Skip to content

P
Paddle

TonyTonyFun / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 0
Fork 0
P
Paddle

体验新版 GitCode，发现更多精彩内容 >>

提交 974183b4 编写于 作者: F fengjiayi
浏览文件
操作

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

上级 773f2f73 b7b5de7f
展开全部 隐藏空白更改
内联 并排
Showing with 910 addition and 2680 deletion
CODE_OF_CONDUCT_cn.md
浏览文件 @ 974183b4
# 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：
* 使用歡迎和包容性語言
* 尊重不同的觀點和經驗
* 優雅地接受建設性批評
* 關注在對於社群最好的事情上
* 對其他社群成員的表現友善
有助于创造正面环境的行为包括但不限于：
* 使用友好和包容性语言
* 尊重不同的观点和经历
* 耐心地接受建设性批评
* 关注对社区最有利的事情
* 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：
* 使用與性有關的言語或是圖像，以及不受歡迎的性騷擾
* 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
* 公開或私下的騷擾
* 未經許可地發布他人的個人資料，例如住址或是電子地址
* 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
身为参与者不能接受的行为包括但不限于：
* 使用与性有关的言语或是图像，以及不受欢迎的性骚扰
* 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
* 公开或私下的骚扰
* 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
可在此觀看https://www.contributor-covenant.org/zh-tw/version/1/4/code-of-conduct.html
本行为标准改编自[贡献者公约][主页]，版本 1.4
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 @ 773f2f73
# 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 @ 773f2f73
# 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 @ 773f2f73
# 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 @ 773f2f73
# 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 @ 773f2f73
# 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 @ 773f2f73
# 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 @ 773f2f73
# 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 @ 773f2f73
# 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/design/error_clip.md
浏览文件 @ 974183b4
......@@ -46,12 +46,12 @@ class ErrorClipByValue(BaseErrorClipAttr):
self.min = min
def append_clip_op(self, block, grad_name):
block.append_op(
type="clip",
inputs={"X": grad_name},
outputs={"Out": grad_name},
attrs={"min": self.min,
"max": self.max})
clip_op_desc = block.desc.append_op()
clip_op_desc.set_type("clip")
clip_op_desc.set_input("X", [grad_name])
clip_op_desc.set_output("Out", [grad_name])
clip_op_desc.set_attr("min", self.min)
clip_op_desc.set_attr("max", self.max)
```
The `BaseErrorClipAttr` have one main member functions: `append_clip_op(self, block, grad_name)`.
......@@ -80,6 +80,11 @@ def error_clip_callback(block, context):
op_desc.output_arg_names()):
fwd_var = block.var_recursive(grad_to_var[grad_n])
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:
error_clip.append_clip_op(block, grad_n)
```
......
doc/design/switch_kernel.md
浏览文件 @ 974183b4
## 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:
```
struct KernelType {
```cpp
struct OpKernelType {
Place place_;
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
......@@ -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.
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 $KT$.
## 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 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
using DataTransformationFN = std::function<void(const Tensor& in, Tensor* out)>;
using KernelTypePair = std::pair<KernelType, KernelType>;
map<KernelTypePair, DataTransformationFN> g_data_transformation_;
void OpWithKernel::Run() {
vec<Tensor> inputs = ...
auto actual_kernel_type = GetActualKernelType(inputs);
// The expected kernel type is related to actual kernel type.
// For the most operators, the expected kernel type is as same as
// actual kernel type.
//
// So we pass `actual_kernel_type` as a parameter of
// GetExpectedKernelType
auto expect_kernel_type = GetExpectedKernelType(actual_kernel_type);
auto trans = g_data_transformation_[{actual_kernel_type, expect_kernel_type}];
kernel.run(trans(inputs));
void OperatorWithKernel::Run(
const Scope& scope,
const platform::Place& place) const {
ExecutionContext ctx(...);
auto expected_kernel_key = this->GetExpectedKernelType(ctx);
Scope& new_scope = scope.NewScope();
for (auto& var_name : this->Inputs()) {
auto* tensor_in = GetTensor(var_name);
auto kernel_type_for_var = this->GetKernelTypeForVar(...);
if (kernel_type_for_var.place_ != expected_kernel_key.place_) {
auto* trans_var = new_scope.Var(var_name);
auto* out = DataTransform(expected_kernel_key,
kernel_type_for_var,
*tensor_in);
CopyVariableWithTensor(...);
}
}
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/usage/cluster/fluid_cluster_train_en.md
浏览文件 @ 974183b4
......@@ -103,7 +103,7 @@ t.transpile(optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2)
# in pserver, run this
exe.run(fluid.default_startup_program())
#current_endpoint here means current pserver IP:PORT you wish to run on
exe.run(t.get_pserver_program(current_endpoint, optimize_ops))
exe.run(t.get_pserver_program(current_endpoint))
# in trainer, run this
... # define data reader
......
doc/v1_api_tutorials/README.md 已删除 100644 → 0
浏览文件 @ 773f2f73
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 @ 773f2f73
# 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 @ 773f2f73
# 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 @ 773f2f73
# 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 @ 773f2f73
# 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 @ 773f2f73
# 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 @ 773f2f73
=============
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 @ 773f2f73
此差异已折叠。
paddle/inference/CMakeLists.txt
浏览文件 @ 974183b4
......@@ -8,27 +8,6 @@ cc_library(paddle_fluid_api
# Merge all modules into a simgle static library
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)
if(APPLE)
set(OPTIONAL_LINK_FLAGS)
......
paddle/inference/example.cc
浏览文件 @ 974183b4
......@@ -18,33 +18,21 @@ limitations under the License. */
#include "paddle/inference/inference.h"
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) {
google::ParseCommandLineFlags(&argc, &argv, true);
if (FLAGS_dirname.empty() || FLAGS_feed_var_names.empty() ||
FLAGS_fetch_var_names.empty()) {
if (FLAGS_dirname.empty()) {
// Example:
// ./example --dirname=recognize_digits_mlp.inference.model
// --feed_var_names="x"
// --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;
std::cout << "Usage: ./example --dirname=path/to/your/model" << std::endl;
exit(1);
}
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::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();
engine->LoadInferenceModel(dirname, feed_var_names, fetch_var_names);
engine->LoadInferenceModel(dirname);
paddle::framework::LoDTensor input;
srand(time(0));
......
paddle/inference/inference.cc
浏览文件 @ 974183b4
......@@ -25,19 +25,37 @@ limitations under the License. */
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(
const std::string& dirname,
const std::vector<std::string>& feed_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";
LOG(INFO) << "loading model from " << model_filename;
std::ifstream inputfs(model_filename, std::ios::in | std::ios::binary);
......@@ -48,7 +66,7 @@ void InferenceEngine::LoadInferenceModel(
LOG(INFO) << "program_desc_str's size: " << program_desc_str.size();
inputfs.read(&program_desc_str[0], program_desc_str.size());
inputfs.close();
#endif
program_ = new framework::ProgramDesc(program_desc_str);
GenerateLoadProgram(dirname);
......@@ -62,7 +80,7 @@ void InferenceEngine::LoadInferenceModel(
}
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
for (size_t i = 0; i < program_->Size(); ++i) {
const framework::BlockDesc& block = program_->Block(i);
......
paddle/inference/inference.h
浏览文件 @ 974183b4
......@@ -28,6 +28,7 @@ public:
delete load_program_;
}
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>& fetch_var_names);
......
paddle/operators/CMakeLists.txt
浏览文件 @ 974183b4
......@@ -178,14 +178,13 @@ foreach(src ${GENERAL_OPS})
endforeach()
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")
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(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_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)
if(WITH_GPU)
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 @ 974183b4
......@@ -29,7 +29,7 @@ void BeamSearch::operator()(const framework::LoDTensor &pre_ids,
PruneEndidCandidates(pre_ids, &selected_items);
// calculate the output tensor's height
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(); });
// the output tensor shape should be [num_instances, 1]
auto dims = framework::make_ddim(
......@@ -48,12 +48,20 @@ void BeamSearch::operator()(const framework::LoDTensor &pre_ids,
size_t low_offset = 0;
for (auto &items : selected_items) {
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) {
ids_data[low_offset] = item.id;
scores_data[low_offset] = item.score;
low_offset++;
}
}
low_level.push_back(low_offset);
// fill lod
auto abs_lod = framework::ToAbsOffset(ids_->lod());
auto &high_level = abs_lod[lod_level_];
......@@ -64,16 +72,21 @@ void BeamSearch::operator()(const framework::LoDTensor &pre_ids,
selected_scores->set_lod(lod);
}
void BeamSearch::PruneEndidCandidates(const framework::LoDTensor &pre_ids,
std::vector<std::vector<Item>> *items) {
int BeamSearch::PruneEndidCandidates(const framework::LoDTensor &pre_ids,
std::vector<std::vector<Item>> *items) {
auto *pre_ids_data = pre_ids.data<int64_t>();
int res = 0;
for (size_t offset = 0; offset < items->size(); offset++) {
auto prefix_id = pre_ids_data[offset];
if (prefix_id == end_id_) {
items->at(offset).clear();
} else {
res++;
}
}
return res;
}
std::vector<std::vector<BeamSearch::Item>> BeamSearch::ToMap(
......@@ -121,11 +134,7 @@ bool BeamSearch::NextItemSet(std::vector<BeamSearch::Item> *items) {
auto ids = *ids_;
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());
PADDLE_ENFORCE_GE(source_abs_two_level_lod.size(), 2UL);
auto *ids_data = ids.data<int64_t>();
auto *scores_data = scores.data<float>();
......
paddle/operators/beam_search_op.h
浏览文件 @ 974183b4
......@@ -73,7 +73,15 @@ namespace operators {
* second level:
* [0, 2, 4]
*
* tensor's data
* id tensor's data
* [[
* 4,
* 1,
* 3,
* 8,
* ]]
*
* score tensor's data
* [[
* 0.5,
* 0.3,
......@@ -137,16 +145,21 @@ class BeamSearch {
Item() {}
Item(size_t offset, size_t id, float score)
: offset(offset), id(id), score(score) {}
// offset in the lod_level_+1
// offset in the higher lod level.
size_t offset;
// // prefix id in the lower lod level.
// size_t prefix;
// the candidate id
id_t id;
// the corresponding 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.
......
paddle/operators/beam_search_op_test.cc 0 → 100644
浏览文件 @ 974183b4
/* 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/conv_op.cc
浏览文件 @ 974183b4
......@@ -70,6 +70,13 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType ConvOp::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const {
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_;
if (use_cudnn) {
library_ = framework::LibraryType::kCUDNN;
......@@ -283,6 +290,14 @@ void ConvOpGrad::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const {
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_;
if (use_cudnn) {
library_ = framework::LibraryType::kCUDNN;
......
paddle/operators/conv_transpose_op.cc
浏览文件 @ 974183b4
......@@ -61,6 +61,13 @@ void ConvTransposeOp::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType ConvTransposeOp::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const {
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_;
if (use_cudnn) {
library_ = framework::LibraryType::kCUDNN;
......@@ -263,6 +270,13 @@ void ConvTransposeOpGrad::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType ConvTransposeOpGrad::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const {
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_;
if (use_cudnn) {
library_ = framework::LibraryType::kCUDNN;
......
paddle/operators/math/sequence_padding.cc
浏览文件 @ 974183b4
......@@ -32,7 +32,8 @@ class PaddingLoDTensorFunctor<platform::CPUDeviceContext, T> {
framework::LoD abs_offset_lod = framework::ToAbsOffset(lod);
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 "
"equal to the sum of all sequences's length.");
......@@ -41,32 +42,32 @@ class PaddingLoDTensorFunctor<platform::CPUDeviceContext, T> {
"The input padding should be a 3-D Tensor of shape "
"[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,
"The first dimension of Tensor padding should be 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,
"The second dimension of Tensor padding should be 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,
"The third dimension of Tensor padding should be the "
"width of sequence in LoDTensor seq.");
const T* seq_data = seq.data<T>();
T* padding_data = padding.data<T>();
for (size_t i = 0; i < max_sequence_length; ++i) {
for (size_t j = 0; j < num_sequences; ++j) {
size_t start_pos = abs_offset_lod[level][j];
size_t sequence_length = abs_offset_lod[level][j + 1] - start_pos;
for (int64_t i = 0; i < max_sequence_length; ++i) {
for (int64_t j = 0; j < num_sequences; ++j) {
int64_t start_pos = abs_offset_lod[level][j];
int64_t sequence_length = abs_offset_lod[level][j + 1] - start_pos;
if (i < sequence_length) {
// i > 0 => sequence_length > 0
T scale =
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] =
seq_data[(start_pos + i) * sequence_width + k] * scale;
}
......@@ -93,7 +94,8 @@ class UnpaddingLoDTensorFunctor<platform::CPUDeviceContext, T> {
framework::LoD abs_offset_lod = framework::ToAbsOffset(lod);
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 "
"equal to the sum of all sequences's length.");
......@@ -102,31 +104,31 @@ class UnpaddingLoDTensorFunctor<platform::CPUDeviceContext, T> {
"The input padding should be a 3-D Tensor of shape "
"[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,
"The first dimension of Tensor padding should be "
"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,
"The second dimension of Tensor padding should be "
"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,
"The third dimension of Tensor padding should be the "
"width of sequence in LoDTensor seq.");
const T* padding_data = padding.data<T>();
T* seq_data = seq.data<T>();
for (size_t i = 0; i < num_sequences; ++i) {
size_t start_pos = abs_offset_lod[level][i];
size_t sequence_length = abs_offset_lod[level][i + 1] - start_pos;
for (size_t j = 0; j < sequence_length; ++j) {
for (int64_t i = 0; i < num_sequences; ++i) {
int64_t start_pos = abs_offset_lod[level][i];
int64_t sequence_length = abs_offset_lod[level][i + 1] - start_pos;
for (int64_t j = 0; j < sequence_length; ++j) {
// sequence_width > j > 0
T scale =
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] =
padding_data[(j * num_sequences + i) * sequence_width + k] *
scale;
......
paddle/operators/math/sequence_padding.cu
浏览文件 @ 974183b4
......@@ -71,7 +71,8 @@ class PaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
framework::LoD abs_offset_lod = framework::ToAbsOffset(lod);
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 "
"equal to the sum of all sequences's length.");
......@@ -80,17 +81,17 @@ class PaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
"The input padding should be a 3-D Tensor of shape "
"[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,
"The first dimension of Tensor padding should be 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,
"The second dimension of Tensor padding should be 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,
"The third dimension of Tensor padding should be the "
"width of sequence in LoDTensor seq.");
......@@ -101,7 +102,7 @@ class PaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
return;
}
const size_t kBlockSize = 512;
const int64_t kBlockSize = 512;
/* At least use 32 threads to copy sequence_width elements,
* and at least 8 elements for each thread.
......@@ -143,7 +144,8 @@ class UnpaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
framework::LoD abs_offset_lod = framework::ToAbsOffset(lod);
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 "
"equal to the sum of all sequences's length.");
......@@ -152,17 +154,17 @@ class UnpaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
"The input padding should be a 3-D Tensor of shape "
"[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,
"The first dimension of Tensor padding should be "
"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,
"The second dimension of Tensor padding should be "
"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,
"The third dimension of Tensor padding should be the "
"width of sequence in LoDTensor seq.");
......@@ -173,7 +175,7 @@ class UnpaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
return;
}
const size_t kBlockSize = 512;
const int64_t kBlockSize = 512;
/* At least use 32 threads to copy sequence_width elements,
* and at least 8 elements for each thread.
......
paddle/operators/math/sequence_padding_test.cc
浏览文件 @ 974183b4
......@@ -31,7 +31,7 @@ void TestSequencePadding(const paddle::framework::LoD& lod,
cpu_seq.set_lod(lod);
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);
}
......@@ -69,7 +69,7 @@ void TestSequencePadding(const paddle::framework::LoD& lod,
EXPECT_EQ(cpu_seq.numel(), cpu_seq_back.numel());
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]);
}
......
paddle/operators/pool_op.cc
浏览文件 @ 974183b4
......@@ -64,6 +64,13 @@ void PoolOp::InferShape(framework::InferShapeContext *ctx) const {
framework::OpKernelType PoolOp::GetExpectedKernelType(
const framework::ExecutionContext &ctx) const {
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_;
if (use_cudnn) {
library_ = framework::LibraryType::kCUDNN;
......@@ -88,6 +95,13 @@ void PoolOpGrad::InferShape(framework::InferShapeContext *ctx) const {
framework::OpKernelType PoolOpGrad::GetExpectedKernelType(
const framework::ExecutionContext &ctx) const {
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_;
if (use_cudnn) {
library_ = framework::LibraryType::kCUDNN;
......
paddle/operators/split_selected_rows_op.cc 0 → 100644
浏览文件 @ 974183b4
/* 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 @ 974183b4
/* 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 @ 974183b4
/* 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
python/paddle/v2/fluid/distribute_transpiler.py
浏览文件 @ 974183b4
......@@ -407,7 +407,7 @@ class DistributeTranspiler:
outputs=opt_op.outputs,
attrs=opt_op.attrs)
def get_pserver_program(self, endpoint, optimize_ops):
def get_pserver_program(self, endpoint):
"""
get pserver side program by endpoint
......@@ -422,9 +422,9 @@ class DistributeTranspiler:
self._clone_var(pserver_program.global_block(), v)
# step6
optimize_sub_program = Program()
for idx, opt_op in enumerate(optimize_ops):
is_op_on_pserver = self._is_op_on_pserver(endpoint, optimize_ops,
idx)
for idx, opt_op in enumerate(self.optimize_ops):
is_op_on_pserver = self._is_op_on_pserver(endpoint,
self.optimize_ops, idx)
if not is_op_on_pserver:
continue
if opt_op.inputs.has_key("Grad"):
......
python/paddle/v2/fluid/io.py
浏览文件 @ 974183b4
......@@ -15,6 +15,7 @@ import os
import cPickle as pickle
from paddle.v2.fluid.framework import Program, Parameter, default_main_program, Variable
from . import core
__all__ = [
'save_vars',
......@@ -191,6 +192,33 @@ def get_inference_program(target_vars, main_program=None):
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,
feeded_var_names,
target_vars,
......@@ -241,6 +269,9 @@ def save_inference_model(dirname,
"fetch_var_names": fetch_var_names
}, 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
# in another file: __model__.dat
with open(model_file_name + ".dat", "wb") as fp:
......
python/paddle/v2/fluid/layers/nn.py
浏览文件 @ 974183b4
......@@ -22,36 +22,13 @@ from ..param_attr import ParamAttr
from tensor import concat
__all__ = [
'fc',
'embedding',
'dynamic_lstm',
'gru_unit',
'linear_chain_crf',
'crf_decoding',
'cos_sim',
'cross_entropy',
'square_error_cost',
'accuracy',
'chunk_eval',
'sequence_conv',
'conv2d',
'sequence_pool',
'pool2d',
'batch_norm',
'beam_search_decode',
'conv2d_transpose',
'sequence_expand',
'lstm_unit',
'reduce_sum',
'reduce_mean',
'reduce_max',
'reduce_min',
'sequence_first_step',
'sequence_last_step',
'dropout',
'split',
'l2_normalize',
'matmul',
'fc', 'embedding', 'dynamic_lstm', 'gru_unit', 'linear_chain_crf',
'crf_decoding', 'cos_sim', 'cross_entropy', 'square_error_cost', 'accuracy',
'chunk_eval', 'sequence_conv', 'conv2d', 'sequence_pool', 'pool2d',
'batch_norm', 'beam_search_decode', 'conv2d_transpose', 'sequence_expand',
'lstm_unit', 'reduce_sum', 'reduce_mean', 'reduce_max', 'reduce_min',
'sequence_first_step', 'sequence_last_step', 'dropout', 'split',
'l2_normalize', 'matmul', 'warpctc'
]
......@@ -676,6 +653,7 @@ def conv2d(input,
groups=None,
param_attr=None,
bias_attr=None,
use_cudnn=True,
act=None):
"""
**Convlution2D Layer**
......@@ -739,6 +717,8 @@ def conv2d(input,
connected to the second half of the input channels. Default: groups=1
param_attr(ParamAttr): The parameters to 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
Returns:
......@@ -774,6 +754,8 @@ def conv2d(input,
stride = [stride, stride]
if isinstance(padding, int):
padding = [padding, padding]
if not isinstance(use_cudnn, bool):
raise ValueError("use_cudnn should be True or False")
input_shape = input.shape
filter_shape = [num_filters, num_filter_channels] + filter_size
......@@ -797,9 +779,12 @@ def conv2d(input,
'Filter': filter_param,
},
outputs={"Output": pre_bias},
attrs={'strides': stride,
'paddings': padding,
'groups': groups})
attrs={
'strides': stride,
'paddings': padding,
'groups': groups,
'use_cudnn': use_cudnn
})
pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2)
......@@ -948,6 +933,7 @@ def pool2d(input,
pool_stride=None,
pool_padding=None,
global_pooling=False,
use_cudnn=True,
name=None):
"""
This function adds the operator for pooling in 2 dimensions, using the
......@@ -967,6 +953,8 @@ def pool2d(input,
pool_stride = [pool_stride, pool_stride]
if isinstance(pool_padding, int):
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())
dtype = helper.input_dtype()
......@@ -981,7 +969,8 @@ def pool2d(input,
"ksize": pool_size,
"global_pooling": global_pooling,
"strides": pool_stride,
"paddings": pool_padding
"paddings": pool_padding,
"use_cudnn": use_cudnn
})
return pool_out
......@@ -1096,6 +1085,7 @@ def conv2d_transpose(input,
stride=None,
dilation=None,
param_attr=None,
use_cudnn=True,
name=None):
"""
The transpose of conv2d layer.
......@@ -1123,6 +1113,8 @@ def conv2d_transpose(input,
contain two integers, (dilation_H, dilation_W). Otherwise, the
dilation_H = dilation_W = dilation.
param_attr: Parameter Attribute.
use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn
library is installed. Default: True
name(str|None): A name for this layer(optional). If set None, the layer
will be named automatically.
......@@ -1151,6 +1143,10 @@ def conv2d_transpose(input,
elif dilation is not None:
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 output_size is None:
raise ValueError("output_size must be set when filter_size is None")
......@@ -1702,29 +1698,29 @@ def l2_normalize(x, axis, epsilon=1e-12, 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.
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:
- 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
: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
opposite: It is treated as :math:`[D, 1]` in nontransposed form and as
- 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
: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
opposite: It is treated as :math:`[D, 1]` in nontransposed form and as
: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.
- 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
last two dimensions and a batched matrix multiply supporting broadcast
- 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
applies on the two tensors.
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
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
removed after matrix multipication.
Args:
......@@ -1732,7 +1728,7 @@ def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
y (Variable): The input variable which is a Tensor or LoDTensor.
transpose_x (bool): Whether to transpose :math:`x` 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.
Returns:
......@@ -1769,3 +1765,56 @@ def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
attrs={'transpose_X': transpose_x,
'transpose_Y': transpose_y})
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
python/paddle/v2/fluid/nets.py
浏览文件 @ 974183b4
......@@ -28,19 +28,22 @@ def simple_img_conv_pool(input,
pool_stride,
act,
param_attr=None,
pool_type='max'):
pool_type='max',
use_cudnn=True):
conv_out = layers.conv2d(
input=input,
num_filters=num_filters,
filter_size=filter_size,
param_attr=param_attr,
act=act)
act=act,
use_cudnn=use_cudnn)
pool_out = layers.pool2d(
input=conv_out,
pool_size=pool_size,
pool_type=pool_type,
pool_stride=pool_stride)
pool_stride=pool_stride,
use_cudnn=use_cudnn)
return pool_out
......@@ -54,7 +57,8 @@ def img_conv_group(input,
conv_with_batchnorm=False,
conv_batchnorm_drop_rate=None,
pool_stride=1,
pool_type=None):
pool_type=None,
use_cudnn=True):
"""
Image Convolution Group, Used for vgg net.
"""
......@@ -85,7 +89,8 @@ def img_conv_group(input,
filter_size=conv_filter_size[i],
padding=conv_padding[i],
param_attr=param_attr[i],
act=local_conv_act)
act=local_conv_act,
use_cudnn=use_cudnn)
if conv_with_batchnorm[i]:
tmp = layers.batch_norm(input=tmp, act=conv_act)
......@@ -97,7 +102,8 @@ def img_conv_group(input,
input=tmp,
pool_size=pool_size,
pool_type=pool_type,
pool_stride=pool_stride)
pool_stride=pool_stride,
use_cudnn=use_cudnn)
return pool_out
......
python/paddle/v2/fluid/registry.py
浏览文件 @ 974183b4
# 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
# 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.
# 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 re
import cStringIO
import warnings
......@@ -167,13 +167,18 @@ def register_layer(op_type):
inputs[ipt.name] = val
outputs = dict()
out = helper.create_tmp_variable(dtype=dtype)
outputs[o_name] = [out]
out = kwargs.pop(_convert_(o_name), [])
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:
outputs[name] = [helper.create_tmp_variable(dtype=dtype)]
helper.append_op(
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.__doc__ = _generate_doc_string_(op_proto)
......
python/paddle/v2/fluid/tests/book_distribute/notest_dist_fit_a_line.py
浏览文件 @ 974183b4
......@@ -53,7 +53,7 @@ if training_role == "PSERVER":
if not current_endpoint:
print("need env SERVER_ENDPOINT")
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())
exe.run(pserver_prog)
else:
......
python/paddle/v2/fluid/tests/book_distribute/notest_dist_label_semantic_roles.py
浏览文件 @ 974183b4
......@@ -197,7 +197,7 @@ def main():
if not current_endpoint:
print("need env SERVER_ENDPOINT")
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())
exe.run(pserver_prog)
elif training_role == "TRAINER":
......
python/paddle/v2/fluid/tests/book_distribute/notest_dist_word2vec.py
浏览文件 @ 974183b4
......@@ -87,7 +87,7 @@ if training_role == "PSERVER":
if not current_endpoint:
print("need env SERVER_ENDPOINT")
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())
exe.run(pserver_prog)
elif training_role == "TRAINER":
......
python/paddle/v2/fluid/tests/book_distribute/notest_recognize_digits_conv_dist.py
浏览文件 @ 974183b4
......@@ -66,7 +66,7 @@ if training_role == "PSERVER":
if not current_endpoint:
print("need env SERVER_ENDPOINT")
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())
exe.run(pserver_prog)
elif training_role == "TRAINER":
......
adversarial/fluid_mnist.py → python/paddle/v2/fluid/tests/book_distribute/notest_recognize_digits_mlp_dist.py
浏览文件 @ 974183b4
......@@ -11,89 +11,77 @@
#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.
"""
CNN on mnist data using fluid api of paddlepaddle
"""
from __future__ import print_function
import numpy as np
import paddle.v2 as paddle
import paddle.v2.fluid as fluid
import os
BATCH_SIZE = 128
PASS_NUM = 100
def mnist_cnn_model(img):
"""
Mnist cnn model
images = fluid.layers.data(name='x', shape=[784], dtype='float32')
Args:
img(Varaible): the input image to be recognized
# TODO(aroraabhinav) Add regularization and error clipping after
# 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:
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')
label = fluid.layers.data(name='y', shape=[1], dtype='int64')
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')
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(x=cost)
logits = fluid.layers.fc(input=conv_pool_2, size=10, act='softmax')
return logits
optimizer = fluid.optimizer.Momentum(learning_rate=0.001, momentum=0.9)
optimize_ops, params_grads = optimizer.minimize(avg_cost)
accuracy = fluid.evaluator.Accuracy(input=predict, label=label)
def main():
"""
Train the cnn model on mnist datasets
"""
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)
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.mnist.train(), buf_size=8192),
batch_size=BATCH_SIZE)
accuracy = fluid.evaluator.Accuracy(input=logits, label=label)
place = fluid.CPUPlace()
exe = fluid.Executor(place)
BATCH_SIZE = 50
PASS_NUM = 3
ACC_THRESHOLD = 0.98
LOSS_THRESHOLD = 10.0
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.mnist.train(), buf_size=500),
batch_size=BATCH_SIZE)
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=2)
place = fluid.CPUPlace()
exe = fluid.Executor(place)
feeder = fluid.DataFeeder(feed_list=[img, label], place=place)
if training_role == "PSERVER":
if not current_endpoint:
print("need env SERVER_ENDPOINT")
exit(1)
pserver_prog = t.get_pserver_program(current_endpoint)
exe.run(fluid.default_startup_program())
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())
for pass_id in range(PASS_NUM):
accuracy.reset(exe)
batch_id = 0
for data in train_reader():
loss, acc = exe.run(fluid.default_main_program(),
loss, acc = exe.run(trainer_prog,
feed=feeder.feed(data),
fetch_list=[avg_cost] + accuracy.metrics)
pass_acc = accuracy.eval(exe)
print("pass_id=" + str(pass_id) + " acc=" + str(acc) + " pass_acc="
+ str(pass_acc))
if loss < LOSS_THRESHOLD and pass_acc > ACC_THRESHOLD:
break
if batch_id % 100 == 0:
print("batch_id %d, loss: %f, acc: %f" %
(batch_id, loss, pass_acc))
batch_id += 1
pass_acc = accuracy.eval(exe)
print("pass_id=" + str(pass_id) + " pass_acc=" + str(pass_acc))
fluid.io.save_params(
exe, dirname='./mnist', main_program=fluid.default_main_program())
print('train mnist done')
if __name__ == '__main__':
main()
else:
print("environment var TRAINER_ROLE should be TRAINER os PSERVER")
python/paddle/v2/fluid/tests/book_distribute/notest_understand_sentiment_conv_dist.py
浏览文件 @ 974183b4
......@@ -98,7 +98,7 @@ def main():
if not current_endpoint:
print("need env SERVER_ENDPOINT")
exit(1)
pserver_prog = t.get_pserver_program(current_endpoint, optimize_ops)
pserver_prog = t.get_pserver_program(current_endpoint)
exe.run(pserver_prog)
elif training_role == "TRAINER":
trainer_prog = t.get_trainer_program()
......
python/paddle/v2/fluid/tests/test_split_selected_rows_op.py 0 → 100644
浏览文件 @ 974183b4
# 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