Y": rsquare_expanded},
outputs={"Out": out})
return out
def matmul(x, y, transpose_x=False, transpose_y=False, name=None): def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
""" """
Applies matrix multipication to two tensors. Currently only rank 1 to rank Applies matrix multipication to two tensors. Currently only rank 1 to rank
3 input tensors are supported. 3 input tensors are supported.
The actual behavior depends on the shapes of :math:`x`, :math:`y` and the The actual behavior depends on the shapes of :math:`x`, :math:`y` and the
flag values of :attr:`transpose_x`, :attr:`transpose_y`. Specifically: flag values of :attr:`transpose_x`, :attr:`transpose_y`. Specifically:
- If a transpose flag is specified, the last two dimensions of the tensor - If a transpose flag is specified, the last two dimensions of the tensor
are transposed. If the tensor is rank-1 of shape :math:`[D]`, then for are transposed. If the tensor is rank-1 of shape :math:`[D]`, then for
:math:`x` it is treated as :math:`[1, D]` in nontransposed form and as :math:`x` it is treated as :math:`[1, D]` in nontransposed form and as
:math:`[D, 1]` in transposed form, whereas for :math:`y` it is the :math:`[D, 1]` in transposed form, whereas for :math:`y` it is the
opposite: It is treated as :math:`[D, 1]` in nontransposed form and as opposite: It is treated as :math:`[D, 1]` in nontransposed form and as
:math:`[1, D]` in transposed form. :math:`[1, D]` in transposed form.
- After transpose, the two tensors are 2-D or 3-D and matrix multipication - After transpose, the two tensors are 2-D or 3-D and matrix multipication
performs in the following way. performs in the following way.
- If both are 2-D, they are multiplied like conventional matrices. - If both are 2-D, they are multiplied like conventional matrices.
- If either is 3-D, it is treated as a stack of matrices residing in the - If either is 3-D, it is treated as a stack of matrices residing in the
last two dimensions and a batched matrix multiply supporting broadcast last two dimensions and a batched matrix multiply supporting broadcast
applies on the two tensors. applies on the two tensors.
Also note that if the raw tensor :math:`x` or :math:`y` is rank-1 and Also note that if the raw tensor :math:`x` or :math:`y` is rank-1 and
nontransposed, the prepended or appended dimension :math:`1` will be nontransposed, the prepended or appended dimension :math:`1` will be
removed after matrix multipication. removed after matrix multipication.
Args: Args:
...@@ -1632,7 +1824,7 @@ def matmul(x, y, transpose_x=False, transpose_y=False, name=None): ...@@ -1632,7 +1824,7 @@ def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
y (Variable): The input variable which is a Tensor or LoDTensor. y (Variable): The input variable which is a Tensor or LoDTensor.
transpose_x (bool): Whether to transpose :math:`x` before multiplication. transpose_x (bool): Whether to transpose :math:`x` before multiplication.
transpose_y (bool): Whether to transpose :math:`y` before multiplication. transpose_y (bool): Whether to transpose :math:`y` before multiplication.
name(str|None): A name for this layer(optional). If set None, the layer name(str|None): A name for this layer(optional). If set None, the layer
will be named automatically. will be named automatically.
Returns: Returns:
...@@ -1653,6 +1845,7 @@ def matmul(x, y, transpose_x=False, transpose_y=False, name=None): ...@@ -1653,6 +1845,7 @@ def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
# x: [K], y: [K] # x: [K], y: [K]
fluid.layers.matmul(x, y) # out: [1] fluid.layers.matmul(x, y) # out: [1]
# x: [M], y: [N] # x: [M], y: [N]
fluid.layers.matmul(x, y, True, True) # out: [M, N] fluid.layers.matmul(x, y, True, True) # out: [M, N]
""" """
helper = LayerHelper('matmul', **locals()) helper = LayerHelper('matmul', **locals())
...@@ -1668,3 +1861,110 @@ def matmul(x, y, transpose_x=False, transpose_y=False, name=None): ...@@ -1668,3 +1861,110 @@ def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
attrs={'transpose_X': transpose_x, attrs={'transpose_X': transpose_x,
'transpose_Y': transpose_y}) 'transpose_Y': transpose_y})
return out return out
def warpctc(input, label, blank=0, norm_by_times=False, **kwargs):
"""
An operator integrating the open source Warp-CTC library
(https://github.com/baidu-research/warp-ctc)
to compute Connectionist Temporal Classification (CTC) loss.
It can be aliased as softmax with CTC, since a native softmax activation is
interated to the Warp-CTC library, to to normlize values for each row of the
input tensor.
Args:
input(Variable): (LodTensor, default: LoDTensor<float>),
the unscaled 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).
label(Variable): (LodTensor, default: LoDTensor<int>), the ground truth
of variable-length sequence, which is a 2-D Tensor with LoD
information. It is of the shape [Lg, 1], where Lg is th sum of
all labels' length.
blank: (int, default: 0), the blank label index of Connectionist
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
sequence's length. There is no need to normalize the gradients
if warpctc layer was follewed by a mean_op.
Returns:
Variable: The Connectionist Temporal Classification (CTC) loss,
which is a 2-D Tensor of the shape [batch_size, 1].
Examples:
.. code-block:: python
y = layers.data(name='y', shape=[11, 8], dtype='float32', lod_level=1)
y_predict = layers.data(name='y_predict', shape=[11, 1], dtype='float32')
cost = layers.warpctc(input=y_predict, label=y)
"""
helper = LayerHelper('warpctc', **kwargs)
loss_out = helper.create_tmp_variable(dtype=input.dtype)
grad_out = helper.create_tmp_variable(dtype=input.dtype)
helper.append_op(
type='warpctc',
inputs={'Logits': [input],
'Label': [label]},
outputs={'WarpCTCGrad': [grad_out],
'Loss': [loss_out]},
attrs={'blank': blank,
'norm_by_times': norm_by_times})
return loss_out
def sequence_reshape(input, new_dim):
"""
**Sequence Reshape Layer**
This layer will rearrange the input sequences. The new dimension is set by
user. Length of each sequence is computed according to original length,
original dimension and new dimension. The following example will help to
illustrate the function of this layer:
.. code-block:: text
x is a LoDTensor:
x.lod = [[0, 2, 6]]
x.data = [[1, 2], [3, 4],
[5, 6], [7, 8], [9, 10], [11, 12]]
x.dims = [6, 2]
set new_dim = 4
then out is a LoDTensor:
out.lod = [[0, 1, 3]]
out.data = [[1, 2, 3, 4],
[5, 6, 7, 8], [9, 10, 11, 12]]
out.dims = [3, 4]
Currently, only 1-level LoDTensor is supported and please make sure
(original length * original dimension) can be divided by new dimension with
no remainder for each sequence.
Args:
input (Variable): (LodTensor, default: LoDTensor<float>), a 2-D LoDTensor
with shape being [N, M] where M for dimension.
new_dim (int): New dimension which the input LoDTensor is reshaped to.
Returns:
Variable: Reshaped LoDTensor according to new dimension.
Examples:
.. code-block:: python
x = fluid.layers.data(name='x', shape=[5, 20],
dtype='float32', lod_level=1)
x_reshaped = layers.sequence_reshape(input=x, new_dim=10)
"""
helper = LayerHelper('sequence_reshape', **locals())
out = helper.create_tmp_variable(helper.input_dtype())
helper.append_op(
type='sequence_reshape',
inputs={'X': [input]},
outputs={'Out': [out]},
attrs={'new_dim': new_dim})
return out
python/paddle/v2/fluid/layers/ops.py
浏览文件 @ e92fe9e6
...@@ -55,6 +55,8 @@ __all__ = [ ...@@ -55,6 +55,8 @@ __all__ = [
'elementwise_div', 'elementwise_div',
'elementwise_sub', 'elementwise_sub',
'elementwise_mul', 'elementwise_mul',
'elementwise_max',
'elementwise_min',
'clip', 'clip',
'sequence_softmax', 'sequence_softmax',
] + __activations__ ] + __activations__
......
python/paddle/v2/fluid/nets.py
浏览文件 @ e92fe9e6
...@@ -28,19 +28,22 @@ def simple_img_conv_pool(input, ...@@ -28,19 +28,22 @@ def simple_img_conv_pool(input,
pool_stride, pool_stride,
act, act,
param_attr=None, param_attr=None,
pool_type='max'): pool_type='max',
use_cudnn=True):
conv_out = layers.conv2d( conv_out = layers.conv2d(
input=input, input=input,
num_filters=num_filters, num_filters=num_filters,
filter_size=filter_size, filter_size=filter_size,
param_attr=param_attr, param_attr=param_attr,
act=act) act=act,
use_cudnn=use_cudnn)
pool_out = layers.pool2d( pool_out = layers.pool2d(
input=conv_out, input=conv_out,
pool_size=pool_size, pool_size=pool_size,
pool_type=pool_type, pool_type=pool_type,
pool_stride=pool_stride) pool_stride=pool_stride,
use_cudnn=use_cudnn)
return pool_out return pool_out
...@@ -54,7 +57,8 @@ def img_conv_group(input, ...@@ -54,7 +57,8 @@ def img_conv_group(input,
conv_with_batchnorm=False, conv_with_batchnorm=False,
conv_batchnorm_drop_rate=None, conv_batchnorm_drop_rate=None,
pool_stride=1, pool_stride=1,
pool_type=None): pool_type=None,
use_cudnn=True):
""" """
Image Convolution Group, Used for vgg net. Image Convolution Group, Used for vgg net.
""" """
...@@ -85,7 +89,8 @@ def img_conv_group(input, ...@@ -85,7 +89,8 @@ def img_conv_group(input,
filter_size=conv_filter_size[i], filter_size=conv_filter_size[i],
padding=conv_padding[i], padding=conv_padding[i],
param_attr=param_attr[i], param_attr=param_attr[i],
act=local_conv_act) act=local_conv_act,
use_cudnn=use_cudnn)
if conv_with_batchnorm[i]: if conv_with_batchnorm[i]:
tmp = layers.batch_norm(input=tmp, act=conv_act) tmp = layers.batch_norm(input=tmp, act=conv_act)
...@@ -97,7 +102,8 @@ def img_conv_group(input, ...@@ -97,7 +102,8 @@ def img_conv_group(input,
input=tmp, input=tmp,
pool_size=pool_size, pool_size=pool_size,
pool_type=pool_type, pool_type=pool_type,
pool_stride=pool_stride) pool_stride=pool_stride,
use_cudnn=use_cudnn)
return pool_out return pool_out
......
python/paddle/v2/fluid/registry.py
浏览文件 @ e92fe9e6
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. # Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
# #
#Licensed under the Apache License, Version 2.0 (the "License"); # Licensed under the Apache License, Version 2.0 (the "License");
#you may not use this file except in compliance with the License. # you may not use this file except in compliance with the License.
#You may obtain a copy of the License at # You may obtain a copy of the License at
# #
# http://www.apache.org/licenses/LICENSE-2.0 # http://www.apache.org/licenses/LICENSE-2.0
# #
#Unless required by applicable law or agreed to in writing, software # Unless required by applicable law or agreed to in writing, software
#distributed under the License is distributed on an "AS IS" BASIS, # distributed under the License is distributed on an "AS IS" BASIS,
#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
#limitations under the License. # limitations under the License.
import re import re
import cStringIO import cStringIO
import warnings import warnings
...@@ -167,13 +167,18 @@ def register_layer(op_type): ...@@ -167,13 +167,18 @@ def register_layer(op_type):
inputs[ipt.name] = val inputs[ipt.name] = val
outputs = dict() outputs = dict()
out = helper.create_tmp_variable(dtype=dtype) out = kwargs.pop(_convert_(o_name), [])
outputs[o_name] = [out] if out:
out_var = out[0] if (isinstance(out, list) or
isinstance(out, tuple)) else out
else:
out_var = helper.create_tmp_variable(dtype=dtype)
outputs[o_name] = [out_var]
for name in intermediate_output_names: for name in intermediate_output_names:
outputs[name] = [helper.create_tmp_variable(dtype=dtype)] outputs[name] = [helper.create_tmp_variable(dtype=dtype)]
helper.append_op( helper.append_op(
type=op_type, inputs=inputs, outputs=outputs, attrs=kwargs) type=op_type, inputs=inputs, outputs=outputs, attrs=kwargs)
return helper.append_activation(out) return helper.append_activation(out_var)
func.__name__ = op_type func.__name__ = op_type
func.__doc__ = _generate_doc_string_(op_proto) func.__doc__ = _generate_doc_string_(op_proto)
......
python/paddle/v2/fluid/tests/book_distribute/notest_dist_fit_a_line.py
浏览文件 @ e92fe9e6
...@@ -53,8 +53,9 @@ if training_role == "PSERVER": ...@@ -53,8 +53,9 @@ if training_role == "PSERVER":
if not current_endpoint: if not current_endpoint:
print("need env SERVER_ENDPOINT") print("need env SERVER_ENDPOINT")
exit(1) exit(1)
pserver_prog = t.get_pserver_program(current_endpoint, optimize_ops) pserver_prog = t.get_pserver_program(current_endpoint)
exe.run(fluid.default_startup_program()) pserver_startup = t.get_startup_program(current_endpoint, pserver_prog)
exe.run(pserver_startup)
exe.run(pserver_prog) exe.run(pserver_prog)
else: else:
trainer_prog = t.get_trainer_program() trainer_prog = t.get_trainer_program()
......
python/paddle/v2/fluid/tests/book_distribute/notest_dist_image_classification.py 0 → 100644
浏览文件 @ e92fe9e6
#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 __future__ import print_function
import sys
import paddle.v2 as paddle
import paddle.v2.fluid as fluid
import os
import sys
TRAINERS = 5
BATCH_SIZE = 128
PASS_NUM = 100
def resnet_cifar10(input, depth=32):
def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'):
tmp = fluid.layers.conv2d(
input=input,
filter_size=filter_size,
num_filters=ch_out,
stride=stride,
padding=padding,
act=None,
bias_attr=False)
return fluid.layers.batch_norm(input=tmp, act=act)
def shortcut(input, ch_in, ch_out, stride):
if ch_in != ch_out:
return conv_bn_layer(input, ch_out, 1, stride, 0, None)
else:
return input
def basicblock(input, ch_in, ch_out, stride):
tmp = conv_bn_layer(input, ch_out, 3, stride, 1)
tmp = conv_bn_layer(tmp, ch_out, 3, 1, 1, act=None)
short = shortcut(input, ch_in, ch_out, stride)
return fluid.layers.elementwise_add(x=tmp, y=short, act='relu')
def layer_warp(block_func, input, ch_in, ch_out, count, stride):
tmp = block_func(input, ch_in, ch_out, stride)
for i in range(1, count):
tmp = block_func(tmp, ch_out, ch_out, 1)
return tmp
assert (depth - 2) % 6 == 0
n = (depth - 2) / 6
conv1 = conv_bn_layer(
input=input, ch_out=16, filter_size=3, stride=1, padding=1)
res1 = layer_warp(basicblock, conv1, 16, 16, n, 1)
res2 = layer_warp(basicblock, res1, 16, 32, n, 2)
res3 = layer_warp(basicblock, res2, 32, 64, n, 2)
pool = fluid.layers.pool2d(
input=res3, pool_size=8, pool_type='avg', pool_stride=1)
return pool
def vgg16_bn_drop(input):
def conv_block(input, num_filter, groups, dropouts):
return fluid.nets.img_conv_group(
input=input,
pool_size=2,
pool_stride=2,
conv_num_filter=[num_filter] * groups,
conv_filter_size=3,
conv_act='relu',
conv_with_batchnorm=True,
conv_batchnorm_drop_rate=dropouts,
pool_type='max')
conv1 = conv_block(input, 64, 2, [0.3, 0])
conv2 = conv_block(conv1, 128, 2, [0.4, 0])
conv3 = conv_block(conv2, 256, 3, [0.4, 0.4, 0])
conv4 = conv_block(conv3, 512, 3, [0.4, 0.4, 0])
conv5 = conv_block(conv4, 512, 3, [0.4, 0.4, 0])
drop = fluid.layers.dropout(x=conv5, dropout_prob=0.5)
fc1 = fluid.layers.fc(input=drop, size=512, act=None)
bn = fluid.layers.batch_norm(input=fc1, act='relu')
drop2 = fluid.layers.dropout(x=bn, dropout_prob=0.5)
fc2 = fluid.layers.fc(input=drop2, size=512, act=None)
return fc2
classdim = 10
data_shape = [3, 32, 32]
images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
net_type = "vgg"
if len(sys.argv) >= 2:
net_type = sys.argv[1]
if net_type == "vgg":
print("train vgg net")
net = vgg16_bn_drop(images)
elif net_type == "resnet":
print("train resnet")
net = resnet_cifar10(images, 32)
else:
raise ValueError("%s network is not supported" % net_type)
predict = fluid.layers.fc(input=net, size=classdim, act='softmax')
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(x=cost)
optimizer = fluid.optimizer.Adam(learning_rate=0.001)
optimize_ops, params_grads = optimizer.minimize(avg_cost)
accuracy = fluid.evaluator.Accuracy(input=predict, label=label)
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.cifar.train10(), buf_size=128 * 10),
batch_size=BATCH_SIZE)
place = fluid.CPUPlace()
exe = fluid.Executor(place)
t = fluid.DistributeTranspiler()
# all parameter server endpoints list for spliting parameters
pserver_endpoints = os.getenv("PSERVERS")
# server endpoint for current node
current_endpoint = os.getenv("SERVER_ENDPOINT")
# run as trainer or parameter server
training_role = os.getenv("TRAINING_ROLE",
"TRAINER") # get the training role: trainer/pserver
t.transpile(
optimize_ops, params_grads, pservers=pserver_endpoints, trainers=TRAINERS)
if training_role == "PSERVER":
if not current_endpoint:
print("need env SERVER_ENDPOINT")
exit(1)
print("start pserver at:", current_endpoint)
pserver_prog = t.get_pserver_program(current_endpoint)
pserver_startup = t.get_startup_program(current_endpoint, pserver_prog)
exe.run(pserver_startup)
exe.run(pserver_prog)
print("pserver run end")
elif training_role == "TRAINER":
print("start trainer")
trainer_prog = t.get_trainer_program()
feeder = fluid.DataFeeder(place=place, feed_list=[images, label])
exe.run(fluid.default_startup_program())
for pass_id in range(PASS_NUM):
accuracy.reset(exe)
for data in train_reader():
loss, acc = exe.run(trainer_prog,
feed=feeder.feed(data),
fetch_list=[avg_cost] + accuracy.metrics)
pass_acc = accuracy.eval(exe)
print("loss:" + str(loss) + " acc:" + str(acc) + " pass_acc:" + str(
pass_acc))
# this model is slow, so if we can train two mini batch, we think it works properly.
print("trainer run end")
else:
print("environment var TRAINER_ROLE should be TRAINER os PSERVER")
exit(1)
python/paddle/v2/fluid/tests/book_distribute/notest_dist_label_semantic_roles.py
浏览文件 @ e92fe9e6
...@@ -197,8 +197,9 @@ def main(): ...@@ -197,8 +197,9 @@ def main():
if not current_endpoint: if not current_endpoint:
print("need env SERVER_ENDPOINT") print("need env SERVER_ENDPOINT")
exit(1) exit(1)
pserver_prog = t.get_pserver_program(current_endpoint, optimize_ops) pserver_prog = t.get_pserver_program(current_endpoint)
exe.run(fluid.default_startup_program()) pserver_startup = t.get_startup_program(current_endpoint, pserver_prog)
exe.run(pserver_startup)
exe.run(pserver_prog) exe.run(pserver_prog)
elif training_role == "TRAINER": elif training_role == "TRAINER":
trainer_prog = t.get_trainer_program() trainer_prog = t.get_trainer_program()
......
python/paddle/v2/fluid/tests/book_distribute/notest_dist_word2vec.py
浏览文件 @ e92fe9e6
...@@ -87,8 +87,9 @@ if training_role == "PSERVER": ...@@ -87,8 +87,9 @@ if training_role == "PSERVER":
if not current_endpoint: if not current_endpoint:
print("need env SERVER_ENDPOINT") print("need env SERVER_ENDPOINT")
exit(1) exit(1)
pserver_prog = t.get_pserver_program(current_endpoint, optimize_ops) pserver_prog = t.get_pserver_program(current_endpoint)
exe.run(fluid.default_startup_program()) pserver_startup = t.get_startup_program(current_endpoint, pserver_prog)
exe.run(pserver_startup)
exe.run(pserver_prog) exe.run(pserver_prog)
elif training_role == "TRAINER": elif training_role == "TRAINER":
feeder = fluid.DataFeeder( feeder = fluid.DataFeeder(
......
python/paddle/v2/fluid/tests/book_distribute/notest_recognize_digits_conv_dist.py
浏览文件 @ e92fe9e6
...@@ -52,26 +52,27 @@ train_reader = paddle.batch( ...@@ -52,26 +52,27 @@ train_reader = paddle.batch(
place = fluid.CPUPlace() place = fluid.CPUPlace()
exe = fluid.Executor(place) exe = fluid.Executor(place)
t = fluid.DistributeTranspiler() pserver_endpoints = os.getenv("PSERVERS") # all pserver endpoints
# all parameter server endpoints list for spliting parameters trainers = int(os.getenv("TRAINERS")) # total trainer count
pserver_endpoints = os.getenv("PSERVERS") current_endpoint = os.getenv("SERVER_ENDPOINT") # current pserver endpoint
# server endpoint for current node
current_endpoint = os.getenv("SERVER_ENDPOINT")
# run as trainer or parameter server
training_role = os.getenv("TRAINING_ROLE", training_role = os.getenv("TRAINING_ROLE",
"TRAINER") # get the training role: trainer/pserver "TRAINER") # get the training role: trainer/pserver
t.transpile(optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2) t = fluid.DistributeTranspiler()
t.transpile(
optimize_ops, params_grads, pservers=pserver_endpoints, trainers=trainers)
if training_role == "PSERVER": if training_role == "PSERVER":
if not current_endpoint: if not current_endpoint:
print("need env SERVER_ENDPOINT") print("need env SERVER_ENDPOINT")
exit(1) exit(1)
pserver_prog = t.get_pserver_program(current_endpoint, optimize_ops) pserver_prog = t.get_pserver_program(current_endpoint)
exe.run(fluid.default_startup_program()) pserver_startup = t.get_startup_program(current_endpoint, pserver_prog)
exe.run(pserver_startup)
exe.run(pserver_prog) exe.run(pserver_prog)
elif training_role == "TRAINER": elif training_role == "TRAINER":
trainer_prog = t.get_trainer_program() trainer_prog = t.get_trainer_program()
feeder = fluid.DataFeeder(feed_list=[images, label], place=place) feeder = fluid.DataFeeder(feed_list=[images, label], place=place)
# TODO(typhoonzero): change trainer startup program to fetch parameters from pserver
exe.run(fluid.default_startup_program()) exe.run(fluid.default_startup_program())
for pass_id in range(PASS_NUM): for pass_id in range(PASS_NUM):
......
adversarial/fluid_mnist.py → python/paddle/v2/fluid/tests/book_distribute/notest_recognize_digits_mlp_dist.py
浏览文件 @ e92fe9e6
...@@ -11,89 +11,78 @@ ...@@ -11,89 +11,78 @@
#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. #WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#See the License for the specific language governing permissions and #See the License for the specific language governing permissions and
#limitations under the License. #limitations under the License.
""" from __future__ import print_function
CNN on mnist data using fluid api of paddlepaddle import numpy as np
"""
import paddle.v2 as paddle import paddle.v2 as paddle
import paddle.v2.fluid as fluid import paddle.v2.fluid as fluid
import os
BATCH_SIZE = 128
PASS_NUM = 100
def mnist_cnn_model(img): images = fluid.layers.data(name='x', shape=[784], dtype='float32')
"""
Mnist cnn model
Args: # TODO(aroraabhinav) Add regularization and error clipping after
img(Varaible): the input image to be recognized # Issue 7432(https://github.com/PaddlePaddle/Paddle/issues/7432) is resolved.
hidden1 = fluid.layers.fc(input=images, size=128, act='relu')
hidden2 = fluid.layers.fc(input=hidden1, size=64, act='relu')
predict = fluid.layers.fc(input=hidden2, size=10, act='softmax')
Returns: label = fluid.layers.data(name='y', shape=[1], dtype='int64')
Variable: the label prediction
"""
conv_pool_1 = fluid.nets.simple_img_conv_pool(
input=img,
num_filters=20,
filter_size=5,
pool_size=2,
pool_stride=2,
act='relu')
conv_pool_2 = fluid.nets.simple_img_conv_pool( cost = fluid.layers.cross_entropy(input=predict, label=label)
input=conv_pool_1, avg_cost = fluid.layers.mean(x=cost)
num_filters=50,
filter_size=5,
pool_size=2,
pool_stride=2,
act='relu')
logits = fluid.layers.fc(input=conv_pool_2, size=10, act='softmax') optimizer = fluid.optimizer.Momentum(learning_rate=0.001, momentum=0.9)
return logits optimize_ops, params_grads = optimizer.minimize(avg_cost)
accuracy = fluid.evaluator.Accuracy(input=predict, label=label)
def main(): train_reader = paddle.batch(
""" paddle.reader.shuffle(
Train the cnn model on mnist datasets paddle.dataset.mnist.train(), buf_size=8192),
""" batch_size=BATCH_SIZE)
img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
logits = mnist_cnn_model(img)
cost = fluid.layers.cross_entropy(input=logits, label=label)
avg_cost = fluid.layers.mean(x=cost)
optimizer = fluid.optimizer.Adam(learning_rate=0.01)
optimizer.minimize(avg_cost)
accuracy = fluid.evaluator.Accuracy(input=logits, label=label) place = fluid.CPUPlace()
exe = fluid.Executor(place)
BATCH_SIZE = 50 t = fluid.DistributeTranspiler()
PASS_NUM = 3 # all parameter server endpoints list for spliting parameters
ACC_THRESHOLD = 0.98 pserver_endpoints = os.getenv("PSERVERS")
LOSS_THRESHOLD = 10.0 # server endpoint for current node
train_reader = paddle.batch( current_endpoint = os.getenv("SERVER_ENDPOINT")
paddle.reader.shuffle( # run as trainer or parameter server
paddle.dataset.mnist.train(), buf_size=500), training_role = os.getenv("TRAINING_ROLE",
batch_size=BATCH_SIZE) "TRAINER") # get the training role: trainer/pserver
t.transpile(optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2)
place = fluid.CPUPlace() if training_role == "PSERVER":
exe = fluid.Executor(place) if not current_endpoint:
feeder = fluid.DataFeeder(feed_list=[img, label], place=place) print("need env SERVER_ENDPOINT")
exit(1)
pserver_prog = t.get_pserver_program(current_endpoint)
pserver_startup = t.get_startup_program(current_endpoint, pserver_prog)
exe.run(pserver_startup)
exe.run(pserver_prog)
elif training_role == "TRAINER":
trainer_prog = t.get_trainer_program()
feeder = fluid.DataFeeder(feed_list=[images, label], place=place)
exe.run(fluid.default_startup_program()) exe.run(fluid.default_startup_program())
for pass_id in range(PASS_NUM): for pass_id in range(PASS_NUM):
accuracy.reset(exe) accuracy.reset(exe)
batch_id = 0
for data in train_reader(): for data in train_reader():
loss, acc = exe.run(fluid.default_main_program(), loss, acc = exe.run(trainer_prog,
feed=feeder.feed(data), feed=feeder.feed(data),
fetch_list=[avg_cost] + accuracy.metrics) fetch_list=[avg_cost] + accuracy.metrics)
pass_acc = accuracy.eval(exe) pass_acc = accuracy.eval(exe)
print("pass_id=" + str(pass_id) + " acc=" + str(acc) + " pass_acc=" if batch_id % 100 == 0:
+ str(pass_acc)) print("batch_id %d, loss: %f, acc: %f" %
if loss < LOSS_THRESHOLD and pass_acc > ACC_THRESHOLD: (batch_id, loss, pass_acc))
break batch_id += 1
pass_acc = accuracy.eval(exe) pass_acc = accuracy.eval(exe)
print("pass_id=" + str(pass_id) + " pass_acc=" + str(pass_acc)) print("pass_id=" + str(pass_id) + " pass_acc=" + str(pass_acc))
fluid.io.save_params( else:
exe, dirname='./mnist', main_program=fluid.default_main_program()) print("environment var TRAINER_ROLE should be TRAINER os PSERVER")
print('train mnist done')
if __name__ == '__main__':
main()
python/paddle/v2/fluid/tests/book_distribute/notest_understand_sentiment_conv_dist.py
浏览文件 @ e92fe9e6
...@@ -92,15 +92,16 @@ def main(): ...@@ -92,15 +92,16 @@ def main():
t.transpile( t.transpile(
optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2) optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2)
exe.run(fluid.default_startup_program())
if training_role == "PSERVER": if training_role == "PSERVER":
if not current_endpoint: if not current_endpoint:
print("need env SERVER_ENDPOINT") print("need env SERVER_ENDPOINT")
exit(1) exit(1)
pserver_prog = t.get_pserver_program(current_endpoint, optimize_ops) pserver_prog = t.get_pserver_program(current_endpoint)
pserver_startup = t.get_startup_program(current_endpoint, pserver_prog)
exe.run(pserver_startup)
exe.run(pserver_prog) exe.run(pserver_prog)
elif training_role == "TRAINER": elif training_role == "TRAINER":
exe.run(fluid.default_startup_program())
trainer_prog = t.get_trainer_program() trainer_prog = t.get_trainer_program()
feeder = fluid.DataFeeder(feed_list=[data, label], place=place) feeder = fluid.DataFeeder(feed_list=[data, label], place=place)
......
python/paddle/v2/fluid/tests/test_ctc_align.py 0 → 100644
浏览文件 @ e92fe9e6
# 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 sys
import unittest
import numpy as np
from op_test import OpTest
from test_softmax_op import stable_softmax
def CTCAlign(input, lod, blank, merge_repeated):
lod0 = lod[0]
result = []
for i in range(len(lod0) - 1):
prev_token = -1
for j in range(lod0[i], lod0[i + 1]):
token = input[j][0]
if (token != blank) and not (merge_repeated and
token == prev_token):
result.append(token)
prev_token = token
result = np.array(result).reshape([len(result), 1]).astype("int32")
return result
class TestCTCAlignOp(OpTest):
def config(self):
self.op_type = "ctc_align"
self.input_lod = [[0, 11, 18]]
self.blank = 0
self.merge_repeated = False
self.input = np.array(
[0, 1, 2, 2, 0, 4, 0, 4, 5, 0, 6, 6, 0, 0, 7, 7, 7, 0]).reshape(
[18, 1]).astype("int32")
def setUp(self):
self.config()
output = CTCAlign(self.input, self.input_lod, self.blank,
self.merge_repeated)
self.inputs = {"Input": (self.input, self.input_lod), }
self.outputs = {"Output": output}
self.attrs = {
"blank": self.blank,
"merge_repeated": self.merge_repeated
}
def test_check_output(self):
self.check_output()
pass
class TestCTCAlignOpCase1(TestCTCAlignOp):
def config(self):
self.op_type = "ctc_align"
self.input_lod = [[0, 11, 19]]
self.blank = 0
self.merge_repeated = True
self.input = np.array(
[0, 1, 2, 2, 0, 4, 0, 4, 5, 0, 6, 6, 0, 0, 7, 7, 7, 0, 0]).reshape(
[19, 1]).astype("int32")
if __name__ == "__main__":
unittest.main()
python/paddle/v2/fluid/tests/test_edit_distance_op.py
浏览文件 @ e92fe9e6
# 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 unittest
import numpy as np import numpy as np
from op_test import OpTest from op_test import OpTest
...@@ -38,8 +51,8 @@ class TestEditDistanceOp(OpTest): ...@@ -38,8 +51,8 @@ class TestEditDistanceOp(OpTest):
def setUp(self): def setUp(self):
self.op_type = "edit_distance" self.op_type = "edit_distance"
normalized = False normalized = False
x1 = np.array([[0, 12, 3, 5, 8, 2]]).astype("int32") x1 = np.array([[0, 12, 3, 5, 8, 2]]).astype("int64")
x2 = np.array([[0, 12, 4, 7, 8]]).astype("int32") x2 = np.array([[0, 12, 4, 7, 8]]).astype("int64")
x1 = np.transpose(x1) x1 = np.transpose(x1)
x2 = np.transpose(x2) x2 = np.transpose(x2)
x1_lod = [0, 1, 5] x1_lod = [0, 1, 5]
...@@ -66,8 +79,8 @@ class TestEditDistanceOpNormalized(OpTest): ...@@ -66,8 +79,8 @@ class TestEditDistanceOpNormalized(OpTest):
def setUp(self): def setUp(self):
self.op_type = "edit_distance" self.op_type = "edit_distance"
normalized = True normalized = True
x1 = np.array([[0, 10, 3, 6, 5, 8, 2]]).astype("int32") x1 = np.array([[0, 10, 3, 6, 5, 8, 2]]).astype("int64")
x2 = np.array([[0, 10, 4, 6, 7, 8]]).astype("int32") x2 = np.array([[0, 10, 4, 6, 7, 8]]).astype("int64")
x1 = np.transpose(x1) x1 = np.transpose(x1)
x2 = np.transpose(x2) x2 = np.transpose(x2)
x1_lod = [0, 1, 3, 6] x1_lod = [0, 1, 3, 6]
......
python/paddle/v2/fluid/tests/test_elementwise_add_op.py
浏览文件 @ e92fe9e6
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. # Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
# #
#Licensed under the Apache License, Version 2.0 (the "License"); # Licensed under the Apache License, Version 2.0 (the "License");
#you may not use this file except in compliance with the License. # you may not use this file except in compliance with the License.
#You may obtain a copy of the License at # You may obtain a copy of the License at
# #
# http://www.apache.org/licenses/LICENSE-2.0 # http://www.apache.org/licenses/LICENSE-2.0
# #
#Unless required by applicable law or agreed to in writing, software # Unless required by applicable law or agreed to in writing, software
#distributed under the License is distributed on an "AS IS" BASIS, # distributed under the License is distributed on an "AS IS" BASIS,
#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
#limitations under the License. # limitations under the License.
import unittest import unittest
import numpy as np import numpy as np
from op_test import OpTest from op_test import OpTest
...@@ -40,6 +40,16 @@ class TestElementwiseOp(OpTest): ...@@ -40,6 +40,16 @@ class TestElementwiseOp(OpTest):
['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y')) ['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y'))
class TestElementwiseAddOp_scalar(TestElementwiseOp):
def setUp(self):
self.op_type = "elementwise_add"
self.inputs = {
'X': np.random.rand(2, 3, 4).astype(np.float32),
'Y': np.random.rand(1).astype(np.float32)
}
self.outputs = {'Out': self.inputs['X'] + self.inputs['Y']}
class TestElementwiseAddOp_Vector(TestElementwiseOp): class TestElementwiseAddOp_Vector(TestElementwiseOp):
def setUp(self): def setUp(self):
self.op_type = "elementwise_add" self.op_type = "elementwise_add"
......
python/paddle/v2/fluid/tests/test_elementwise_div_op.py
浏览文件 @ e92fe9e6
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. # Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
# #
#Licensed under the Apache License, Version 2.0 (the "License"); # Licensed under the Apache License, Version 2.0 (the "License");
#you may not use this file except in compliance with the License. # you may not use this file except in compliance with the License.
#You may obtain a copy of the License at # You may obtain a copy of the License at
# #
# http://www.apache.org/licenses/LICENSE-2.0 # http://www.apache.org/licenses/LICENSE-2.0
# #
#Unless required by applicable law or agreed to in writing, software # Unless required by applicable law or agreed to in writing, software
#distributed under the License is distributed on an "AS IS" BASIS, # distributed under the License is distributed on an "AS IS" BASIS,
#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
#limitations under the License. # limitations under the License.
import unittest import unittest
import numpy as np import numpy as np
from op_test import OpTest from op_test import OpTest
...@@ -45,6 +45,16 @@ class ElementwiseDivOp(OpTest): ...@@ -45,6 +45,16 @@ class ElementwiseDivOp(OpTest):
['X'], 'Out', max_relative_error=0.05, no_grad_set=set('Y')) ['X'], 'Out', max_relative_error=0.05, no_grad_set=set('Y'))
class TestElementwiseDivOp_scalar(ElementwiseDivOp):
def setUp(self):
self.op_type = "elementwise_div"
self.inputs = {
'X': np.random.uniform(0.1, 1, [2, 3, 4]).astype(np.float32),
'Y': np.random.uniform(0.1, 1, [1]).astype(np.float32)
}
self.outputs = {'Out': self.inputs['X'] / self.inputs['Y']}
class TestElementwiseDivOp_Vector(ElementwiseDivOp): class TestElementwiseDivOp_Vector(ElementwiseDivOp):
def setUp(self): def setUp(self):
self.op_type = "elementwise_div" self.op_type = "elementwise_div"
......
python/paddle/v2/fluid/tests/test_elementwise_max_op.py 0 → 100644
浏览文件 @ e92fe9e6
# 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
class TestElementwiseOp(OpTest):
def setUp(self):
self.op_type = "elementwise_max"
# If x and y have the same value, the max() is not differentiable.
# So we generate test data by the following method
# to avoid them being too close to each other.
x = np.random.uniform(0.1, 1, [13, 17]).astype("float32")
sgn = np.random.choice([-1, 1], [13, 17]).astype("float32")
y = x + sgn * np.random.uniform(0.1, 1, [13, 17]).astype("float32")
self.inputs = {'X': x, 'Y': y}
self.outputs = {'Out': np.maximum(self.inputs['X'], self.inputs['Y'])}
def test_check_output(self):
self.check_output()
def test_check_grad_normal(self):
self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.005)
def test_check_grad_ingore_x(self):
self.check_grad(
['Y'], 'Out', max_relative_error=0.005, no_grad_set=set("X"))
def test_check_grad_ingore_y(self):
self.check_grad(
['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y'))
class TestElementwiseMaxOp_scalar(TestElementwiseOp):
def setUp(self):
self.op_type = "elementwise_max"
x = np.random.random_integers(-5, 5, [2, 3, 4]).astype("float32")
y = np.array([0.5]).astype("float32")
self.inputs = {'X': x, 'Y': y}
self.outputs = {'Out': np.maximum(self.inputs['X'], self.inputs['Y'])}
class TestElementwiseMaxOp_Vector(TestElementwiseOp):
def setUp(self):
self.op_type = "elementwise_max"
x = np.random.random((32, )).astype("float32")
sgn = np.random.choice([-1, 1], (32, )).astype("float32")
y = x + sgn * np.random.uniform(0.1, 1, (32, )).astype("float32")
self.inputs = {'X': x, 'Y': y}
self.outputs = {'Out': np.maximum(self.inputs['X'], self.inputs['Y'])}
class TestElementwiseMaxOp_broadcast_0(TestElementwiseOp):
def setUp(self):
self.op_type = "elementwise_max"
x = np.random.uniform(0.5, 1, (2, 3, 4)).astype(np.float32)
sgn = np.random.choice([-1, 1], (2, )).astype(np.float32)
y = x[:, 0, 0] + sgn * \
np.random.uniform(1, 2, (2, )).astype(np.float32)
self.inputs = {'X': x, 'Y': y}
self.attrs = {'axis': 0}
self.outputs = {
'Out':
np.maximum(self.inputs['X'], self.inputs['Y'].reshape(2, 1, 1))
}
class TestElementwiseMaxOp_broadcast_1(TestElementwiseOp):
def setUp(self):
self.op_type = "elementwise_max"
x = np.random.uniform(0.5, 1, (2, 3, 4)).astype(np.float32)
sgn = np.random.choice([-1, 1], (3, )).astype(np.float32)
y = x[0, :, 0] + sgn * \
np.random.uniform(1, 2, (3, )).astype(np.float32)
self.inputs = {'X': x, 'Y': y}
self.attrs = {'axis': 1}
self.outputs = {
'Out':
np.maximum(self.inputs['X'], self.inputs['Y'].reshape(1, 3, 1))
}
class TestElementwiseMaxOp_broadcast_2(TestElementwiseOp):
def setUp(self):
self.op_type = "elementwise_max"
x = np.random.uniform(0.5, 1, (2, 3, 4)).astype(np.float32)
sgn = np.random.choice([-1, 1], (4, )).astype(np.float32)
y = x[0, 0, :] + sgn * \
np.random.uniform(1, 2, (4, )).astype(np.float32)
self.inputs = {'X': x, 'Y': y}
self.outputs = {
'Out':
np.maximum(self.inputs['X'], self.inputs['Y'].reshape(1, 1, 4))
}
class TestElementwiseMaxOp_broadcast_3(TestElementwiseOp):
def setUp(self):
self.op_type = "elementwise_max"
x = np.random.uniform(0.5, 1, (2, 3, 4, 5)).astype(np.float32)
sgn = np.random.choice([-1, 1], (3, 4)).astype(np.float32)
y = x[0, :, :, 0] + sgn * \
np.random.uniform(1, 2, (3, 4)).astype(np.float32)
self.inputs = {'X': x, 'Y': y}
self.attrs = {'axis': 1}
self.outputs = {
'Out':
np.maximum(self.inputs['X'], self.inputs['Y'].reshape(1, 3, 4, 1))
}
if __name__ == '__main__':
unittest.main()
python/paddle/v2/fluid/tests/test_elementwise_min_op.py 0 → 100644
浏览文件 @ e92fe9e6
# 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
class TestElementwiseOp(OpTest):
def setUp(self):
self.op_type = "elementwise_min"
# If x and y have the same value, the min() is not differentiable.
# So we generate test data by the following method
# to avoid them being too close to each other.
x = np.random.uniform(0.1, 1, [13, 17]).astype("float32")
sgn = np.random.choice([-1, 1], [13, 17]).astype("float32")
y = x + sgn * np.random.uniform(0.1, 1, [13, 17]).astype("float32")
self.inputs = {'X': x, 'Y': y}
self.outputs = {'Out': np.minimum(self.inputs['X'], self.inputs['Y'])}
def test_check_output(self):
self.check_output()
def test_check_grad_normal(self):
self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.005)
def test_check_grad_ingore_x(self):
self.check_grad(
['Y'], 'Out', max_relative_error=0.005, no_grad_set=set("X"))
def test_check_grad_ingore_y(self):
self.check_grad(
['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y'))
class TestElementwiseMinOp_scalar(TestElementwiseOp):
def setUp(self):
self.op_type = "elementwise_min"
x = np.random.random_integers(-5, 5, [2, 3, 4]).astype("float32")
y = np.array([0.5]).astype("float32")
self.inputs = {'X': x, 'Y': y}
self.outputs = {'Out': np.minimum(self.inputs['X'], self.inputs['Y'])}
class TestElementwiseMaxOp_Vector(TestElementwiseOp):
def setUp(self):
self.op_type = "elementwise_min"
x = np.random.random((32, )).astype("float32")
sgn = np.random.choice([-1, 1], (32, )).astype("float32")
y = x + sgn * np.random.uniform(0.1, 1, (32, )).astype("float32")
self.inputs = {'X': x, 'Y': y}
self.outputs = {'Out': np.minimum(self.inputs['X'], self.inputs['Y'])}
class TestElementwiseMaxOp_broadcast_0(TestElementwiseOp):
def setUp(self):
self.op_type = "elementwise_min"
x = np.random.uniform(0.5, 1, (2, 3, 4)).astype(np.float32)
sgn = np.random.choice([-1, 1], (2, )).astype(np.float32)
y = x[:, 0, 0] + sgn * \
np.random.uniform(1, 2, (2, )).astype(np.float32)
self.inputs = {'X': x, 'Y': y}
self.attrs = {'axis': 0}
self.outputs = {
'Out':
np.minimum(self.inputs['X'], self.inputs['Y'].reshape(2, 1, 1))
}
class TestElementwiseMaxOp_broadcast_1(TestElementwiseOp):
def setUp(self):
self.op_type = "elementwise_min"
x = np.random.uniform(0.5, 1, (2, 3, 4)).astype(np.float32)
sgn = np.random.choice([-1, 1], (3, )).astype(np.float32)
y = x[0, :, 0] + sgn * \
np.random.uniform(1, 2, (3, )).astype(np.float32)
self.inputs = {'X': x, 'Y': y}
self.attrs = {'axis': 1}
self.outputs = {
'Out':
np.minimum(self.inputs['X'], self.inputs['Y'].reshape(1, 3, 1))
}
class TestElementwiseMaxOp_broadcast_2(TestElementwiseOp):
def setUp(self):
self.op_type = "elementwise_min"
x = np.random.uniform(0.5, 1, (2, 3, 4)).astype(np.float32)
sgn = np.random.choice([-1, 1], (4, )).astype(np.float32)
y = x[0, 0, :] + sgn * \
np.random.uniform(1, 2, (4, )).astype(np.float32)
self.inputs = {'X': x, 'Y': y}
self.outputs = {
'Out':
np.minimum(self.inputs['X'], self.inputs['Y'].reshape(1, 1, 4))
}
class TestElementwiseMaxOp_broadcast_3(TestElementwiseOp):
def setUp(self):
self.op_type = "elementwise_min"
x = np.random.uniform(0.5, 1, (2, 3, 4, 5)).astype(np.float32)
sgn = np.random.choice([-1, 1], (3, 4)).astype(np.float32)
y = x[0, :, :, 0] + sgn * \
np.random.uniform(1, 2, (3, 4)).astype(np.float32)
self.inputs = {'X': x, 'Y': y}
self.attrs = {'axis': 1}
self.outputs = {
'Out':
np.minimum(self.inputs['X'], self.inputs['Y'].reshape(1, 3, 4, 1))
}
if __name__ == '__main__':
unittest.main()
python/paddle/v2/fluid/tests/test_elementwise_mul_op.py
浏览文件 @ e92fe9e6
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. # Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
# #
#Licensed under the Apache License, Version 2.0 (the "License"); # Licensed under the Apache License, Version 2.0 (the "License");
#you may not use this file except in compliance with the License. # you may not use this file except in compliance with the License.
#You may obtain a copy of the License at # You may obtain a copy of the License at
# #
# http://www.apache.org/licenses/LICENSE-2.0 # http://www.apache.org/licenses/LICENSE-2.0
# #
#Unless required by applicable law or agreed to in writing, software # Unless required by applicable law or agreed to in writing, software
#distributed under the License is distributed on an "AS IS" BASIS, # distributed under the License is distributed on an "AS IS" BASIS,
#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
#limitations under the License. # limitations under the License.
import unittest import unittest
import numpy as np import numpy as np
from op_test import OpTest from op_test import OpTest
...@@ -38,6 +38,16 @@ class ElementwiseMulOp(OpTest): ...@@ -38,6 +38,16 @@ class ElementwiseMulOp(OpTest):
self.check_grad(['X'], 'Out', no_grad_set=set('Y')) self.check_grad(['X'], 'Out', no_grad_set=set('Y'))
class TestElementwiseMulOp_scalar(ElementwiseMulOp):
def setUp(self):
self.op_type = "elementwise_mul"
self.inputs = {
'X': np.random.rand(2, 3, 4).astype(np.float32),
'Y': np.random.rand(1).astype(np.float32)
}
self.outputs = {'Out': self.inputs['X'] * self.inputs['Y']}
class TestElementwiseMulOp_Vector(ElementwiseMulOp): class TestElementwiseMulOp_Vector(ElementwiseMulOp):
def setUp(self): def setUp(self):
self.op_type = "elementwise_mul" self.op_type = "elementwise_mul"
......
python/paddle/v2/fluid/tests/test_elementwise_sub_op.py
浏览文件 @ e92fe9e6
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. # Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
# #
#Licensed under the Apache License, Version 2.0 (the "License"); # Licensed under the Apache License, Version 2.0 (the "License");
#you may not use this file except in compliance with the License. # you may not use this file except in compliance with the License.
#You may obtain a copy of the License at # You may obtain a copy of the License at
# #
# http://www.apache.org/licenses/LICENSE-2.0 # http://www.apache.org/licenses/LICENSE-2.0
# #
#Unless required by applicable law or agreed to in writing, software # Unless required by applicable law or agreed to in writing, software
#distributed under the License is distributed on an "AS IS" BASIS, # distributed under the License is distributed on an "AS IS" BASIS,
#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
#limitations under the License. # limitations under the License.
import unittest import unittest
import numpy as np import numpy as np
from op_test import OpTest from op_test import OpTest
...@@ -40,6 +40,16 @@ class TestElementwiseOp(OpTest): ...@@ -40,6 +40,16 @@ class TestElementwiseOp(OpTest):
['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y')) ['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y'))
class TestElementwiseSubOp_scalar(TestElementwiseOp):
def setUp(self):
self.op_type = "elementwise_sub"
self.inputs = {
'X': np.random.rand(2, 3, 4).astype(np.float32),
'Y': np.random.rand(1).astype(np.float32)
}
self.outputs = {'Out': self.inputs['X'] - self.inputs['Y']}
class TestElementwiseSubOp_Vector(TestElementwiseOp): class TestElementwiseSubOp_Vector(TestElementwiseOp):
def setUp(self): def setUp(self):
self.op_type = "elementwise_sub" self.op_type = "elementwise_sub"
......
python/paddle/v2/fluid/tests/test_layers.py
浏览文件 @ e92fe9e6
...@@ -216,6 +216,14 @@ class TestBook(unittest.TestCase): ...@@ -216,6 +216,14 @@ class TestBook(unittest.TestCase):
self.assertIsNotNone(x) self.assertIsNotNone(x)
print(str(program)) print(str(program))
def test_sequence_reshape(self):
program = Program()
with program_guard(program):
x = layers.data(name='x', shape=[8], dtype='float32', lod_level=1)
out = layers.sequence_reshape(input=x, new_dim=16)
self.assertIsNotNone(out)
print(str(program))
if __name__ == '__main__': if __name__ == '__main__':
unittest.main() unittest.main()
python/paddle/v2/fluid/tests/test_normalization_wrapper.py 0 → 100644
浏览文件 @ e92fe9e6
# 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 paddle.v2.fluid as fluid
import paddle.v2.fluid.core as core
import numpy as np
class TestNormalization(unittest.TestCase):
data_desc = {"name": "input", "shape": (2, 3, 7)}
def gen_random_input(self):
"""Generate random input data.
"""
self.data = np.random.random(
size=self.data_desc["shape"]).astype("float32")
def set_program(self, axis, epsilon):
"""Build the test program.
"""
data = fluid.layers.data(
name=self.data_desc["name"],
shape=self.data_desc["shape"],
dtype="float32",
append_batch_size=False)
data.stop_gradient = False
l2_norm = fluid.layers.l2_normalize(x=data, axis=axis, epsilon=epsilon)
out = fluid.layers.reduce_sum(l2_norm, dim=None)
fluid.backward.append_backward(loss=out)
self.fetch_list = [l2_norm]
def run_program(self):
"""Run the test program.
"""
places = [core.CPUPlace()]
if core.is_compile_gpu():
places.append(core.CUDAPlace(0))
for place in places:
self.set_inputs(place)
exe = fluid.Executor(place)
output = exe.run(fluid.default_main_program(),
feed=self.inputs,
fetch_list=self.fetch_list,
return_numpy=True)
self.op_output = output
def set_inputs(self, place):
"""Set the randomly generated data to the test program.
"""
self.inputs = {}
tensor = fluid.Tensor()
tensor.set(self.data, place)
self.inputs[self.data_desc["name"]] = tensor
def l2_normalize(self, data, axis, epsilon):
""" Compute the groundtruth.
"""
output = data * np.reciprocal(
np.sum(np.square(data), axis=axis, keepdims=True))
return output
def test_l2_normalize(self):
""" Test the python wrapper for l2_normalize.
"""
axis = 1
#TODO(caoying) epsilon is not supported due to lack of a maximum_op.
epsilon = 1e-6
self.gen_random_input()
self.set_program(axis, epsilon)
self.run_program()
expect_output = self.l2_normalize(self.data, axis, epsilon)
# check output
self.assertTrue(np.allclose(self.op_output, expect_output, atol=0.001))
if __name__ == '__main__':
unittest.main()
python/paddle/v2/fluid/tests/test_parallel_op.py
浏览文件 @ e92fe9e6
...@@ -15,9 +15,6 @@ import unittest ...@@ -15,9 +15,6 @@ import unittest
import paddle.v2.fluid as fluid import paddle.v2.fluid as fluid
import numpy import numpy
import sys
# TODO(dzhwinter): get places op check need to be enhanced.
sys.exit(0)
class BaseParallelForTest(unittest.TestCase): class BaseParallelForTest(unittest.TestCase):
...@@ -151,23 +148,27 @@ class BaseParallelForTest(unittest.TestCase): ...@@ -151,23 +148,27 @@ class BaseParallelForTest(unittest.TestCase):
class ParallelOpTest(BaseParallelForTest): class ParallelOpTest(BaseParallelForTest):
def test_simple_fc(self): @staticmethod
def __network__(): def __network__():
x = fluid.layers.data(shape=[784], dtype='float32', name='img') x = fluid.layers.data(shape=[784], dtype='float32', name='img')
# FIXME: This is a bug of parallel.do x = yield x
x.stop_gradient = False hidden = fluid.layers.fc(input=x, size=200, param_attr='fc1.w')
x = yield x loss = fluid.layers.mean(x=hidden)
hidden = fluid.layers.fc(input=x, size=200, param_attr='fc1.w') yield loss
loss = fluid.layers.mean(x=hidden)
yield loss
def test_simple_fc(self):
self.run_test( self.run_test(
callback=__network__, callback=ParallelOpTest.__network__,
feed={ feed={
'img': 'img': numpy.random.random(size=(51, 784)).astype('float32')
numpy.random.random(size=(128 * 3, 784)).astype('float32')
}, },
fetch='fc1.w@GRAD') fetch=['fc1.w@GRAD'])
def test_fc_with_tiny_data(self):
self.run_test(
callback=ParallelOpTest.__network__,
feed={'img': numpy.random.random(size=(1, 784)).astype('float32')},
fetch=['fc1.w@GRAD'])
if __name__ == '__main__': if __name__ == '__main__':
......
python/paddle/v2/fluid/tests/test_sequence_erase_op.py
浏览文件 @ e92fe9e6
...@@ -29,7 +29,7 @@ def sequence_erase(in_seq, lod0, tokens): ...@@ -29,7 +29,7 @@ def sequence_erase(in_seq, lod0, tokens):
return np.array(out_seq).astype("int32"), new_lod0 return np.array(out_seq).astype("int32"), new_lod0
class TestSequenceEraseOp(OpTest): class TestSequenceEraseOpInt32(OpTest):
def setUp(self): def setUp(self):
self.op_type = "sequence_erase" self.op_type = "sequence_erase"
in_seq = np.random.randint(0, 10, (30, 1)).astype("int32") in_seq = np.random.randint(0, 10, (30, 1)).astype("int32")
...@@ -44,5 +44,35 @@ class TestSequenceEraseOp(OpTest): ...@@ -44,5 +44,35 @@ class TestSequenceEraseOp(OpTest):
self.check_output() self.check_output()
class TestSequenceEraseOpInt64(OpTest):
def setUp(self):
self.op_type = "sequence_erase"
in_seq = np.random.randint(0, 10, (30, 1)).astype("int64")
lod = [[0, 9, 13, 24, 30]]
tokens = [2, 3, 5]
out_seq, new_lod0 = sequence_erase(in_seq, lod[0], tokens)
self.attrs = {'tokens': tokens}
self.inputs = {'X': (in_seq, lod)}
self.outputs = {'Out': (out_seq, [new_lod0])}
def test_check_output(self):
self.check_output()
class TestSequenceEraseOpEmpty(OpTest):
def setUp(self):
self.op_type = "sequence_erase"
in_seq = np.random.randint(0, 10, (30, 1)).astype("int32")
lod = [[0, 9, 13, 24, 30]]
tokens = []
out_seq, new_lod0 = sequence_erase(in_seq, lod[0], tokens)
self.attrs = {'tokens': tokens}
self.inputs = {'X': (in_seq, lod)}
self.outputs = {'Out': (out_seq, [new_lod0])}
def test_check_output(self):
self.check_output()
if __name__ == '__main__': if __name__ == '__main__':
unittest.main() unittest.main()
python/paddle/v2/fluid/tests/test_sequence_reshape.py 0 → 100644
浏览文件 @ e92fe9e6
# 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
import math
from op_test import OpTest
class TestSequenceReshape(OpTest):
def setUp(self):
self.op_type = 'sequence_reshape'
dimension = 12
x_lod = [[0, 4, 5, 8, 11]]
x = np.random.uniform(0.1, 1, [11, 24]).astype('float32')
self.inputs = {'X': (x, x_lod)}
self.attrs = {'new_dim': dimension}
out, out_lod = self.compute_output(x, x_lod, dimension)
self.outputs = {'Out': (out, out_lod)}
def compute_output(self, x, x_lod, dimension):
x_width = x.shape[1]
out_lod = [[0]]
for i in xrange(len(x_lod[0]) - 1):
seq_len = x_lod[0][i + 1] - x_lod[0][i]
offset = (seq_len * x_width) / dimension
assert int(offset) * dimension == seq_len * x_width
out_lod[0].append(out_lod[0][-1] + int(offset))
out = np.zeros(shape=(out_lod[0][-1], dimension)).astype('float32')
out.ravel()[:] = x.ravel()[:]
return out, out_lod
def test_check_output(self):
self.check_output()
def test_check_grad(self):
self.check_grad(["X"], "Out")
class TestSequenceReshape_reduce(TestSequenceReshape):
def setUp(self):
self.op_type = 'sequence_reshape'
dimension = 24
x_lod = [[0, 4, 6, 8, 12]]
x = np.random.uniform(0.1, 1, [12, 12]).astype('float32')
self.inputs = {'X': (x, x_lod)}
self.attrs = {'new_dim': dimension}
out, out_lod = self.compute_output(x, x_lod, dimension)
self.outputs = {'Out': (out, out_lod)}
class TestSequenceReshape_same(TestSequenceReshape):
def setUp(self):
self.op_type = 'sequence_reshape'
dimension = 12
x_lod = [[0, 4, 6, 8, 12]]
x = np.random.uniform(0.1, 1, [12, 12]).astype('float32')
self.inputs = {'X': (x, x_lod)}
self.attrs = {'new_dim': dimension}
out, out_lod = self.compute_output(x, x_lod, dimension)
self.outputs = {'Out': (out, out_lod)}
if __name__ == '__main__':
unittest.main()
python/paddle/v2/fluid/tests/test_split_selected_rows_op.py 0 → 100644
浏览文件 @ e92fe9e6
# 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 paddle.v2.fluid.core as core
import numpy as np
from paddle.v2.fluid.op import Operator
class TestSpliteSelectedRows(unittest.TestCase):
def get_places(self):
places = [core.CPUPlace()]
if core.is_compile_gpu():
places.append(core.CUDAPlace(0))
return places
def test_check_output(self):
for place in self.get_places():
self.check_with_place(place)
def test_check_grad(self):
for place in self.get_places():
self.check_grad_with_place(place)
def check_with_place(self, place):
scope = core.Scope()
rows = [0, 5, 7, 4]
height = 10
row_numel = 2
# initialize input variable X
x = scope.var('X').get_selected_rows()
x.set_rows(rows)
x.set_height(height)
np_array = np.ones((len(rows), row_numel)).astype("float32")
np_array[0, 0] = 2.0
np_array[2, 1] = 4.0
x_tensor = x.get_tensor()
x_tensor.set(np_array, place)
rows_sections = [2, 2]
height_sections = []
# initialize output variables [out0, out1]
out0 = scope.var('out0').get_selected_rows()
out1 = scope.var('out1').get_selected_rows()
# expected output selected rows
expected_out0_rows = [0, 5]
expected_out1_rows = [7, 4]
expected_height = height
op = Operator(
"split_selected_rows",
X="X",
Out=["out0", "out1"],
rows_sections=rows_sections,
height_sections=height_sections)
op.run(scope, place)
self.assertEqual(out0.rows(), expected_out0_rows)
self.assertEqual(out1.rows(), expected_out1_rows)
self.assertEqual(out0.height(), expected_height)
self.assertEqual(out1.height(), expected_height)
self.assertAlmostEqual(2.0, np.array(out0.get_tensor())[0, 0])
self.assertAlmostEqual(4.0, np.array(out1.get_tensor())[0, 1])
def check_grad_with_place(self, place):
scope = core.Scope()
height = 10
row_numel = 2
# attr
rows_sections = [2, 2]
height_sections = []
# initialize input variable X
out0_grad = scope.var("out0@GRAD").get_selected_rows()
rows0 = [0, 5]
out0_grad.set_rows(rows0)
out0_grad.set_height(height)
out0_grad_tensor = out0_grad.get_tensor()
np_array = np.ones((len(rows0), row_numel)).astype("float32")
np_array[0, 0] = 2.0
out0_grad_tensor.set(np_array, place)
out1_grad = scope.var("out1@GRAD").get_selected_rows()
rows1 = [7, 5]
out1_grad.set_rows(rows1)
out1_grad.set_height(height)
out1_grad_tensor = out1_grad.get_tensor()
np_array = np.ones((len(rows1), row_numel)).astype("float32")
np_array[0, 1] = 4.0
out1_grad_tensor.set(np_array, place)
x_grad = scope.var("X@GRAD").get_selected_rows()
grad_op = Operator(
"sum",
X=["out0@GRAD", "out1@GRAD"],
Out="X@GRAD",
rows_sections=rows_sections,
height_sections=height_sections)
grad_op.run(scope, place)
self.assertEqual(x_grad.rows(), rows0 + rows1)
self.assertEqual(x_grad.height(), height)
self.assertAlmostEqual(2.0, np.array(x_grad.get_tensor())[0, 0])
self.assertAlmostEqual(4.0, np.array(x_grad.get_tensor())[2, 1])
if __name__ == "__main__":
unittest.main()
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