diff --git a/.copyright.hook b/.copyright.hook index e28e88e2664b25b5aecce6660470c485b20a1a6e..2446e27248125134ab624ed557823993c90fafc5 100644 --- a/.copyright.hook +++ b/.copyright.hook @@ -49,12 +49,12 @@ def generate_copyright(template, lang='C'): LANG_COMMENT_MARK = "//" lines = template.split(NEW_LINE_MARK) - ans = LANG_COMMENT_MARK + COPYRIGHT_HEADER + NEW_LINE_MARK + ans = LANG_COMMENT_MARK + " " + COPYRIGHT_HEADER + NEW_LINE_MARK for lino, line in enumerate(lines): if lino == 0 or lino == 1 or lino == len(lines) - 1: continue - ans += LANG_COMMENT_MARK + line + NEW_LINE_MARK + ans += LANG_COMMENT_MARK + " " + line + NEW_LINE_MARK - return ans + return ans + "\n" def lang_type(filename): @@ -90,7 +90,7 @@ def main(argv=None): retv = 0 for filename in args.filenames: first_line = io.open(filename).readline() - if "Copyright" in first_line: continue + if "COPYRIGHT" in first_line.upper() : continue original_contents = io.open(filename).read() new_contents = generate_copyright( COPYRIGHT, lang_type(filename)) + original_contents diff --git a/CODE_OF_CONDUCT.md b/CODE_OF_CONDUCT.md new file mode 100644 index 0000000000000000000000000000000000000000..54131b48eca463aef817a4b96ba1b64de4b60aab --- /dev/null +++ b/CODE_OF_CONDUCT.md @@ -0,0 +1,46 @@ +# Contributor Covenant Code of Conduct + +## Our Pledge + +In the interest of fostering an open and welcoming environment, we as contributors and maintainers pledge to making participation in our project and our community a harassment-free experience for everyone, regardless of age, body size, disability, ethnicity, gender identity and expression, level of experience, nationality, personal appearance, race, religion, or sexual identity and orientation. + +## Our Standards + +Examples of behavior that contributes to creating a positive environment include: + +* Using welcoming and inclusive language +* Being respectful of differing viewpoints and experiences +* Gracefully accepting constructive criticism +* Focusing on what is best for the community +* Showing empathy towards other community members + +Examples of unacceptable behavior by participants include: + +* The use of sexualized language or imagery and unwelcome sexual attention or advances +* Trolling, insulting/derogatory comments, and personal or political attacks +* Public or private harassment +* Publishing others' private information, such as a physical or electronic address, without explicit permission +* Other conduct which could reasonably be considered inappropriate in a professional setting + +## Our Responsibilities + +Project maintainers are responsible for clarifying the standards of acceptable behavior and are expected to take appropriate and fair corrective action in response to any instances of unacceptable behavior. + +Project maintainers have the right and responsibility to remove, edit, or reject comments, commits, code, wiki edits, issues, and other contributions that are not aligned to this Code of Conduct, or to ban temporarily or permanently any contributor for other behaviors that they deem inappropriate, threatening, offensive, or harmful. + +## Scope + +This Code of Conduct applies both within project spaces and in public spaces when an individual is representing the project or its community. Examples of representing a project or community include using an official project e-mail address, posting via an official social media account, or acting as an appointed representative at an online or offline event. Representation of a project may be further defined and clarified by project maintainers. + +## Enforcement + +Instances of abusive, harassing, or otherwise unacceptable behavior may be reported by contacting the project team at paddle-dev@baidu.com. The project team will review and investigate all complaints, and will respond in a way that it deems appropriate to the circumstances. The project team is obligated to maintain confidentiality with regard to the reporter of an incident. Further details of specific enforcement policies may be posted separately. + +Project maintainers who do not follow or enforce the Code of Conduct in good faith may face temporary or permanent repercussions as determined by other members of the project's leadership. + +## Attribution + +This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4, available at [http://contributor-covenant.org/version/1/4][version] + +[homepage]: http://contributor-covenant.org +[version]: http://contributor-covenant.org/version/1/4/ diff --git a/CODE_OF_CONDUCT_cn.md b/CODE_OF_CONDUCT_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..2be794f1f324cf9b6bc304d4e5812076b56f4551 --- /dev/null +++ b/CODE_OF_CONDUCT_cn.md @@ -0,0 +1,50 @@ +# 参与者公约 + +## 我们的保证 + +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。 + +## 我们的标准 + +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 + +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 + +## 我们的责任 + +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。 + +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。 + +## 使用范围 + +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。 + +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。 + +## 强制执行 + +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。 + +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。 + +## 来源 + +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 + +[主页]: https://www.contributor-covenant.org diff --git a/adversarial/README.md b/adversarial/README.md deleted file mode 100644 index 51da21918a9d6e2192a2e03eabef4fde97896bc5..0000000000000000000000000000000000000000 --- a/adversarial/README.md +++ /dev/null @@ -1,9 +0,0 @@ -# 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. diff --git a/adversarial/advbox/__init__.py b/adversarial/advbox/__init__.py deleted file mode 100644 index f56f14f18dafdfe1e712cea178a63f09a087b587..0000000000000000000000000000000000000000 --- a/adversarial/advbox/__init__.py +++ /dev/null @@ -1,16 +0,0 @@ -# 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 -""" diff --git a/adversarial/advbox/attacks/base.py b/adversarial/advbox/attacks/base.py deleted file mode 100644 index 000baa48f626c7dddce49502d20c499f6424cd06..0000000000000000000000000000000000000000 --- a/adversarial/advbox/attacks/base.py +++ /dev/null @@ -1,52 +0,0 @@ -# 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 diff --git a/adversarial/advbox/attacks/gradientsign.py b/adversarial/advbox/attacks/gradientsign.py deleted file mode 100644 index 77d93bd793936abbaae0302050e2bcf714adfa1a..0000000000000000000000000000000000000000 --- a/adversarial/advbox/attacks/gradientsign.py +++ /dev/null @@ -1,51 +0,0 @@ -# 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 diff --git a/adversarial/advbox/models/__init__.py b/adversarial/advbox/models/__init__.py deleted file mode 100644 index eee0f6efd4774b42fcd082eb06d1398d2ee51bc4..0000000000000000000000000000000000000000 --- a/adversarial/advbox/models/__init__.py +++ /dev/null @@ -1,16 +0,0 @@ -# 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 -""" diff --git a/adversarial/advbox/models/base.py b/adversarial/advbox/models/base.py deleted file mode 100644 index 084e563f7b423f54f350b90649b04acc17b2db97..0000000000000000000000000000000000000000 --- a/adversarial/advbox/models/base.py +++ /dev/null @@ -1,103 +0,0 @@ -# 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 diff --git a/adversarial/advbox/models/paddle.py b/adversarial/advbox/models/paddle.py deleted file mode 100644 index 4048b47f897000c1b004cb05f8bbca985d5bbbb8..0000000000000000000000000000000000000000 --- a/adversarial/advbox/models/paddle.py +++ /dev/null @@ -1,114 +0,0 @@ -# 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 diff --git a/adversarial/fluid_mnist.py b/adversarial/fluid_mnist.py deleted file mode 100644 index f8c7fe8d0ef6a6b0756ef73c14d8937b9cd1a738..0000000000000000000000000000000000000000 --- a/adversarial/fluid_mnist.py +++ /dev/null @@ -1,99 +0,0 @@ -# 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. -""" -CNN on mnist data using fluid api of paddlepaddle -""" -import paddle.v2 as paddle -import paddle.v2.fluid as fluid - - -def mnist_cnn_model(img): - """ - Mnist cnn model - - Args: - img(Varaible): the input image to be recognized - - 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') - - 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(): - """ - 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) - - accuracy = fluid.evaluator.Accuracy(input=logits, label=label) - - 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) - - place = fluid.CPUPlace() - exe = fluid.Executor(place) - feeder = fluid.DataFeeder(feed_list=[img, label], place=place) - exe.run(fluid.default_startup_program()) - - for pass_id in range(PASS_NUM): - accuracy.reset(exe) - for data in train_reader(): - loss, acc = exe.run(fluid.default_main_program(), - 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 - - 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() diff --git a/adversarial/mnist_tutorial_fgsm.py b/adversarial/mnist_tutorial_fgsm.py deleted file mode 100644 index c63e030cd826abe24eacab21394c612b7c2d9495..0000000000000000000000000000000000000000 --- a/adversarial/mnist_tutorial_fgsm.py +++ /dev/null @@ -1,100 +0,0 @@ -# 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() diff --git a/doc/api/v2/fluid/layers.rst b/doc/api/v2/fluid/layers.rst index 62c154e65dcff1bdfb00109bf1b724c34731652e..0eb531cf021bd63c5ad93d9a1d12e9be811e4407 100644 --- a/doc/api/v2/fluid/layers.rst +++ b/doc/api/v2/fluid/layers.rst @@ -364,6 +364,12 @@ split .. autofunction:: paddle.v2.fluid.layers.split :noindex: + +matmul +------ +.. autofunction:: paddle.v2.fluid.layers.matmul + :noindex: + logsigmoid ---------- .. autofunction:: paddle.v2.fluid.layers.logsigmoid @@ -493,3 +499,8 @@ swish ------ .. autofunction:: paddle.v2.fluid.layers.swish :noindex: + +l2_normalize +------------ +.. autofunction:: paddle.v2.fluid.layers.l2_normalize + :noindex: diff --git a/doc/api/v2/fluid/nets.rst b/doc/api/v2/fluid/nets.rst index cca0dcdf082c0d809fab1aebba2c0b6c7b8efa2a..f6b1cb4ba10659fb336899f08376c265c67290f1 100644 --- a/doc/api/v2/fluid/nets.rst +++ b/doc/api/v2/fluid/nets.rst @@ -25,3 +25,9 @@ glu .. autofunction:: paddle.v2.fluid.nets.glu :noindex: + +dot_product_attention +--------------------- +.. autofunction:: paddle.v2.fluid.nets.dot_product_attention + :noindex: + diff --git a/doc/design/error_clip.md b/doc/design/error_clip.md index 8e845462cce2a29556bcb6010b08f00fbc3d99d7..58aa73b8cd38d01e2426278a3479714e4fb6a3b0 100644 --- a/doc/design/error_clip.md +++ b/doc/design/error_clip.md @@ -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) ``` diff --git a/doc/design/switch_kernel.md b/doc/design/switch_kernel.md index 1846e5d9f99dd433b44ac6b5ae52893ec8f0d451..9719e031c70979cd95400701efd30879662e19bc 100644 --- a/doc/design/switch_kernel.md +++ b/doc/design/switch_kernel.md @@ -1,21 +1,24 @@ ## 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; -using KernelTypePair = std::pair; - -map g_data_transformation_; - -void OpWithKernel::Run() { - vec 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) +``` diff --git a/doc/howto/dev/new_op_en.md b/doc/howto/dev/new_op_en.md index 7175d8370d6ce08c6d502eb42b8e53252db89bbb..da8b1bdd1082e439456daf25e9b3a1e8eb534375 100644 --- a/doc/howto/dev/new_op_en.md +++ b/doc/howto/dev/new_op_en.md @@ -4,7 +4,8 @@ - [Implementing C++ Types](#implementing-c-types) - [Defining ProtoMaker](#defining-protomaker) - [Defining Operator](#defining-operator) - - [Registering Operator](#registering-operator) + - [Defining OpKernel](#defining-opkernel) + - [Registering Operator and OpKernel](#registering-operator-and-opkernel) - [Compilation](#compilation) - [Python Binding](#python-binding) - [Unit Tests](#unit-tests) @@ -16,12 +17,13 @@ Here are the base types needed. For details, please refer to the design docs. -- `framework::OperatorBase`: Operator (Op)base class. -- `framework::OpKernel`: Base class for Op computation. -- `framework::OperatorWithKernel`: Inherited from OperatorBase, describing an operator with computation. - `class OpProtoAndCheckerMaker`: Describes an Operator's input, output, attributes and description, mainly used to interface with Python API. +- `framework::OperatorBase`: Operator (Op)base class. +- `framework::OpKernel`: Base class for Op computation kernel. +- `framework::OperatorWithKernel`: Inherited from OperatorBase, describing an operator with computation kernels. + -An operator can be differentiated by whether in has kernel methods. An operator with kernel inherits from `OperatorWithKernel` while the ones without inherit from `OperatorBase`. This tutorial focuses on implementing operators with kernels. In short, an operator includes the following information: +Operators can be categorized into two groups: operator with kernel(s) and operator without kernel(s). An operator with kernel(s) inherits from `OperatorWithKernel` while the one without kernel(s) inherits from `OperatorBase`. This tutorial focuses on implementing operators with kernels. In short, an operator includes the following information: Information | Where is it defined @@ -32,7 +34,7 @@ Kernel implementation | The kernel methods shared between CPU and CUDA are Registering the Op | Ops are registered in `.cc` files; For Kernel registration, `.cc` files contain the CPU implementation, while `.cu` files contain the CUDA implementation. -New Operator implementations are added to the list [paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators), with file names in the format `*_op.h` (if applicable), `*_op.cc`, `*_op.cu` (if applicable).** The system will use the naming scheme to automatically build operators and their corresponding Python extensions. ** +New Operator implementations are added to the list [paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators), with file names in the format `*_op.h` (if applicable), `*_op.cc`, `*_op.cu` (if applicable).** The system will use the naming scheme to automatically build operators and their corresponding Python extensions.** Let's take matrix multiplication operator, [MulOp](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc), as an example to introduce the writing of an Operator with Kernel. @@ -156,7 +158,8 @@ Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, w - `typename T` denotes data type, such as `float` or `double`. `MulKernel` types need to rewrite the interface for `Compute`. -- `Compute` takes one input variable `const framework::ExecutionContext& context`. + +- `Compute` takes one input parameter: `const framework::ExecutionContext& context`. - Compared with `InferShapeContext`, `ExecutionContext` includes device types, and can similarly extract input, output, and attribute variables. - `Compute` implements the computation logics of an `OpKernel`. @@ -177,7 +180,7 @@ Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, w }; ``` -Note that **different devices (CPU, CUDA)share an Op definition; whether or not they share the same `OpKernel` depends on whether `Compute` calls functions that support both devices.** +Note that **different devices (CPU, CUDA)share one Op definition; whether or not they share the same `OpKernel` depends on whether `Compute` calls functions can support both devices.** `MulOp`'s CPU and CUDA share the same `Kernel`. A non-sharing `OpKernel` example can be seen in [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43). @@ -188,13 +191,14 @@ This concludes the forward implementation of an operator. Next its operation and The definition of its corresponding backward operator, if applicable, is similar to that of an forward operator. **Note that a backward operator does not include a `ProtoMaker`**. -### Registering Operator +### Registering Operator and OpKernel - In `.cc` files, register forward and backward operator classes and the CPU kernel. ```cpp namespace ops = paddle::operators; REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad); + REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel); REGISTER_OP_CPU_KERNEL(mul_grad, ops::MulGradKernel); @@ -204,6 +208,7 @@ The definition of its corresponding backward operator, if applicable, is similar - `REGISTER_OP` registers the `ops::MulOp` class, type named `mul`, its type `ProtoMaker` is `ops::MulOpMaker`, registering `ops::MulOpGrad` as `mul_grad`. - `REGISTER_OP_WITHOUT_GRADIENT` registers an operator without gradient. + - `REGISTER_OP_CPU_KERNEL` registers `ops::MulKernel` class and specialized template types `paddle::platform::CPUPlace` and `float`, which also registers `ops::MulGradKernel`. @@ -225,6 +230,7 @@ The definition of its corresponding backward operator, if applicable, is similar Run the following commands to compile. ``` +# maybe you need to rerun cmake make mul_op ``` diff --git a/doc/howto/usage/capi/organization_of_the_inputs_cn.md b/doc/howto/usage/capi/organization_of_the_inputs_cn.md index 563ec5ca21ec5d75800fa201943d65e6d6fe51ea..a889ae4ffab7be02468b4a5ac5a18e3cc77803c9 100644 --- a/doc/howto/usage/capi/organization_of_the_inputs_cn.md +++ b/doc/howto/usage/capi/organization_of_the_inputs_cn.md @@ -19,7 +19,7 @@ ### 基本使用概念 -- 在PaddlePaddle内部，神经网络中一个计算层的输入/输出被组织为一个 `Argument` 结构体，如果神经网络有多个输入或者多个输入，每一个输入/输入都会对应有自己的`Argument`。 +- 在PaddlePaddle内部，神经网络中一个计算层的输入/输出被组织为一个 `Argument` 结构体，如果神经网络有多个输入或者多个输出，每一个输入/输出都会对应有自己的`Argument`。 - `Argument` 并不真正“存储”数据，而是将输入/输出信息有机地组织在一起。 - 在`Argument`内部由`IVector`（对应着上文提到的一维整型数组）和`Matrix`（对应着上文提到的二维浮点型矩阵）来实际存储数据；由 `Sequence Start Positions` (下文详细解释) 来描述输入/输出的序列信息。 diff --git a/doc/howto/usage/cluster/fluid_cluster_train_en.md b/doc/howto/usage/cluster/fluid_cluster_train_en.md new file mode 100644 index 0000000000000000000000000000000000000000..a64004a7c4ea12bc0d949d7f11f3e26af62bf912 --- /dev/null +++ b/doc/howto/usage/cluster/fluid_cluster_train_en.md @@ -0,0 +1,140 @@ +# Fluid Distributed Training + +## Introduction + +In this article, we'll explain how to config and run distributed training jobs with PaddlePaddle Fluid in a bare metal cluster. + +## Preparations + +### Get your cluster ready + +Prepare your computer nodes in the cluster. Nodes in this cluster can be of any specification that runs PaddlePaddle, and with a unique IP address assigned to it. Make sure they can communicate with each other. + +### Have PaddlePaddle installed + +PaddlePaddle must be installed on all nodes. If you have GPU cards on your nodes, be sure to properly install drivers and CUDA libraries. + +PaddlePaddle build and installation guide can be found from [here](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/index_en.html). + +### Update training script + +#### Non-cluster training script + +Let's take [Deep Learning 101](http://www.paddlepaddle.org/docs/develop/book/01.fit_a_line/index.html)'s first chapter: "fit a line" as an example. + +This demo's non-cluster version with fluid API is as follows: + +``` python +import paddle.v2 as paddle +import paddle.v2.fluid as fluid + +x = fluid.layers.data(name='x', shape=[13], dtype='float32') +y_predict = fluid.layers.fc(input=x, size=1, act=None) +y = fluid.layers.data(name='y', shape=[1], dtype='float32') + +cost = fluid.layers.square_error_cost(input=y_predict, label=y) +avg_cost = fluid.layers.mean(x=cost) + +sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001) +sgd_optimizer.minimize(avg_cost) + +BATCH_SIZE = 20 + +train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.uci_housing.train(), buf_size=500), + batch_size=BATCH_SIZE) + +place = fluid.CPUPlace() +feeder = fluid.DataFeeder(place=place, feed_list=[x, y]) +exe = fluid.Executor(place) + +exe.run(fluid.default_startup_program()) + +PASS_NUM = 100 +for pass_id in range(PASS_NUM): + fluid.io.save_persistables(exe, "./fit_a_line.model/") + fluid.io.load_persistables(exe, "./fit_a_line.model/") + for data in train_reader(): + avg_loss_value, = exe.run(fluid.default_main_program(), + feed=feeder.feed(data), + fetch_list=[avg_cost]) + + if avg_loss_value[0] < 10.0: + exit(0) # if avg cost less than 10.0, we think our code is good. +exit(1) +``` + +We created a simple fully connected neural networks training program and handed it to the fluid executor to run for 100 passes. + +Now let's try to convert it to a distributed version to run in a cluster. + +#### Introducing parameter server + +As you see from the non-cluster version of training script, there is only one role in it: the trainer, who does the computing as well as holding parameters. In cluster training, since multi-trainers are working on the same task, they need one centralized place to hold and distribute parameters. This centralized place is called the Parameter Server in PaddlePaddle. + +![parameter server architect](src/trainer.png) + +Parameter Server in fluid does not only hold parameters but is also assigned with a part of the program. Trainers communicate with parameter servers via send/receive OPs. For more tech detail, please refer to this [document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/dist_refactor/distributed_architecture.md). + +Now we need to create program for both trainers and parameter servers, the question is how? + +#### Slice the program + +Fluid provides a tool called "Distribute Transpiler" to automatically convert the non-cluster program into cluster program. + +The idea behind this tool is to find optimize OPs and gradient parameters, slice the program into 2 pieces and connect them with send/receive OP. + +Optimize OPs and gradient parameters can be found from the return values of optimizer's minimize function. + +To put them together: + +``` python +... #define the program, cost, and create sgd optimizer + +optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) #get optimize OPs and gradient parameters + +t = fluid.DistributeTranspiler() # create transpiler instance +# slice the program into 2 pieces with optimizer_ops and gradient parameters list, as well as pserver_endpoints, which is a comma separated list of [IP:PORT] and number of trainers +t.transpile(optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2) + +... #create executor + +# in pserver, run this +#current_endpoint here means current pserver IP:PORT you wish to run on +pserver_prog = t.get_pserver_program(current_endpoint) +pserver_startup = t.get_startup_program(current_endpoint, pserver_prog) +exe.run(pserver_startup) +exe.run(pserver_prog) + +# in trainer, run this +... # define data reader +exe.run(fluid.default_startup_program()) +for pass_id in range(100): + for data in train_reader(): + exe.run(t.get_trainer_program()) + + +``` + +### E2E demo + +Please find the complete demo from [here](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/tests/book_distribute/notest_dist_fit_a_line.py). In parameter server node run this in the command line: + +``` bash +PSERVERS=192.168.1.2:6174 SERVER_ENDPOINT=192.168.1.2:6174 TRAINING_ROLE=PSERVER python notest_dist_fit_a_line.py +``` + +*please note we assume that your parameter server runs at 192.168.1.2:6174* + +Wait until the prompt `Server listening on 192.168.1.2:6174` + +Then in 2 of your trainer node run this: + +``` bash +PSERVERS=192.168.1.2:6174 SERVER_ENDPOINT=192.168.1.2:6174 TRAINING_ROLE=TRAINER python notest_dist_fit_a_line.py +``` + +*the reason you need to run this command twice in 2 nodes is: in the script we set the trainer count to be 2. You can change this setting on line 50* + +Now you have 2 trainers and 1 parameter server up and running. diff --git a/doc/v1_api_tutorials/README.md b/doc/v1_api_tutorials/README.md deleted file mode 100644 index 071b8da61fbcab3e88819273008b4526546202ad..0000000000000000000000000000000000000000 --- a/doc/v1_api_tutorials/README.md +++ /dev/null @@ -1,5 +0,0 @@ -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. diff --git a/doc/v1_api_tutorials/embedding_model/index_cn.md b/doc/v1_api_tutorials/embedding_model/index_cn.md deleted file mode 100644 index 2b4a79fbbfc0c4af74aa73c540919f5d9cf2635b..0000000000000000000000000000000000000000 --- a/doc/v1_api_tutorials/embedding_model/index_cn.md +++ /dev/null @@ -1,139 +0,0 @@ -# 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： - - ``: 分词序列的开始 - - ``: 分词序列的结束 - - `PALCEHOLDER_JUST_IGNORE_THE_EMBEDDING`: 占位符，没有实际意义 - - ``: 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。 -
![](./neural-n-gram-model.png)
-
Figure 1. neural-n-gram-model
- -### 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 diff --git a/doc/v1_api_tutorials/embedding_model/index_en.md b/doc/v1_api_tutorials/embedding_model/index_en.md deleted file mode 100644 index 9525f64f9b5384c8e44690fb0887fb2293108e0a..0000000000000000000000000000000000000000 --- a/doc/v1_api_tutorials/embedding_model/index_en.md +++ /dev/null @@ -1,140 +0,0 @@ -# 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: - - ``: the start of a sequence - - ``: the end of a sequence - - `PALCEHOLDER_JUST_IGNORE_THE_EMBEDDING`: a placeholder, just ignore it and its embedding - - ``: 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. -
![](./neural-n-gram-model.png)
-
Figure 1. neural-n-gram-model
- -### 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 `` = 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 ',' diff --git a/doc/v1_api_tutorials/embedding_model/neural-n-gram-model.png b/doc/v1_api_tutorials/embedding_model/neural-n-gram-model.png deleted file mode 100644 index f70b765b3fd69816345a79fc59adfea46008dbfd..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/embedding_model/neural-n-gram-model.png and /dev/null differ diff --git a/doc/v1_api_tutorials/gan/gan.png b/doc/v1_api_tutorials/gan/gan.png deleted file mode 100644 index 0eafd7cb49b545f412f8e775804bcd0b22c42454..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/gan/gan.png and /dev/null differ diff --git a/doc/v1_api_tutorials/gan/index_en.md b/doc/v1_api_tutorials/gan/index_en.md deleted file mode 100644 index ac9ed37b2264778869f92c0910b1cb946fb4427f..0000000000000000000000000000000000000000 --- a/doc/v1_api_tutorials/gan/index_en.md +++ /dev/null @@ -1,137 +0,0 @@ -# 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. - -
![](./gan.png)
-

- Figure 1. GAN-Model-Structure - figure credit -

- -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. - -
![](./uniform_sample.png)
-

- Figure 2. Uniform Sample -

- -## 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. -
![](./mnist_sample.png)
-

- Figure 3. MNIST Sample -

diff --git a/doc/v1_api_tutorials/gan/mnist_sample.png b/doc/v1_api_tutorials/gan/mnist_sample.png deleted file mode 100644 index f9c7bf7ddd7f148eac4fe347e9c38afaa8876760..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/gan/mnist_sample.png and /dev/null differ diff --git a/doc/v1_api_tutorials/gan/uniform_sample.png b/doc/v1_api_tutorials/gan/uniform_sample.png deleted file mode 100644 index e716c48e782019a757bed0cb443f2ed97386cbe2..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/gan/uniform_sample.png and /dev/null differ diff --git a/doc/v1_api_tutorials/imagenet_model/resnet_block.jpg b/doc/v1_api_tutorials/imagenet_model/resnet_block.jpg deleted file mode 100644 index e16bd3c624030c4c09b358a015b491141b42d8f1..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/imagenet_model/resnet_block.jpg and /dev/null differ diff --git a/doc/v1_api_tutorials/imagenet_model/resnet_model_cn.md b/doc/v1_api_tutorials/imagenet_model/resnet_model_cn.md deleted file mode 100644 index 82ec9d70b345c11aba3aa86f8206eedc8072bb88..0000000000000000000000000000000000000000 --- a/doc/v1_api_tutorials/imagenet_model/resnet_model_cn.md +++ /dev/null @@ -1,284 +0,0 @@ -# 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。 - -
![resnet_block](./resnet_block.jpg)
-
图 1. ResNet 网络模块
- -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。 -
- ----- - - - - - - - - - - - - - - - - - - - - - - - - - - -
ResNetTop-1Model Size
ResNet-5024.9%99M
ResNet-10123.7%173M
ResNet-15223.2%234M
-
- -## 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]() 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。 -
- ----- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
参数名尺寸含义
_res2_1_branch1_bn.w0256gamma, 缩放参数
_res2_1_branch1_bn.w1256特征图均值
_res2_1_branch1_bn.w2256特征图方差
_res2_1_branch1_bn.wbias256beta, 偏置参数
-
- -### 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 -``` diff --git a/doc/v1_api_tutorials/imagenet_model/resnet_model_en.md b/doc/v1_api_tutorials/imagenet_model/resnet_model_en.md deleted file mode 100644 index 478ad06193b14ba7fe02238df621db1f7b0804d4..0000000000000000000000000000000000000000 --- a/doc/v1_api_tutorials/imagenet_model/resnet_model_en.md +++ /dev/null @@ -1,284 +0,0 @@ -# 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 . - -
![resnet_block](./resnet_block.jpg)
-
Figure 1. ResNet Block
- -We present three ResNet models, which are converted from the models provided by the authors . 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. -
- ----- - - - - - - - - - - - - - - - - - - - - - - - - - - -
ResNetTop-1Model Size
ResNet-5024.9%99M
ResNet-10123.7%173M
ResNet-15223.2%234M
-
- -## 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]() 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. -
- ----- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Parameter NameNumberMeaning
_res2_1_branch1_bn.w0256gamma, scale parameter
_res2_1_branch1_bn.w1256mean value of feature map
_res2_1_branch1_bn.w2256variance of feature map
_res2_1_branch1_bn.wbias256beta, shift parameter
-
- -### 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 -``` diff --git a/doc/v1_api_tutorials/quick_start/index_cn.rst b/doc/v1_api_tutorials/quick_start/index_cn.rst deleted file mode 100644 index d565fcf95ef8489eb22a5a1b5a552b5336f4e371..0000000000000000000000000000000000000000 --- a/doc/v1_api_tutorials/quick_start/index_cn.rst +++ /dev/null @@ -1,397 +0,0 @@ -============= -快速入门教程 -============= - -我们将以 `文本分类问题 `_ 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 `_ 分为好评(正样本)和差评(负样本)两种类别。 -`源代码 `_ 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 `_ 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 `_ 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 `_, `GRU模型 `_ 和 `LSTM模型 `_ 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 `_ 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 - =========================================== ============================================================== diff --git a/doc/v1_api_tutorials/quick_start/index_en.md b/doc/v1_api_tutorials/quick_start/index_en.md deleted file mode 100644 index ca110431cf921ae0480d3fb2b17c58f90a84cc0e..0000000000000000000000000000000000000000 --- a/doc/v1_api_tutorials/quick_start/index_en.md +++ /dev/null @@ -1,562 +0,0 @@ -# Quick Start - -This tutorial will teach the basics of deep learning (DL), including how to implement many different models in PaddlePaddle. You will learn how to: - - Prepare data into the standardized format that PaddlePaddle accepts. - - Write data providers that read data into PaddlePaddle. - - Configure neural networks in PaddlePaddle layer by layer. - - Train models. - - Perform inference with trained models. - - -## Install - -To get started, please install PaddlePaddle on your computer. Throughout this tutorial, you will learn by implementing different DL models for text classification. - -To install PaddlePaddle, please follow the instructions here: Build and Install. - -## Overview -For the first step, you will use PaddlePaddle to build a **text classification** system. For example, suppose you run an e-commence website, and you want to analyze the sentiment of user reviews to evaluate product quality. - -For example, given the input - -``` -This monitor is fantastic. -``` - -Your classifier should output “positive”, since this text snippet shows that the user is satisfied with the product. Given this input: - -``` -The monitor breaks down two months after purchase. -``` - -the classifier should output “negative“. - -To build your text classification system, your code will need to perform five steps: -
![](./src/Pipeline_en.jpg)
- - - Preprocess data into a standardized format. - - Provide data to the learning model. - - Specify the neural network structure. - - Train the model. - - Inference (make prediction on test examples). - - -1. Preprocess data into standardized format - - In the text classification example, you will start with a text file with one training example per line. Each line contains category id (in machine learning, often denoted the target y), followed by the input text (often denoted x); these two elements are separated by a Tab. For example: ```positive [tab] This monitor is fantastic```. You will preprocess this raw data into a format that Paddle can use. - -2. Provide data to the learning model. - - You can write data providers in Python. For any required data preprocessing step, you can add the preprocessing code to the PyDataProvider Python file. - - In our text classification example, every word or character will be converted into an integer id, specified in a dictionary file. It perform a dictionary lookup in PyDataProvider to get the id. -3. Specify neural network structure. (From easy to hard, we provide 4 kinds of network configurations) - - A logistic regression model. - - A word embedding model. - - A convolutional neural network model. - - A sequential recurrent neural network model. - - You will also learn different learning algorithms. -4. Training model. -5. Inference. - -## Preprocess data into standardized format -In this example, you are going to use [Amazon electronic product review dataset](http://jmcauley.ucsd.edu/data/amazon/) to build a bunch of deep neural network models for text classification. Each text in this dataset is a product review. This dataset has two categories: “positive” and “negative”. Positive means the reviewer likes the product, while negative means the reviewer does not like the product. - -`demo/quick_start` in the [source code](https://github.com/PaddlePaddle/Paddle) provides script for downloading the preprocessed data as shown below. (If you want to process the raw data, you can use the script `demo/quick_start/data/proc_from_raw_data/get_data.sh`). - -```bash -cd demo/quick_start -./data/get_data.sh -``` - -## Transfer Data to Model -### Write Data Provider with Python -The following `dataprovider_bow.py` gives a complete example of writing data provider with Python. It includes the following parts: - -* initalizer： define the additional meta-data of the data provider and the types of the input data. -* process： Each `yield` returns a data sample. In this case, it return the text representation and category id. The order of features in the returned result needs to be consistent with the definition of the input types in `initalizer`. - -```python -from paddle.trainer.PyDataProvider2 import * - -# id of the word not in dictionary -UNK_IDX = 0 - -# initializer is called by the framework during initialization. -# It allows the user to describe the data types and setup the -# necessary data structure for later use. -# `settings` is an object. initializer need to properly fill settings.input_types. -# initializer can also store other data structures needed to be used at process(). -# In this example, dictionary is stored in settings. -# `dictionay` and `kwargs` are arguments passed from trainer_config.lr.py -def initializer(settings, dictionary, **kwargs): - # Put the word dictionary into settings - settings.word_dict = dictionary - - # setting.input_types specifies what the data types the data provider - # generates. - settings.input_types = [ - # The first input is a sparse_binary_vector, - # which means each dimension of the vector is either 0 or 1. It is the - # bag-of-words (BOW) representation of the texts. - sparse_binary_vector(len(dictionary)), - # The second input is an integer. It represents the category id of the - # sample. 2 means there are two labels in the dataset. - # (1 for positive and 0 for negative) - integer_value(2)] - -# Delaring a data provider. It has an initializer 'data_initialzer'. -# It will cache the generated data of the first pass in memory, so that -# during later pass, no on-the-fly data generation will be needed. -# `setting` is the same object used by initializer() -# `file_name` is the name of a file listed train_list or test_list file given -# to define_py_data_sources2(). See trainer_config.lr.py. -@provider(init_hook=initializer, cache=CacheType.CACHE_PASS_IN_MEM) -def process(settings, file_name): - # Open the input data file. - with open(file_name, 'r') as f: - # Read each line. - for line in f: - # Each line contains the label and text of the comment, separated by \t. - label, comment = line.strip().split('\t') - - # Split the words into a list. - words = comment.split() - - # convert the words into a list of ids by looking them up in word_dict. - word_vector = [settings.word_dict.get(w, UNK_IDX) for w in words] - - # Return the features for the current comment. The first is a list - # of ids representing a 0-1 binary sparse vector of the text, - # the second is the integer id of the label. - yield word_vector, int(label) -``` - -### Define Python Data Provider in Configuration files. -You need to add a data provider definition `define_py_data_sources2` in our network configuration. This definition specifies: - -- The path of the training and testing data (`data/train.list`, `data/test.list`). -- The location of the data provider file (`dataprovider_bow`). -- The function to call to get data. (`process`). -- Additional arguments or data. Here it passes the path of word dictionary. - -```python -from paddle.trainer_config_helpers import * - -file = "data/dict.txt" -word_dict = dict() -with open(dict_file, 'r') as f: - for i, line in enumerate(f): - w = line.strip().split()[0] - word_dict[w] = i -# define the data sources for the model. -# We need to use different process for training and prediction. -# For training, the input data includes both word IDs and labels. -# For prediction, the input data only includs word Ids. -define_py_data_sources2(train_list='data/train.list', - test_list='data/test.list', - module="dataprovider_bow", - obj="process", - args={"dictionary": word_dict}) -``` -You can refer to the following link for more detailed examples and data formats: PyDataProvider2. - -## Network Architecture -We will describe four kinds of network architectures in this section. -
![](./src/PipelineNetwork_en.jpg)
- -First, you will build a logistic regression model. Later, you will also get chance to build other more powerful network architectures. -For more detailed documentation, you could refer to: layer documentation. All configuration files are in `demo/quick_start` directory. - -### Logistic Regression -The architecture is illustrated in the following picture: -
![](./src/NetLR_en.png)
- -- You need define the data for text features. The size of the data layer is the number of words in the dictionary. - -```python -word = data_layer(name="word", size=voc_dim) -``` - -- You also need to define the category id for each example. The size of the data layer is the number of labels. - -```python -label = data_layer(name="label", size=label_dim) -``` - -- It uses logistic regression model to classify the vector, and it will output the classification error during training. - - Each layer has an *input* argument that specifies its input layer. Some layers can have multiple input layers. You can use a list of the input layers as input in that case. - - *size* for each layer means the number of neurons of the layer. - - *act_type* means activation function applied to the output of each neuron independently. - - Some layers can have additional special inputs. For example, `classification_cost` needs ground truth label as input to compute classification loss and error. -```python -# Define a fully connected layer with logistic activation (also called softmax activation). -output = fc_layer(input=word, - size=label_dim, - act_type=SoftmaxActivation()) -# Define cross-entropy classification loss and error. -classification_cost(input=output, label=label) -``` - -Performance summary: You can refer to the training and testing scripts later. In order to compare different network architectures, the model complexity and test classification error are listed in the following table: - - -
- - - - - - - - - - - - - - - - - -
Network nameNumber of parametersTest error
Logistic regression252 KB8.652%
- -
- -### Word Embedding Model -In order to use the word embedding model, you need to change the data provider a little bit to make the input words as a sequence of word IDs. The revised data provider `dataprovider_emb.py` is listed below. You only need to change initializer() for the type of the first input. It is changed from sparse_binary_vector to sequence of intergers. process() remains the same. This data provider can also be used for later sequence models. - -```python -def initializer(settings, dictionary, **kwargs): - # Put the word dictionary into settings - settings.word_dict = dictionary - settings.input_types = [ - # Define the type of the first input as a sequence of integers. - integer_value_sequence(len(dictionary)), - # Define the second input for label id - integer_value(2)] - -@provider(init_hook=initializer) -def process(settings, file_name): - ... - # omitted, it is same as the data provider for LR model -``` - -This model is very similar to the framework of logistic regression, but it uses word embedding vectors instead of a sparse vectors to represent words. -
![](./src/NetContinuous_en.png)
- -- It can look up the dense word embedding vector in the dictionary (its words embedding vector is `word_dim`). The input is a sequence of N words, the output is N word_dim dimensional vectors. - -```python -emb = embedding_layer(input=word, dim=word_dim) -``` - -- It averages all the word embedding in a sentence to get its sentence representation. - -```python -avg = pooling_layer(input=emb, pooling_type=AvgPooling()) -``` - -The other parts of the model are the same as logistic regression network. - -The performance is summarized in the following table: - - -
- - - - - - - - - - - - - - - - - -
Network nameNumber of parametersTest error
Word embedding model15 MB8.484%
-
-
- -### Convolutional Neural Network Model -Convolutional neural network converts a sequence of word embeddings into a sentence representation using temporal convolutions. You will transform the fully connected layer of the word embedding model to 3 new sub-steps. -
![](./src/NetConv_en.png)
- - -Text convolution has 3 steps: -1. Get K nearest neighbor context of each word in a sentence, stack them into a 2D vector representation. -2. Apply temporal convolution to this representation to produce a new hidden_dim dimensional vector. -3. Apply max-pooling to the new vectors at all the time steps in a sentence to get a sentence representation. - -```python -# context_len means convolution kernel size. -# context_start means the start of the convolution. It can be negative. In that case, zero padding is applied. -text_conv = sequence_conv_pool(input=emb, - context_start=k, - context_len=2 * k + 1) -``` - -The performance is summarized in the following table： - - -
- - - - - - - - - - - - - - - - - -
Network nameNumber of parametersTest error
Convolutional model16 MB5.628%
-
- -### Recurrent Model -
![](./src/NetRNN_en.png)
- -You can use Recurrent neural network as our time sequence model, including simple RNN model, GRU model, and LSTM model。 - -- GRU model can be specified via： - -```python -gru = simple_gru(input=emb, size=gru_size) -``` - -- LSTM model can be specified via： - -```python -lstm = simple_lstm(input=emb, size=lstm_size) -``` - -You can use single layer LSTM model with Dropout for our text classification problem. The performance is summarized in the following table: - - -
- - - - - - - - - - - - - - - - - -
Network nameNumber of parametersTest error
Recurrent model16 MB4.812%
- -
- -## Optimization Algorithm -Optimization algorithms include Momentum, RMSProp, AdaDelta, AdaGrad, Adam, and Adamax. You can use Adam optimization method here, with L2 regularization and gradient clipping, because Adam has been proved to work very well for training recurrent neural network. - -```python -settings(batch_size=128, - learning_rate=2e-3, - learning_method=AdamOptimizer(), - regularization=L2Regularization(8e-4), - gradient_clipping_threshold=25) -``` - -## Training Model -After completing data preparation and network architecture specification, you will run the training script. -
![](./src/PipelineTrain_en.png)
- -Training script: our training script is in `train.sh` file. The training arguments are listed below: - -```bash -paddle train \ ---config=trainer_config.py \ ---log_period=20 \ ---save_dir=./output \ ---num_passes=15 \ ---use_gpu=false -``` - -We do not provide examples on how to train on clusters here. If you want to train on clusters, please follow the distributed training documentation or other demos for more details. - -## Inference -You can use the trained model to perform prediction on the dataset with no labels. You can also evaluate the model on dataset with labels to obtain its test accuracy. -
![](./src/PipelineTest_en.png)
- -The test script is listed below. PaddlePaddle can evaluate a model on the data with labels specified in `test.list`. - -```bash -paddle train \ ---config=trainer_config.lstm.py \ ---use_gpu=false \ ---job=test \ ---init_model_path=./output/pass-0000x -``` - -We will give an example of performing prediction using Recurrent model on a dataset with no labels. You can refer to Python Prediction API tutorial，or other demo for the prediction process using Python. You can also use the following script for inference or evaluation. - -inference script (predict.sh)： - -```bash -model="output/pass-00003" -paddle train \ - --config=trainer_config.lstm.py \ - --use_gpu=false \ - --job=test \ - --init_model_path=$model \ - --config_args=is_predict=1 \ - --predict_output_dir=. \ - -mv rank-00000 result.txt -``` -User can choose the best model base on the training log instead of model `output/pass-00003`. There are several differences between training and inference network configurations. -- You do not need labels during inference. -- Outputs need to be specified to the classification probability layer (the output of softmax layer), or the id of maximum probability (`max_id` layer). An example to output the id and probability is given in the code snippet. -- batch_size = 1. -- You need to specify the location of `test_list` in the test data. - -The results in `result.txt` is as follows, each line is one sample. - -``` -predicted_label_id;probability_of_label_0 probability_of_label_1 # the first sample -predicted_label_id;probability_of_label_0 probability_of_label_1 # the second sample -``` - - -```python -is_predict = get_config_arg('is_predict', bool, False) -trn = 'data/train.list' if not is_predict else None -tst = 'data/test.list' if not is_predict else 'data/pred.list' -obj = 'process' if not is_predict else 'process_pre' -batch_size = 128 if not is_predict else 1 -if is_predict: - maxid = maxid_layer(output) - outputs([maxid,output]) -else: - label = data_layer(name="label", size=2) - cls = classification_cost(input=output, label=label) outputs(cls) -``` - -## Summary -The scripts of data downloading, network configurations, and training scrips are in `/demo/quick_start`. The following table summarizes the performance of our network architecture on Amazon-Elec dataset(25k): - -
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Network nameNumber of parametersError rateConfiguration file name
Logistic regression model(BOW) 252KB 8.652%trainer_config.lr.py
Word embedding 15MB 8.484%trainer_config.emb.py
Convolution model 16MB 5.628%trainer_config.cnn.py
Time sequence model 16MB 4.812%trainer_config.lstm.py
-
-
- -## Appendix -### Command Line Argument - -* \--config：network architecture path. -* \--save_dir：model save directory. -* \--log_period：the logging period per batch. -* \--num_passes：number of training passes. One pass means the training would go over the whole training dataset once. -* \--config_args：Other configuration arguments. -* \--init_model_path：The path of the initial model parameter. - -By default, the trainer will save model every pass. You can also specify `saving_period_by_batches` to set the frequency of batch saving. You can use `show_parameter_stats_period` to print the statistics of the parameters, which are very useful for tuning parameters. Other command line arguments can be found in command line argument documentation。 - -### Log - -``` -TrainerInternal.cpp:160] Batch=20 samples=2560 AvgCost=0.628761 CurrentCost=0.628761 Eval: classification_error_evaluator=0.304297 CurrentEval: classification_error_evaluator=0.304297 -``` -During model training, you will see the log like the examples above: -
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
NameExplanation
Batch=20 You have trained 20 batches.
samples=2560 You have trained 2560 examples.
AvgCost The average cost from the first batch to the current batch.
CurrentCost the average cost of the last log_period batches
Eval: classification_error_evaluator The average classification error from the first batch to the current batch.
CurrentEval: classification_error_evaluator The average error rate of the last log_period batches
-
-
diff --git a/doc/v1_api_tutorials/quick_start/src/NetContinuous_cn.jpg b/doc/v1_api_tutorials/quick_start/src/NetContinuous_cn.jpg deleted file mode 100755 index b18e452a481232f47c453fc30521f8c78ad2da0b..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/quick_start/src/NetContinuous_cn.jpg and /dev/null differ diff --git a/doc/v1_api_tutorials/quick_start/src/NetContinuous_en.png b/doc/v1_api_tutorials/quick_start/src/NetContinuous_en.png deleted file mode 100644 index 7bdef1aa366711806585d35c8653c987fd63d59e..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/quick_start/src/NetContinuous_en.png and /dev/null differ diff --git a/doc/v1_api_tutorials/quick_start/src/NetConv_cn.jpg b/doc/v1_api_tutorials/quick_start/src/NetConv_cn.jpg deleted file mode 100755 index 0f5ebfa52ff2abcfc63464d48f00e406d4cbfe86..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/quick_start/src/NetConv_cn.jpg and /dev/null differ diff --git a/doc/v1_api_tutorials/quick_start/src/NetConv_en.png b/doc/v1_api_tutorials/quick_start/src/NetConv_en.png deleted file mode 100644 index ad618d1d6f8f4839f566f5f5cb5db37a4b7d9093..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/quick_start/src/NetConv_en.png and /dev/null differ diff --git a/doc/v1_api_tutorials/quick_start/src/NetLR_cn.jpg b/doc/v1_api_tutorials/quick_start/src/NetLR_cn.jpg deleted file mode 100755 index ee65d1f4127a347e991b4de5e5208347b2b0f51f..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/quick_start/src/NetLR_cn.jpg and /dev/null differ diff --git a/doc/v1_api_tutorials/quick_start/src/NetLR_en.png b/doc/v1_api_tutorials/quick_start/src/NetLR_en.png deleted file mode 100644 index 9d514bf1b18a0c330f98c28785e5d008f409fc1d..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/quick_start/src/NetLR_en.png and /dev/null differ diff --git a/doc/v1_api_tutorials/quick_start/src/NetRNN_cn.jpg b/doc/v1_api_tutorials/quick_start/src/NetRNN_cn.jpg deleted file mode 100755 index f8bc081827f3a50af80495b4561813c5efe9c9dc..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/quick_start/src/NetRNN_cn.jpg and /dev/null differ diff --git a/doc/v1_api_tutorials/quick_start/src/NetRNN_en.png b/doc/v1_api_tutorials/quick_start/src/NetRNN_en.png deleted file mode 100644 index 180f273d32ea59dc8ececa69c08e249f79f9d4f7..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/quick_start/src/NetRNN_en.png and /dev/null differ diff --git a/doc/v1_api_tutorials/quick_start/src/PipelineNetwork_cn.jpg b/doc/v1_api_tutorials/quick_start/src/PipelineNetwork_cn.jpg deleted file mode 100755 index 7e68891d7a536b06027ada6b40cb13dbc3c2d72e..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/quick_start/src/PipelineNetwork_cn.jpg and /dev/null differ diff --git a/doc/v1_api_tutorials/quick_start/src/PipelineNetwork_en.jpg b/doc/v1_api_tutorials/quick_start/src/PipelineNetwork_en.jpg deleted file mode 100644 index e779aed06d5cdb2b442754e7915e79b72946418e..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/quick_start/src/PipelineNetwork_en.jpg and /dev/null differ diff --git a/doc/v1_api_tutorials/quick_start/src/PipelineTest_cn.jpg b/doc/v1_api_tutorials/quick_start/src/PipelineTest_cn.jpg deleted file mode 100755 index 01715db886d811a776c9e7d47f32ba8b1ffe6230..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/quick_start/src/PipelineTest_cn.jpg and /dev/null differ diff --git a/doc/v1_api_tutorials/quick_start/src/PipelineTest_en.png b/doc/v1_api_tutorials/quick_start/src/PipelineTest_en.png deleted file mode 100644 index 7e7ef520b5effa2f43fd2964048f05c42f2ea890..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/quick_start/src/PipelineTest_en.png and /dev/null differ diff --git a/doc/v1_api_tutorials/quick_start/src/PipelineTrain_cn.jpg b/doc/v1_api_tutorials/quick_start/src/PipelineTrain_cn.jpg deleted file mode 100755 index 8049d3e53c483a4d4956a4840b9a0dce5c2990f1..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/quick_start/src/PipelineTrain_cn.jpg and /dev/null differ diff --git a/doc/v1_api_tutorials/quick_start/src/PipelineTrain_en.png b/doc/v1_api_tutorials/quick_start/src/PipelineTrain_en.png deleted file mode 100644 index 132d29bfd5d678d2518161d0b5ed2e16a233a048..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/quick_start/src/PipelineTrain_en.png and /dev/null differ diff --git a/doc/v1_api_tutorials/quick_start/src/Pipeline_cn.jpg b/doc/v1_api_tutorials/quick_start/src/Pipeline_cn.jpg deleted file mode 100755 index d5d99253ea528ebd2d8af6cddc6a9f00e20de8cd..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/quick_start/src/Pipeline_cn.jpg and /dev/null differ diff --git a/doc/v1_api_tutorials/quick_start/src/Pipeline_en.jpg b/doc/v1_api_tutorials/quick_start/src/Pipeline_en.jpg deleted file mode 100644 index 21a7a7bb6a1af746120e6f4f51f797b6aaafb9d8..0000000000000000000000000000000000000000 Binary files a/doc/v1_api_tutorials/quick_start/src/Pipeline_en.jpg and /dev/null differ diff --git a/paddle/framework/lod_tensor.cc b/paddle/framework/lod_tensor.cc index 87a57d095141cc456af2cbabbc227715a02375e9..b29f528f3f749efa3463125c774c2f4d4ebcbc7c 100644 --- a/paddle/framework/lod_tensor.cc +++ b/paddle/framework/lod_tensor.cc @@ -135,6 +135,65 @@ bool operator==(const LoD &a, const LoD &b) { return true; } +bool CheckLoD(const LoD &in, int tensor_height) { + if (in.empty()) return true; + for (const auto &level : in) { + // check: there should be more than 2 offsets existing in each level. + if (level.size() < 2) return false; + // check: the first offset(the begin offset) of each level should be 0. + if (level.front() != 0) return false; + // check: all the offsets in a level should be ascending(no same items + // allows). + if (!std::is_sorted(level.begin(), level.begin(), [](size_t a, size_t b) { + if (a < b) return true; + return false; + })) { + LOG(INFO) << "ascending error"; + return false; + } + } + // check: the lowest level's last offset should equals `tensor_height` if + // tensor_height>0. + if (tensor_height > 0 && (size_t)tensor_height != in.back().back()) + return false; + + // check: the higher level's last offset should equals the lower level's + // size-1. + // NOTE LoD store the levels from top to bottom, so the higher level goes + // first. + for (size_t level = 0; level < in.size() - 1; level++) { + if (in[level].back() != in[level + 1].size() - 1) return false; + } + return true; +} + +bool CheckAbsLoD(const LoD &in, int tensor_height) { + if (in.empty()) return true; + for (const auto &level : in) { + // check: all the offsets in a level should be ascending(no same items + // allows). + if (!std::is_sorted(level.begin(), level.begin(), [](size_t a, size_t b) { + if (a < b) return true; + return false; + })) { + return false; + } + + // check: there should be more than 2 offsets existing in each level. + if (level.size() < 2) return false; + + // check: the first offset of each level should be 0, and the last should be + // the same(the height of underlying tensor). + if (level.front() != 0) return false; + if (tensor_height < 0) { + tensor_height = level.back(); + } else if ((size_t)tensor_height != level.back()) { + return false; + } + } + return true; +} + using LoDAndOffset = std::pair>; LoDAndOffset GetSubLoDAndAbsoluteOffset(const LoD &lod, size_t start_idx, size_t end_idx, size_t start_level) { @@ -227,48 +286,86 @@ void DeserializeFromStream(std::istream &is, LoDTensor *tensor, DeserializeFromStream(is, static_cast(tensor), dev_ctx); } -// TODO(tonyyang-svail): make this function support LoD std::vector LoDTensor::SplitLoDTensor( const std::vector places) const { check_memory_size(); - PADDLE_ENFORCE(lod().empty(), "Disable parallel lod for now"); - PADDLE_ENFORCE(dims()[0] % places.size() == 0, - "Batch size should be divided by places size"); - - std::vector lods; - for (size_t place_idx = 0; place_idx < places.size(); ++place_idx) { - int begin = place_idx * dims()[0] / places.size(); - int end = (place_idx + 1) * dims()[0] / places.size(); + int batch_size = + lod().empty() ? dims()[0] : static_cast(lod()[0].size()) - 1; + size_t result_size = std::min(static_cast(batch_size), places.size()); + size_t remainder = batch_size % places.size(); + + std::vector results; + results.reserve(result_size); + + int step_width = static_cast(batch_size / result_size); + for (size_t i = 0; i < result_size; ++i) { + int begin = static_cast(i * step_width); + int end = static_cast((i + 1) * step_width); + if (i + 1 == places.size()) { // last + end += remainder; + } - auto src = Slice(begin, end); - auto &dst_place = places[place_idx]; LoDTensor dst; - framework::Copy(src, dst_place, &dst); - - lods.emplace_back(dst); + if (lod().empty()) { + auto src = Slice(begin, end); + auto &dst_place = places[i]; + framework::Copy(src, dst_place, &dst); + } else { + auto lod_and_offset = GetSubLoDAndAbsoluteOffset(lod(), begin, end, 0); + + auto &offset = lod_and_offset.second; + auto src = Slice(offset.first, offset.second); + auto &dst_place = places[i]; + framework::Copy(src, dst_place, &dst); + + LoD my_lod; + for (auto &l : lod_and_offset.first) { + std::vector v{0}; + for (auto &ll : l) { + v.push_back(ll + v.back()); + } + my_lod.emplace_back(v); + } + dst.set_lod(my_lod); + } + results.emplace_back(dst); } - return lods; + return results; } -// TODO(tonyyang-svail): make this function support LoD void LoDTensor::MergeLoDTensor( const std::vector &lod_tensors, platform::Place dst_place) { PADDLE_ENFORCE(!lod_tensors.empty()); + framework::DDim new_dim = lod_tensors[0]->dims(); std::type_index new_type = lod_tensors[0]->type(); - auto new_layout = lod_tensors[0]->layout(); - for (auto *lod : lod_tensors) { - PADDLE_ENFORCE(new_dim == lod->dims()); - PADDLE_ENFORCE(new_type == lod->type()); - PADDLE_ENFORCE(new_layout == lod->layout()); + framework::DataLayout new_layout = lod_tensors[0]->layout(); + LoD new_lod = lod_tensors[0]->lod(); + for (size_t i = 1; i < lod_tensors.size(); ++i) { + auto *t = lod_tensors[i]; + PADDLE_ENFORCE_EQ(new_type.hash_code(), t->type().hash_code()); + PADDLE_ENFORCE_EQ(new_layout, t->layout()); + + PADDLE_ENFORCE_EQ(framework::product(new_dim) / new_dim[0], + framework::product(t->dims()) / t->dims()[0]); + new_dim[0] += t->dims()[0]; + + auto &lod = t->lod(); + for (size_t j = 0; j < lod.size(); ++j) { + auto &sub_lod = new_lod[j]; + auto &offset = sub_lod.back(); + for (size_t k = 1; k < lod[j].size(); ++k) { + sub_lod.push_back(lod[j][k] + offset); + } + } } - new_dim[0] *= lod_tensors.size(); Resize(new_dim); set_layout(new_layout); - + set_lod(new_lod); mutable_data(dst_place, new_type); + int begin = 0; for (auto *src : lod_tensors) { int end = begin + src->dims()[0]; diff --git a/paddle/framework/lod_tensor.h b/paddle/framework/lod_tensor.h index 88ea78f2682b2ffc962c9663f6b3c636dedb931d..9d1294fdeb9bd76bf944f7ec3687e3c5bb333241 100644 --- a/paddle/framework/lod_tensor.h +++ b/paddle/framework/lod_tensor.h @@ -71,6 +71,38 @@ LoD ToAbsOffset(const LoD& in); bool operator==(const LoD& a, const LoD& b); +/* + * Check whether this lod's format is valid. + * + * ATTENTION: + * - Empty lod is treated as valid. + * + * It will check two things: + * + * 1. all the offsets in a level should be ascending(no same items allows). + * 2. there should be more than 2 offsets existing in each level. + * 3. the higher level's last offset should equals the lower level's size-1. + * 4. the first offset(the begin offset) of each level should be 0. + * 5. the lowest level's last offset should equals `tensor_height` if + * tensor_height>0. + */ + +bool CheckLoD(const LoD& in, int tensor_height = -1); +/* + * Check whether this absolute lod's format is valid. + * + * ATTENTION: + * - Empty lod is treated as valid. + * + * It will check two things: + * 1. all the offsets in a level should be ascending(no same items allows) + * 2. there should be more than 2 offsets existing in each level. + * 3. the first offset of each level should be 0, and the last should be the + * same(the height of underlying tensor) or `tensor_height` if + * tensor_height>0. + */ +bool CheckAbsLoD(const LoD& in, int tensor_height = -1); + /* * LoDTensor (Level of details Tensor) * see https://en.wikipedia.org/wiki/Level_of_details for reference. diff --git a/paddle/framework/lod_tensor_test.cc b/paddle/framework/lod_tensor_test.cc index 19ae7815cc1008f9f468821d61b26f221e266d5e..2c4d2e10e016da8253410dd241cf8790bd453d72 100644 --- a/paddle/framework/lod_tensor_test.cc +++ b/paddle/framework/lod_tensor_test.cc @@ -1,28 +1,16 @@ -// 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. -/* - 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. -*/ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ #include "paddle/framework/lod_tensor.h" @@ -35,38 +23,6 @@ namespace paddle { namespace framework { -const int kLodTensorSize = 20 * 128; - -class LoDTensorTester : public ::testing::Test { - public: - virtual void SetUp() override { - // tensor's batch_size: 30 - // 3 levels - // 0 10 20 - // 0 5 10 15 20 - // 0 2 5 7 10 12 15 20 - LoD lod; - lod.push_back(std::vector{0, 2, 3}); - lod.push_back(std::vector{0, 2, 5, 8}); - lod.push_back(std::vector{0, 2, 5, 7, 10, 12, 15, 17, 20}); - - ASSERT_EQ(lod.size(), 3UL); - - lod_tensor_.Resize({20 /*batch size*/, 128 /*dim*/}); - // malloc memory - float* dst_ptr = lod_tensor_.mutable_data(place); - for (int i = 0; i < kLodTensorSize; ++i) { - dst_ptr[i] = i; - } - - lod_tensor_.set_lod(lod); - } - - protected: - platform::CPUPlace place; - LoDTensor lod_tensor_; -}; - TEST(LodExpand, test) { LoD lod{{0, 2}}; LoDTensor tensor; @@ -144,5 +100,118 @@ TEST(LoD, ToAbsOffset) { EXPECT_EQ(abs_lod, expected); } +TEST(LoD, SplitLoDTensor) { + LoD lod; + lod.push_back(std::vector({0, 2, 4, 5, 6})); + lod.push_back(std::vector({0, 1, 6, 8, 13, 15, 20})); + + platform::CPUPlace place; + LoDTensor lod_tensor; + lod_tensor.Resize({20, 1}); + float* dst_ptr = lod_tensor.mutable_data(place); + for (int i = 0; i < lod_tensor.numel(); ++i) { + dst_ptr[i] = i; + } + lod_tensor.set_lod(lod); + + std::vector places{platform::CPUPlace(), + platform::CPUPlace()}; + LoD lod0; + lod0.push_back(std::vector({0, 2, 4})); + lod0.push_back(std::vector({0, 1, 6, 8, 13})); + LoD lod1; + lod1.push_back(std::vector({0, 1, 2})); + lod1.push_back(std::vector({0, 2, 7})); + + auto lods = lod_tensor.SplitLoDTensor(places); + EXPECT_EQ(lods[0].lod(), lod0); + EXPECT_EQ(lods[1].lod(), lod1); +} + +TEST(LoD, MergeLoDTensor) { + LoD lod; + lod.push_back(std::vector({0, 2, 4, 5, 6})); + lod.push_back(std::vector({0, 1, 6, 8, 13, 15, 20})); + + platform::CPUPlace place; + + LoDTensor lod_tensor0; + LoD lod0; + lod0.push_back(std::vector({0, 2, 4})); + lod0.push_back(std::vector({0, 1, 6, 8, 13})); + lod_tensor0.set_lod(lod0); + + lod_tensor0.Resize({13, 1}); + float* dst_ptr = lod_tensor0.mutable_data(place); + for (int i = 0; i < lod_tensor0.numel(); ++i) { + dst_ptr[i] = i; + } + + LoDTensor lod_tensor1; + LoD lod1; + lod1.push_back(std::vector({0, 1, 2})); + lod1.push_back(std::vector({0, 2, 7})); + lod_tensor1.set_lod(lod1); + lod_tensor1.Resize({7, 1}); + dst_ptr = lod_tensor1.mutable_data(place); + for (int i = 0; i < lod_tensor1.numel(); ++i) { + dst_ptr[i] = i; + } + + std::vector lods{&lod_tensor0, &lod_tensor1}; + + LoDTensor lod_tensor; + lod_tensor.MergeLoDTensor(lods, place); + EXPECT_EQ(lod_tensor.lod(), lod); +} + +TEST(LoD, CheckLoD) { + LoD relative_lod; + relative_lod.push_back(std::vector({0, 2})); + relative_lod.push_back(std::vector({0, 1, 3})); + relative_lod.push_back(std::vector({0, 2, 4, 5})); + + // check compatible + ASSERT_TRUE(CheckLoD(relative_lod)); + relative_lod[1].back()++; + ASSERT_FALSE(CheckLoD(relative_lod)); + relative_lod[1].back()--; // recover it + + // check empty + LoD empty_lod; + ASSERT_TRUE(CheckLoD(empty_lod)); + + // check less than 2 offsets in a level + LoD some_lod0; + some_lod0.push_back(std::vector({0})); + ASSERT_FALSE(CheckLoD(some_lod0)); + + // check with underlying tensor storage. + ASSERT_TRUE(CheckLoD(relative_lod, 5)); + ASSERT_FALSE(CheckLoD(relative_lod, 9)); +} + +TEST(LoD, CheckAbsLoD) { + LoD relative_lod; + relative_lod.push_back(std::vector({0, 2})); + relative_lod.push_back(std::vector({0, 1, 3})); + relative_lod.push_back(std::vector({0, 2, 4, 5})); + + auto abs_lod = ToAbsOffset(relative_lod); + + ASSERT_TRUE(CheckAbsLoD(abs_lod)); + + // check less than 2 offsets in a level. + + // check the last item should be compatible with tensor height. + abs_lod.back().back()++; + ASSERT_FALSE(CheckAbsLoD(abs_lod)); + abs_lod.back().back()--; // restore + + // check less than 2 offsets in a lod. + LoD abs_lod0; + abs_lod0.push_back(std::vector({0})); + ASSERT_FALSE(CheckAbsLoD(abs_lod0)); +} } // namespace framework } // namespace paddle diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h index d75c0233e8e0134ddf4edc50c07490a234b65cd0..5de9ae559c435439f30931c7840e54e0d2bb744c 100644 --- a/paddle/framework/op_registry.h +++ b/paddle/framework/op_registry.h @@ -177,16 +177,16 @@ class OpKernelRegistrar : public Registrar { /** * Macro to register OperatorKernel. */ -#define REGISTER_OP_KERNEL(op_type, DEVICE_TYPE, place_class, ...) \ - STATIC_ASSERT_GLOBAL_NAMESPACE( \ - __reg_op_kernel_##op_type##_##DEVICE_TYPE##__, \ - "REGISTER_OP_KERNEL must be called in global namespace"); \ - static ::paddle::framework::OpKernelRegistrar \ - __op_kernel_registrar_##op_type##_##DEVICE_TYPE##__(#op_type, \ - #DEVICE_TYPE); \ - int TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE() { \ - __op_kernel_registrar_##op_type##_##DEVICE_TYPE##__.Touch(); \ - return 0; \ +#define REGISTER_OP_KERNEL(op_type, LIBRARY_TYPE, place_class, ...) \ + STATIC_ASSERT_GLOBAL_NAMESPACE( \ + __reg_op_kernel_##op_type##_##LIBRARY_TYPE##__, \ + "REGISTER_OP_KERNEL must be called in global namespace"); \ + static ::paddle::framework::OpKernelRegistrar \ + __op_kernel_registrar_##op_type##_##LIBRARY_TYPE##__(#op_type, \ + #LIBRARY_TYPE); \ + int TouchOpKernelRegistrar_##op_type##_##LIBRARY_TYPE() { \ + __op_kernel_registrar_##op_type##_##LIBRARY_TYPE##__.Touch(); \ + return 0; \ } #define REGISTER_OP_CUDA_KERNEL(op_type, ...) \ @@ -208,14 +208,14 @@ class OpKernelRegistrar : public Registrar { static int use_op_itself_##op_type##_ __attribute__((unused)) = \ TouchOpRegistrar_##op_type() -#define USE_OP_DEVICE_KERNEL(op_type, DEVICE_TYPE) \ - STATIC_ASSERT_GLOBAL_NAMESPACE( \ - __use_op_kernel_##op_type##_##DEVICE_TYPE##__, \ - "USE_OP_DEVICE_KERNEL must be in global namespace"); \ - extern int TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE(); \ - static int use_op_kernel_##op_type##_##DEVICE_TYPE##_ \ - __attribute__((unused)) = \ - TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE() +#define USE_OP_DEVICE_KERNEL(op_type, LIBRARY_TYPE) \ + STATIC_ASSERT_GLOBAL_NAMESPACE( \ + __use_op_kernel_##op_type##_##LIBRARY_TYPE##__, \ + "USE_OP_DEVICE_KERNEL must be in global namespace"); \ + extern int TouchOpKernelRegistrar_##op_type##_##LIBRARY_TYPE(); \ + static int use_op_kernel_##op_type##_##LIBRARY_TYPE##_ \ + __attribute__((unused)) = \ + TouchOpKernelRegistrar_##op_type##_##LIBRARY_TYPE() // TODO(fengjiayi): The following macros // seems ugly, do we have better method? diff --git a/paddle/gserver/layers/MKLDNNConcatLayer.cpp b/paddle/gserver/layers/MKLDNNConcatLayer.cpp index 44bb0883b89c712d70e2d4fdfe16bdfde86f81b7..520ccc1a995e966de73080b61a8c20cbee722267 100644 --- a/paddle/gserver/layers/MKLDNNConcatLayer.cpp +++ b/paddle/gserver/layers/MKLDNNConcatLayer.cpp @@ -43,7 +43,7 @@ void MKLDNNConcatLayer::reshape( channels_[0] = ic; oc = ic; for (size_t i = 1; i < inputLayers_.size(); i++) { - int batchsize, height, witdh; + int batchsize = 0, height = 0, witdh = 0; reshapeInput(batchsize, height, witdh, i); CHECK_EQ(bs, batchsize); CHECK_EQ(ih, height); @@ -84,6 +84,7 @@ void MKLDNNConcatLayer::resetFwdBuffers(std::vector& inputs, bool has8c = false, has16c = false, hasnc = false; for (size_t i = 0; i < inputs.size(); i++) { resetInValue(inputs[i], nullptr, i, channels_[i]); + inputs[i]->downSpatial(); CHECK(inputs[i]); auto dm = inputs[i]->getDims(); // inputs format can be different, but ndims must equal diff --git a/paddle/gserver/tests/img_conv_cudnn.py b/paddle/gserver/tests/img_conv_cudnn.py index e424261bda2c0fea4b1796ce443810e8e4e4599b..0ea6d6bae66b0a307748bd0d0fa9a53ed5f7927d 100644 --- a/paddle/gserver/tests/img_conv_cudnn.py +++ b/paddle/gserver/tests/img_conv_cudnn.py @@ -1,4 +1,4 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# 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. @@ -11,20 +11,6 @@ #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. -#edit-mode: -*- python -*- -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. from paddle.trainer_config_helpers import * diff --git a/paddle/gserver/tests/img_conv_exconv.py b/paddle/gserver/tests/img_conv_exconv.py index 3b59cd80b8b8349f022b774b36d4e4db2d517b73..c618cdab27c52d70042b0a118f7f6fe935a6b9d7 100644 --- a/paddle/gserver/tests/img_conv_exconv.py +++ b/paddle/gserver/tests/img_conv_exconv.py @@ -1,4 +1,4 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# 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. @@ -11,20 +11,6 @@ #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. -#edit-mode: -*- python -*- -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. from paddle.trainer_config_helpers import * diff --git a/paddle/gserver/tests/pyDataProvider.py b/paddle/gserver/tests/pyDataProvider.py index 7235a239439b7544805d1bd06dfb1a72c2e0e937..d2ad5888b5a4c79d8b663ce8c2f313184151beb6 100644 --- a/paddle/gserver/tests/pyDataProvider.py +++ b/paddle/gserver/tests/pyDataProvider.py @@ -1,17 +1,16 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# 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 +#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 +# 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 numpy import struct import traceback diff --git a/paddle/gserver/tests/rnn_data_provider.py b/paddle/gserver/tests/rnn_data_provider.py index 913365a5a4037d14fcba1e1546508ba89668e0d6..063a4127e542d23012359a2eac0045bf69a51356 100644 --- a/paddle/gserver/tests/rnn_data_provider.py +++ b/paddle/gserver/tests/rnn_data_provider.py @@ -1,17 +1,16 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# 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 +#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 +# 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. from paddle.trainer.PyDataProvider2 import * # Note that each config should has an independent provider diff --git a/paddle/gserver/tests/sequenceGen.py b/paddle/gserver/tests/sequenceGen.py index fd725727c04677b5ea8918f6721f0c007e80915d..04a1732d61c8618984d16550acf7c94da1bd3578 100644 --- a/paddle/gserver/tests/sequenceGen.py +++ b/paddle/gserver/tests/sequenceGen.py @@ -1,17 +1,16 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# 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 +#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 +# 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 os import sys diff --git a/paddle/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py b/paddle/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py index 7303d088043d5096a3491d3b3b32b231bde09a0a..aeaaa221f9fab981af88cfd63c30349e1b02a0ee 100644 --- a/paddle/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py +++ b/paddle/gserver/tests/sequence_nest_rnn_multi_unequalength_inputs.py @@ -1,18 +1,16 @@ -# edit-mode: -*- python -*- -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# 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 +#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 +# 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. from paddle.trainer_config_helpers import * ######################## data source ################################ diff --git a/paddle/gserver/tests/sequence_recurrent.py b/paddle/gserver/tests/sequence_recurrent.py index d20d3331ae6887aa3bca9df7e7817e08ebb91a48..8786a5465db82d786d3772357b02ab837073a576 100644 --- a/paddle/gserver/tests/sequence_recurrent.py +++ b/paddle/gserver/tests/sequence_recurrent.py @@ -1,4 +1,4 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# 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. @@ -11,20 +11,6 @@ #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. -#!/usr/bin/env python -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. from paddle.trainer_config_helpers import * diff --git a/paddle/gserver/tests/sequence_recurrent_group.py b/paddle/gserver/tests/sequence_recurrent_group.py index a1d54542e3bc4e89f70d31d5e89c0f44953c9f90..8b5a3d49838c9bb49321a9d7514fc0241e6d67cd 100644 --- a/paddle/gserver/tests/sequence_recurrent_group.py +++ b/paddle/gserver/tests/sequence_recurrent_group.py @@ -1,18 +1,16 @@ -#!/usr/bin/env python -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# 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 +# http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. - from paddle.trainer_config_helpers import * ######################## data source ################################ diff --git a/paddle/gserver/tests/sequence_rnn_matched_inputs.py b/paddle/gserver/tests/sequence_rnn_matched_inputs.py index b156fa9baaf386a8625d63020d72828047ddca2f..0c55f2cf9d07b194aa06f88892f831f1a9ce6436 100644 --- a/paddle/gserver/tests/sequence_rnn_matched_inputs.py +++ b/paddle/gserver/tests/sequence_rnn_matched_inputs.py @@ -1,4 +1,4 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# 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. @@ -11,20 +11,6 @@ #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. -# edit-mode: -*- python -*- -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. from paddle.trainer_config_helpers import * diff --git a/paddle/gserver/tests/sequence_rnn_mixed_inputs.py b/paddle/gserver/tests/sequence_rnn_mixed_inputs.py index c2c98aae5f5c5f73c61b309382f2e3d7d649ef47..22b376b91aa4736d16fead698105466d679dd248 100644 --- a/paddle/gserver/tests/sequence_rnn_mixed_inputs.py +++ b/paddle/gserver/tests/sequence_rnn_mixed_inputs.py @@ -1,4 +1,4 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# 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. @@ -11,20 +11,6 @@ #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. -# edit-mode: -*- python -*- -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. from paddle.trainer_config_helpers import * diff --git a/paddle/gserver/tests/sequence_rnn_multi_unequalength_inputs.py b/paddle/gserver/tests/sequence_rnn_multi_unequalength_inputs.py index c812c6ced2458b6c40e141606655aa45e2bbc55a..3ce87490bbd0f30a3c42b947b073adb2a6c5b51c 100644 --- a/paddle/gserver/tests/sequence_rnn_multi_unequalength_inputs.py +++ b/paddle/gserver/tests/sequence_rnn_multi_unequalength_inputs.py @@ -1,4 +1,4 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# 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. @@ -11,20 +11,6 @@ #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. -#edit-mode: -*- python -*- -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. from paddle.trainer_config_helpers import * diff --git a/paddle/gserver/tests/test_PyDataProvider2.py b/paddle/gserver/tests/test_PyDataProvider2.py index 0d0fe476ff5eac8bf8ad1c9fe09b32c1a8f73ebc..044aede98e684a432c48b3ea5bb82a4a677682d4 100644 --- a/paddle/gserver/tests/test_PyDataProvider2.py +++ b/paddle/gserver/tests/test_PyDataProvider2.py @@ -1,17 +1,16 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# 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 +#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 +# 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 random from paddle.trainer.PyDataProvider2 import * diff --git a/paddle/inference/CMakeLists.txt b/paddle/inference/CMakeLists.txt index 8437b2b21942ead544dab8636db1b355b7cf7bd5..02ca8a45a851d262eed6962a9a227b5009ef03a5 100644 --- a/paddle/inference/CMakeLists.txt +++ b/paddle/inference/CMakeLists.txt @@ -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) diff --git a/paddle/inference/example.cc b/paddle/inference/example.cc index 9711b20e6fb4099a2cc497029468ebd1fd0b3456..0c18b45624dedcb5839d4b771e044b4a7b32af52 100644 --- a/paddle/inference/example.cc +++ b/paddle/inference/example.cc @@ -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 feed_var_names = {FLAGS_feed_var_names}; - std::vector 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)); diff --git a/paddle/inference/inference.cc b/paddle/inference/inference.cc index 37b8b20ddfcf2566b8410f950308309e5b2b2a7c..49001778808173b82865a4b6632a6b175ef96242 100644 --- a/paddle/inference/inference.cc +++ b/paddle/inference/inference.cc @@ -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& feed_var_names, const std::vector& 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); diff --git a/paddle/inference/inference.h b/paddle/inference/inference.h index a3f3ef4b440036a0b27353cc092eed1bbf96eeb3..7fc09cb9e539a65a8cd3cceb1543bc7d111c22b3 100644 --- a/paddle/inference/inference.h +++ b/paddle/inference/inference.h @@ -28,6 +28,7 @@ public: delete load_program_; } + void LoadInferenceModel(const std::string& dirname); void LoadInferenceModel(const std::string& dirname, const std::vector& feed_var_names, const std::vector& fetch_var_names); diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt index 2412ebd82a02c872e73fd310c56221309441f630..6745a8da17723d663913a29f28e5ea9eedc0372a 100644 --- a/paddle/operators/CMakeLists.txt +++ b/paddle/operators/CMakeLists.txt @@ -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) diff --git a/paddle/operators/beam_search_op.cc b/paddle/operators/beam_search_op.cc index ed2e7b738acd81140467ff22ad077155bba2fde1..4c71d66d22899d2cf6418935bf9358a0f73cec27 100644 --- a/paddle/operators/beam_search_op.cc +++ b/paddle/operators/beam_search_op.cc @@ -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 &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> *items) { +int BeamSearch::PruneEndidCandidates(const framework::LoDTensor &pre_ids, + std::vector> *items) { auto *pre_ids_data = pre_ids.data(); + 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> BeamSearch::ToMap( @@ -121,11 +134,7 @@ bool BeamSearch::NextItemSet(std::vector *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(); auto *scores_data = scores.data(); diff --git a/paddle/operators/beam_search_op.h b/paddle/operators/beam_search_op.h index 08b551ef9bd63106ed222d3a956a912294f827ec..45d14d68fe8d1c4a84aa826e68e76692444765a8 100644 --- a/paddle/operators/beam_search_op.h +++ b/paddle/operators/beam_search_op.h @@ -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>* items); + /* + * Delete all the records that follows the end token. + */ + int PruneEndidCandidates(const framework::LoDTensor& pre_ids, + std::vector>* items); /* * Transform the items into a map whose key is offset, value is the items. diff --git a/paddle/operators/beam_search_op_test.cc b/paddle/operators/beam_search_op_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..d4beb64a85a1645b3fed22c3325bd8c0b7cd12b1 --- /dev/null +++ b/paddle/operators/beam_search_op_test.cc @@ -0,0 +1,86 @@ +/* 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 +#include + +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 level0({0, 1, 4}); + vector 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({4, 3})); + ids->Resize(dims); + scores->Resize(dims); + CPUPlace place; + + auto* ids_data = ids->mutable_data(place); + auto* scores_data = scores->mutable_data(place); + vector _ids({4, 2, 5, 2, 1, 3, 3, 5, 2, 8, 2, 1}); + vector _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(4, 1))); + for (int i = 0; i < 4; i++) { + pre_ids.mutable_data(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 tids({2, 4, 3, 8}); + vector tscores({0.3, 0.5, 0.9, 0.7}); + + for (int i = 0; i < 4; i++) { + ASSERT_EQ(tids[i], sids.data()[i]); + ASSERT_EQ(tscores[i], sscores.data()[i]); + } +} + +} // namespace test +} // namespace paddle diff --git a/paddle/operators/clip_op.cc b/paddle/operators/clip_op.cc index 573bb9c7dfdac2366c2458dd9f27a035a9f9b813..7adb74eab78dcdd0251b8db60781f6e24e348634 100644 --- a/paddle/operators/clip_op.cc +++ b/paddle/operators/clip_op.cc @@ -51,8 +51,8 @@ class ClipOpMaker : public framework::OpProtoAndCheckerMaker { AddComment(R"DOC( Clip Operator. -The clip operator limits the value of given input within an interval. The interval is -specified with arguments 'min' and 'max': +The clip operator limits the value of given input within an interval. The +interval is specified with arguments 'min' and 'max': $$ Out = \min(\max(X, min), max) diff --git a/paddle/operators/conv_op.cc b/paddle/operators/conv_op.cc index 424eccdb7dc57195d20f75460032a99c76e6adcd..d6882b275b22b9a2a2b6ff8cfb53a3462bbdbefe 100644 --- a/paddle/operators/conv_op.cc +++ b/paddle/operators/conv_op.cc @@ -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("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(); + 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("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(); + use_cudnn &= dev_ctx.cudnn_handle() != nullptr; + } +#endif + framework::LibraryType library_; if (use_cudnn) { library_ = framework::LibraryType::kCUDNN; diff --git a/paddle/operators/conv_transpose_op.cc b/paddle/operators/conv_transpose_op.cc index cf4e8c0a303d6888c3b1f2475a483c4bfc90981b..a2382a7e42eb9c5c6a8f13265b0e6173e6b05f76 100644 --- a/paddle/operators/conv_transpose_op.cc +++ b/paddle/operators/conv_transpose_op.cc @@ -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("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(); + 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("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(); + use_cudnn &= dev_ctx.cudnn_handle() != nullptr; + } +#endif framework::LibraryType library_; if (use_cudnn) { library_ = framework::LibraryType::kCUDNN; diff --git a/paddle/operators/detail/grpc_client.cc b/paddle/operators/detail/grpc_client.cc index aee56ffe018aa8d0d2106df24bd9358c930a02ca..1e41587c418fb0ce4e452d5c6735c54e2d42f798 100644 --- a/paddle/operators/detail/grpc_client.cc +++ b/paddle/operators/detail/grpc_client.cc @@ -63,9 +63,6 @@ bool RPCClient::AsyncGetVariable(const std::string& ep, sendrecv::VariableMessage req; req.set_varname(var_name); - auto* var = scope.FindVar(var_name); - SerializeToMessage(var_name, var, ctx, &req); - // varhandle VarHandle var_h; var_h.ep = ep; diff --git a/paddle/operators/detail/grpc_server.cc b/paddle/operators/detail/grpc_server.cc index c0b94746a0b7f6ffb657bbf5af18360426933858..3ddcd839bdd23547216465dfaf44a3cd8285fe6d 100644 --- a/paddle/operators/detail/grpc_server.cc +++ b/paddle/operators/detail/grpc_server.cc @@ -36,7 +36,10 @@ class RequestBase { CallStatus Status() { return status_; } void SetStatus(CallStatus status) { status_ = status; } - virtual std::string GetReqName() { assert(false); } + virtual std::string GetReqName() { + assert(false); + return ""; + } protected: grpc::ServerContext ctx_; @@ -80,11 +83,13 @@ class RequestGet final : public RequestBase { public: explicit RequestGet(sendrecv::SendRecvService::AsyncService* service, grpc::ServerCompletionQueue* cq, framework::Scope* scope, - const platform::DeviceContext* dev_ctx) + const platform::DeviceContext* dev_ctx, + SimpleBlockQueue* queue) : RequestBase(service, cq), responder_(&ctx_), scope_(scope), - dev_ctx_(dev_ctx) { + dev_ctx_(dev_ctx), + queue_(queue) { service_->RequestGetVariable(&ctx_, &request_, &responder_, cq_, cq_, this); } @@ -100,6 +105,7 @@ class RequestGet final : public RequestBase { // TODO(gongwb): check var's info. responder_.Finish(reply_, grpc::Status::OK, this); status_ = FINISH; + queue_->Push('c'); } protected: @@ -108,8 +114,15 @@ class RequestGet final : public RequestBase { ServerAsyncResponseWriter responder_; framework::Scope* scope_; const platform::DeviceContext* dev_ctx_; + SimpleBlockQueue* queue_; }; +void AsyncGRPCServer::WaitClientGet(int count) { + for (int i = 0; i < count; ++i) { + var_get_queue_.Pop(); + } +} + void AsyncGRPCServer::RunSyncUpdate() { grpc::ServerBuilder builder; builder.AddListeningPort(address_, grpc::InsecureServerCredentials()); @@ -149,7 +162,6 @@ void AsyncGRPCServer::ShutdownQueue() { } // This URL explains why shutdown is complicate: -// https://stackoverflow.com/questions/35708348/grpc-what-is-the-recommended-way-to-shut-down-an-asynchronous-server-in-c void AsyncGRPCServer::ShutDown() { server_->Shutdown(); ShutdownQueue(); @@ -170,10 +182,12 @@ void AsyncGRPCServer::TryToRegisterNewGetOne() { if (is_shut_down_) { return; } - RequestGet* get = new RequestGet(&service_, cq_get_.get(), scope_, dev_ctx_); + RequestGet* get = new RequestGet(&service_, cq_get_.get(), scope_, dev_ctx_, + &var_get_queue_); VLOG(4) << "create Requestget status:" << get->Status(); } +// FIXME(typhoonzero): remove wait argument and change cq_name to enum. void AsyncGRPCServer::HandleRequest(bool wait, grpc::ServerCompletionQueue* cq, std::string cq_name, std::function TryToRegisterNewOne) { @@ -188,9 +202,9 @@ void AsyncGRPCServer::HandleRequest(bool wait, grpc::ServerCompletionQueue* cq, } PADDLE_ENFORCE(tag); - if (wait && !done_) { - Wait(); - } + // FIXME(typhoonzero): de-couple the barriers with recv_op + if (cq_name == "cq_get") WaitCond(1); + if (cq_name == "cq_send") WaitCond(0); RequestBase* base = (RequestBase*)tag; // reference: @@ -222,22 +236,18 @@ void AsyncGRPCServer::HandleRequest(bool wait, grpc::ServerCompletionQueue* cq, } } -void AsyncGRPCServer::Wait() { - std::unique_lock lock(this->mutex_); - condition_.wait(lock, [=] { return this->done_ == true; }); -} - -void AsyncGRPCServer::Reset() { - std::lock_guard lock(this->mutex_); - done_ = false; +void AsyncGRPCServer::WaitCond(int cond) { + std::unique_lock lock(this->barrier_mutex_); + barrier_condition_.wait(lock, + [=] { return this->barrier_cond_step_ == cond; }); } -void AsyncGRPCServer::Done() { +void AsyncGRPCServer::SetCond(int cond) { { - std::lock_guard lock(this->mutex_); - done_ = true; + std::lock_guard lock(this->barrier_mutex_); + barrier_cond_step_ = cond; } - condition_.notify_all(); + barrier_condition_.notify_all(); } } // namespace detail diff --git a/paddle/operators/detail/grpc_server.h b/paddle/operators/detail/grpc_server.h index 2c078b77771656dc7fc0342ecf21b8d33dc11817..1ca9086c744c558fd05fb4fc1a7280729afbec28 100644 --- a/paddle/operators/detail/grpc_server.h +++ b/paddle/operators/detail/grpc_server.h @@ -41,9 +41,10 @@ class AsyncGRPCServer final : public sendrecv::SendRecvService::Service { void RunSyncUpdate(); - void Reset(); - - void Done(); + // functions to sync server barrier status. + void WaitCond(int cond); + void SetCond(int cond); + void WaitClientGet(int count); void SetScope(framework::Scope *scope) { scope_ = scope; } @@ -56,7 +57,6 @@ class AsyncGRPCServer final : public sendrecv::SendRecvService::Service { void ShutDown(); protected: - void Wait(); void HandleRequest(bool wait, grpc::ServerCompletionQueue *cq, std::string cq_name, std::function TryToRegisterNewOne); @@ -78,11 +78,12 @@ class AsyncGRPCServer final : public sendrecv::SendRecvService::Service { const platform::DeviceContext *dev_ctx_; // received variable from RPC, operators fetch variable from this queue. SimpleBlockQueue var_recv_queue_; + SimpleBlockQueue var_get_queue_; // condition of the sub program - std::mutex mutex_; - volatile mutable bool done_; - std::condition_variable condition_; + std::mutex barrier_mutex_; + mutable int barrier_cond_step_; + std::condition_variable barrier_condition_; std::unique_ptr t_send_; std::unique_ptr t_get_; diff --git a/paddle/operators/elementwise_add_op.h b/paddle/operators/elementwise_add_op.h index 6478e1e0c2e1cfc8a1be5e8842113ec8ca33d762..a8389429f26c17ceab1db22175c90888546ead6f 100644 --- a/paddle/operators/elementwise_add_op.h +++ b/paddle/operators/elementwise_add_op.h @@ -28,39 +28,7 @@ template class ElementwiseAddKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - using Tensor = framework::Tensor; - - auto* x = ctx.Input("X"); - auto* y = ctx.Input("Y"); - auto* z = ctx.Output("Out"); - z->mutable_data(ctx.GetPlace()); - TransformFunctor, T, DeviceContext> functor( - x, y, z, ctx.template device_context(), AddFunctor()); - - auto x_dims = x->dims(); - auto y_dims = y->dims(); - PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(), - "Rank of first input must >= rank of second input."); - - if (x_dims == y_dims) { - functor.Run(); - return; - } - - int axis = ctx.Attr("axis"); - axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis); - PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(), - "Axis should be in range [0, x_dims)"); - - int pre, n, post; - get_mid_dims(x_dims, y_dims, axis, pre, n, post); - if (post == 1) { - functor.RunRowWise(n, pre); - return; - } else { - functor.RunMidWise(n, pre, post); - return; - } + ElementwiseComputeEx, DeviceContext, T>(ctx); } }; diff --git a/paddle/operators/elementwise_div_op.h b/paddle/operators/elementwise_div_op.h index 7783875e24ef0d94e50f49b3598c93ff8bd5d73f..ef26cb6c914f50ded07cc9d0d8de3f49f2151129 100644 --- a/paddle/operators/elementwise_div_op.h +++ b/paddle/operators/elementwise_div_op.h @@ -19,11 +19,16 @@ limitations under the License. */ namespace paddle { namespace operators { +template +struct DivFunctor { + inline HOSTDEVICE T operator()(T a, T b) const { return a / b; } +}; + template class ElementwiseDivKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - ElementwiseCompute(ctx); + ElementwiseComputeEx, DeviceContext, T>(ctx); } }; diff --git a/paddle/operators/elementwise_max_op.cc b/paddle/operators/elementwise_max_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..53c27ae5be4cbfe85ce61aa27196594ae152eea4 --- /dev/null +++ b/paddle/operators/elementwise_max_op.cc @@ -0,0 +1,45 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/elementwise_max_op.h" +#include "paddle/operators/elementwise_op.h" + +namespace paddle { +namespace operators { +class ElementwiseMaxOpMaker : public ElementwiseOpMaker { + public: + ElementwiseMaxOpMaker(OpProto* proto, OpAttrChecker* op_checker) + : ElementwiseOpMaker(proto, op_checker) { + SetComment("Max", "Out = max(X, Y)"); + AddComment(comment_); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(elementwise_max, ops::ElementwiseOp, ops::ElementwiseMaxOpMaker, + elementwise_max_grad, ops::ElementwiseOpGrad); +REGISTER_OP_CPU_KERNEL( + elementwise_max, + ops::ElementwiseMaxKernel, + ops::ElementwiseMaxKernel, + ops::ElementwiseMaxKernel, + ops::ElementwiseMaxKernel); +REGISTER_OP_CPU_KERNEL( + elementwise_max_grad, + ops::ElementwiseMaxGradKernel, + ops::ElementwiseMaxGradKernel, + ops::ElementwiseMaxGradKernel, + ops::ElementwiseMaxGradKernel); diff --git a/paddle/operators/elementwise_max_op.cu b/paddle/operators/elementwise_max_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..5ff4af17477cbd35b765cc00d46c95fda620e2df --- /dev/null +++ b/paddle/operators/elementwise_max_op.cu @@ -0,0 +1,32 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/elementwise_max_op.h" + +namespace ops = paddle::operators; + +REGISTER_OP_CUDA_KERNEL( + elementwise_max, + ops::ElementwiseMaxKernel, + ops::ElementwiseMaxKernel, + ops::ElementwiseMaxKernel, + ops::ElementwiseMaxKernel); +REGISTER_OP_CUDA_KERNEL( + elementwise_max_grad, + ops::ElementwiseMaxGradKernel, + ops::ElementwiseMaxGradKernel, + ops::ElementwiseMaxGradKernel, + ops::ElementwiseMaxGradKernel); diff --git a/paddle/operators/elementwise_max_op.h b/paddle/operators/elementwise_max_op.h new file mode 100644 index 0000000000000000000000000000000000000000..255728e8e620665a7de225b228c19d6c510da1c8 --- /dev/null +++ b/paddle/operators/elementwise_max_op.h @@ -0,0 +1,120 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/operators/elementwise_op_function.h" + +namespace paddle { +namespace operators { + +template +struct MaxFunctor { + inline HOSTDEVICE T operator()(T a, T b) const { return a > b ? a : b; } +}; + +template +class ElementwiseMaxKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + ElementwiseComputeEx, DeviceContext, T>(ctx); + } +}; + +template +struct ElementwiseMaxGradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) { + auto x_e = framework::EigenVector::Flatten(*x); + auto y_e = framework::EigenVector::Flatten(*y); + auto dz_e = framework::EigenVector::Flatten(*dz); + + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = (x_e > y_e).template cast() * dz_e; + } + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = (x_e <= y_e).template cast() * dz_e; + } + } +}; + +template +struct ElementwiseMaxBroadCastGradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) { + auto x_e = framework::EigenVector::Flatten(*x); + auto y_e = framework::EigenVector::Flatten(*y); + auto dz_e = framework::EigenVector::Flatten(*dz); + + auto y_e_bcast = y_e.reshape(Eigen::DSizes(1, n)) + .broadcast(Eigen::DSizes(pre, 1)) + .reshape(Eigen::DSizes(x_e.size())); + + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = (x_e > y_e_bcast).template cast() * dz_e; + } + + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = ((x_e <= y_e_bcast).template cast() * dz_e) + .reshape(Eigen::DSizes(pre, n)) + .sum(Eigen::array{{0}}); + } + } +}; + +template +struct ElementwiseMaxBroadCast2GradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n, + Post post) { + auto x_e = framework::EigenVector::Flatten(*x); + auto y_e = framework::EigenVector::Flatten(*y); + auto dz_e = framework::EigenVector::Flatten(*dz); + + auto y_e_bcast = y_e.reshape(Eigen::DSizes(1, n, 1)) + .broadcast(Eigen::DSizes(pre, 1, post)) + .reshape(Eigen::DSizes(x_e.size())); + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = (x_e > y_e_bcast).template cast() * dz_e; + } + + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = ((x_e <= y_e_bcast).template cast() * dz_e) + .reshape(Eigen::DSizes(pre, n, post)) + .sum(Eigen::array{{0, 2}}); + } + } +}; + +template +class ElementwiseMaxGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + ElementwiseGradCompute, + ElementwiseMaxBroadCastGradFunctor, + ElementwiseMaxBroadCast2GradFunctor>(ctx); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/elementwise_min_op.cc b/paddle/operators/elementwise_min_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..99482e1bf60c88062087c5fe0105e90aa0a8677c --- /dev/null +++ b/paddle/operators/elementwise_min_op.cc @@ -0,0 +1,45 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/operators/elementwise_min_op.h" +#include "paddle/operators/elementwise_op.h" + +namespace paddle { +namespace operators { +class ElementwiseMinOpMaker : public ElementwiseOpMaker { + public: + ElementwiseMinOpMaker(OpProto* proto, OpAttrChecker* op_checker) + : ElementwiseOpMaker(proto, op_checker) { + SetComment("Max", "Out = min(X, Y)"); + AddComment(comment_); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(elementwise_min, ops::ElementwiseOp, ops::ElementwiseMinOpMaker, + elementwise_min_grad, ops::ElementwiseOpGrad); +REGISTER_OP_CPU_KERNEL( + elementwise_min, + ops::ElementwiseMinKernel, + ops::ElementwiseMinKernel, + ops::ElementwiseMinKernel, + ops::ElementwiseMinKernel); +REGISTER_OP_CPU_KERNEL( + elementwise_min_grad, + ops::ElementwiseMinGradKernel, + ops::ElementwiseMinGradKernel, + ops::ElementwiseMinGradKernel, + ops::ElementwiseMinGradKernel); diff --git a/paddle/operators/elementwise_min_op.cu b/paddle/operators/elementwise_min_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..3547e6ccb77177002b1ecbee4e4604b602f72209 --- /dev/null +++ b/paddle/operators/elementwise_min_op.cu @@ -0,0 +1,32 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#define EIGEN_USE_GPU +#include "paddle/operators/elementwise_min_op.h" + +namespace ops = paddle::operators; + +REGISTER_OP_CUDA_KERNEL( + elementwise_min, + ops::ElementwiseMinKernel, + ops::ElementwiseMinKernel, + ops::ElementwiseMinKernel, + ops::ElementwiseMinKernel); +REGISTER_OP_CUDA_KERNEL( + elementwise_min_grad, + ops::ElementwiseMinGradKernel, + ops::ElementwiseMinGradKernel, + ops::ElementwiseMinGradKernel, + ops::ElementwiseMinGradKernel); diff --git a/paddle/operators/elementwise_min_op.h b/paddle/operators/elementwise_min_op.h new file mode 100644 index 0000000000000000000000000000000000000000..e6627a0f1bb468c8e4661b83489cb964b72dddb0 --- /dev/null +++ b/paddle/operators/elementwise_min_op.h @@ -0,0 +1,120 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/operators/elementwise_op_function.h" + +namespace paddle { +namespace operators { + +template +struct MinFunctor { + inline HOSTDEVICE T operator()(T a, T b) const { return a < b ? a : b; } +}; + +template +class ElementwiseMinKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + ElementwiseComputeEx, DeviceContext, T>(ctx); + } +}; + +template +struct ElementwiseMinGradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) { + auto x_e = framework::EigenVector::Flatten(*x); + auto y_e = framework::EigenVector::Flatten(*y); + auto dz_e = framework::EigenVector::Flatten(*dz); + + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = (x_e < y_e).template cast() * dz_e; + } + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = (x_e >= y_e).template cast() * dz_e; + } + } +}; + +template +struct ElementwiseMinBroadCastGradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) { + auto x_e = framework::EigenVector::Flatten(*x); + auto y_e = framework::EigenVector::Flatten(*y); + auto dz_e = framework::EigenVector::Flatten(*dz); + + auto y_e_bcast = y_e.reshape(Eigen::DSizes(1, n)) + .broadcast(Eigen::DSizes(pre, 1)) + .reshape(Eigen::DSizes(x_e.size())); + + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = (x_e < y_e_bcast).template cast() * dz_e; + } + + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = ((x_e >= y_e_bcast).template cast() * dz_e) + .reshape(Eigen::DSizes(pre, n)) + .sum(Eigen::array{{0}}); + } + } +}; + +template +struct ElementwiseMinBroadCast2GradFunctor { + template + void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n, + Post post) { + auto x_e = framework::EigenVector::Flatten(*x); + auto y_e = framework::EigenVector::Flatten(*y); + auto dz_e = framework::EigenVector::Flatten(*dz); + + auto y_e_bcast = y_e.reshape(Eigen::DSizes(1, n, 1)) + .broadcast(Eigen::DSizes(pre, 1, post)) + .reshape(Eigen::DSizes(x_e.size())); + if (dx) { + auto dx_e = framework::EigenVector::Flatten(*dx); + dx_e.device(d) = (x_e < y_e_bcast).template cast() * dz_e; + } + + if (dy) { + auto dy_e = framework::EigenVector::Flatten(*dy); + dy_e.device(d) = ((x_e >= y_e_bcast).template cast() * dz_e) + .reshape(Eigen::DSizes(pre, n, post)) + .sum(Eigen::array{{0, 2}}); + } + } +}; + +template +class ElementwiseMinGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + ElementwiseGradCompute, + ElementwiseMinBroadCastGradFunctor, + ElementwiseMinBroadCast2GradFunctor>(ctx); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/elementwise_mul_op.h b/paddle/operators/elementwise_mul_op.h index 0e6559eacc0f90fec78397ce66372071370291e5..4b86b00b5a095ae898f9ce0c17cde2cc91060ba9 100644 --- a/paddle/operators/elementwise_mul_op.h +++ b/paddle/operators/elementwise_mul_op.h @@ -18,11 +18,16 @@ limitations under the License. */ namespace paddle { namespace operators { +template +struct MulFunctor { + inline HOSTDEVICE T operator()(T a, T b) const { return a * b; } +}; + template class ElementwiseMulKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - ElementwiseCompute(ctx); + ElementwiseComputeEx, DeviceContext, T>(ctx); } }; diff --git a/paddle/operators/elementwise_op.h b/paddle/operators/elementwise_op.h index a342595b546bfca1a344cf8a549597df6a29adec..1a0131d8b943da3deebd0c461f78cb02b34e6dc2 100644 --- a/paddle/operators/elementwise_op.h +++ b/paddle/operators/elementwise_op.h @@ -26,9 +26,9 @@ class ElementwiseOp : public framework::OperatorWithKernel { using Tensor = framework::Tensor; void InferShape(framework::InferShapeContext* ctx) const override { PADDLE_ENFORCE(ctx->HasInput("X"), - "Input(X) of elementwise op should not be null"); + "Input(X) of elementwise op should not be null."); PADDLE_ENFORCE(ctx->HasInput("Y"), - "Input(Y) of elementwise op should not be null"); + "Input(Y) of elementwise op should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of elementwise op should not be null."); @@ -45,12 +45,12 @@ class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker { public: ElementwiseOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "(Tensor) The first input tensor of elementwise op"); - AddInput("Y", "(Tensor) The second input tensor of elementwise op"); - AddOutput("Out", "The output of elementwise op"); + AddInput("X", "(Tensor), The first input tensor of elementwise op."); + AddInput("Y", "(Tensor), The second input tensor of elementwise op."); + AddOutput("Out", "The output of elementwise op."); AddAttr("axis", - "(int, default -1) The starting dimension index " - "for broadcasting Y onto X") + "(int, default -1). The start dimension index " + "for broadcasting Y onto X.") .SetDefault(-1) .EqualGreaterThan(-1); comment_ = R"DOC( @@ -58,19 +58,18 @@ Limited Elementwise {name} Operator. The equation is: -.. math:: - {equation} +$${equation}$$ -X is a tensor of any dimension and the dimensions of tensor Y must be smaller than -or equal to the dimensions of X. +$X$ is a tensor of any dimension and the dimensions of tensor $Y$ must be +smaller than or equal to the dimensions of $X$. There are two cases for this operator: -1. The shape of Y is same with X; -2. The shape of Y is a subset of X. +1. The shape of $Y$ is same with $X$; +2. The shape of $Y$ is a subset of $X$. For case 2: -Y will be broadcasted to match the shape of X and axis should be -the starting dimension index for broadcasting Y onto X. +$Y$ will be broadcasted to match the shape of $X$ and axis should be +set to index of the start dimension to broadcast $Y$ onto $X$. For example .. code-block:: python @@ -81,7 +80,8 @@ For example shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1 shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0 -Either of the inputs X and Y or none can carry the LoD (Level of Details) information. However, the output only shares the LoD information with input X. +Either of the inputs $X$ and $Y$ or none can carry the LoD (Level of Details) +information. However, the output only shares the LoD information with input $X$. )DOC"; AddComment(comment_); diff --git a/paddle/operators/elementwise_op_function.h b/paddle/operators/elementwise_op_function.h index 0c75276b03140473cee4b57a4022ff3c6989ab4c..db5d30c1af286913f8decd7ab74058fd732ead65 100644 --- a/paddle/operators/elementwise_op_function.h +++ b/paddle/operators/elementwise_op_function.h @@ -340,6 +340,13 @@ void ElementwiseGradCompute(const framework::ExecutionContext& ctx) { return; } + if (y_dims.size() == 1 && y_dims[0] == 1) { + // y is a scalar + auto extended_dims = framework::vectorize(x_dims); + extended_dims.push_back(1); + x_dims = framework::make_ddim(extended_dims); + } + int axis = ctx.Attr("axis"); axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis); @@ -356,5 +363,50 @@ void ElementwiseGradCompute(const framework::ExecutionContext& ctx) { return; } } + +template +void ElementwiseComputeEx(const framework::ExecutionContext& ctx) { + using Tensor = framework::Tensor; + + auto* x = ctx.Input("X"); + auto* y = ctx.Input("Y"); + auto* z = ctx.Output("Out"); + z->mutable_data(ctx.GetPlace()); + TransformFunctor functor( + x, y, z, ctx.template device_context(), Functor()); + + auto x_dims = x->dims(); + auto y_dims = y->dims(); + PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(), + "Rank of first input must >= rank of second input."); + + if (x_dims == y_dims) { + functor.Run(); + return; + } + + if (y_dims.size() == 1 && y_dims[0] == 1) { + // y is a scalar + auto extended_dims = framework::vectorize(x_dims); + extended_dims.push_back(1); + x_dims = framework::make_ddim(extended_dims); + } + + int axis = ctx.Attr("axis"); + axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis); + PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(), + "Axis should be in range [0, x_dims)"); + + int pre, n, post; + get_mid_dims(x_dims, y_dims, axis, pre, n, post); + if (post == 1) { + functor.RunRowWise(n, pre); + return; + } else { + functor.RunMidWise(n, pre, post); + return; + } +} + } // namespace operators } // namespace paddle diff --git a/paddle/operators/elementwise_sub_op.h b/paddle/operators/elementwise_sub_op.h index 347e92f87c7a46d3935f69864718d63661bef05f..a2aca793026189ec87e00b52d7c351689f870400 100644 --- a/paddle/operators/elementwise_sub_op.h +++ b/paddle/operators/elementwise_sub_op.h @@ -18,11 +18,16 @@ limitations under the License. */ namespace paddle { namespace operators { +template +struct SubFunctor { + inline HOSTDEVICE T operator()(T a, T b) const { return a - b; } +}; + template class ElementwiseSubKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - ElementwiseCompute(ctx); + ElementwiseComputeEx, DeviceContext, T>(ctx); } }; diff --git a/paddle/operators/expand_op.cc b/paddle/operators/expand_op.cc index 08fa91ed72aa41ed2f513c090b9085410bb5cc47..043c93654d33f7c105c89960e18ec72d3557237d 100644 --- a/paddle/operators/expand_op.cc +++ b/paddle/operators/expand_op.cc @@ -58,21 +58,21 @@ class ExpandOpMaker : public framework::OpProtoAndCheckerMaker { ExpandOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", - "(Tensor, default Tensor) A tensor with rank in [1, 6]." - "X is the input tensor to be expanded."); + "(Tensor, default Tensor). A tensor with rank in [1, 6]." + "X is the input to be expanded."); AddOutput("Out", - "(Tensor, default Tensor) A tensor with rank in [1, 6]." - "The rank of Output(Out) is same as Input(X) except that each " - "dimension size of Output(Out) is equal to corresponding " - "dimension size of Input(X) multiplying corresponding value of " - "Attr(expand_times)."); + "(Tensor, default Tensor). A tensor with rank in [1, 6]." + "The rank of Output(Out) have the same with Input(X). " + "After expanding, size of each dimension of Output(Out) is equal " + "to size of the corresponding dimension of Input(X) multiplying " + "the corresponding value given by Attr(expand_times)."); AddAttr>("expand_times", "Expand times number for each dimension."); AddComment(R"DOC( Expand operator tiles the input by given times number. You should set times number for each dimension by providing attribute 'expand_times'. The rank of X -should be in [1, 6]. Please notice that size of 'expand_times' must be same with -X's rank. Following is a using case: +should be in [1, 6]. Please note that size of 'expand_times' must be the same +with X's rank. Following is a using case: Input(X) is a 3-D tensor with shape [2, 3, 1]: diff --git a/paddle/operators/lstm_op.cc b/paddle/operators/lstm_op.cc index 3b90b64b4effacf7240fb1bee8c0aa44251ad727..afb095a04e73c2f09b828c01630ef2347ff49613 100644 --- a/paddle/operators/lstm_op.cc +++ b/paddle/operators/lstm_op.cc @@ -117,7 +117,7 @@ class LSTMOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("C0", "(Tensor, optional) the initial cell state is an optional " "input. This is a tensor with shape (N x D), where N is the " - "batch size. `H0` and `C0` can be NULL but only at the same time") + "batch size. `H0` and `C0` can be NULL but only at the same time.") .AsDispensable(); AddInput("Weight", "(Tensor) the learnable hidden-hidden weights." diff --git a/paddle/operators/math/sequence_padding.cc b/paddle/operators/math/sequence_padding.cc index fd66455eaef60209b9ca334480951a9f7687729b..2e69aa47eb8a060a6cac6588b2f37960898aba92 100644 --- a/paddle/operators/math/sequence_padding.cc +++ b/paddle/operators/math/sequence_padding.cc @@ -32,7 +32,8 @@ class PaddingLoDTensorFunctor { 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(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 { "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* padding_data = padding.data(); - 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(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 { 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(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 { "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* seq_data = seq.data(); - 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(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; diff --git a/paddle/operators/math/sequence_padding.cu b/paddle/operators/math/sequence_padding.cu index e4be178f81581dea2e84cf488b01d5f7f4cc0030..a38df26f59569c4fd54a1ba5691b2cd5f3245344 100644 --- a/paddle/operators/math/sequence_padding.cu +++ b/paddle/operators/math/sequence_padding.cu @@ -71,7 +71,8 @@ class PaddingLoDTensorFunctor { 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(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 { "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 { 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 { 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(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 { "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 { 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. diff --git a/paddle/operators/math/sequence_padding_test.cc b/paddle/operators/math/sequence_padding_test.cc index 9799bcd65dc65d5741813374c68a2640eaf4556c..3e504f4a15c2cb4e2380f5ff8a39d83626dae062 100644 --- a/paddle/operators/math/sequence_padding_test.cc +++ b/paddle/operators/math/sequence_padding_test.cc @@ -31,7 +31,7 @@ void TestSequencePadding(const paddle::framework::LoD& lod, cpu_seq.set_lod(lod); cpu_seq.mutable_data(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()[i] = static_cast(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()[i], cpu_seq_back.data()[i]); } diff --git a/paddle/operators/op_documentation/batch_norm_op.md b/paddle/operators/op_documentation/batch_norm_op.md index 80948adf2b9047a9685dbdd90b2296b5a955f9c1..d1392619c42d9206bf4bddcd33ad11b033e6cbdb 100644 --- a/paddle/operators/op_documentation/batch_norm_op.md +++ b/paddle/operators/op_documentation/batch_norm_op.md @@ -66,7 +66,7 @@ As most C++ operators do, `batch_norm_op` is defined by inputs, outputs, attribu The following graph showes the training computational process of `batch_norm_op`: - + cudnn provides APIs to finish the whole series of computation, we can use them in our GPU kernel. @@ -124,7 +124,7 @@ for pass_id in range(PASS_NUM): `is_infer` is an attribute. Once an operator is created, its attributes can not be changed. It suggests us that we shall maintain two `batch_norm_op` in the model, one's `is_infer` is `True`(we call it `infer_batch_norm_op`) and the other one's is `False`(we call it `train_batch_norm_op`). They share all parameters and variables, but be placed in two different branches. That is to say, if a network contains a `batch_norm_op`, it will fork into two branches, one go through `train_batch_norm_op` and the other one go through `infer_batch_norm_op`:
- +
Just like what is shown in the above graph, the net forks before `batch_norm_op` and will never merge again. All the operators after `batch_norm_op` will duplicate. diff --git a/paddle/operators/parallel_do_op.cc b/paddle/operators/parallel_do_op.cc index e1bec0421e76143bef669a4f6fa373cdf01226b2..a00458ea068dd703d2c7f362511ed08bc212d2a8 100644 --- a/paddle/operators/parallel_do_op.cc +++ b/paddle/operators/parallel_do_op.cc @@ -30,16 +30,13 @@ static constexpr char kParallelScopes[] = "parallel_scopes"; static constexpr char kParallelBlock[] = "sub_block"; -// using ParallelScopeVar = std::vector; using LoDTensor = framework::LoDTensor; -using OperatorBase = framework::OperatorBase; -void SplitTensorAndMoveTensorToScopes( - const framework::Scope &scope, - const std::vector &sub_scopes, +static void SplitTensorAndMoveTensorToScopes( + const framework::Scope &scope, std::vector *sub_scopes, const std::vector &places, const std::vector &names) { - PADDLE_ENFORCE_EQ(sub_scopes.size(), places.size()); + size_t num_sub_scopes = 0; for (auto &argu : names) { auto *var = scope.FindVar(argu); const auto &tensor = var->Get(); @@ -48,13 +45,31 @@ void SplitTensorAndMoveTensorToScopes( for (auto &lod : lod_tensors) { VLOG(3) << lod.dims(); } + if (num_sub_scopes == 0) { + num_sub_scopes = lod_tensors.size(); + } else { + PADDLE_ENFORCE_EQ(num_sub_scopes, lod_tensors.size()); + } + PADDLE_ENFORCE_NE(num_sub_scopes, 0); + if (sub_scopes->size() == 0) { + sub_scopes->reserve(num_sub_scopes); + for (size_t i = 0; i < num_sub_scopes; ++i) { + sub_scopes->emplace_back(&scope.NewScope()); + } + } - for (size_t i = 0; i < sub_scopes.size(); ++i) { - *sub_scopes[i]->Var(argu)->GetMutable() = lod_tensors[i]; + for (size_t i = 0; i < lod_tensors.size(); ++i) { + *(*sub_scopes)[i]->Var(argu)->GetMutable() = lod_tensors[i]; } } } +void WaitOnPlace(const platform::Place place) { + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + dev_ctx.Wait(); +} + void WaitOnPlaces(const std::vector places) { platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); @@ -70,7 +85,7 @@ class ParallelDoOp : public framework::OperatorBase { const framework::VariableNameMap &inputs, const framework::VariableNameMap &outputs, const framework::AttributeMap &attrs) - : OperatorBase(type, inputs, outputs, attrs) {} + : framework::OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, const platform::Place &place) const override { @@ -85,19 +100,17 @@ class ParallelDoOp : public framework::OperatorBase { auto &sub_scopes = *scope.FindVar(Output(kParallelScopes)) ->GetMutable>(); - for (size_t place_idx = 0; place_idx < places.size(); ++place_idx) { - sub_scopes.push_back(&scope.NewScope()); - } // split input - SplitTensorAndMoveTensorToScopes(scope, sub_scopes, places, + SplitTensorAndMoveTensorToScopes(scope, &sub_scopes, places, Inputs(kInputs)); + // copy parameter for (auto ¶m : Inputs(kParameters)) { PADDLE_ENFORCE(scope.FindVar(param)->IsType(), "Only support parameter type as LoDTensor"); auto &src = scope.FindVar(param)->Get(); - for (size_t i = 0; i < places.size(); ++i) { + for (size_t i = 0; i < sub_scopes.size(); ++i) { auto &place = places[i]; auto *sub_scope = sub_scopes[i]; auto *dst = sub_scope->Var(param)->GetMutable(); @@ -108,9 +121,7 @@ class ParallelDoOp : public framework::OperatorBase { std::vector> workers; workers.reserve(places.size()); - for (size_t place_idx = 0; place_idx < places.size(); ++place_idx) { - VLOG(3) << "Run " << place_idx; - + for (size_t place_idx = 0; place_idx < sub_scopes.size(); ++place_idx) { auto &place = places[place_idx]; auto *cur_scope = sub_scopes[place_idx]; @@ -157,21 +168,16 @@ ParallelDo Operator. } }; -class ParallelDoGradOp : public OperatorBase { +class ParallelDoGradOp : public framework::OperatorBase { public: ParallelDoGradOp(const std::string &type, const framework::VariableNameMap &inputs, const framework::VariableNameMap &outputs, const framework::AttributeMap &attrs) - : OperatorBase(type, inputs, outputs, attrs) {} + : framework::OperatorBase(type, inputs, outputs, attrs) {} void Run(const framework::Scope &scope, const platform::Place &place) const override { - // // get device context from pool - // platform::DeviceContextPool &pool = - // platform::DeviceContextPool::Instance(); - // auto &dev_ctx = *pool.Get(place); - auto *block = Attr(kParallelBlock); auto *program = block->Program(); @@ -181,26 +187,16 @@ class ParallelDoGradOp : public OperatorBase { auto &places = scope.FindVar(Input(kPlaces))->Get(); // feed output@grad - SplitTensorAndMoveTensorToScopes(scope, sub_scopes, places, - Inputs(framework::GradVarName(kOutputs))); + SplitTensorAndMoveTensorToScopes( + scope, const_cast *>(&sub_scopes), + places, Inputs(framework::GradVarName(kOutputs))); WaitOnPlaces(places); - // for debugging - for (auto &s : Inputs(framework::GradVarName(kOutputs))) { - VLOG(3) << s; - VLOG(3) << scope.FindVar(s)->Get(); - for (auto *sub_scope : sub_scopes) { - VLOG(3) << sub_scope->FindVar(s)->Get(); - } - } - // exe run std::vector> workers; - for (size_t place_idx = 0; place_idx < places.size(); ++place_idx) { - VLOG(3) << "Run " << place_idx; - - auto &place = places[place_idx]; - auto *cur_scope = sub_scopes[place_idx]; + for (size_t i = 0; i < sub_scopes.size(); ++i) { + auto &place = places[i]; + auto *cur_scope = sub_scopes[i]; // execute workers.emplace_back(framework::Async([program, cur_scope, place, block] { @@ -216,33 +212,40 @@ class ParallelDoGradOp : public OperatorBase { // merge grad for (auto &s : Outputs(framework::GradVarName(kParameters))) { - VLOG(3) << "merge grad " << s; - - auto &t = sub_scopes[0]->FindVar(s)->Get(); - VLOG(3) << t; - - std::string s_buf = s + "@BUF"; - auto *t_buf = sub_scopes[0]->Var(s_buf)->GetMutable(); - - for (size_t place_idx = 1; place_idx < places.size(); ++place_idx) { - auto &tt = sub_scopes[place_idx]->FindVar(s)->Get(); - VLOG(3) << place_idx; - VLOG(3) << tt; - framework::Copy(tt, places[0], t_buf); + auto &result = sub_scopes[0]->FindVar(s)->Get(); + std::string tmp_name; + auto *tmp = sub_scopes[0]->Var(&tmp_name)->GetMutable(); + + for (size_t i = 1; i < sub_scopes.size(); ++i) { + auto &tensor_to_merge = sub_scopes[i]->FindVar(s)->Get(); + if (!(places[i] == places[0])) { + framework::Copy(tensor_to_merge, places[0], tmp); + WaitOnPlace(places[0]); + } else { + tmp->ShareDataWith(tensor_to_merge); + } auto sum_op = framework::OpRegistry::CreateOp( - "sum", {{"X", {s, s_buf}}}, {{"Out", {s}}}, + "sum", {{"X", {s, tmp_name}}}, {{"Out", {s}}}, framework::AttributeMap{}); sum_op->Run(*sub_scopes[0], places[0]); - WaitOnPlaces(places); + WaitOnPlace(places[0]); } - VLOG(3) << t; - framework::Copy(t, place, scope.FindVar(s)->GetMutable()); + VLOG(3) << result; + framework::Copy(result, place, scope.FindVar(s)->GetMutable()); } + WaitOnPlaces(places); } }; +std::ostream &operator<<(std::ostream &sout, + const std::vector &strs) { + std::copy(strs.begin(), strs.end(), + std::ostream_iterator(sout, ",")); + return sout; +} + class ParallelDoGradOpDescMaker : public framework::SingleGradOpDescMaker { public: using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; @@ -283,18 +286,30 @@ class ParallelDoGradOpShapeInference : public framework::InferShapeBase { void operator()(framework::InferShapeContext *ctx) const override { std::vector input{kParameters, kInputs}; std::vector output{kOutputs}; - for (auto &s : input) { - PADDLE_ENFORCE(ctx->HasInputs(s)); - PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName(s)), - "Cannot find the gradient variable %s", - framework::GradVarName(s)); - } + + PADDLE_ENFORCE(ctx->HasInputs(kParameters)); + PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName(kParameters))); + PADDLE_ENFORCE(ctx->HasInput(kInputs)); + for (auto &s : output) { PADDLE_ENFORCE(ctx->HasInputs(s)); } - for (auto &s : input) { - ctx->SetOutputsDim(framework::GradVarName(s), ctx->GetInputsDim(s)); + + ctx->SetOutputsDim(framework::GradVarName(kParameters), + ctx->GetInputsDim(kParameters)); + + auto i_dims = ctx->GetInputsDim(kInputs); + auto ig_names = ctx->Outputs(framework::GradVarName(kInputs)); + + for (size_t i = 0; i < ig_names.size(); ++i) { + auto &ig_name = ig_names[i]; + if (ig_name == framework::kEmptyVarName) { + continue; + } + + ctx->SetDims({ig_name}, {i_dims[i]}); } + if (ctx->HasInputs(kParameters)) { PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName(kParameters))); ctx->SetOutputsDim(framework::GradVarName(kParameters), diff --git a/paddle/operators/pool_op.cc b/paddle/operators/pool_op.cc index 3e567efd082ed913ce3c19f87c93a2868ebe8864..b97333bb1a13a0170c325520b86ac73e68282f91 100644 --- a/paddle/operators/pool_op.cc +++ b/paddle/operators/pool_op.cc @@ -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("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(); + 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("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(); + use_cudnn &= dev_ctx.cudnn_handle() != nullptr; + } +#endif framework::LibraryType library_; if (use_cudnn) { library_ = framework::LibraryType::kCUDNN; diff --git a/paddle/operators/recv_op.cc b/paddle/operators/recv_op.cc index f9ed7516826319da422fbb0af4e5c277afa7ae40..8d1479bdd6311709baaf2a6c673db3d0de4610f8 100644 --- a/paddle/operators/recv_op.cc +++ b/paddle/operators/recv_op.cc @@ -27,12 +27,17 @@ limitations under the License. */ #include "paddle/operators/detail/grpc_server.h" #include "paddle/operators/detail/sendrecvop_utils.h" #include "paddle/operators/detail/simple_block_queue.h" +#include "paddle/string/printf.h" #define LISTEN_TERMINATE_MESSAGE "TERMINATE@RECV" namespace paddle { namespace operators { +constexpr int kCondStart = 0; +constexpr int kCondRunning = 1; +constexpr int kCondDone = 2; + void RunServer(std::shared_ptr service) { service->RunSyncUpdate(); VLOG(4) << "RunServer thread end"; @@ -77,42 +82,41 @@ class RecvOp : public framework::OperatorBase { if (grads_counter_.find(varname) == grads_counter_.end()) { grads_counter_[varname] = 0; } - char ret[256]; - snprintf(ret, sizeof(ret), "%s.trainer_%d", varname.c_str(), - grads_counter_[varname]++); - return std::string(ret); + return string::Sprintf("%s.trainer_%d", varname, grads_counter_[varname]++); } void Run(const framework::Scope &scope, const platform::Place &dev_place) const override { - // FIXME(typhoonzero): no new scopes for every run. - framework::Scope &recv_scope = scope.NewScope(); platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); auto &dev_ctx = *pool.Get(dev_place); + framework::Scope &recv_scope = scope.NewScope(); // FIXME(Yancey1989): initialize rpc server with laze mode. rpc_service_->SetScope(&recv_scope); rpc_service_->SetDevCtx(&dev_ctx); auto param_list = Attr>("ParamList"); auto grad_list = Attr>("GradList"); - auto trainer_count = Attr("Trainers"); + auto fan_in = Attr("Fanin"); size_t param_count = param_list.size(); - rpc_service_->Reset(); + std::string program_str = Attr("OptimizeProgram"); + framework::proto::ProgramDesc program_desc; + program_desc.ParseFromString(program_str); + framework::ProgramDesc program(program_desc); + framework::Executor executor(dev_place); + // TODO(typhoonzero): change this to a while_op for every cluster-batch. bool exit_flag = false; - VLOG(4) << "param_count:" << param_count - << " trainer_count:" << trainer_count; + int64_t barrier_size = param_count * fan_in; while (!exit_flag) { - // TODO(gognwb): simply this loop. - // Get from multiple trainers, we don't care about order in which - // the gradient arrives, just add suffix 0~n then average the gradient. - for (size_t i = 0; i < param_count * trainer_count; ++i) { - // blocking get one var from client. + // Get from multiple trainers, we don't care about the order in which + // the gradients arrives, just add suffix 0~n and merge the gradient. + rpc_service_->SetCond(0); + for (size_t i = 0; i < barrier_size; ++i) { const detail::MessageWithName &v = rpc_service_->Get(); auto grad_var_name = v.first; if (grad_var_name == LISTEN_TERMINATE_MESSAGE) { - VLOG(4) << "received LISTEN_TERMINATE_MESSAGE and RunOp.Run() exit"; + LOG(INFO) << "received terminate message and exit"; exit_flag = true; break; } @@ -121,49 +125,31 @@ class RecvOp : public framework::OperatorBase { if (it != grad_list.end()) { param_var_name = param_list[it - grad_list.begin()]; } else { - LOG(ERROR) << "grad have no paired param found!\"" << grad_var_name - << "\""; + LOG(ERROR) << "grad have no paired param:" << grad_var_name; } VLOG(3) << "recved grad: " << grad_var_name << " updating param: " << param_var_name; - - auto *merged_grad = recv_scope.FindVar(grad_var_name); - if (merged_grad == nullptr) { - auto *ptr = recv_scope.Var(grad_var_name); - CreateTensorFromMessageType(ptr, v.second.type()); - VLOG(3) << "Create Variable " << grad_var_name - << " on recv scope, which pointer is " << ptr << " type is " - << v.second.type(); - } - - if (trainer_count > 1) { + if (fan_in > 1) { grad_var_name = this->GetGradVarNameForTrainer(grad_var_name); } - - auto *var = recv_scope.Var(grad_var_name); + auto *var = recv_scope.FindVar(grad_var_name); + if (var == nullptr) { + LOG(ERROR) << "can not find server side var: " << grad_var_name; + PADDLE_THROW("can not find server side var"); + } detail::DeserializeFromMessage(v.second, dev_ctx, var); } - if (exit_flag) { break; } - - rpc_service_->Reset(); - - std::string program_str = Attr("OptimizeProgram"); - framework::proto::ProgramDesc program_desc; - program_desc.ParseFromString(program_str); - framework::ProgramDesc program(program_desc); - framework::Executor executor(dev_place); - // Run sub graph to get optimized tensor try { executor.Run(program, &recv_scope, 0, /*global_block*/ false /*create_local_scope*/, false /*create_vars*/); } catch (std::exception &e) { LOG(ERROR) << "run sub program error " << e.what(); } - - rpc_service_->Done(); + rpc_service_->SetCond(1); + rpc_service_->WaitClientGet(barrier_size); grads_counter_.clear(); } // while(true) } @@ -199,7 +185,7 @@ This operator will recv tensor from send_op "GradList", "type list of string", "grad->param name mapping to find which param to optimize.") .SetDefault({}); - AddAttr("Trainers", "type int", + AddAttr("Fanin", "type int", "Number of trainers in the current cluster job") .SetDefault(1); } diff --git a/paddle/operators/reduce_op.cc b/paddle/operators/reduce_op.cc index 172d28bb3b647901d4de7bc03c9de21e3468a364..09b7091358e65221374a604122b742d763cfbafc 100644 --- a/paddle/operators/reduce_op.cc +++ b/paddle/operators/reduce_op.cc @@ -129,7 +129,7 @@ If reduce_all is true, just reduce along all dimensions and output a scalar. } void SetComment(std::string name, std::string op) { - Replace(comment_, "{ReduceOP}", name); + Replace(comment_, "{ReduceOp}", name); Replace(comment_, "{reduce}", op); } }; diff --git a/paddle/operators/send_op.cc b/paddle/operators/send_op.cc index 7c81a9524d6609a65b3167d95053bf4e85eef0db..a38d8b7a6f24ece35bd65332463f5abac2dc2d74 100644 --- a/paddle/operators/send_op.cc +++ b/paddle/operators/send_op.cc @@ -41,10 +41,13 @@ class SendOp : public framework::OperatorBase { platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); auto& ctx = *pool.Get(place); for (size_t i = 0; i < ins.size(); i++) { + VLOG(3) << "sending " << ins[i]; client_.AsyncSendVariable(epmap[i], ctx, scope, ins[i]); } + PADDLE_ENFORCE(client_.Wait()); for (size_t i = 0; i < outs.size(); i++) { + VLOG(3) << "getting " << outs[i]; client_.AsyncGetVariable(epmap[i], ctx, scope, outs[i]); } diff --git a/paddle/operators/sequence_erase_op.cc b/paddle/operators/sequence_erase_op.cc index d17b2686238b2d2f872331edfdbb095fb8693b87..aa0c00aa6f7854ee5e34aef78970971b78df6514 100644 --- a/paddle/operators/sequence_erase_op.cc +++ b/paddle/operators/sequence_erase_op.cc @@ -86,4 +86,5 @@ REGISTER_OP_WITHOUT_GRADIENT(sequence_erase, ops::SequenceEraseOp, ops::SequenceEraseOpMaker); REGISTER_OP_CPU_KERNEL( sequence_erase, - ops::SequenceEraseKernel); + ops::SequenceEraseKernel, + ops::SequenceEraseKernel); diff --git a/paddle/operators/sequence_erase_op.cu b/paddle/operators/sequence_erase_op.cu index 5da8eba3e1ac1fb85dfc65c2fd801574599e02d9..f1e3b96acd0259de2b3ca1348834bd17e1e174a2 100644 --- a/paddle/operators/sequence_erase_op.cu +++ b/paddle/operators/sequence_erase_op.cu @@ -23,27 +23,22 @@ using platform::PADDLE_CUDA_NUM_THREADS; using LoDTensor = framework::LoDTensor; template -__global__ void LabelErasedIdx(const T* in_dat, const int in_len, - const T* tokens, const int tokens_len, - int* num_erased) { +__global__ void LabelErasedIdx(const T* in_dat, const int64_t in_len, + const int* tokens, const size_t tokens_len, + size_t* num_erased) { int index = blockIdx.x * blockDim.x + threadIdx.x; if (index < in_len) { - int erased = 0; - for (int i = 0; i < tokens_len; ++i) { + for (size_t i = 0; i < tokens_len; ++i) { if (in_dat[index] == tokens[i]) { - erased = 1; + num_erased[index + 1] = 1; + break; } } - num_erased[index + 1] = erased; - if (index == 0) { - num_erased[0] = 0; - } } } -template -__global__ void GetOutLod(const T* num_erased, const int* in_lod, - const int lod_len, int* out_lod0) { +__global__ void GetOutLod(const size_t* num_erased, const size_t* in_lod, + const size_t lod_len, size_t* out_lod0) { int index = blockIdx.x * blockDim.x + threadIdx.x; if (index < lod_len) { out_lod0[index] = in_lod[index] - num_erased[in_lod[index]]; @@ -51,11 +46,11 @@ __global__ void GetOutLod(const T* num_erased, const int* in_lod, } template -__global__ void SetOutput(const T* in_dat, const int in_len, - const int* num_erased, T* out_dat) { +__global__ void SetOutput(const T* in_dat, const int64_t in_len, + const size_t* num_erased, T* out_dat) { int index = blockIdx.x * blockDim.x + threadIdx.x; if (index < in_len) { - if (in_dat[index] != in_dat[index + 1]) { + if (num_erased[index] == num_erased[index + 1]) { out_dat[index - num_erased[index]] = in_dat[index]; } } @@ -72,53 +67,44 @@ class SequenceEraseOpCUDAKernel : public framework::OpKernel { PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now."); PADDLE_ENFORCE_EQ(lod[0].back(), (size_t)in->numel(), "The actual size mismatches with the LoD information."); - auto tokens = ctx.Attr>("tokens"); - auto tokens_len = tokens.size(); + auto tokens = ctx.Attr>("tokens"); auto in_len = in->numel(); auto in_dat = in->data(); - auto lod0 = lod[0]; - - thrust::host_vector host_tokens(tokens_len); - for (size_t i = 0; i < tokens.size(); ++i) { - host_tokens[i] = tokens[i]; - } - thrust::device_vector dev_tokens = host_tokens; - thrust::device_vector num_erased(in_len + 1); - - T* dev_tokens_ptr = thrust::raw_pointer_cast(dev_tokens.data()); - int* num_erased_ptr = thrust::raw_pointer_cast(num_erased.data()); + // Copy tokens to GPU + thrust::device_vector dev_tokens(tokens.begin(), tokens.end()); + int* dev_tokens_ptr = thrust::raw_pointer_cast(dev_tokens.data()); + // Count number of elements to be erased + thrust::device_vector num_erased(in_len + 1, 0); + size_t* num_erased_ptr = thrust::raw_pointer_cast(num_erased.data()); auto stream = ctx.cuda_device_context().stream(); LabelErasedIdx<<<(in_len - 1) / PADDLE_CUDA_NUM_THREADS + 1, PADDLE_CUDA_NUM_THREADS, 0, stream>>>( - in_dat, in_len, dev_tokens_ptr, tokens_len, num_erased_ptr); + in_dat, in_len, dev_tokens_ptr, tokens.size(), num_erased_ptr); thrust::inclusive_scan(num_erased.begin() + 1, num_erased.end(), num_erased.begin() + 1); - // Calc LoD + // Copy LoD to GPU + auto lod0 = lod[0]; auto lod_len = lod0.size(); - thrust::host_vector host_lod(lod_len); - for (size_t i = 0; i < lod_len; ++i) { - host_lod[i] = lod0[i]; - } - thrust::device_vector dev_in_lod = host_lod; - thrust::device_vector dev_out_lod(lod_len); - int* dev_in_lod_ptr = thrust::raw_pointer_cast(dev_in_lod.data()); - int* dev_out_lod_ptr = thrust::raw_pointer_cast(dev_out_lod.data()); + thrust::device_vector dev_in_lod = lod0; + size_t* dev_in_lod_ptr = thrust::raw_pointer_cast(dev_in_lod.data()); + + // Calc output LoD + thrust::device_vector dev_out_lod(lod_len); + size_t* dev_out_lod_ptr = thrust::raw_pointer_cast(dev_out_lod.data()); GetOutLod<<<(lod_len - 1) / PADDLE_CUDA_NUM_THREADS + 1, PADDLE_CUDA_NUM_THREADS, 0, stream>>>( num_erased_ptr, dev_in_lod_ptr, lod_len, dev_out_lod_ptr); - thrust::host_vector host_out_lod = dev_out_lod; - std::vector out_lod0(lod_len, 0); - for (size_t i = 0; i < lod_len; i++) { - out_lod0[i] = host_out_lod[i]; - } + + // Set LoD for output + thrust::host_vector out_lod0 = dev_out_lod; framework::LoD out_lod; out_lod.push_back(out_lod0); out->set_lod(out_lod); // Set output - out->Resize({out_lod0.back(), 1}); + out->Resize({static_cast(out_lod0.back()), 1}); auto out_dat = out->mutable_data(ctx.GetPlace()); SetOutput<<<(in_len - 1) / PADDLE_CUDA_NUM_THREADS + 1, PADDLE_CUDA_NUM_THREADS, 0, stream>>>(in_dat, in_len, @@ -130,4 +116,5 @@ class SequenceEraseOpCUDAKernel : public framework::OpKernel { } // namespace paddle REGISTER_OP_CUDA_KERNEL(sequence_erase, - paddle::operators::SequenceEraseOpCUDAKernel); + paddle::operators::SequenceEraseOpCUDAKernel, + paddle::operators::SequenceEraseOpCUDAKernel); diff --git a/paddle/operators/split_selected_rows_op.cc b/paddle/operators/split_selected_rows_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..d9a023987b3fcdf08a995d66bad65941dfd8d440 --- /dev/null +++ b/paddle/operators/split_selected_rows_op.cc @@ -0,0 +1,114 @@ +/* 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>("rows_sections", "Rows section for output.") + .SetDefault(std::vector({})); + AddAttr>("height_sections", + "Height for each output SelectedRows.") + .SetDefault(std::vector({})); + + 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 height_sections = + ctx->Attrs().Get>("height_sections"); + std::vector rows_sections = + ctx->Attrs().Get>("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 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(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 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(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); diff --git a/paddle/operators/split_selected_rows_op.cu b/paddle/operators/split_selected_rows_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..983285480fd9de7a2a4d2787a9bba72c160b7fae --- /dev/null +++ b/paddle/operators/split_selected_rows_op.cu @@ -0,0 +1,19 @@ +/* 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); diff --git a/paddle/operators/split_selected_rows_op.h b/paddle/operators/split_selected_rows_op.h new file mode 100644 index 0000000000000000000000000000000000000000..1cae53f1af250064833f593bd90ff790a1921fcd --- /dev/null +++ b/paddle/operators/split_selected_rows_op.h @@ -0,0 +1,58 @@ +/* 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 +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class SplitSelectedRowsOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto outs = ctx.MultiOutput("Out"); + + auto rows_sections = ctx.Attr>("rows_sections"); + auto height_sections = ctx.Attr>("height_sections"); + + int64_t n = outs.size(); + int offset = 0; + + for (int64_t i = 0; i < n; ++i) { + framework::Vector 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(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 diff --git a/paddle/string/to_string.h b/paddle/string/to_string.h index 3b3bcc69a478045156225728236174fd601461dd..178edc18951f94291a63566516df36956f25f67b 100644 --- a/paddle/string/to_string.h +++ b/paddle/string/to_string.h @@ -15,9 +15,15 @@ limitations under the License. */ #pragma once #include #include +#include namespace paddle { namespace string { +inline std::ostream& operator<<(std::ostream& s, const std::type_index& t) { + s << t.name(); + return s; +} + template inline std::string to_string(T v) { std::ostringstream sout; @@ -25,6 +31,11 @@ inline std::string to_string(T v) { return sout.str(); } +template <> +inline std::string to_string(std::type_index t) { + return t.name(); +} + // Faster std::string/const char* type template <> inline std::string to_string(std::string v) { diff --git a/python/paddle/trainer_config_helpers/evaluators.py b/python/paddle/trainer_config_helpers/evaluators.py index 95797fba8f67bacb421f5c2813ad6332bc53cbc9..0eeaf7eabb179f19d2af8dafe821f7baa153fead 100644 --- a/python/paddle/trainer_config_helpers/evaluators.py +++ b/python/paddle/trainer_config_helpers/evaluators.py @@ -16,13 +16,22 @@ from paddle.trainer.config_parser import * from default_decorators import * __all__ = [ - "evaluator_base", "classification_error_evaluator", "auc_evaluator", - "pnpair_evaluator", "precision_recall_evaluator", "ctc_error_evaluator", - "chunk_evaluator", "sum_evaluator", "column_sum_evaluator", - "value_printer_evaluator", "gradient_printer_evaluator", - "maxid_printer_evaluator", "maxframe_printer_evaluator", - "seqtext_printer_evaluator", "classification_error_printer_evaluator", - "detection_map_evaluator" + "evaluator_base", + "classification_error_evaluator", + "auc_evaluator", + "pnpair_evaluator", + "precision_recall_evaluator", + "ctc_error_evaluator", + "chunk_evaluator", + "sum_evaluator", + "column_sum_evaluator", + "value_printer_evaluator", + "gradient_printer_evaluator", + "maxid_printer_evaluator", + "maxframe_printer_evaluator", + "seqtext_printer_evaluator", + "classification_error_printer_evaluator", + "detection_map_evaluator", ] diff --git a/python/paddle/v2/fluid/distribute_transpiler.py b/python/paddle/v2/fluid/distribute_transpiler.py index 06a7b6fb02f5e38e0762d112492854f027fe66ad..bd957f88de5d51a2fa3e482284e2d8080f1be76b 100644 --- a/python/paddle/v2/fluid/distribute_transpiler.py +++ b/python/paddle/v2/fluid/distribute_transpiler.py @@ -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 @@ -420,11 +420,24 @@ class DistributeTranspiler: pserver_program = Program() for v in self.param_grad_ep_mapping[endpoint]["params"]: self._clone_var(pserver_program.global_block(), v) + for v in self.param_grad_ep_mapping[endpoint]["grads"]: + # create vars for each trainer in global scope, so + # we don't need to create them when grad arrives. + pserver_program.global_block().create_var( + name=v.name, persistable=True, dtype=v.dtype, shape=v.shape) + for trainer_id in xrange(self.trainers): + print("create variable for program: %s.trainer_%d" % + (v.name, trainer_id)) + pserver_program.global_block().create_var( + name="%s.trainer_%d" % (v.name, trainer_id), + persistable=True, + dtype=v.dtype, + shape=v.shape) # step6 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"): @@ -449,7 +462,7 @@ class DistributeTranspiler: p.name for p in self.param_grad_ep_mapping[endpoint]["grads"] ], - "Trainers": self.trainers + "Fanin": self.trainers }) pserver_program.sync_with_cpp() return pserver_program diff --git a/python/paddle/v2/fluid/framework.py b/python/paddle/v2/fluid/framework.py index 8042febfed7ed7db2f6d1507142b17079aa00fd8..4f8366b64039b2edcd4c273439c87397bdc33595 100644 --- a/python/paddle/v2/fluid/framework.py +++ b/python/paddle/v2/fluid/framework.py @@ -116,8 +116,8 @@ def _debug_string_(proto, throw_on_error=True): """ error_fields = list() if not proto.IsInitialized(error_fields) and throw_on_error: - raise ValueError("{0} are not initialized\nThe message is {1}".format( - error_fields, proto)) + raise ValueError("{0} are not initialized.\nThe message is {1}:\n". + format(error_fields, proto)) return proto.__str__() @@ -374,12 +374,13 @@ class Operator(object): >>> outputs={"Out": [var1]}) Args: - block(Block): The block has the current operator - desc(core.OpDesc): The protobuf description + block(Block): The block has the current operator. + desc(core.OpDesc): The protobuf description. type(str): The type of operator. inputs(dict): The input dictionary. Key is the input parameter name. Value is a list of variables. - outputs(dict): The output dictionary. Has same format with inputs + outputs(dict): The output dictionary which has the same format with + inputs. attrs(dict): The attributes dictionary. Key is attribute name. Value is the attribute value. The attribute type should be as same as the type registered in C++ @@ -436,10 +437,11 @@ class Operator(object): for m in proto.outputs: need.add(m.name) if not given == need: - raise ValueError( - "Incorrect setting for output(s) of operator \"%s\". Need: [%s] Given: [%s]" - % (type, ", ".join(str(e) for e in need), ", ".join( - str(e) for e in given))) + raise ValueError(("Incorrect setting for output(s) of " + "operator \"%s\". Need: [%s] Given: [%s]") % + (type, ", ".join(str(e) + for e in need), ", ".join( + str(e) for e in given))) for out_proto in proto.outputs: out_args = outputs[out_proto.name] @@ -818,9 +820,8 @@ class Program(object): if isinstance(t, Variable): t = t.op else: - raise ValueError( - "All targets of prune() can only be Variable or Operator." - ) + raise ValueError(("All targets of prune() can only be " + "Variable or Operator.")) targets_idx.append([t.block.idx, t.idx]) res = Program() diff --git a/python/paddle/v2/fluid/io.py b/python/paddle/v2/fluid/io.py index 499df05e592855f63f41ec8ceb939edf0e4d435c..e7a06a07145758335986368fbe111de2bb930246 100644 --- a/python/paddle/v2/fluid/io.py +++ b/python/paddle/v2/fluid/io.py @@ -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: diff --git a/python/paddle/v2/fluid/layers/control_flow.py b/python/paddle/v2/fluid/layers/control_flow.py index e3a1437f5ecc91557bd1ea861d37a0864e8e8b33..e72b22c83f65789b9e5d81611bec602d8d78be6b 100644 --- a/python/paddle/v2/fluid/layers/control_flow.py +++ b/python/paddle/v2/fluid/layers/control_flow.py @@ -1343,20 +1343,44 @@ class DynamicRNN(object): else: return self.outputs - def memory(self, init=None, shape=None, value=0.0, dtype='float32'): + def memory(self, + init=None, + shape=None, + value=0.0, + need_reorder=False, + dtype='float32'): self._assert_in_rnn_block_('memory') if init is not None: if not isinstance(init, Variable): raise TypeError( "The input arg `init` of memory() must be a Variable") parent_block = self._parent_block_() + init_tensor = init + if need_reorder == True: + if self.lod_rank_table is None: + raise ValueError( + 'If set need_reorder to True, make sure step_input be ' + 'invoked before ' + 'memory(init=init, need_reordered=True, ...).') + init_reordered = parent_block.create_var( + name=unique_name('dynamic_rnn_mem_init_reordered'), + type=core.VarDesc.VarType.LOD_TENSOR, + dtype=init.dtype) + parent_block.append_op( + type='reorder_lod_tensor_by_rank', + inputs={ + 'X': [init_tensor], + 'RankTable': [self.lod_rank_table] + }, + outputs={'Out': [init_reordered]}) + init_tensor = init_reordered mem_array = parent_block.create_var( name=unique_name('dynamic_rnn_mem_array'), type=core.VarDesc.VarType.LOD_TENSOR_ARRAY, dtype=init.dtype) parent_block.append_op( type='write_to_array', - inputs={'X': init, + inputs={'X': init_tensor, 'I': self.zero_idx}, outputs={'Out': mem_array}) retv = array_read(array=mem_array, i=self.step_idx) diff --git a/python/paddle/v2/fluid/layers/io.py b/python/paddle/v2/fluid/layers/io.py index 6177f0b4d71cdda8637fc40e3e65c72842bf7439..a43e0ee4def668bf7033f37cfa1a3f59d10a88d0 100644 --- a/python/paddle/v2/fluid/layers/io.py +++ b/python/paddle/v2/fluid/layers/io.py @@ -28,9 +28,9 @@ def data(name, **Data Layer** This function takes in the input and based on whether data has - to be returned back as a minibatch, it creates the global variable using + to be returned back as a minibatch, it creates the global variable by using the helper functions. The global variables can be accessed by all the - following operations and layers in the graph. + following operators in the graph. All the input variables of this function are passed in as local variables to the LayerHelper constructor. diff --git a/python/paddle/v2/fluid/layers/nn.py b/python/paddle/v2/fluid/layers/nn.py index 4e8fd407c9983e2827d3137fa4f49a8425d5dce2..7d61d111a8ae033f4b976a2dc45064949ac216f9 100644 --- a/python/paddle/v2/fluid/layers/nn.py +++ b/python/paddle/v2/fluid/layers/nn.py @@ -22,34 +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', + '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' ] @@ -227,6 +206,102 @@ def dynamic_lstm(input, cell_activation='tanh', candidate_activation='tanh', dtype='float32'): + """ + **Dynamic LSTM Layer** + + The defalut implementation is diagonal/peephole connection + (https://arxiv.org/pdf/1402.1128.pdf), the formula is as follows: + + .. math:: + + i_t & = \sigma(W_{ix}x_{t} + W_{ih}h_{t-1} + W_{ic}c_{t-1} + b_i) + + f_t & = \sigma(W_{fx}x_{t} + W_{fh}h_{t-1} + W_{fc}c_{t-1} + b_f) + + \\tilde{c_t} & = act_g(W_{cx}x_t + W_{ch}h_{t-1} + b_c) + + o_t & = \sigma(W_{ox}x_{t} + W_{oh}h_{t-1} + W_{oc}c_t + b_o) + + c_t & = f_t \odot c_{t-1} + i_t \odot \\tilde{c_t} + + h_t & = o_t \odot act_h(c_t) + + where the :math:`W` terms denote weight matrices (e.g. :math:`W_{xi}` is + the matrix of weights from the input gate to the input), :math:`W_{ic}, \ + W_{fc}, W_{oc}` are diagonal weight matrices for peephole connections. In + our implementation, we use vectors to reprenset these diagonal weight + matrices. The :math:`b` terms denote bias vectors (:math:`b_i` is the input + gate bias vector), :math:`\sigma` is the non-line activations, such as + logistic sigmoid function, and :math:`i, f, o` and :math:`c` are the input + gate, forget gate, output gate, and cell activation vectors, respectively, + all of which have the same size as the cell output activation vector :math:`h`. + + The :math:`\odot` is the element-wise product of the vectors. :math:`act_g` + and :math:`act_h` are the cell input and cell output activation functions + and `tanh` is usually used for them. :math:`\\tilde{c_t}` is also called + candidate hidden state, which is computed based on the current input and + the previous hidden state. + + Set `use_peepholes` to `False` to disable peephole connection. The formula + is omitted here, please refer to the paper + http://www.bioinf.jku.at/publications/older/2604.pdf for details. + + Note that these :math:`W_{xi}x_{t}, W_{xf}x_{t}, W_{xc}x_{t}, W_{xo}x_{t}` + operations on the input :math:`x_{t}` are NOT included in this operator. + Users can choose to use fully-connect layer before LSTM layer. + + Args: + input(Variable): The input of dynamic_lstm layer, which supports + variable-time length input sequence. The underlying + tensor in this Variable is a matrix with shape + (T X 4D), where T is the total time steps in this + mini-batch, D is the hidden size. + size(int): 4 * hidden size. + param_attr(ParamAttr): The parameter attribute for the learnable + hidden-hidden weights. + + - The shape is (D x 4D), where D is the hidden + size. + - Weights = {:math:`W_{ch}, W_{ih}, \ + W_{fh}, W_{oh}`} + bias_attr(ParamAttr): The bias attribute for the learnable bias + weights, which contains two parts, input-hidden + bias weights and peephole connections weights if + setting `use_peepholes` to `True`. + + 1. `use_peepholes = False` + - The shape is (1 x 4D). + - Biases = {:math:`b_c, b_i, b_f, b_o`}. + 2. `use_peepholes = True` + - The shape is (1 x 7D). + - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \ + W_{fc}, W_{oc}`}. + use_peepholes(bool): Whether to enable diagonal/peephole connections, + default `True`. + is_reverse(bool): Whether to compute reversed LSTM, default `False`. + gate_activation(str): The activation for input gate, forget gate and + output gate. Choices = ["sigmoid", "tanh", "relu", + "identity"], default "sigmoid". + cell_activation(str): The activation for cell output. Choices = ["sigmoid", + "tanh", "relu", "identity"], default "tanh". + candidate_activation(str): The activation for candidate hidden state. + Choices = ["sigmoid", "tanh", "relu", "identity"], + default "tanh". + dtype(str): Data type. Choices = ["float32", "float64"], default "float32". + + Returns: + tuple: The hidden state, and cell state of LSTM. The shape of both \ + is (T x D), and lod is the same with the `input`. + + Examples: + .. code-block:: python + + hidden_dim = 512 + forward_proj = fluid.layers.fc(input=input_seq, size=hidden_dim * 4, + act=None, bias_attr=None) + forward, _ = fluid.layers.dynamic_lstm( + input=forward_proj, size=hidden_dim * 4, use_peepholes=False) + """ helper = LayerHelper('lstm', **locals()) size = size / 4 weight = helper.create_parameter( @@ -674,6 +749,7 @@ def conv2d(input, groups=None, param_attr=None, bias_attr=None, + use_cudnn=True, act=None): """ **Convlution2D Layer** @@ -737,6 +813,8 @@ 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: @@ -772,6 +850,8 @@ 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 @@ -795,9 +875,12 @@ 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) @@ -945,7 +1028,9 @@ def pool2d(input, pool_type, pool_stride=None, pool_padding=None, - global_pooling=False): + global_pooling=False, + use_cudnn=True, + name=None): """ This function adds the operator for pooling in 2 dimensions, using the pooling configurations mentioned in input parameters. @@ -964,6 +1049,8 @@ 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() @@ -978,7 +1065,8 @@ 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 @@ -991,7 +1079,8 @@ def batch_norm(input, epsilon=1e-05, param_attr=None, bias_attr=None, - data_layout='NCHW'): + data_layout='NCHW', + name=None): """ This function helps create an operator to implement the BatchNorm layer using the configurations from the input parameters. @@ -1067,7 +1156,7 @@ def batch_norm(input, return helper.append_activation(batch_norm_out) -def beam_search_decode(ids, scores): +def beam_search_decode(ids, scores, name=None): helper = LayerHelper('beam_search_decode', **locals()) sentence_ids = helper.create_tmp_variable(dtype=ids.dtype) sentence_scores = helper.create_tmp_variable(dtype=ids.dtype) @@ -1091,7 +1180,9 @@ def conv2d_transpose(input, padding=None, stride=None, dilation=None, - param_attr=None): + param_attr=None, + use_cudnn=True, + name=None): """ The transpose of conv2d layer. @@ -1118,8 +1209,10 @@ def conv2d_transpose(input, contain two integers, (dilation_H, dilation_W). Otherwise, the dilation_H = dilation_W = dilation. param_attr: Parameter Attribute. - main_program(Program): the main program - startup_program(Program): the startup program + use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn + library is installed. Default: True + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: Variable: Output image. @@ -1146,6 +1239,10 @@ 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") @@ -1183,7 +1280,7 @@ def conv2d_transpose(input, return out -def sequence_expand(x, y): +def sequence_expand(x, y, name=None): """Sequence Expand Layer. This layer will expand the input variable **x** according to LoD information of **y**. And the following examples will explain how sequence_expand works: @@ -1227,6 +1324,8 @@ def sequence_expand(x, y): Args: x (Variable): The input variable which is a Tensor or LoDTensor. y (Variable): The input variable which is a LoDTensor. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: Variable: The expanded variable which is a LoDTensor. @@ -1253,7 +1352,8 @@ def lstm_unit(x_t, cell_t_prev, forget_bias=0.0, param_attr=None, - bias_attr=None): + bias_attr=None, + name=None): """Lstm unit layer. The equation of a lstm step is: .. math:: @@ -1300,6 +1400,8 @@ def lstm_unit(x_t, initializer, name etc. bias_attr (ParamAttr): The attributes of bias weights, if not False, bias weights will be created and be set to default value. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: tuple: The hidden value and cell value of lstm unit. @@ -1365,7 +1467,7 @@ def lstm_unit(x_t, return h, c -def reduce_sum(input, dim=None, keep_dim=False): +def reduce_sum(input, dim=None, keep_dim=False, name=None): """ Computes the sum of tensor elements over the given dimension. @@ -1379,6 +1481,8 @@ def reduce_sum(input, dim=None, keep_dim=False): keep_dim (bool): Whether to reserve the reduced dimension in the output Tensor. The result tensor will have one fewer dimension than the :attr:`input` unless :attr:`keep_dim` is true. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: Variable: The reduced Tensor variable. @@ -1409,7 +1513,7 @@ def reduce_sum(input, dim=None, keep_dim=False): return out -def reduce_mean(input, dim=None, keep_dim=False): +def reduce_mean(input, dim=None, keep_dim=False, name=None): """ Computes the mean of tensor elements over the given dimension. @@ -1423,6 +1527,8 @@ def reduce_mean(input, dim=None, keep_dim=False): keep_dim (bool): Whether to reserve the reduced dimension in the output Tensor. The result tensor will have one fewer dimension than the :attr:`input` unless :attr:`keep_dim` is true. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: Variable: The reduced Tensor variable. @@ -1453,7 +1559,7 @@ def reduce_mean(input, dim=None, keep_dim=False): return out -def reduce_max(input, dim=None, keep_dim=False): +def reduce_max(input, dim=None, keep_dim=False, name=None): """ Computes the maximum of tensor elements over the given dimension. @@ -1467,6 +1573,8 @@ def reduce_max(input, dim=None, keep_dim=False): keep_dim (bool): Whether to reserve the reduced dimension in the output Tensor. The result tensor will have one fewer dimension than the :attr:`input` unless :attr:`keep_dim` is true. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: Variable: The reduced Tensor variable. @@ -1497,7 +1605,7 @@ def reduce_max(input, dim=None, keep_dim=False): return out -def reduce_min(input, dim=None, keep_dim=False): +def reduce_min(input, dim=None, keep_dim=False, name=None): """ Computes the minimum of tensor elements over the given dimension. @@ -1511,6 +1619,8 @@ def reduce_min(input, dim=None, keep_dim=False): keep_dim (bool): Whether to reserve the reduced dimension in the output Tensor. The result tensor will have one fewer dimension than the :attr:`input` unless :attr:`keep_dim` is true. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: Variable: The reduced Tensor variable. @@ -1541,20 +1651,22 @@ def reduce_min(input, dim=None, keep_dim=False): return out -def split(input, num_or_sections, dim=-1): +def split(input, num_or_sections, dim=-1, name=None): """ - Splits the tensor into multiple sub-tensors. + Split the input tensor into multiple sub-tensors. Args: input (Variable): The input variable which is a Tensor or LoDTensor. - num_or_sections (int|list): If :attr:`num_or_sections` is an integer, - then the integer indicates the number of equal sized sub-tensors - that the tensor will be divided into. If :attr:`num_or_sections` - is a list of integers, the length of list indicates the number of - sub-tensors and the integers indicate the sizes of sub-tensors' + num_or_sections (int|list): If :attr:`num_or_sections` is an integer, + then the integer indicates the number of equal sized sub-tensors + that the tensor will be divided into. If :attr:`num_or_sections` + is a list of integers, the length of list indicates the number of + sub-tensors and the integers indicate the sizes of sub-tensors' :attr:`dim` dimension orderly. - dim (int): The dimension along which to split. If :math:`dim < 0`, the + dim (int): The dimension along which to split. If :math:`dim < 0`, the dimension to split along is :math:`rank(input) + dim`. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: List: The list of segmented tensor variables. @@ -1597,3 +1709,208 @@ def split(input, num_or_sections, dim=-1): 'axis': dim }) return outs + + +def l2_normalize(x, axis, epsilon=1e-12, name=None): + """ + **L2 normalize Layer** + + The l2 normalize layer normalizes `x` along dimension `axis` using an L2 + norm. For a 1-D tensor (`dim` is fixed to 0), this layer computes + + output = x / sqrt(max(sum(x**2), epsilon)) + + For `x` with more dimensions, this layer independently normalizes each 1-D + slice along dimension `axis`. + + Args: + x(Variable|list): The input tensor to l2_normalize layer. + axis(int): Dimension along which to normalize the input. + epsilon(float): A lower bound value for `x`'s l2 norm. sqrt(epsilon) will + be used as the divisor if the l2 norm of `x` is less than + sqrt(epsilon). + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. + + + Returns: + Variable: The output tensor variable. + + Examples: + .. code-block:: python + + data = fluid.layers.data(name="data", + shape=(3, 17, 13), + dtype="float32") + fc = fluid.layers.l2_normalize(x=data, axis=1) + """ + + if len(x.shape) == 1: axis = 0 + + helper = LayerHelper("l2_normalize", **locals()) + + square = helper.create_tmp_variable(dtype=x.dtype) + helper.append_op(type="square", inputs={"X": x}, outputs={"Out": square}) + + reduced_sum = helper.create_tmp_variable(dtype=x.dtype) + helper.append_op( + type="reduce_sum", + inputs={"X": square}, + outputs={"Out": reduced_sum}, + attrs={ + "dim": 1 if axis is None else axis, + "keep_dim": True, + "reduce_all": False + }) + + # TODO(caoying) A lower bound value epsilon for the norm is needed to + # imporve the numeric stability of reciprocal. This requires a maximum_op. + rsquare = helper.create_tmp_variable(dtype=x.dtype) + helper.append_op( + type="reciprocal", inputs={"X": reduced_sum}, outputs={"Out": rsquare}) + + # TODO(caoying) the current elementwise_mul operator does not support a + # general broadcast rule which broadcasts input(Y) to have the same + # dimension with Input(X) starting from a specified dimension. So this + # exanpsion is requred. Once a general broadcast rule is spported, this + # expanding canbe removed. + rsquare_expanded = helper.create_tmp_variable(dtype=x.dtype) + expand_times = [1] * len(x.shape) + expand_times[axis] = int(x.shape[axis]) + helper.append_op( + type="expand", + inputs={"X": rsquare}, + outputs={"Out": rsquare_expanded}, + attrs={"expand_times": expand_times}) + + out = helper.create_tmp_variable(dtype=x.dtype) + helper.append_op( + type="elementwise_mul", + inputs={"X": x, + "Y": rsquare_expanded}, + outputs={"Out": out}) + return out + + +def matmul(x, y, transpose_x=False, transpose_y=False, name=None): + """ + 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 + 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 + :math:`[1, D]` in transposed form. + + - 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 + 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 + removed after matrix multipication. + + Args: + x (Variable): The input variable which is a Tensor or LoDTensor. + y (Variable): The input variable which is a Tensor or LoDTensor. + transpose_x (bool): Whether to transpose :math:`x` before multiplication. + transpose_y (bool): Whether to transpose :math:`y` before multiplication. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + Variable: The product Tensor variable. + + Examples: + .. code-block:: python + + # Examples to clarify shapes of the inputs and output + # x: [B, M, K], y: [B, K, N] + fluid.layers.matmul(x, y) # out: [B, M, N] + # x: [B, M, K], y: [K, N] + fluid.layers.matmul(x, y) # out: [B, M, N] + # x: [B, M, K], y: [K] + fluid.layers.matmul(x, y) # out: [B, M] + # x: [M, K], y: [K, N] + fluid.layers.matmul(x, y) # out: [M, N] + # x: [K], y: [K] + fluid.layers.matmul(x, y) # out: [1] + # x: [M], y: [N] + + fluid.layers.matmul(x, y, True, True) # out: [M, N] + """ + helper = LayerHelper('matmul', **locals()) + assert max( + len(x.shape), len(y.shape) + ) <= 3, 'Currently only rank 1 to rank 3 input tensors are supported.' + out = helper.create_tmp_variable(dtype=helper.input_dtype()) + helper.append_op( + type='matmul', + inputs={'X': x, + 'Y': y}, + outputs={'Out': out}, + 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), + 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), 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 diff --git a/python/paddle/v2/fluid/layers/ops.py b/python/paddle/v2/fluid/layers/ops.py index 73d7c895806ef28ffc98db88809e317a86762769..21945edf0827e7a86ec2f8ce8f84c9093808c68b 100644 --- a/python/paddle/v2/fluid/layers/ops.py +++ b/python/paddle/v2/fluid/layers/ops.py @@ -55,6 +55,8 @@ __all__ = [ 'elementwise_div', 'elementwise_sub', 'elementwise_mul', + 'elementwise_max', + 'elementwise_min', 'clip', 'sequence_softmax', ] + __activations__ diff --git a/python/paddle/v2/fluid/nets.py b/python/paddle/v2/fluid/nets.py index b5c26e713db4e5c48d6c0851fd68275aeabeea64..c53fbd280fd79dc0d598fda2b5796886abee4ed4 100644 --- a/python/paddle/v2/fluid/nets.py +++ b/python/paddle/v2/fluid/nets.py @@ -17,6 +17,7 @@ __all__ = [ "simple_img_conv_pool", "sequence_conv_pool", "glu", + "dot_product_attention", ] @@ -27,19 +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 @@ -53,7 +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. """ @@ -84,7 +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) @@ -96,7 +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 @@ -150,3 +157,55 @@ def glu(input, dim=-1): act_b = layers.sigmoid(x=b) out = layers.elementwise_mul(x=a, y=act_b) return out + + +def dot_product_attention(querys, keys, values): + """ + The dot-product attention. + + Attention mechanism can be seen as mapping a query and a set of key-value + pairs to an output. The output is computed as a weighted sum of the values, + where the weight assigned to each value is computed by a compatibility + function (dot-product here) of the query with the corresponding key. + + The dot-product attention can be implemented through (batch) matrix + multipication as follows: + + .. math:: + + Attention(Q, K, V)= softmax(QK^\mathrm{T})V + + Refer to `Attention Is All You Need + `_. + + Note that batch data containing sequences with different lengths is not + supported by this because of the (batch) matrix multipication. + + Args: + query (Variable): The input variable which is a Tensor or LoDTensor. + key (Variable): The input variable which is a Tensor or LoDTensor. + value (Variable): The input variable which is a Tensor or LoDTensor. + + Returns: + tuple: The Tensor variables representing the output and attention scores. + + Examples: + .. code-block:: python + + # Suppose q, k, v are tensor variables with the following shape: + # q: [3, 5, 9], k: [3, 6, 9], v: [3, 6, 10] + out, attn_scores = fluid.nets.dot_product_attention(q, k, v) + out.shape # [3, 5, 10] + attn_scores.shape # [3, 5, 6] + """ + assert keys.shape[-2] == values.shape[ + -2], 'The shapes of keys and values mismatch.' + assert querys.shape[-1] == keys.shape[ + -1], 'The shapes of querys and keys mismatch.' + product = layers.matmul(x=querys, y=keys, transpose_y=True) + attn_scores = layers.reshape( + x=layers.reshape( + x=product, shape=[-1, product.shape[-1]], act='softmax'), + shape=product.shape) + out = layers.matmul(attn_scores, values) + return out, attn_scores diff --git a/python/paddle/v2/fluid/registry.py b/python/paddle/v2/fluid/registry.py index 6c0c3a35185391873fe5bb98d1ed5ee1cf13aa15..ff10542d40aabaf31897842754d38b7868472b21 100644 --- a/python/paddle/v2/fluid/registry.py +++ b/python/paddle/v2/fluid/registry.py @@ -1,16 +1,16 @@ # 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) diff --git a/python/paddle/v2/fluid/tests/book_distribute/notest_dist_fit_a_line.py b/python/paddle/v2/fluid/tests/book_distribute/notest_dist_fit_a_line.py index b886071f94768d3373bbf0e0b7655c924b218645..00e0b22c6458d55140eb54cffb3280d868e4218d 100644 --- a/python/paddle/v2/fluid/tests/book_distribute/notest_dist_fit_a_line.py +++ b/python/paddle/v2/fluid/tests/book_distribute/notest_dist_fit_a_line.py @@ -53,8 +53,9 @@ 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) - exe.run(fluid.default_startup_program()) + pserver_prog = t.get_pserver_program(current_endpoint) + pserver_startup = t.get_startup_program(current_endpoint, pserver_prog) + exe.run(pserver_startup) exe.run(pserver_prog) else: trainer_prog = t.get_trainer_program() diff --git a/python/paddle/v2/fluid/tests/book_distribute/notest_dist_label_semantic_roles.py b/python/paddle/v2/fluid/tests/book_distribute/notest_dist_label_semantic_roles.py index 2b5a098ff253b8a96afba7cd03d7f9998ff400af..adc7ae8adf6893cb646c3c82838e95283a851e9f 100644 --- a/python/paddle/v2/fluid/tests/book_distribute/notest_dist_label_semantic_roles.py +++ b/python/paddle/v2/fluid/tests/book_distribute/notest_dist_label_semantic_roles.py @@ -197,8 +197,9 @@ def main(): if not current_endpoint: print("need env SERVER_ENDPOINT") exit(1) - pserver_prog = t.get_pserver_program(current_endpoint, optimize_ops) - exe.run(fluid.default_startup_program()) + pserver_prog = t.get_pserver_program(current_endpoint) + pserver_startup = t.get_startup_program(current_endpoint, pserver_prog) + exe.run(pserver_startup) exe.run(pserver_prog) elif training_role == "TRAINER": trainer_prog = t.get_trainer_program() diff --git a/python/paddle/v2/fluid/tests/book_distribute/notest_dist_word2vec.py b/python/paddle/v2/fluid/tests/book_distribute/notest_dist_word2vec.py index dc04af5b7b6ee143847685d6cf4da91747afd3ec..27ef2cad1de7eaedb9cc116e7451a9b73180f82d 100644 --- a/python/paddle/v2/fluid/tests/book_distribute/notest_dist_word2vec.py +++ b/python/paddle/v2/fluid/tests/book_distribute/notest_dist_word2vec.py @@ -87,8 +87,9 @@ 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) - exe.run(fluid.default_startup_program()) + pserver_prog = t.get_pserver_program(current_endpoint) + pserver_startup = t.get_startup_program(current_endpoint, pserver_prog) + exe.run(pserver_startup) exe.run(pserver_prog) elif training_role == "TRAINER": feeder = fluid.DataFeeder( diff --git a/python/paddle/v2/fluid/tests/book_distribute/notest_recognize_digits_conv_dist.py b/python/paddle/v2/fluid/tests/book_distribute/notest_recognize_digits_conv_dist.py index 27512c4f7812b6b55d5dc6d1a12c3b83df8b3e6f..ee5cd07fb3794efd0c4083e6fe7ac1a092c8c373 100644 --- a/python/paddle/v2/fluid/tests/book_distribute/notest_recognize_digits_conv_dist.py +++ b/python/paddle/v2/fluid/tests/book_distribute/notest_recognize_digits_conv_dist.py @@ -52,26 +52,27 @@ train_reader = paddle.batch( place = fluid.CPUPlace() exe = fluid.Executor(place) -t = fluid.DistributeTranspiler() -# all parameter server endpoints list for spliting parameters -pserver_endpoints = os.getenv("PSERVERS") -# server endpoint for current node -current_endpoint = os.getenv("SERVER_ENDPOINT") -# run as trainer or parameter server +pserver_endpoints = os.getenv("PSERVERS") # all pserver endpoints +trainers = int(os.getenv("TRAINERS")) # total trainer count +current_endpoint = os.getenv("SERVER_ENDPOINT") # current pserver endpoint training_role = os.getenv("TRAINING_ROLE", "TRAINER") # get the training role: trainer/pserver -t.transpile(optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2) +t = fluid.DistributeTranspiler() +t.transpile( + optimize_ops, params_grads, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": if not current_endpoint: print("need env SERVER_ENDPOINT") exit(1) - pserver_prog = t.get_pserver_program(current_endpoint, optimize_ops) - exe.run(fluid.default_startup_program()) + pserver_prog = t.get_pserver_program(current_endpoint) + pserver_startup = t.get_startup_program(current_endpoint, pserver_prog) + exe.run(pserver_startup) exe.run(pserver_prog) elif training_role == "TRAINER": trainer_prog = t.get_trainer_program() feeder = fluid.DataFeeder(feed_list=[images, label], place=place) + # TODO(typhoonzero): change trainer startup program to fetch parameters from pserver exe.run(fluid.default_startup_program()) for pass_id in range(PASS_NUM): diff --git a/python/paddle/v2/fluid/tests/book_distribute/notest_recognize_digits_mlp_dist.py b/python/paddle/v2/fluid/tests/book_distribute/notest_recognize_digits_mlp_dist.py new file mode 100644 index 0000000000000000000000000000000000000000..9cfb4ab8c4cf85843f12d1a10e3433a0688ec9b4 --- /dev/null +++ b/python/paddle/v2/fluid/tests/book_distribute/notest_recognize_digits_mlp_dist.py @@ -0,0 +1,88 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# +#Licensed under the Apache License, Version 2.0 (the "License"); +#you may not use this file except in compliance with the License. +#You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +#Unless required by applicable law or agreed to in writing, software +#distributed under the License is distributed on an "AS IS" BASIS, +#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +#See the License for the specific language governing permissions and +#limitations under the License. +from __future__ import print_function +import numpy as np +import paddle.v2 as paddle +import paddle.v2.fluid as fluid +import os + +BATCH_SIZE = 128 +PASS_NUM = 100 + +images = fluid.layers.data(name='x', shape=[784], dtype='float32') + +# 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') + +label = fluid.layers.data(name='y', shape=[1], dtype='int64') + +cost = fluid.layers.cross_entropy(input=predict, label=label) +avg_cost = fluid.layers.mean(x=cost) + +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) + +train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.mnist.train(), buf_size=8192), + batch_size=BATCH_SIZE) + +place = fluid.CPUPlace() +exe = fluid.Executor(place) + +t = fluid.DistributeTranspiler() +# all parameter server endpoints list for spliting parameters +pserver_endpoints = os.getenv("PSERVERS") +# server endpoint for current node +current_endpoint = os.getenv("SERVER_ENDPOINT") +# run as trainer or parameter server +training_role = os.getenv("TRAINING_ROLE", + "TRAINER") # get the training role: trainer/pserver +t.transpile(optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2) + +if training_role == "PSERVER": + if not current_endpoint: + print("need env SERVER_ENDPOINT") + exit(1) + pserver_prog = t.get_pserver_program(current_endpoint) + pserver_startup = t.get_startup_program(current_endpoint, pserver_prog) + exe.run(pserver_startup) + exe.run(pserver_prog) +elif training_role == "TRAINER": + trainer_prog = t.get_trainer_program() + feeder = fluid.DataFeeder(feed_list=[images, label], place=place) + exe.run(fluid.default_startup_program()) + + for pass_id in range(PASS_NUM): + accuracy.reset(exe) + batch_id = 0 + for data in train_reader(): + loss, acc = exe.run(trainer_prog, + feed=feeder.feed(data), + fetch_list=[avg_cost] + accuracy.metrics) + pass_acc = accuracy.eval(exe) + 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)) +else: + print("environment var TRAINER_ROLE should be TRAINER os PSERVER") diff --git a/python/paddle/v2/fluid/tests/book_distribute/notest_understand_sentiment_conv_dist.py b/python/paddle/v2/fluid/tests/book_distribute/notest_understand_sentiment_conv_dist.py index 74f20f3f4cc8c81a38c1ad8ab33df6a07fbcad44..840afb6376249f10cb78f80b01e112a619032244 100644 --- a/python/paddle/v2/fluid/tests/book_distribute/notest_understand_sentiment_conv_dist.py +++ b/python/paddle/v2/fluid/tests/book_distribute/notest_understand_sentiment_conv_dist.py @@ -92,15 +92,16 @@ def main(): t.transpile( optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2) - exe.run(fluid.default_startup_program()) - 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) + pserver_startup = t.get_startup_program(current_endpoint, pserver_prog) + exe.run(pserver_startup) exe.run(pserver_prog) elif training_role == "TRAINER": + exe.run(fluid.default_startup_program()) trainer_prog = t.get_trainer_program() feeder = fluid.DataFeeder(feed_list=[data, label], place=place) diff --git a/python/paddle/v2/fluid/tests/test_edit_distance_op.py b/python/paddle/v2/fluid/tests/test_edit_distance_op.py index 38e87728b387bb70a8921a2fe73a4e69701aabe9..cf118df634bb8288456009ebd4954f08d5eb4323 100644 --- a/python/paddle/v2/fluid/tests/test_edit_distance_op.py +++ b/python/paddle/v2/fluid/tests/test_edit_distance_op.py @@ -1,3 +1,16 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# +#Licensed under the Apache License, Version 2.0 (the "License"); +#you may not use this file except in compliance with the License. +#You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +#Unless required by applicable law or agreed to in writing, software +#distributed under the License is distributed on an "AS IS" BASIS, +#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +#See the License for the specific language governing permissions and +#limitations under the License. import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/v2/fluid/tests/test_elementwise_add_op.py b/python/paddle/v2/fluid/tests/test_elementwise_add_op.py index 1e88231877f869052d859a30fbbd8f7690b64095..3564772fb52882e9e58ea88caeb12c5e91137525 100644 --- a/python/paddle/v2/fluid/tests/test_elementwise_add_op.py +++ b/python/paddle/v2/fluid/tests/test_elementwise_add_op.py @@ -1,16 +1,16 @@ # 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 unittest import numpy as np from op_test import OpTest @@ -40,6 +40,16 @@ class TestElementwiseOp(OpTest): ['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y')) +class TestElementwiseAddOp_scalar(TestElementwiseOp): + def setUp(self): + self.op_type = "elementwise_add" + self.inputs = { + 'X': np.random.rand(2, 3, 4).astype(np.float32), + 'Y': np.random.rand(1).astype(np.float32) + } + self.outputs = {'Out': self.inputs['X'] + self.inputs['Y']} + + class TestElementwiseAddOp_Vector(TestElementwiseOp): def setUp(self): self.op_type = "elementwise_add" diff --git a/python/paddle/v2/fluid/tests/test_elementwise_div_op.py b/python/paddle/v2/fluid/tests/test_elementwise_div_op.py index fbabc79be2406efa916d2acafe9ce3d69112b160..77b113af7693c4a71a5a13c791cfb3e0420f4ff8 100644 --- a/python/paddle/v2/fluid/tests/test_elementwise_div_op.py +++ b/python/paddle/v2/fluid/tests/test_elementwise_div_op.py @@ -1,16 +1,16 @@ # 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 unittest import numpy as np from op_test import OpTest @@ -45,6 +45,16 @@ class ElementwiseDivOp(OpTest): ['X'], 'Out', max_relative_error=0.05, no_grad_set=set('Y')) +class TestElementwiseDivOp_scalar(ElementwiseDivOp): + def setUp(self): + self.op_type = "elementwise_div" + self.inputs = { + 'X': np.random.uniform(0.1, 1, [2, 3, 4]).astype(np.float32), + 'Y': np.random.uniform(0.1, 1, [1]).astype(np.float32) + } + self.outputs = {'Out': self.inputs['X'] / self.inputs['Y']} + + class TestElementwiseDivOp_Vector(ElementwiseDivOp): def setUp(self): self.op_type = "elementwise_div" diff --git a/python/paddle/v2/fluid/tests/test_elementwise_max_op.py b/python/paddle/v2/fluid/tests/test_elementwise_max_op.py new file mode 100644 index 0000000000000000000000000000000000000000..9526f0199bf2f1a537a18010ef6b87105589ed07 --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_elementwise_max_op.py @@ -0,0 +1,129 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# +#Licensed under the Apache License, Version 2.0 (the "License"); +#you may not use this file except in compliance with the License. +#You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +#Unless required by applicable law or agreed to in writing, software +#distributed under the License is distributed on an "AS IS" BASIS, +#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +#See the License for the specific language governing permissions and +#limitations under the License. +import unittest +import numpy as np +from op_test import OpTest + + +class TestElementwiseOp(OpTest): + def setUp(self): + self.op_type = "elementwise_max" + # If x and y have the same value, the max() is not differentiable. + # So we generate test data by the following method + # to avoid them being too close to each other. + x = np.random.uniform(0.1, 1, [13, 17]).astype("float32") + sgn = np.random.choice([-1, 1], [13, 17]).astype("float32") + y = x + sgn * np.random.uniform(0.1, 1, [13, 17]).astype("float32") + self.inputs = {'X': x, 'Y': y} + self.outputs = {'Out': np.maximum(self.inputs['X'], self.inputs['Y'])} + + def test_check_output(self): + self.check_output() + + def test_check_grad_normal(self): + self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.005) + + def test_check_grad_ingore_x(self): + self.check_grad( + ['Y'], 'Out', max_relative_error=0.005, no_grad_set=set("X")) + + def test_check_grad_ingore_y(self): + self.check_grad( + ['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y')) + + +class TestElementwiseMaxOp_scalar(TestElementwiseOp): + def setUp(self): + self.op_type = "elementwise_max" + x = np.random.random_integers(-5, 5, [2, 3, 4]).astype("float32") + y = np.array([0.5]).astype("float32") + self.inputs = {'X': x, 'Y': y} + self.outputs = {'Out': np.maximum(self.inputs['X'], self.inputs['Y'])} + + +class TestElementwiseMaxOp_Vector(TestElementwiseOp): + def setUp(self): + self.op_type = "elementwise_max" + x = np.random.random((32, )).astype("float32") + sgn = np.random.choice([-1, 1], (32, )).astype("float32") + y = x + sgn * np.random.uniform(0.1, 1, (32, )).astype("float32") + self.inputs = {'X': x, 'Y': y} + self.outputs = {'Out': np.maximum(self.inputs['X'], self.inputs['Y'])} + + +class TestElementwiseMaxOp_broadcast_0(TestElementwiseOp): + def setUp(self): + self.op_type = "elementwise_max" + x = np.random.uniform(0.5, 1, (2, 3, 4)).astype(np.float32) + sgn = np.random.choice([-1, 1], (2, )).astype(np.float32) + y = x[:, 0, 0] + sgn * \ + np.random.uniform(1, 2, (2, )).astype(np.float32) + self.inputs = {'X': x, 'Y': y} + + self.attrs = {'axis': 0} + self.outputs = { + 'Out': + np.maximum(self.inputs['X'], self.inputs['Y'].reshape(2, 1, 1)) + } + + +class TestElementwiseMaxOp_broadcast_1(TestElementwiseOp): + def setUp(self): + self.op_type = "elementwise_max" + x = np.random.uniform(0.5, 1, (2, 3, 4)).astype(np.float32) + sgn = np.random.choice([-1, 1], (3, )).astype(np.float32) + y = x[0, :, 0] + sgn * \ + np.random.uniform(1, 2, (3, )).astype(np.float32) + self.inputs = {'X': x, 'Y': y} + + self.attrs = {'axis': 1} + self.outputs = { + 'Out': + np.maximum(self.inputs['X'], self.inputs['Y'].reshape(1, 3, 1)) + } + + +class TestElementwiseMaxOp_broadcast_2(TestElementwiseOp): + def setUp(self): + self.op_type = "elementwise_max" + x = np.random.uniform(0.5, 1, (2, 3, 4)).astype(np.float32) + sgn = np.random.choice([-1, 1], (4, )).astype(np.float32) + y = x[0, 0, :] + sgn * \ + np.random.uniform(1, 2, (4, )).astype(np.float32) + self.inputs = {'X': x, 'Y': y} + + self.outputs = { + 'Out': + np.maximum(self.inputs['X'], self.inputs['Y'].reshape(1, 1, 4)) + } + + +class TestElementwiseMaxOp_broadcast_3(TestElementwiseOp): + def setUp(self): + self.op_type = "elementwise_max" + x = np.random.uniform(0.5, 1, (2, 3, 4, 5)).astype(np.float32) + sgn = np.random.choice([-1, 1], (3, 4)).astype(np.float32) + y = x[0, :, :, 0] + sgn * \ + np.random.uniform(1, 2, (3, 4)).astype(np.float32) + self.inputs = {'X': x, 'Y': y} + + self.attrs = {'axis': 1} + self.outputs = { + 'Out': + np.maximum(self.inputs['X'], self.inputs['Y'].reshape(1, 3, 4, 1)) + } + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_elementwise_min_op.py b/python/paddle/v2/fluid/tests/test_elementwise_min_op.py new file mode 100644 index 0000000000000000000000000000000000000000..b9007282335a32f85defcd1e2b190745abd38794 --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_elementwise_min_op.py @@ -0,0 +1,129 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# +#Licensed under the Apache License, Version 2.0 (the "License"); +#you may not use this file except in compliance with the License. +#You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +#Unless required by applicable law or agreed to in writing, software +#distributed under the License is distributed on an "AS IS" BASIS, +#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +#See the License for the specific language governing permissions and +#limitations under the License. +import unittest +import numpy as np +from op_test import OpTest + + +class TestElementwiseOp(OpTest): + def setUp(self): + self.op_type = "elementwise_min" + # If x and y have the same value, the min() is not differentiable. + # So we generate test data by the following method + # to avoid them being too close to each other. + x = np.random.uniform(0.1, 1, [13, 17]).astype("float32") + sgn = np.random.choice([-1, 1], [13, 17]).astype("float32") + y = x + sgn * np.random.uniform(0.1, 1, [13, 17]).astype("float32") + self.inputs = {'X': x, 'Y': y} + self.outputs = {'Out': np.minimum(self.inputs['X'], self.inputs['Y'])} + + def test_check_output(self): + self.check_output() + + def test_check_grad_normal(self): + self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.005) + + def test_check_grad_ingore_x(self): + self.check_grad( + ['Y'], 'Out', max_relative_error=0.005, no_grad_set=set("X")) + + def test_check_grad_ingore_y(self): + self.check_grad( + ['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y')) + + +class TestElementwiseMinOp_scalar(TestElementwiseOp): + def setUp(self): + self.op_type = "elementwise_min" + x = np.random.random_integers(-5, 5, [2, 3, 4]).astype("float32") + y = np.array([0.5]).astype("float32") + self.inputs = {'X': x, 'Y': y} + self.outputs = {'Out': np.minimum(self.inputs['X'], self.inputs['Y'])} + + +class TestElementwiseMaxOp_Vector(TestElementwiseOp): + def setUp(self): + self.op_type = "elementwise_min" + x = np.random.random((32, )).astype("float32") + sgn = np.random.choice([-1, 1], (32, )).astype("float32") + y = x + sgn * np.random.uniform(0.1, 1, (32, )).astype("float32") + self.inputs = {'X': x, 'Y': y} + self.outputs = {'Out': np.minimum(self.inputs['X'], self.inputs['Y'])} + + +class TestElementwiseMaxOp_broadcast_0(TestElementwiseOp): + def setUp(self): + self.op_type = "elementwise_min" + x = np.random.uniform(0.5, 1, (2, 3, 4)).astype(np.float32) + sgn = np.random.choice([-1, 1], (2, )).astype(np.float32) + y = x[:, 0, 0] + sgn * \ + np.random.uniform(1, 2, (2, )).astype(np.float32) + self.inputs = {'X': x, 'Y': y} + + self.attrs = {'axis': 0} + self.outputs = { + 'Out': + np.minimum(self.inputs['X'], self.inputs['Y'].reshape(2, 1, 1)) + } + + +class TestElementwiseMaxOp_broadcast_1(TestElementwiseOp): + def setUp(self): + self.op_type = "elementwise_min" + x = np.random.uniform(0.5, 1, (2, 3, 4)).astype(np.float32) + sgn = np.random.choice([-1, 1], (3, )).astype(np.float32) + y = x[0, :, 0] + sgn * \ + np.random.uniform(1, 2, (3, )).astype(np.float32) + self.inputs = {'X': x, 'Y': y} + + self.attrs = {'axis': 1} + self.outputs = { + 'Out': + np.minimum(self.inputs['X'], self.inputs['Y'].reshape(1, 3, 1)) + } + + +class TestElementwiseMaxOp_broadcast_2(TestElementwiseOp): + def setUp(self): + self.op_type = "elementwise_min" + x = np.random.uniform(0.5, 1, (2, 3, 4)).astype(np.float32) + sgn = np.random.choice([-1, 1], (4, )).astype(np.float32) + y = x[0, 0, :] + sgn * \ + np.random.uniform(1, 2, (4, )).astype(np.float32) + self.inputs = {'X': x, 'Y': y} + + self.outputs = { + 'Out': + np.minimum(self.inputs['X'], self.inputs['Y'].reshape(1, 1, 4)) + } + + +class TestElementwiseMaxOp_broadcast_3(TestElementwiseOp): + def setUp(self): + self.op_type = "elementwise_min" + x = np.random.uniform(0.5, 1, (2, 3, 4, 5)).astype(np.float32) + sgn = np.random.choice([-1, 1], (3, 4)).astype(np.float32) + y = x[0, :, :, 0] + sgn * \ + np.random.uniform(1, 2, (3, 4)).astype(np.float32) + self.inputs = {'X': x, 'Y': y} + + self.attrs = {'axis': 1} + self.outputs = { + 'Out': + np.minimum(self.inputs['X'], self.inputs['Y'].reshape(1, 3, 4, 1)) + } + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_elementwise_mul_op.py b/python/paddle/v2/fluid/tests/test_elementwise_mul_op.py index ef3a829abc65f3af60b899ba50927424fd563607..12dfa6599cd634e1d806980f89e7c013b8eb8754 100644 --- a/python/paddle/v2/fluid/tests/test_elementwise_mul_op.py +++ b/python/paddle/v2/fluid/tests/test_elementwise_mul_op.py @@ -1,16 +1,16 @@ # 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 unittest import numpy as np from op_test import OpTest @@ -38,6 +38,16 @@ class ElementwiseMulOp(OpTest): self.check_grad(['X'], 'Out', no_grad_set=set('Y')) +class TestElementwiseMulOp_scalar(ElementwiseMulOp): + def setUp(self): + self.op_type = "elementwise_mul" + self.inputs = { + 'X': np.random.rand(2, 3, 4).astype(np.float32), + 'Y': np.random.rand(1).astype(np.float32) + } + self.outputs = {'Out': self.inputs['X'] * self.inputs['Y']} + + class TestElementwiseMulOp_Vector(ElementwiseMulOp): def setUp(self): self.op_type = "elementwise_mul" diff --git a/python/paddle/v2/fluid/tests/test_elementwise_sub_op.py b/python/paddle/v2/fluid/tests/test_elementwise_sub_op.py index db24db7b304e2ef8b946e128b22ed62d71260ddc..cf53d85bbad81f393a6263f8742fc942d357135f 100644 --- a/python/paddle/v2/fluid/tests/test_elementwise_sub_op.py +++ b/python/paddle/v2/fluid/tests/test_elementwise_sub_op.py @@ -1,16 +1,16 @@ # 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 unittest import numpy as np from op_test import OpTest @@ -40,6 +40,16 @@ class TestElementwiseOp(OpTest): ['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y')) +class TestElementwiseSubOp_scalar(TestElementwiseOp): + def setUp(self): + self.op_type = "elementwise_sub" + self.inputs = { + 'X': np.random.rand(2, 3, 4).astype(np.float32), + 'Y': np.random.rand(1).astype(np.float32) + } + self.outputs = {'Out': self.inputs['X'] - self.inputs['Y']} + + class TestElementwiseSubOp_Vector(TestElementwiseOp): def setUp(self): self.op_type = "elementwise_sub" diff --git a/python/paddle/v2/fluid/tests/test_matmul_op.py b/python/paddle/v2/fluid/tests/test_matmul_op.py index 0220cfe1287ab6908f379f9f7451cd257dd7f0a6..f7dc4e053217dcceaf9c64e3605286fc0698593b 100644 --- a/python/paddle/v2/fluid/tests/test_matmul_op.py +++ b/python/paddle/v2/fluid/tests/test_matmul_op.py @@ -96,18 +96,18 @@ class Generator(object): self.outputs = {'Out': Out} def test_check_output(self): - self.check_output(atol=1e-2) + self.check_output(atol=1e-3) def test_check_grad_normal(self): - self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5) + self.check_grad(['X', 'Y'], 'Out', max_relative_error=1e-3) def test_check_grad_ignore_x(self): self.check_grad( - ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X")) + ['Y'], 'Out', max_relative_error=1e-3, no_grad_set=set("X")) def test_check_grad_ignore_y(self): self.check_grad( - ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y')) + ['X'], 'Out', max_relative_error=1e-3, no_grad_set=set('Y')) # Generate test cases for all possibilities diff --git a/python/paddle/v2/fluid/tests/test_normalization_wrapper.py b/python/paddle/v2/fluid/tests/test_normalization_wrapper.py new file mode 100644 index 0000000000000000000000000000000000000000..caff63011d038d785472cb38a26a51f3f4cc9288 --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_normalization_wrapper.py @@ -0,0 +1,95 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. +# +#Licensed under the Apache License, Version 2.0 (the "License"); +#you may not use this file except in compliance with the License. +#You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +#Unless required by applicable law or agreed to in writing, software +#distributed under the License is distributed on an "AS IS" BASIS, +#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +#See the License for the specific language governing permissions and +#limitations under the License. +import unittest +import paddle.v2.fluid as fluid +import paddle.v2.fluid.core as core +import numpy as np + + +class TestNormalization(unittest.TestCase): + data_desc = {"name": "input", "shape": (2, 3, 7)} + + def gen_random_input(self): + """Generate random input data. + """ + self.data = np.random.random( + size=self.data_desc["shape"]).astype("float32") + + def set_program(self, axis, epsilon): + """Build the test program. + """ + data = fluid.layers.data( + name=self.data_desc["name"], + shape=self.data_desc["shape"], + dtype="float32", + append_batch_size=False) + data.stop_gradient = False + l2_norm = fluid.layers.l2_normalize(x=data, axis=axis, epsilon=epsilon) + out = fluid.layers.reduce_sum(l2_norm, dim=None) + + fluid.backward.append_backward(loss=out) + self.fetch_list = [l2_norm] + + def run_program(self): + """Run the test program. + """ + places = [core.CPUPlace()] + if core.is_compile_gpu(): + places.append(core.CUDAPlace(0)) + + for place in places: + self.set_inputs(place) + exe = fluid.Executor(place) + + output = exe.run(fluid.default_main_program(), + feed=self.inputs, + fetch_list=self.fetch_list, + return_numpy=True) + self.op_output = output + + def set_inputs(self, place): + """Set the randomly generated data to the test program. + """ + self.inputs = {} + tensor = fluid.Tensor() + tensor.set(self.data, place) + self.inputs[self.data_desc["name"]] = tensor + + def l2_normalize(self, data, axis, epsilon): + """ Compute the groundtruth. + """ + output = data * np.reciprocal( + np.sum(np.square(data), axis=axis, keepdims=True)) + return output + + def test_l2_normalize(self): + """ Test the python wrapper for l2_normalize. + """ + axis = 1 + #TODO(caoying) epsilon is not supported due to lack of a maximum_op. + epsilon = 1e-6 + + self.gen_random_input() + + self.set_program(axis, epsilon) + self.run_program() + + expect_output = self.l2_normalize(self.data, axis, epsilon) + + # check output + self.assertTrue(np.allclose(self.op_output, expect_output, atol=0.001)) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_parallel_op.py b/python/paddle/v2/fluid/tests/test_parallel_op.py index 3c190477d167c6aa48078869a9abb15488d21dd3..d36f7d07ac381d835618af4b420525ff0e607651 100644 --- a/python/paddle/v2/fluid/tests/test_parallel_op.py +++ b/python/paddle/v2/fluid/tests/test_parallel_op.py @@ -15,9 +15,6 @@ import unittest import paddle.v2.fluid as fluid import numpy -import sys -# TODO(dzhwinter): get places op check need to be enhanced. -sys.exit(0) class BaseParallelForTest(unittest.TestCase): @@ -151,23 +148,27 @@ class BaseParallelForTest(unittest.TestCase): class ParallelOpTest(BaseParallelForTest): - def test_simple_fc(self): - def __network__(): - x = fluid.layers.data(shape=[784], dtype='float32', name='img') - # FIXME: This is a bug of parallel.do - x.stop_gradient = False - x = yield x - hidden = fluid.layers.fc(input=x, size=200, param_attr='fc1.w') - loss = fluid.layers.mean(x=hidden) - yield loss + @staticmethod + def __network__(): + x = fluid.layers.data(shape=[784], dtype='float32', name='img') + x = yield x + hidden = fluid.layers.fc(input=x, size=200, param_attr='fc1.w') + loss = fluid.layers.mean(x=hidden) + yield loss + def test_simple_fc(self): self.run_test( - callback=__network__, + callback=ParallelOpTest.__network__, feed={ - 'img': - numpy.random.random(size=(128 * 3, 784)).astype('float32') + 'img': numpy.random.random(size=(51, 784)).astype('float32') }, - fetch='fc1.w@GRAD') + fetch=['fc1.w@GRAD']) + + def test_fc_with_tiny_data(self): + self.run_test( + callback=ParallelOpTest.__network__, + feed={'img': numpy.random.random(size=(1, 784)).astype('float32')}, + fetch=['fc1.w@GRAD']) if __name__ == '__main__': diff --git a/python/paddle/v2/fluid/tests/test_sequence_erase_op.py b/python/paddle/v2/fluid/tests/test_sequence_erase_op.py index 650984009a76a56fb65811f7bb805ca656194a35..4cc2613cf9c26845cef988160405b632706c4b11 100644 --- a/python/paddle/v2/fluid/tests/test_sequence_erase_op.py +++ b/python/paddle/v2/fluid/tests/test_sequence_erase_op.py @@ -29,7 +29,7 @@ def sequence_erase(in_seq, lod0, tokens): return np.array(out_seq).astype("int32"), new_lod0 -class TestSequenceEraseOp(OpTest): +class TestSequenceEraseOpInt32(OpTest): def setUp(self): self.op_type = "sequence_erase" in_seq = np.random.randint(0, 10, (30, 1)).astype("int32") @@ -44,5 +44,35 @@ class TestSequenceEraseOp(OpTest): self.check_output() +class TestSequenceEraseOpInt64(OpTest): + def setUp(self): + self.op_type = "sequence_erase" + in_seq = np.random.randint(0, 10, (30, 1)).astype("int64") + lod = [[0, 9, 13, 24, 30]] + tokens = [2, 3, 5] + out_seq, new_lod0 = sequence_erase(in_seq, lod[0], tokens) + self.attrs = {'tokens': tokens} + self.inputs = {'X': (in_seq, lod)} + self.outputs = {'Out': (out_seq, [new_lod0])} + + def test_check_output(self): + self.check_output() + + +class TestSequenceEraseOpEmpty(OpTest): + def setUp(self): + self.op_type = "sequence_erase" + in_seq = np.random.randint(0, 10, (30, 1)).astype("int32") + lod = [[0, 9, 13, 24, 30]] + tokens = [] + out_seq, new_lod0 = sequence_erase(in_seq, lod[0], tokens) + self.attrs = {'tokens': tokens} + self.inputs = {'X': (in_seq, lod)} + self.outputs = {'Out': (out_seq, [new_lod0])} + + def test_check_output(self): + self.check_output() + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/v2/fluid/tests/test_split_selected_rows_op.py b/python/paddle/v2/fluid/tests/test_split_selected_rows_op.py new file mode 100644 index 0000000000000000000000000000000000000000..a6cc4f6c6db803d4e3e8c57a2f2fbf8a24abc00f --- /dev/null +++ b/python/paddle/v2/fluid/tests/test_split_selected_rows_op.py @@ -0,0 +1,128 @@ +# 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() diff --git a/v1_api_demo/README.md b/v1_api_demo/README.md deleted file mode 100644 index 0460a85fae078800332982751a5d4a9644c50bd6..0000000000000000000000000000000000000000 --- a/v1_api_demo/README.md +++ /dev/null @@ -1,5 +0,0 @@ -The examples in v1_api_demo are using v1_api currently, and will be upgraded to v2_api later. -Thus, v1_api_demo 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. diff --git a/v1_api_demo/gan/.gitignore b/v1_api_demo/gan/.gitignore deleted file mode 100644 index 93a6f5080a16a601cffb0bff51af9aef3ba3bae7..0000000000000000000000000000000000000000 --- a/v1_api_demo/gan/.gitignore +++ /dev/null @@ -1,11 +0,0 @@ -output/ -uniform_params/ -cifar_params/ -mnist_params/ -*.png -.pydevproject -.project -*.log -*.pyc -data/mnist_data/ -data/cifar-10-batches-py/ diff --git a/v1_api_demo/gan/README.md b/v1_api_demo/gan/README.md deleted file mode 100644 index 1908b534b0c1f63904d5503399b961d74ce0037c..0000000000000000000000000000000000000000 --- a/v1_api_demo/gan/README.md +++ /dev/null @@ -1,13 +0,0 @@ -# Generative Adversarial Networks (GAN) - -This demo implements GAN training described in the original GAN paper (https://arxiv.org/abs/1406.2661) and DCGAN (https://arxiv.org/abs/1511.06434). - -The general training procedures are implemented in gan_trainer.py. The neural network configurations are specified in gan_conf.py (for synthetic data) and gan_conf_image.py (for image data). - -In order to run the model, first download the corresponding data by running the shell script in ./data. -Then you can run 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). - -$python gan_trainer.py -d cifar --use_gpu 1 - -The generated images will be stored in ./cifar_samples/ -The corresponding models will be stored in ./cifar_params/ diff --git a/v1_api_demo/gan/data/download_cifar.sh b/v1_api_demo/gan/data/download_cifar.sh deleted file mode 100755 index bbadc7c10c73e45a0948018b8812f79040d14bc4..0000000000000000000000000000000000000000 --- a/v1_api_demo/gan/data/download_cifar.sh +++ /dev/null @@ -1,18 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -set -e -wget https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz -tar zxf cifar-10-python.tar.gz -rm cifar-10-python.tar.gz diff --git a/v1_api_demo/gan/data/get_mnist_data.sh b/v1_api_demo/gan/data/get_mnist_data.sh deleted file mode 100755 index a77c81bf5af9ddb6634ff89460797ca543c5e517..0000000000000000000000000000000000000000 --- a/v1_api_demo/gan/data/get_mnist_data.sh +++ /dev/null @@ -1,17 +0,0 @@ -#!/usr/bin/env sh -# This script downloads the mnist data and unzips it. -set -e -DIR="$( cd "$(dirname "$0")" ; pwd -P )" -rm -rf "$DIR/mnist_data" -mkdir "$DIR/mnist_data" -cd "$DIR/mnist_data" - -echo "Downloading..." - -for fname in train-images-idx3-ubyte train-labels-idx1-ubyte t10k-images-idx3-ubyte t10k-labels-idx1-ubyte -do - if [ ! -e $fname ]; then - wget --no-check-certificate http://yann.lecun.com/exdb/mnist/${fname}.gz - gunzip ${fname}.gz - fi -done diff --git a/v1_api_demo/gan/gan_conf.py b/v1_api_demo/gan/gan_conf.py deleted file mode 100644 index 86ac2dffe5f4490a88e12d1fa5e8cd9fa61a69f4..0000000000000000000000000000000000000000 --- a/v1_api_demo/gan/gan_conf.py +++ /dev/null @@ -1,151 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -from paddle.trainer_config_helpers import * - -mode = get_config_arg("mode", str, "generator") -assert mode in set([ - "generator", "discriminator", "generator_training", "discriminator_training" -]) - -is_generator_training = mode == "generator_training" -is_discriminator_training = mode == "discriminator_training" -is_generator = mode == "generator" -is_discriminator = mode == "discriminator" - -# The network structure below follows the ref https://arxiv.org/abs/1406.2661 -# Here we used two hidden layers and batch_norm - -print('mode=%s' % mode) -# the dim of the noise (z) as the input of the generator network -noise_dim = 10 -# the dim of the hidden layer -hidden_dim = 10 -# the dim of the generated sample -sample_dim = 2 - -settings( - batch_size=128, - learning_rate=1e-4, - learning_method=AdamOptimizer(beta1=0.5)) - - -def discriminator(sample): - """ - discriminator ouputs the probablity of a sample is from generator - or real data. - The output has two dimenstional: dimension 0 is the probablity - of the sample is from generator and dimension 1 is the probabblity - of the sample is from real data. - """ - param_attr = ParamAttr(is_static=is_generator_training) - bias_attr = ParamAttr( - is_static=is_generator_training, initial_mean=1.0, initial_std=0) - - hidden = fc_layer( - input=sample, - name="dis_hidden", - size=hidden_dim, - bias_attr=bias_attr, - param_attr=param_attr, - act=ReluActivation()) - - hidden2 = fc_layer( - input=hidden, - name="dis_hidden2", - size=hidden_dim, - bias_attr=bias_attr, - param_attr=param_attr, - act=LinearActivation()) - - hidden_bn = batch_norm_layer( - hidden2, - act=ReluActivation(), - name="dis_hidden_bn", - bias_attr=bias_attr, - param_attr=ParamAttr( - is_static=is_generator_training, initial_mean=1.0, - initial_std=0.02), - use_global_stats=False) - - return fc_layer( - input=hidden_bn, - name="dis_prob", - size=2, - bias_attr=bias_attr, - param_attr=param_attr, - act=SoftmaxActivation()) - - -def generator(noise): - """ - generator generates a sample given noise - """ - param_attr = ParamAttr(is_static=is_discriminator_training) - bias_attr = ParamAttr( - is_static=is_discriminator_training, initial_mean=1.0, initial_std=0) - - hidden = fc_layer( - input=noise, - name="gen_layer_hidden", - size=hidden_dim, - bias_attr=bias_attr, - param_attr=param_attr, - act=ReluActivation()) - - hidden2 = fc_layer( - input=hidden, - name="gen_hidden2", - size=hidden_dim, - bias_attr=bias_attr, - param_attr=param_attr, - act=LinearActivation()) - - hidden_bn = batch_norm_layer( - hidden2, - act=ReluActivation(), - name="gen_layer_hidden_bn", - bias_attr=bias_attr, - param_attr=ParamAttr( - is_static=is_discriminator_training, - initial_mean=1.0, - initial_std=0.02), - use_global_stats=False) - - return fc_layer( - input=hidden_bn, - name="gen_layer1", - size=sample_dim, - bias_attr=bias_attr, - param_attr=param_attr, - act=LinearActivation()) - - -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)) diff --git a/v1_api_demo/gan/gan_conf_image.py b/v1_api_demo/gan/gan_conf_image.py deleted file mode 100644 index c469227994c1a84d1aa73e03bbc74ebeac41d30e..0000000000000000000000000000000000000000 --- a/v1_api_demo/gan/gan_conf_image.py +++ /dev/null @@ -1,298 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -from paddle.trainer_config_helpers import * - -mode = get_config_arg("mode", str, "generator") -dataSource = get_config_arg("data", str, "mnist") -assert mode in set([ - "generator", "discriminator", "generator_training", "discriminator_training" -]) - -is_generator_training = mode == "generator_training" -is_discriminator_training = mode == "discriminator_training" -is_generator = mode == "generator" -is_discriminator = mode == "discriminator" - -# The network structure below follows the dcgan paper -# (https://arxiv.org/abs/1511.06434) - -print('mode=%s' % mode) -# the dim of the noise (z) as the input of the generator network -noise_dim = 100 -# the number of filters in the layer in generator/discriminator that is -# closet to the image -gf_dim = 64 -df_dim = 64 -if dataSource == "mnist": - sample_dim = 28 # image dim - c_dim = 1 # image color -else: - sample_dim = 32 - c_dim = 3 -s2, s4 = int(sample_dim / 2), int(sample_dim / 4), -s8, s16 = int(sample_dim / 8), int(sample_dim / 16) - -settings( - batch_size=128, - learning_rate=2e-4, - learning_method=AdamOptimizer(beta1=0.5)) - - -def conv_bn(input, - channels, - imgSize, - num_filters, - output_x, - stride, - name, - param_attr, - bias_attr, - param_attr_bn, - bn, - trans=False, - act=ReluActivation()): - """ - conv_bn is a utility function that constructs a convolution/deconv layer - with an optional batch_norm layer - - :param bn: whether to use batch_norm_layer - :type bn: bool - :param trans: whether to use conv (False) or deconv (True) - :type trans: bool - """ - - # calculate the filter_size and padding size based on the given - # imgSize and ouput size - tmp = imgSize - (output_x - 1) * stride - if tmp <= 1 or tmp > 5: - raise ValueError("conv input-output dimension does not fit") - elif tmp <= 3: - filter_size = tmp + 2 - padding = 1 - else: - filter_size = tmp - padding = 0 - - print(imgSize, output_x, stride, filter_size, padding) - - if trans: - nameApx = "_convt" - else: - nameApx = "_conv" - - if bn: - conv = img_conv_layer( - input, - filter_size=filter_size, - num_filters=num_filters, - name=name + nameApx, - num_channels=channels, - act=LinearActivation(), - groups=1, - stride=stride, - padding=padding, - bias_attr=bias_attr, - param_attr=param_attr, - shared_biases=True, - layer_attr=None, - filter_size_y=None, - stride_y=None, - padding_y=None, - trans=trans) - - conv_bn = batch_norm_layer( - conv, - act=act, - name=name + nameApx + "_bn", - bias_attr=bias_attr, - param_attr=param_attr_bn, - use_global_stats=False) - - return conv_bn - else: - conv = img_conv_layer( - input, - filter_size=filter_size, - num_filters=num_filters, - name=name + nameApx, - num_channels=channels, - act=act, - groups=1, - stride=stride, - padding=padding, - bias_attr=bias_attr, - param_attr=param_attr, - shared_biases=True, - layer_attr=None, - filter_size_y=None, - stride_y=None, - padding_y=None, - trans=trans) - return conv - - -def generator(noise): - """ - generator generates a sample given noise - """ - param_attr = ParamAttr( - is_static=is_discriminator_training, initial_mean=0.0, initial_std=0.02) - bias_attr = ParamAttr( - is_static=is_discriminator_training, initial_mean=0.0, initial_std=0.0) - - param_attr_bn = ParamAttr( - is_static=is_discriminator_training, initial_mean=1.0, initial_std=0.02) - - h1 = fc_layer( - input=noise, - name="gen_layer_h1", - size=s8 * s8 * gf_dim * 4, - bias_attr=bias_attr, - param_attr=param_attr, - act=LinearActivation()) - - h1_bn = batch_norm_layer( - h1, - act=ReluActivation(), - name="gen_layer_h1_bn", - bias_attr=bias_attr, - param_attr=param_attr_bn, - use_global_stats=False) - - h2_bn = conv_bn( - h1_bn, - channels=gf_dim * 4, - output_x=s8, - num_filters=gf_dim * 2, - imgSize=s4, - stride=2, - name="gen_layer_h2", - param_attr=param_attr, - bias_attr=bias_attr, - param_attr_bn=param_attr_bn, - bn=True, - trans=True) - - h3_bn = conv_bn( - h2_bn, - channels=gf_dim * 2, - output_x=s4, - num_filters=gf_dim, - imgSize=s2, - stride=2, - name="gen_layer_h3", - param_attr=param_attr, - bias_attr=bias_attr, - param_attr_bn=param_attr_bn, - bn=True, - trans=True) - - return conv_bn( - h3_bn, - channels=gf_dim, - output_x=s2, - num_filters=c_dim, - imgSize=sample_dim, - stride=2, - name="gen_layer_h4", - param_attr=param_attr, - bias_attr=bias_attr, - param_attr_bn=param_attr_bn, - bn=False, - trans=True, - act=TanhActivation()) - - -def discriminator(sample): - """ - discriminator ouputs the probablity of a sample is from generator - or real data. - The output has two dimenstional: dimension 0 is the probablity - of the sample is from generator and dimension 1 is the probabblity - of the sample is from real data. - """ - param_attr = ParamAttr( - is_static=is_generator_training, initial_mean=0.0, initial_std=0.02) - bias_attr = ParamAttr( - is_static=is_generator_training, initial_mean=0.0, initial_std=0.0) - - param_attr_bn = ParamAttr( - is_static=is_generator_training, initial_mean=1.0, initial_std=0.02) - - h0 = conv_bn( - sample, - channels=c_dim, - imgSize=sample_dim, - num_filters=df_dim, - output_x=s2, - stride=2, - name="dis_h0", - param_attr=param_attr, - bias_attr=bias_attr, - param_attr_bn=param_attr_bn, - bn=False) - - h1_bn = conv_bn( - h0, - channels=df_dim, - imgSize=s2, - num_filters=df_dim * 2, - output_x=s4, - stride=2, - name="dis_h1", - param_attr=param_attr, - bias_attr=bias_attr, - param_attr_bn=param_attr_bn, - bn=True) - - h2_bn = conv_bn( - h1_bn, - channels=df_dim * 2, - imgSize=s4, - num_filters=df_dim * 4, - output_x=s8, - stride=2, - name="dis_h2", - param_attr=param_attr, - bias_attr=bias_attr, - param_attr_bn=param_attr_bn, - bn=True) - - return fc_layer( - input=h2_bn, - name="dis_prob", - size=2, - bias_attr=bias_attr, - param_attr=param_attr, - act=SoftmaxActivation()) - - -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 * sample_dim * c_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)) diff --git a/v1_api_demo/gan/gan_trainer.py b/v1_api_demo/gan/gan_trainer.py deleted file mode 100644 index 4a26c230f7a21cc6dd4a3cdb52e32730b1ce73ca..0000000000000000000000000000000000000000 --- a/v1_api_demo/gan/gan_trainer.py +++ /dev/null @@ -1,349 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import argparse -import random -import numpy -import cPickle -import sys, os -from PIL import Image - -from paddle.trainer.config_parser import parse_config -from paddle.trainer.config_parser import logger -import py_paddle.swig_paddle as api -import matplotlib.pyplot as plt - - -def plot2DScatter(data, outputfile): - ''' - Plot the data as a 2D scatter plot and save to outputfile - data needs to be two dimensinoal - ''' - x = data[:, 0] - y = data[:, 1] - logger.info("The mean vector is %s" % numpy.mean(data, 0)) - logger.info("The std vector is %s" % numpy.std(data, 0)) - - heatmap, xedges, yedges = numpy.histogram2d(x, y, bins=50) - extent = [xedges[0], xedges[-1], yedges[0], yedges[-1]] - - plt.clf() - plt.scatter(x, y) - plt.savefig(outputfile, bbox_inches='tight') - - -def CHECK_EQ(a, b): - assert a == b, "a=%s, b=%s" % (a, b) - - -def copy_shared_parameters(src, dst): - ''' - copy the parameters from src to dst - :param src: the source of the parameters - :type src: GradientMachine - :param dst: the destination of the parameters - :type dst: GradientMachine - ''' - src_params = [src.getParameter(i) for i in xrange(src.getParameterSize())] - src_params = dict([(p.getName(), p) for p in src_params]) - - for i in xrange(dst.getParameterSize()): - dst_param = dst.getParameter(i) - src_param = src_params.get(dst_param.getName(), None) - if src_param is None: - continue - src_value = src_param.getBuf(api.PARAMETER_VALUE) - dst_value = dst_param.getBuf(api.PARAMETER_VALUE) - CHECK_EQ(len(src_value), len(dst_value)) - dst_value.copyFrom(src_value) - dst_param.setValueUpdated() - - -def print_parameters(src): - src_params = [src.getParameter(i) for i in xrange(src.getParameterSize())] - - print "***************" - for p in src_params: - print "Name is %s" % p.getName() - print "value is %s \n" % p.getBuf(api.PARAMETER_VALUE).copyToNumpyArray( - ) - - -def load_mnist_data(imageFile): - f = open(imageFile, "rb") - f.read(16) - - # Define number of samples for train/test - if "train" in imageFile: - n = 60000 - else: - n = 10000 - - data = numpy.fromfile(f, 'ubyte', count=n * 28 * 28).reshape((n, 28 * 28)) - data = data / 255.0 * 2.0 - 1.0 - - f.close() - return data.astype('float32') - - -def load_cifar_data(cifar_path): - batch_size = 10000 - data = numpy.zeros((5 * batch_size, 32 * 32 * 3), dtype="float32") - for i in range(1, 6): - file = cifar_path + "/data_batch_" + str(i) - fo = open(file, 'rb') - dict = cPickle.load(fo) - fo.close() - data[(i - 1) * batch_size:(i * batch_size), :] = dict["data"] - - data = data / 255.0 * 2.0 - 1.0 - return data - - -# synthesize 2-D uniform data -def load_uniform_data(): - data = numpy.random.rand(1000000, 2).astype('float32') - return data - - -def merge(images, size): - if images.shape[1] == 28 * 28: - h, w, c = 28, 28, 1 - else: - h, w, c = 32, 32, 3 - img = numpy.zeros((h * size[0], w * size[1], c)) - for idx in xrange(size[0] * size[1]): - i = idx % size[1] - j = idx // size[1] - img[j*h:j*h+h, i*w:i*w+w, :] = \ - ((images[idx, :].reshape((h, w, c), order="F").transpose(1, 0, 2) + 1.0) / 2.0 * 255.0) - return img.astype('uint8') - - -def save_images(images, path): - merged_img = merge(images, [8, 8]) - if merged_img.shape[2] == 1: - im = Image.fromarray(numpy.squeeze(merged_img)).convert('RGB') - else: - im = Image.fromarray(merged_img, mode="RGB") - im.save(path) - - -def get_real_samples(batch_size, data_np): - return data_np[numpy.random.choice( - data_np.shape[0], batch_size, replace=False), :] - - -def get_noise(batch_size, noise_dim): - return numpy.random.normal(size=(batch_size, noise_dim)).astype('float32') - - -def get_fake_samples(generator_machine, batch_size, noise): - gen_inputs = api.Arguments.createArguments(1) - gen_inputs.setSlotValue(0, api.Matrix.createDenseFromNumpy(noise)) - gen_outputs = api.Arguments.createArguments(0) - generator_machine.forward(gen_inputs, gen_outputs, api.PASS_TEST) - fake_samples = gen_outputs.getSlotValue(0).copyToNumpyMat() - return fake_samples - - -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) - - -def prepare_discriminator_data_batch_pos(batch_size, data_np): - real_samples = get_real_samples(batch_size, data_np) - labels = numpy.ones(batch_size, dtype='int32') - inputs = api.Arguments.createArguments(2) - inputs.setSlotValue(0, api.Matrix.createDenseFromNumpy(real_samples)) - inputs.setSlotIds(1, api.IVector.createVectorFromNumpy(labels)) - return inputs - - -def prepare_discriminator_data_batch_neg(generator_machine, batch_size, noise): - fake_samples = get_fake_samples(generator_machine, batch_size, noise) - labels = numpy.zeros(batch_size, dtype='int32') - inputs = api.Arguments.createArguments(2) - inputs.setSlotValue(0, api.Matrix.createDenseFromNumpy(fake_samples)) - inputs.setSlotIds(1, api.IVector.createVectorFromNumpy(labels)) - return inputs - - -def prepare_generator_data_batch(batch_size, noise): - label = numpy.ones(batch_size, dtype='int32') - inputs = api.Arguments.createArguments(2) - inputs.setSlotValue(0, api.Matrix.createDenseFromNumpy(noise)) - inputs.setSlotIds(1, api.IVector.createVectorFromNumpy(label)) - return inputs - - -def find(iterable, cond): - for item in iterable: - if cond(item): - return item - return None - - -def get_layer_size(model_conf, layer_name): - layer_conf = find(model_conf.layers, lambda x: x.name == layer_name) - assert layer_conf is not None, "Cannot find '%s' layer" % layer_name - return layer_conf.size - - -def main(): - parser = argparse.ArgumentParser() - parser.add_argument("-d", "--data_source", help="mnist or cifar or uniform") - parser.add_argument( - "--use_gpu", default="1", help="1 means use gpu for training") - parser.add_argument("--gpu_id", default="0", help="the gpu_id parameter") - args = parser.parse_args() - data_source = args.data_source - use_gpu = args.use_gpu - assert data_source in ["mnist", "cifar", "uniform"] - assert use_gpu in ["0", "1"] - - if not os.path.exists("./%s_samples/" % data_source): - os.makedirs("./%s_samples/" % data_source) - - if not os.path.exists("./%s_params/" % data_source): - os.makedirs("./%s_params/" % data_source) - - api.initPaddle('--use_gpu=' + use_gpu, '--dot_period=10', - '--log_period=100', '--gpu_id=' + args.gpu_id, - '--save_dir=' + "./%s_params/" % data_source) - - if data_source == "uniform": - conf = "gan_conf.py" - num_iter = 10000 - else: - conf = "gan_conf_image.py" - num_iter = 1000 - - 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) - batch_size = dis_conf.opt_config.batch_size - noise_dim = get_layer_size(gen_conf.model_config, "noise") - - if data_source == "mnist": - data_np = load_mnist_data("./data/mnist_data/train-images-idx3-ubyte") - elif data_source == "cifar": - data_np = load_cifar_data("./data/cifar-10-batches-py/") - else: - data_np = load_uniform_data() - - # this creates a gradient machine for discriminator - dis_training_machine = api.GradientMachine.createFromConfigProto( - dis_conf.model_config) - # this create a gradient machine for generator - gen_training_machine = api.GradientMachine.createFromConfigProto( - gen_conf.model_config) - - # generator_machine is used to generate data only, which is used for - # training discriminator - logger.info(str(generator_conf.model_config)) - generator_machine = api.GradientMachine.createFromConfigProto( - generator_conf.model_config) - - dis_trainer = api.Trainer.create(dis_conf, dis_training_machine) - - gen_trainer = api.Trainer.create(gen_conf, gen_training_machine) - - dis_trainer.startTrain() - gen_trainer.startTrain() - - # Sync parameters between networks (GradientMachine) at the beginning - copy_shared_parameters(gen_training_machine, dis_training_machine) - copy_shared_parameters(gen_training_machine, generator_machine) - - # constrain that either discriminator or generator can not be trained - # consecutively more than MAX_strike times - curr_train = "dis" - curr_strike = 0 - MAX_strike = 5 - - for train_pass in xrange(100): - dis_trainer.startTrainPass() - gen_trainer.startTrainPass() - for i in xrange(num_iter): - # Do forward pass in discriminator to get the dis_loss - noise = get_noise(batch_size, noise_dim) - data_batch_dis_pos = prepare_discriminator_data_batch_pos( - batch_size, data_np) - dis_loss_pos = get_training_loss(dis_training_machine, - data_batch_dis_pos) - - data_batch_dis_neg = prepare_discriminator_data_batch_neg( - generator_machine, batch_size, noise) - dis_loss_neg = get_training_loss(dis_training_machine, - data_batch_dis_neg) - - dis_loss = (dis_loss_pos + dis_loss_neg) / 2.0 - - # Do forward pass in generator to get the gen_loss - data_batch_gen = prepare_generator_data_batch(batch_size, noise) - gen_loss = get_training_loss(gen_training_machine, data_batch_gen) - - if i % 100 == 0: - print "d_pos_loss is %s d_neg_loss is %s" % (dis_loss_pos, - dis_loss_neg) - print "d_loss is %s g_loss is %s" % (dis_loss, gen_loss) - - # Decide which network to train based on the training history - # And the relative size of the loss - if (not (curr_train == "dis" and curr_strike == MAX_strike)) and \ - ((curr_train == "gen" and curr_strike == MAX_strike) or dis_loss > gen_loss): - if curr_train == "dis": - curr_strike += 1 - else: - curr_train = "dis" - curr_strike = 1 - dis_trainer.trainOneDataBatch(batch_size, data_batch_dis_neg) - dis_trainer.trainOneDataBatch(batch_size, data_batch_dis_pos) - copy_shared_parameters(dis_training_machine, - gen_training_machine) - - else: - if curr_train == "gen": - curr_strike += 1 - else: - curr_train = "gen" - curr_strike = 1 - gen_trainer.trainOneDataBatch(batch_size, data_batch_gen) - # TODO: add API for paddle to allow true parameter sharing between different GradientMachines - # so that we do not need to copy shared parameters. - copy_shared_parameters(gen_training_machine, - dis_training_machine) - copy_shared_parameters(gen_training_machine, generator_machine) - - dis_trainer.finishTrainPass() - gen_trainer.finishTrainPass() - # At the end of each pass, save the generated samples/images - fake_samples = get_fake_samples(generator_machine, batch_size, noise) - if data_source == "uniform": - plot2DScatter(fake_samples, "./%s_samples/train_pass%s.png" % - (data_source, train_pass)) - else: - save_images(fake_samples, "./%s_samples/train_pass%s.png" % - (data_source, train_pass)) - dis_trainer.finishTrain() - gen_trainer.finishTrain() - - -if __name__ == '__main__': - main() diff --git a/v1_api_demo/mnist/.gitignore b/v1_api_demo/mnist/.gitignore deleted file mode 100644 index 7e61d5e3a0cabd46d4185454d46610ac2ee2e63f..0000000000000000000000000000000000000000 --- a/v1_api_demo/mnist/.gitignore +++ /dev/null @@ -1,10 +0,0 @@ -data/raw_data -data/*.list -mnist_vgg_model -plot.png -train.log -*pyc -.ipynb_checkpoints -params.pkl -params.tar -params.tar.gz diff --git a/v1_api_demo/mnist/api_train.py b/v1_api_demo/mnist/api_train.py deleted file mode 100644 index e42c6cbb7e0eed4f3a3625f18d79b3de64fd8e26..0000000000000000000000000000000000000000 --- a/v1_api_demo/mnist/api_train.py +++ /dev/null @@ -1,209 +0,0 @@ -# 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. -""" -A very basic example for how to use current Raw SWIG API to train mnist network. - -Current implementation uses Raw SWIG, which means the API call is directly \ -passed to C++ side of Paddle. - -The user api could be simpler and carefully designed. -""" -import random - -import numpy as np -import paddle.v2 as paddle_v2 -import py_paddle.swig_paddle as api -from paddle.trainer_config_helpers import * -from py_paddle import DataProviderConverter - -from mnist_util import read_from_mnist - - -def init_parameter(network): - assert isinstance(network, api.GradientMachine) - for each_param in network.getParameters(): - assert isinstance(each_param, api.Parameter) - array_size = len(each_param) - array = np.random.uniform(-1.0, 1.0, array_size).astype('float32') - each_param.getBuf(api.PARAMETER_VALUE).copyFromNumpyArray(array) - - -def generator_to_batch(generator, batch_size): - ret_val = list() - for each_item in generator: - ret_val.append(each_item) - if len(ret_val) == batch_size: - yield ret_val - ret_val = list() - if len(ret_val) != 0: - yield ret_val - - -class BatchPool(object): - def __init__(self, generator, batch_size): - self.data = list(generator) - self.batch_size = batch_size - - def __call__(self): - random.shuffle(self.data) - for offset in xrange(0, len(self.data), self.batch_size): - limit = min(offset + self.batch_size, len(self.data)) - yield self.data[offset:limit] - - -def input_order_converter(generator): - for each_item in generator: - yield each_item['pixel'], each_item['label'] - - -def main(): - api.initPaddle("-use_gpu=false", "-trainer_count=4") # use 4 cpu cores - - optimizer = paddle_v2.optimizer.Adam( - learning_rate=1e-4, - batch_size=1000, - model_average=ModelAverage(average_window=0.5), - regularization=L2Regularization(rate=0.5)) - - # Create Local Updater. Local means not run in cluster. - # For a cluster training, here we can change to createRemoteUpdater - # in future. - updater = optimizer.create_local_updater() - assert isinstance(updater, api.ParameterUpdater) - - # define network - images = paddle_v2.layer.data( - name='pixel', type=paddle_v2.data_type.dense_vector(784)) - label = paddle_v2.layer.data( - name='label', type=paddle_v2.data_type.integer_value(10)) - hidden1 = paddle_v2.layer.fc(input=images, size=200) - hidden2 = paddle_v2.layer.fc(input=hidden1, size=200) - inference = paddle_v2.layer.fc(input=hidden2, - size=10, - act=paddle_v2.activation.Softmax()) - cost = paddle_v2.layer.classification_cost(input=inference, label=label) - - # Create Simple Gradient Machine. - model_config = paddle_v2.layer.parse_network(cost) - m = api.GradientMachine.createFromConfigProto(model_config, - api.CREATE_MODE_NORMAL, - optimizer.enable_types()) - - # This type check is not useful. Only enable type hint in IDE. - # Such as PyCharm - assert isinstance(m, api.GradientMachine) - - # Initialize Parameter by numpy. - init_parameter(network=m) - - # Initialize ParameterUpdater. - updater.init(m) - - # DataProvider Converter is a utility convert Python Object to Paddle C++ - # Input. The input format is as same as Paddle's DataProvider. - converter = DataProviderConverter(input_types=[images.type, label.type]) - - train_file = './data/raw_data/train' - test_file = './data/raw_data/t10k' - - # start gradient machine. - # the gradient machine must be started before invoke forward/backward. - # not just for training, but also for inference. - m.start() - - # evaluator can print error rate, etc. It is a C++ class. - batch_evaluator = m.makeEvaluator() - test_evaluator = m.makeEvaluator() - - # Get Train Data. - # TrainData will stored in a data pool. Currently implementation is not care - # about memory, speed. Just a very naive implementation. - train_data_generator = input_order_converter(read_from_mnist(train_file)) - train_data = BatchPool(train_data_generator, 512) - - # outArgs is Neural Network forward result. Here is not useful, just passed - # to gradient_machine.forward - outArgs = api.Arguments.createArguments(0) - - for pass_id in xrange(2): # we train 2 passes. - updater.startPass() - - for batch_id, data_batch in enumerate(train_data()): - # data_batch is input images. - # here, for online learning, we could get data_batch from network. - - # Start update one batch. - pass_type = updater.startBatch(len(data_batch)) - - # Start BatchEvaluator. - # batch_evaluator can be used between start/finish. - batch_evaluator.start() - - # forwardBackward is a shortcut for forward and backward. - # It is sometimes faster than invoke forward/backward separately, - # because in GradientMachine, it may be async. - m.forwardBackward(converter(data_batch), outArgs, pass_type) - - for each_param in m.getParameters(): - updater.update(each_param) - - # Get cost. We use numpy to calculate total cost for this batch. - cost_vec = outArgs.getSlotValue(0) - cost_vec = cost_vec.copyToNumpyMat() - cost = cost_vec.sum() / len(data_batch) - - # Make evaluator works. - m.eval(batch_evaluator) - - # Print logs. - print 'Pass id', pass_id, 'Batch id', batch_id, 'with cost=', \ - cost, batch_evaluator - - batch_evaluator.finish() - # Finish batch. - # * will clear gradient. - # * ensure all values should be updated. - updater.finishBatch(cost) - - # testing stage. use test data set to test current network. - updater.apply() - test_evaluator.start() - test_data_generator = input_order_converter(read_from_mnist(test_file)) - for data_batch in generator_to_batch(test_data_generator, 512): - # in testing stage, only forward is needed. - m.forward(converter(data_batch), outArgs, api.PASS_TEST) - m.eval(test_evaluator) - - # print error rate for test data set - print 'Pass', pass_id, ' test evaluator: ', test_evaluator - test_evaluator.finish() - updater.restore() - - updater.catchUpWith() - params = m.getParameters() - for each_param in params: - assert isinstance(each_param, api.Parameter) - value = each_param.getBuf(api.PARAMETER_VALUE) - value = value.copyToNumpyArray() - - # Here, we could save parameter to every where you want - print each_param.getName(), value - - updater.finishPass() - - m.finish() - - -if __name__ == '__main__': - main() diff --git a/v1_api_demo/mnist/data/generate_list.py b/v1_api_demo/mnist/data/generate_list.py deleted file mode 100644 index 49981cc7a93308bc96ad5097eba749440e958525..0000000000000000000000000000000000000000 --- a/v1_api_demo/mnist/data/generate_list.py +++ /dev/null @@ -1,21 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -o = open("./" + "train.list", "w") -o.write("./data/raw_data/train" + "\n") -o.close() - -o = open("./" + "test.list", "w") -o.write("./data/raw_data/t10k" + "\n") -o.close() diff --git a/v1_api_demo/mnist/data/get_mnist_data.sh b/v1_api_demo/mnist/data/get_mnist_data.sh deleted file mode 100755 index 5a2e34026d4fe7f8315d4f5453bec7c4ee4f6885..0000000000000000000000000000000000000000 --- a/v1_api_demo/mnist/data/get_mnist_data.sh +++ /dev/null @@ -1,21 +0,0 @@ -#!/usr/bin/env sh -# This scripts downloads the mnist data and unzips it. -set -e -DIR="$( cd "$(dirname "$0")" ; pwd -P )" -rm -rf "$DIR/raw_data" -mkdir "$DIR/raw_data" -cd "$DIR/raw_data" - -echo "Downloading..." - -for fname in train-images-idx3-ubyte train-labels-idx1-ubyte t10k-images-idx3-ubyte t10k-labels-idx1-ubyte -do - if [ ! -e $fname ]; then - wget --no-check-certificate http://yann.lecun.com/exdb/mnist/${fname}.gz - gunzip ${fname}.gz - fi -done - -cd $DIR -rm -f *.list -python generate_list.py diff --git a/v1_api_demo/mnist/light_mnist.py b/v1_api_demo/mnist/light_mnist.py deleted file mode 100644 index 33409054357d2f0c6a765b3ab3164eb2e584467e..0000000000000000000000000000000000000000 --- a/v1_api_demo/mnist/light_mnist.py +++ /dev/null @@ -1,79 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -is_predict = get_config_arg("is_predict", bool, False) - -####################Data Configuration ################## - -if not is_predict: - data_dir = './data/' - define_py_data_sources2( - train_list=data_dir + 'train.list', - test_list=data_dir + 'test.list', - module='mnist_provider', - obj='process') - -######################Algorithm Configuration ############# -settings(batch_size=50, learning_rate=0.001, learning_method=AdamOptimizer()) - -#######################Network Configuration ############# - -data_size = 1 * 28 * 28 -label_size = 10 -img = data_layer(name='pixel', size=data_size) - - -# light cnn -# A shallower cnn model: [CNN, BN, ReLU, Max-Pooling] x4 + FC x1 -# Easier to train for mnist dataset and quite efficient -# Final performance is close to deeper ones on tasks such as digital and character classification -def light_cnn(input_image, num_channels, num_classes): - def __light__(ipt, - num_filter=128, - times=1, - conv_filter_size=3, - dropouts=0, - num_channels_=None): - return img_conv_group( - input=ipt, - num_channels=num_channels_, - pool_size=2, - pool_stride=2, - conv_padding=0, - conv_num_filter=[num_filter] * times, - conv_filter_size=conv_filter_size, - conv_act=ReluActivation(), - conv_with_batchnorm=True, - conv_batchnorm_drop_rate=dropouts, - pool_type=MaxPooling()) - - tmp = __light__(input_image, num_filter=128, num_channels_=num_channels) - tmp = __light__(tmp, num_filter=128) - tmp = __light__(tmp, num_filter=128) - tmp = __light__(tmp, num_filter=128, conv_filter_size=1) - - tmp = fc_layer(input=tmp, size=num_classes, act=SoftmaxActivation()) - return tmp - - -predict = light_cnn(input_image=img, num_channels=1, num_classes=label_size) - -if not is_predict: - lbl = data_layer(name="label", size=label_size) - inputs(img, lbl) - outputs(classification_cost(input=predict, label=lbl)) -else: - outputs(predict) diff --git a/v1_api_demo/mnist/mnist_provider.py b/v1_api_demo/mnist/mnist_provider.py deleted file mode 100644 index 4192339837620aada84b64a92fef3e05953971c2..0000000000000000000000000000000000000000 --- a/v1_api_demo/mnist/mnist_provider.py +++ /dev/null @@ -1,25 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. -# -#Licensed under the Apache License, Version 2.0 (the "License"); -#you may not use this file except in compliance with the License. -#You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -#Unless required by applicable law or agreed to in writing, software -#distributed under the License is distributed on an "AS IS" BASIS, -#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -#See the License for the specific language governing permissions and -#limitations under the License. -from paddle.trainer.PyDataProvider2 import * -from mnist_util import read_from_mnist - - -# Define a py data provider -@provider( - input_types={'pixel': dense_vector(28 * 28), - 'label': integer_value(10)}, - cache=CacheType.CACHE_PASS_IN_MEM) -def process(settings, filename): # settings is not used currently. - for each in read_from_mnist(filename): - yield each diff --git a/v1_api_demo/mnist/mnist_util.py b/v1_api_demo/mnist/mnist_util.py deleted file mode 100644 index 3fd88ae7edc821296ca0accbf6dedc083e411744..0000000000000000000000000000000000000000 --- a/v1_api_demo/mnist/mnist_util.py +++ /dev/null @@ -1,30 +0,0 @@ -import numpy - -__all__ = ['read_from_mnist'] - - -def read_from_mnist(filename): - imgf = filename + "-images-idx3-ubyte" - labelf = filename + "-labels-idx1-ubyte" - f = open(imgf, "rb") - l = open(labelf, "rb") - - f.read(16) - l.read(8) - - # Define number of samples for train/test - if "train" in filename: - n = 60000 - else: - n = 10000 - - images = numpy.fromfile( - f, 'ubyte', count=n * 28 * 28).reshape((n, 28 * 28)).astype('float32') - images = images / 255.0 * 2.0 - 1.0 - labels = numpy.fromfile(l, 'ubyte', count=n).astype("int") - - for i in xrange(n): - yield {"pixel": images[i, :], 'label': labels[i]} - - f.close() - l.close() diff --git a/v1_api_demo/mnist/train.sh b/v1_api_demo/mnist/train.sh deleted file mode 100755 index ca2b1ad9eb960685b95b0f294a9b929e1a4acab1..0000000000000000000000000000000000000000 --- a/v1_api_demo/mnist/train.sh +++ /dev/null @@ -1,32 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -set -e -config=vgg_16_mnist.py -output=./mnist_vgg_model -log=train.log - -paddle train \ ---config=$config \ ---dot_period=10 \ ---log_period=100 \ ---test_all_data_in_one_period=1 \ ---use_gpu=0 \ ---trainer_count=1 \ ---num_passes=100 \ ---save_dir=$output \ -2>&1 | tee $log -paddle usage -l $log -e $? -n "mnist_train" >/dev/null 2>&1 - -python -m paddle.utils.plotcurve -i $log > plot.png diff --git a/v1_api_demo/mnist/vgg_16_mnist.py b/v1_api_demo/mnist/vgg_16_mnist.py deleted file mode 100644 index a819b391c690fb473801eb2e7ba3161cc31b5b4b..0000000000000000000000000000000000000000 --- a/v1_api_demo/mnist/vgg_16_mnist.py +++ /dev/null @@ -1,50 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -is_predict = get_config_arg("is_predict", bool, False) - -####################Data Configuration ################## - -if not is_predict: - data_dir = './data/' - define_py_data_sources2( - train_list=data_dir + 'train.list', - test_list=data_dir + 'test.list', - module='mnist_provider', - obj='process') - -######################Algorithm Configuration ############# -settings( - batch_size=128, - learning_rate=0.1 / 128.0, - learning_method=MomentumOptimizer(0.9), - regularization=L2Regularization(0.0005 * 128)) - -#######################Network Configuration ############# - -data_size = 1 * 28 * 28 -label_size = 10 -img = data_layer(name='pixel', size=data_size) - -# small_vgg is predined in trainer_config_helpers.network -predict = small_vgg(input_image=img, num_channels=1, num_classes=label_size) - -if not is_predict: - lbl = data_layer(name="label", size=label_size) - inputs(img, lbl) - outputs(classification_cost(input=predict, label=lbl)) -else: - outputs(predict) diff --git a/v1_api_demo/model_zoo/embedding/.gitignore b/v1_api_demo/model_zoo/embedding/.gitignore deleted file mode 100644 index 908f5a3fb2f7c34368ea24d0fc3ac9cac29a4fdb..0000000000000000000000000000000000000000 --- a/v1_api_demo/model_zoo/embedding/.gitignore +++ /dev/null @@ -1,2 +0,0 @@ -baidu.dict -model_*.emb diff --git a/v1_api_demo/model_zoo/embedding/extract_para.py b/v1_api_demo/model_zoo/embedding/extract_para.py deleted file mode 100755 index 570b90c1f772c8f6abfc6cda02560fd3471ef0b6..0000000000000000000000000000000000000000 --- a/v1_api_demo/model_zoo/embedding/extract_para.py +++ /dev/null @@ -1,113 +0,0 @@ -#!/bin/env python -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -""" -Example: - python extract_para.py --preModel PREMODEL --preDict PREDICT \ - --usrModel USRMODEL --usrDict USRDICT -d DIM - -Options: - -h, --help show this help message and exit - --preModel PREMODEL the name of pretrained embedding model - --preDict PREDICT the name of pretrained dictionary - --usrModel usrModel the name of output usr embedding model - --usrDict usrDict the name of user specified dictionary - -d DIM dimension of parameter -""" -from optparse import OptionParser -import struct - - -def get_row_index(preDict, usrDict): - """ - Get the row positions for all words in user dictionary from pre-trained dictionary. - return: a list of row positions - Example: preDict='a\nb\nc\n', usrDict='a\nc\n', then return [0,2] - """ - pos = [] - index = dict() - with open(preDict, "r") as f: - for line_index, line in enumerate(f): - word = line.strip().split()[0] - index[word] = line_index - with open(usrDict, "r") as f: - for line in f: - word = line.strip().split()[0] - pos.append(index[word]) - return pos - - -def extract_parameters_by_usrDict(preModel, preDict, usrModel, usrDict, - paraDim): - """ - Extract desired parameters from a pretrained embedding model based on user dictionary - """ - if paraDim not in [32, 64, 128, 256]: - raise RuntimeError("We only support 32, 64, 128, 256 dimensions now") - - fi = open(preModel, "rb") - fo = open(usrModel, "wb") - - # write filehead - rowIndex = get_row_index(preDict, usrDict) - newHead = struct.pack("iil", 0, 4, len(rowIndex) * paraDim) - fo.write(newHead) - bytes = 4 * paraDim - for i in range(0, len(rowIndex)): - # find the absolute position of input file - fi.seek(rowIndex[i] * bytes + 16, 0) - fo.write(fi.read(bytes)) - - print "extract parameters finish, total", len(rowIndex), "lines" - fi.close() - - -def main(): - """ - Main entry for running paraconvert.py - """ - usage = "usage: \n" \ - "python %prog --preModel PREMODEL --preDict PREDICT" \ - " --usrModel USRMODEL --usrDict USRDICT -d DIM" - parser = OptionParser(usage) - parser.add_option( - "--preModel", - action="store", - dest="preModel", - help="the name of pretrained embedding model") - parser.add_option( - "--preDict", - action="store", - dest="preDict", - help="the name of pretrained dictionary") - parser.add_option( - "--usrModel", - action="store", - dest="usrModel", - help="the name of output usr embedding model") - parser.add_option( - "--usrDict", - action="store", - dest="usrDict", - help="the name of user specified dictionary") - parser.add_option( - "-d", action="store", dest="dim", help="dimension of parameter") - (options, args) = parser.parse_args() - extract_parameters_by_usrDict(options.preModel, options.preDict, - options.usrModel, options.usrDict, - int(options.dim)) - - -if __name__ == '__main__': - main() diff --git a/v1_api_demo/model_zoo/embedding/paraconvert.py b/v1_api_demo/model_zoo/embedding/paraconvert.py deleted file mode 100755 index ce7a70efc43d7f85708f1e12bb94739f3588370c..0000000000000000000000000000000000000000 --- a/v1_api_demo/model_zoo/embedding/paraconvert.py +++ /dev/null @@ -1,159 +0,0 @@ -#!/bin/env python -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -""" -Example: - python paraconvert.py --b2t -i INPUT -o OUTPUT -d DIM - python paraconvert.py --t2b -i INPUT -o OUTPUT - -Options: - -h, --help show this help message and exit - --b2t convert parameter file of embedding model from binary to text - --t2b convert parameter file of embedding model from text to binary - -i INPUT input parameter file name - -o OUTPUT output parameter file name - -d DIM dimension of parameter -""" -from optparse import OptionParser -import struct - - -def binary2text(input, output, paraDim): - """ - Convert a binary parameter file of embedding model to be a text file. - input: the name of input binary parameter file, the format is: - 1) the first 16 bytes is filehead: - version(4 bytes): version of paddle, default = 0 - floatSize(4 bytes): sizeof(float) = 4 - paraCount(8 bytes): total number of parameter - 2) the next (paraCount * 4) bytes is parameters, each has 4 bytes - output: the name of output text parameter file, for example: - 0,4,32156096 - -0.7845433,1.1937413,-0.1704215,... - 0.0000909,0.0009465,-0.0008813,... - ... - the format is: - 1) the first line is filehead: - version=0, floatSize=4, paraCount=32156096 - 2) other lines print the paramters - a) each line prints paraDim paramters splitted by ',' - b) there is paraCount/paraDim lines (embedding words) - paraDim: dimension of parameters - """ - fi = open(input, "rb") - fo = open(output, "w") - """ - """ - version, floatSize, paraCount = struct.unpack("iil", fi.read(16)) - newHead = ','.join([str(version), str(floatSize), str(paraCount)]) - print >> fo, newHead - - bytes = 4 * int(paraDim) - format = "%df" % int(paraDim) - context = fi.read(bytes) - line = 0 - - while context: - numbers = struct.unpack(format, context) - lst = [] - for i in numbers: - lst.append('%8.7f' % i) - print >> fo, ','.join(lst) - context = fi.read(bytes) - line += 1 - fi.close() - fo.close() - print "binary2text finish, total", line, "lines" - - -def get_para_count(input): - """ - Compute the total number of embedding parameters in input text file. - input: the name of input text file - """ - numRows = 1 - paraDim = 0 - with open(input) as f: - line = f.readline() - paraDim = len(line.split(",")) - for line in f: - numRows += 1 - return numRows * paraDim - - -def text2binary(input, output, paddle_head=True): - """ - Convert a text parameter file of embedding model to be a binary file. - input: the name of input text parameter file, for example: - -0.7845433,1.1937413,-0.1704215,... - 0.0000909,0.0009465,-0.0008813,... - ... - the format is: - 1) it doesn't have filehead - 2) each line stores the same dimension of parameters, - the separator is commas ',' - output: the name of output binary parameter file, the format is: - 1) the first 16 bytes is filehead: - version(4 bytes), floatSize(4 bytes), paraCount(8 bytes) - 2) the next (paraCount * 4) bytes is parameters, each has 4 bytes - """ - fi = open(input, "r") - fo = open(output, "wb") - - newHead = struct.pack("iil", 0, 4, get_para_count(input)) - fo.write(newHead) - - count = 0 - for line in fi: - line = line.strip().split(",") - for i in range(0, len(line)): - binary_data = struct.pack("f", float(line[i])) - fo.write(binary_data) - count += 1 - fi.close() - fo.close() - print "text2binary finish, total", count, "lines" - - -def main(): - """ - Main entry for running paraconvert.py - """ - usage = "usage: \n" \ - "python %prog --b2t -i INPUT -o OUTPUT -d DIM \n" \ - "python %prog --t2b -i INPUT -o OUTPUT" - parser = OptionParser(usage) - parser.add_option( - "--b2t", - action="store_true", - help="convert parameter file of embedding model from binary to text") - parser.add_option( - "--t2b", - action="store_true", - help="convert parameter file of embedding model from text to binary") - parser.add_option( - "-i", action="store", dest="input", help="input parameter file name") - parser.add_option( - "-o", action="store", dest="output", help="output parameter file name") - parser.add_option( - "-d", action="store", dest="dim", help="dimension of parameter") - (options, args) = parser.parse_args() - if options.b2t: - binary2text(options.input, options.output, options.dim) - if options.t2b: - text2binary(options.input, options.output) - - -if __name__ == '__main__': - main() diff --git a/v1_api_demo/model_zoo/embedding/pre_DictAndModel.sh b/v1_api_demo/model_zoo/embedding/pre_DictAndModel.sh deleted file mode 100755 index f61c65a935c76032a06613cfe0b50f1c90bc50d9..0000000000000000000000000000000000000000 --- a/v1_api_demo/model_zoo/embedding/pre_DictAndModel.sh +++ /dev/null @@ -1,32 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -set -e -set -x -BASE_URL='http://paddlepaddle.cdn.bcebos.com/model_zoo/embedding' - -DOWNLOAD_ITEMS=(baidu.dict model_32.emb model_64.emb model_128.emb model_256.emb) -ITEM_MD5=(fa03a12321eaab6c30a8fcc9442eaea3 - f88c8325ee6da6187f1080e8fe66c1cd - 927cf70f27f860aff1a5703ebf7f1584 - a52e43655cd25d279777ed509a1ae27b - b92c67fe9ff70fea53596080e351ac80) - -for ((i=0; i<${#ITEM_MD5[@]}; i++)) -do - FILENAME=${DOWNLOAD_ITEMS[${i}]} - REAL_MD5=`wget ${BASE_URL}/${FILENAME} -O - | tee ${FILENAME} | md5sum | cut -d ' ' -f 1` - EXPECTED_MD5=${ITEM_MD5[${i}]} - [ "${EXPECTED_MD5}" = "${REAL_MD5}" ] -done diff --git a/v1_api_demo/model_zoo/resnet/.gitignore b/v1_api_demo/model_zoo/resnet/.gitignore deleted file mode 100644 index 7a64209b62340a5c5a51626821028e63ed5e588e..0000000000000000000000000000000000000000 --- a/v1_api_demo/model_zoo/resnet/.gitignore +++ /dev/null @@ -1,5 +0,0 @@ -fea_output/ -features/ -model.list -ResNet_50.dot -ResNet_50.png diff --git a/v1_api_demo/model_zoo/resnet/classify.py b/v1_api_demo/model_zoo/resnet/classify.py deleted file mode 100755 index 6074cc1d3a85e13e3e8d336d81e22104f9d8e7cf..0000000000000000000000000000000000000000 --- a/v1_api_demo/model_zoo/resnet/classify.py +++ /dev/null @@ -1,312 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import os -import sys -import cPickle -import logging -from PIL import Image -import numpy as np -from optparse import OptionParser - -import paddle.utils.image_util as image_util - -from py_paddle import swig_paddle, DataProviderConverter -from paddle.trainer.PyDataProvider2 import dense_vector -from paddle.trainer.config_parser import parse_config - -logging.basicConfig( - format='[%(levelname)s %(asctime)s %(filename)s:%(lineno)s] %(message)s') -logging.getLogger().setLevel(logging.INFO) - - -class ImageClassifier(): - def __init__(self, - train_conf, - model_dir=None, - resize_dim=256, - crop_dim=224, - use_gpu=True, - mean_file=None, - output_layer=None, - oversample=False, - is_color=True): - """ - train_conf: network configure. - model_dir: string, directory of model. - resize_dim: int, resized image size. - crop_dim: int, crop size. - mean_file: string, image mean file. - oversample: bool, oversample means multiple crops, namely five - patches (the four corner patches and the center - patch) as well as their horizontal reflections, - ten crops in all. - """ - self.train_conf = train_conf - self.model_dir = model_dir - if model_dir is None: - self.model_dir = os.path.dirname(train_conf) - - self.resize_dim = resize_dim - self.crop_dims = [crop_dim, crop_dim] - self.oversample = oversample - self.is_color = is_color - - self.output_layer = output_layer - if self.output_layer: - assert isinstance(self.output_layer, basestring) - self.output_layer = self.output_layer.split(",") - - self.transformer = image_util.ImageTransformer(is_color=is_color) - self.transformer.set_transpose((2, 0, 1)) - self.transformer.set_channel_swap((2, 1, 0)) - - self.mean_file = mean_file - if self.mean_file is not None: - mean = np.load(self.mean_file)['data_mean'] - mean = mean.reshape(3, self.crop_dims[0], self.crop_dims[1]) - self.transformer.set_mean(mean) # mean pixel - else: - # if you use three mean value, set like: - # this three mean value is calculated from ImageNet. - self.transformer.set_mean(np.array([103.939, 116.779, 123.68])) - - conf_args = "is_test=1,use_gpu=%d,is_predict=1" % (int(use_gpu)) - conf = parse_config(train_conf, conf_args) - swig_paddle.initPaddle("--use_gpu=%d" % (int(use_gpu))) - self.network = swig_paddle.GradientMachine.createFromConfigProto( - conf.model_config) - assert isinstance(self.network, swig_paddle.GradientMachine) - self.network.loadParameters(self.model_dir) - - data_size = 3 * self.crop_dims[0] * self.crop_dims[1] - slots = [dense_vector(data_size)] - self.converter = DataProviderConverter(slots) - - def get_data(self, img_path): - """ - 1. load image from img_path. - 2. resize or oversampling. - 3. transformer data: transpose, channel swap, sub mean. - return K x H x W ndarray. - - img_path: image path. - """ - image = image_util.load_image(img_path, self.is_color) - # Another way to extract oversampled features is that - # cropping and averaging from large feature map which is - # calculated by large size of image. - # This way reduces the computation. - if self.oversample: - # image_util.resize_image: short side is self.resize_dim - image = image_util.resize_image(image, self.resize_dim) - image = np.array(image) - input = np.zeros( - (1, image.shape[0], image.shape[1], 3), dtype=np.float32) - input[0] = image.astype(np.float32) - input = image_util.oversample(input, self.crop_dims) - else: - image = image.resize(self.crop_dims, Image.ANTIALIAS) - input = np.zeros( - (1, self.crop_dims[0], self.crop_dims[1], 3), dtype=np.float32) - input[0] = np.array(image).astype(np.float32) - - data_in = [] - for img in input: - img = self.transformer.transformer(img).flatten() - data_in.append([img.tolist()]) - # paddle input: [[[]],[[]],...], [[]] is one sample. - return data_in - - def forward(self, input_data): - """ - return output arguments which are the Outputs() in network configure. - - input_data: py_paddle input data. - call forward. - """ - in_arg = self.converter(input_data) - return self.network.forwardTest(in_arg) - - def forward(self, data, output_layer): - """ - return output arguments which are the Outputs() in network configure. - - input_data: py_paddle input data. - call forward. - """ - input = self.converter(data) - self.network.forwardTest(input) - output = self.network.getLayerOutputs(output_layer) - res = {} - if isinstance(output_layer, basestring): - output_layer = [output_layer] - for name in output_layer: - # For oversampling, average predictions across crops. - # If not, the shape of output[name]: (1, class_number), - # the mean is also applicable. - res[name] = output[name]['value'].mean(0) - - return res - - def predict(self, data_file): - """ - call forward and predicting. - - data_file: input image list. - """ - image_files = open(data_file, 'rb').readlines() - results = {} - if self.output_layer is None: - self.output_layer = ["output"] - for line in image_files: - image = line.split()[0] - data = self.get_data(image) - prob = self.forward(data, self.output_layer) - lab = np.argsort(-prob[self.output_layer[0]]) - results[image] = lab[0] - logging.info("Label of %s is: %d", image, lab[0]) - return results - - def extract(self, data_file, output_dir, batch_size=10000): - """ - extract and save features of output layers, which are - specify in Outputs() in network configure. - - data_file: file name of input data. - output_dir: saved directory of extracted features. - batch_size: sample number of one batch file. - """ - if not os.path.exists(output_dir): - os.mkdir(output_dir) - - sample_num = 0 - batch_num = 0 - image_feature = {} - image_files = open(data_file, 'rb').readlines() - for idx, line in enumerate(image_files): - image = line.split()[0] - data = self.get_data(image) - feature = self.forward(data, self.output_layer) - # save extracted features - file_name = image.split("/")[-1] - image_feature[file_name] = feature - sample_num += 1 - if sample_num == batch_size: - batch_name = os.path.join(output_dir, 'batch_%d' % (batch_num)) - self.save_file(image_feature, batch_name) - logging.info('Finish batch %d', batch_num) - batch_num += 1 - sample_num = 0 - image_feature = {} - if idx % 1000 == 0: - logging.info('%d/%d, %s', idx, len(image_files), file_name) - if sample_num > 0: - batch_name = os.path.join(output_dir, 'batch_%d' % (batch_num)) - self.save_file(image_feature, batch_name) - logging.info('Finish batch %d', batch_num) - logging.info('Done: make image feature batch') - - def save_file(self, data, file): - of = open(file, 'wb') - cPickle.dump(data, of, protocol=cPickle.HIGHEST_PROTOCOL) - - -def option_parser(): - """ - Main entry for predciting - """ - usage = "%prog -c config -i data_list -w model_dir [options]" - parser = OptionParser(usage="usage: %s" % usage) - parser.add_option( - "-j", - "--job", - action="store", - dest="job_type", - help="job type: predict, extract\ - predict: predicting,\ - extract: extract features") - parser.add_option( - "-c", - "--conf", - action="store", - dest="train_conf", - help="network config") - parser.add_option( - "-i", "--data", action="store", dest="data_file", help="image list") - parser.add_option( - "-w", - "--model", - action="store", - dest="model_path", - default=None, - help="model path") - parser.add_option( - "-g", - "--use_gpu", - action="store", - dest="use_gpu", - default=True, - help="Whether to use gpu mode.") - parser.add_option( - "-o", - "--output_dir", - action="store", - dest="output_dir", - default="output", - help="output path") - parser.add_option( - "-m", - "--mean", - action="store", - dest="mean", - default=None, - help="mean file.") - parser.add_option( - "-p", - "--multi_crop", - action="store_true", - dest="multi_crop", - default=False, - help="Wether to use multiple crops on image.") - parser.add_option("-l", "--output_layer", action="store", - dest="output_layer", default=None, - help="--job=extract, specify layers to extract "\ - "features, --job=predict, specify layer of " - "classification probability, output in resnet.py.") - return parser.parse_args() - - -def main(): - """ - 1. parse input arguments. - 2. predicting or extract features according job type. - """ - options, args = option_parser() - obj = ImageClassifier( - options.train_conf, - options.model_path, - use_gpu=options.use_gpu, - mean_file=options.mean, - output_layer=options.output_layer, - oversample=options.multi_crop) - if options.job_type == "predict": - obj.predict(options.data_file) - - elif options.job_type == "extract": - obj.extract(options.data_file, options.output_dir) - - -if __name__ == '__main__': - main() diff --git a/v1_api_demo/model_zoo/resnet/example/.gitignore b/v1_api_demo/model_zoo/resnet/example/.gitignore deleted file mode 100644 index 4a2b5962a6800f251cba655c026331f14648c86e..0000000000000000000000000000000000000000 --- a/v1_api_demo/model_zoo/resnet/example/.gitignore +++ /dev/null @@ -1 +0,0 @@ -*image_list_provider_copy_1.py diff --git a/v1_api_demo/model_zoo/resnet/example/__init__.py b/v1_api_demo/model_zoo/resnet/example/__init__.py deleted file mode 100644 index f662d6826321eb840739382558f76327d27b5847..0000000000000000000000000000000000000000 --- a/v1_api_demo/model_zoo/resnet/example/__init__.py +++ /dev/null @@ -1,13 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. diff --git a/v1_api_demo/model_zoo/resnet/example/cat.jpg b/v1_api_demo/model_zoo/resnet/example/cat.jpg deleted file mode 100644 index 47b01db90eddc46ff845f10bc2accaf2364c272d..0000000000000000000000000000000000000000 Binary files a/v1_api_demo/model_zoo/resnet/example/cat.jpg and /dev/null differ diff --git a/v1_api_demo/model_zoo/resnet/example/dog.jpg b/v1_api_demo/model_zoo/resnet/example/dog.jpg deleted file mode 100644 index b9cc33cf069da5c453b97dbb7383838edd07c199..0000000000000000000000000000000000000000 Binary files a/v1_api_demo/model_zoo/resnet/example/dog.jpg and /dev/null differ diff --git a/v1_api_demo/model_zoo/resnet/example/image_list_provider.py b/v1_api_demo/model_zoo/resnet/example/image_list_provider.py deleted file mode 100644 index 2cd8eb8bf850f41282ed5db2885dc0b7218c79f7..0000000000000000000000000000000000000000 --- a/v1_api_demo/model_zoo/resnet/example/image_list_provider.py +++ /dev/null @@ -1,102 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.utils.image_util import * -from paddle.trainer.PyDataProvider2 import * - - -def hook(settings, image_size, crop_size, color, file_list, is_train, **kwargs): - """ - Description: Init with a list of data file - file_list is the name list of input files. - kwargs["load_data_args"] is the value of 'load_data_args' - which can be set in config. - Each args is separated by a column. - image_size: the crop image size. - mean_meta: the path of the meta file to store the mean image. - mean_value: can be mean value, not a file. - can not set mean_meta and mean_value at the same time. - color: 'color' means a color image. Otherwise, it means a gray image. - is_train: whether the data provider is used for training. - Data argumentation might be different for training and testing. - """ - settings.img_size = image_size - settings.crop_size = crop_size - settings.mean_img_size = settings.crop_size - settings.color = color # default is color - settings.is_train = is_train - - settings.is_swap_channel = kwargs.get('swap_channel', None) - if settings.is_swap_channel is not None: - settings.swap_channel = settings.is_swap_channel - settings.is_swap_channel = True - - if settings.color: - settings.img_input_size = settings.crop_size * settings.crop_size * 3 - else: - settings.img_input_size = settings.crop_size * settings.crop_size - - settings.file_list = file_list - settings.mean_meta = kwargs.get('mean_meta', None) - settings.mean_value = kwargs.get('mean_value', None) - # can not specify both mean_meta and mean_value. - assert not (settings.mean_meta and settings.mean_value) - if not settings.mean_meta: - settings.mean_value = kwargs.get('mean_value') - sz = settings.crop_size * settings.crop_size - settings.img_mean = np.zeros(sz * 3, dtype=np.single) - for idx, value in enumerate(settings.mean_value): - settings.img_mean[idx * sz:(idx + 1) * sz] = value - settings.img_mean = settings.img_mean.reshape(3, settings.crop_size, - settings.crop_size) - - else: - settings.img_mean = load_meta(settings.mean_meta, - settings.mean_img_size, - settings.crop_size, settings.color) - - settings.input_types = [ - dense_vector(settings.img_input_size), # image feature - integer_value(1) - ] # labels - - settings.logger.info('Image short side: %s', settings.img_size) - settings.logger.info('Crop size: %s', settings.crop_size) - settings.logger.info('Meta path: %s', settings.mean_meta) - if settings.is_swap_channel: - settings.logger.info('swap channel: %s', settings.swap_channel) - settings.logger.info('DataProvider Initialization finished') - - -@provider(init_hook=hook, should_shuffle=False) -def processData(settings, file_list): - """ - The main function for loading data. - Load the batch, iterate all the images and labels in this batch. - file_name: the batch file name. - """ - img_path, lab = file_list.strip().split(' ') - img = Image.open(img_path) - img.load() - img = img.resize((settings.img_size, settings.img_size), Image.ANTIALIAS) - img = np.array(img).astype(np.float32) - if len(img.shape) == 3: - img = np.swapaxes(img, 1, 2) - img = np.swapaxes(img, 1, 0) - # swap channel - if settings.is_swap_channel: - img = img[settings.swap_channel, :, :] - img_feat = preprocess_img(img, settings.img_mean, settings.crop_size, - settings.is_train, settings.color) - yield img_feat.tolist(), int(lab.strip()) diff --git a/v1_api_demo/model_zoo/resnet/example/test.list b/v1_api_demo/model_zoo/resnet/example/test.list deleted file mode 100644 index 30bbf630b640a26239fc104c9c08f6ebc9dfaa82..0000000000000000000000000000000000000000 --- a/v1_api_demo/model_zoo/resnet/example/test.list +++ /dev/null @@ -1,2 +0,0 @@ -example/dog.jpg 0 -example/cat.jpg 0 diff --git a/v1_api_demo/model_zoo/resnet/extract_fea_c++.sh b/v1_api_demo/model_zoo/resnet/extract_fea_c++.sh deleted file mode 100755 index 5447aa92dfb5facd3433eb4a1893e96e3c786c73..0000000000000000000000000000000000000000 --- a/v1_api_demo/model_zoo/resnet/extract_fea_c++.sh +++ /dev/null @@ -1,40 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -set -e - -#set names of layer which you want to extract feature -#in Outputs() of resnet.py -#like: Outputs("res5_3_branch2c_conv", "res5_3_branch2c_bn") -layer_num=50 -configure=./resnet.py -model_path=./model/resnet_$layer_num -fea_dir=fea_output -#Output is text file. -#Each line is one sample's features. -#If you set N layer names in Outputs() -#each line contains N features sperated by ";". - -# create model list file. -model_list=./model.list -touch $model_list | echo $model_path > $model_list - -paddle train \ - --local=true \ - --job=test \ - --config=$configure \ - --model_list=$model_list \ - --use_gpu=1 \ - --predict_output_dir=$fea_dir \ - --config_args=is_test=1,layer_num=$layer_num diff --git a/v1_api_demo/model_zoo/resnet/extract_fea_py.sh b/v1_api_demo/model_zoo/resnet/extract_fea_py.sh deleted file mode 100755 index 2e87152f7f8598f487870291271cdee646105044..0000000000000000000000000000000000000000 --- a/v1_api_demo/model_zoo/resnet/extract_fea_py.sh +++ /dev/null @@ -1,29 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -set -e - -#Note if you use CPU mode, you need to set use_gpu=0 in classify.py. like this: -#conf_args = "is_test=0,use_gpu=1,is_predict=1" -#conf = parse_config(train_conf, conf_args) -#swig_paddle.initPaddle("--use_gpu=0") -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 diff --git a/v1_api_demo/model_zoo/resnet/get_model.sh b/v1_api_demo/model_zoo/resnet/get_model.sh deleted file mode 100755 index b33d8178ab7859fc0b0d514fb19bec2c28a77c3d..0000000000000000000000000000000000000000 --- a/v1_api_demo/model_zoo/resnet/get_model.sh +++ /dev/null @@ -1,32 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -set -e - -DIR="$( cd "$(dirname "$0")" ; pwd -P )" -cd $DIR - -mkdir model -cd model - -echo "Downloading ResNet models..." - -for file in resnet_50.tar.gz resnet_101.tar.gz resnet_152.tar.gz mean_meta_224.tar.gz -do - wget http://paddlepaddle.bj.bcebos.com/model_zoo/imagenet/$file - tar -xvf $file - rm $file -done - -echo "Done." diff --git a/v1_api_demo/model_zoo/resnet/load_feature.py b/v1_api_demo/model_zoo/resnet/load_feature.py deleted file mode 100644 index 5d3d0c0d30ef710c37c98e93a51b2f813d636b59..0000000000000000000000000000000000000000 --- a/v1_api_demo/model_zoo/resnet/load_feature.py +++ /dev/null @@ -1,63 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import os -import sys -import cPickle -import logging - -logging.basicConfig( - format='[%(levelname)s %(asctime)s %(filename)s:%(lineno)s] %(message)s') -logging.getLogger().setLevel(logging.INFO) - - -def load_feature_c(file): - """ - Load feature extracted by C++ interface. - Return a list. - file: feature file. - """ - features = [] - f = open(file, 'r') - for line in f: - sample = [] - for slot in line.strip().split(";"): - fea = [float(val) for val in slot.strip().split()] - if fea: - sample.append(fea) - features.append(sample) - f.close() - return features - - -def load_feature_py(feature_dir): - """ - Load feature extracted by python interface. - Return a dictionary. - feature_dir: directory of feature file. - """ - file_list = os.listdir(feature_dir) - file_list = [os.path.join(feature_dir, f) for f in file_list] - features = {} - for file_name in file_list: - with open(file_name, 'rb') as f: - feature = cPickle.load(f) - features.update(feature) - logging.info('Load feature file %s', file_name) - return features - - -if __name__ == '__main__': - print load_feature_py(sys.argv[1]) - #print load_feature_c(sys.argv[1]) diff --git a/v1_api_demo/model_zoo/resnet/net_diagram.sh b/v1_api_demo/model_zoo/resnet/net_diagram.sh deleted file mode 100755 index 1b06ffa44eec8a0f312420c35699d3902f9a6400..0000000000000000000000000000000000000000 --- a/v1_api_demo/model_zoo/resnet/net_diagram.sh +++ /dev/null @@ -1,39 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -:' -Visual deep residual network -1. Using make_model_diagram.py to generate dot file. -2. Using graphviz to convert dot file. - -Usage: -./net_diagram.sh -' - -set -e - -DIR="$( cd "$(dirname "$0")" ; pwd -P )" -cd $DIR - -img_type=png -img_fileprefix=ResNet_50 -conf_filename=resnet.py -dot_filename=ResNet_50.dot -config_str="layer_num=50,data_provider=0" - -python -m paddle.utils.make_model_diagram $conf_filename $dot_filename $config_str - -# If you have installed graphviz, running like this: -# dot -Tpng -o ResNet.png ResNet.dot diff --git a/v1_api_demo/model_zoo/resnet/predict.sh b/v1_api_demo/model_zoo/resnet/predict.sh deleted file mode 100755 index 2b67b17c48c60cc8a7b7c46a1c80a3f2bf281870..0000000000000000000000000000000000000000 --- a/v1_api_demo/model_zoo/resnet/predict.sh +++ /dev/null @@ -1,23 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -set -e - -python classify.py \ - --job=predict \ - --conf=resnet.py\ - --model=model/resnet_50 \ - --multi_crop \ - --use_gpu=1 \ - --data=./example/test.list diff --git a/v1_api_demo/model_zoo/resnet/resnet.py b/v1_api_demo/model_zoo/resnet/resnet.py deleted file mode 100644 index 6fdd97fefc62392c93ecffae0fc918e8dc4b18c5..0000000000000000000000000000000000000000 --- a/v1_api_demo/model_zoo/resnet/resnet.py +++ /dev/null @@ -1,271 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * -""" -paper: https://arxiv.org/abs/1512.03385 -""" -is_test = get_config_arg("is_test", bool, False) -is_predict = get_config_arg("is_predict", bool, False) -data_provider = get_config_arg("data_provider", bool, True) -layer_num = get_config_arg("layer_num", int, 50) - -if not is_predict and data_provider: - 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) - -batch_size = 1 -learning_rate = 0.1 / batch_size -momentum = 0.9 -weight_decay = 0.0001 * batch_size -default_momentum(momentum) -default_decay_rate(weight_decay) - -Settings( - algorithm='sgd', - batch_size=batch_size, - learning_rate=learning_rate, - - # set the appropriate parameters according your schedule - learning_method='momentum', - learning_rate_decay_a=0.5, - learning_rate_decay_b=1200000 * 10, - learning_rate_schedule="discexp", ) - - -def conv_bn_layer(name, - input, - filter_size, - num_filters, - stride, - padding, - channels=None, - active_type=ReluActivation()): - """ - A wrapper for conv layer with batch normalization layers. - Note: - conv layer has no activation. - """ - - tmp = img_conv_layer( - name=name + "_conv", - input=input, - filter_size=filter_size, - num_channels=channels, - num_filters=num_filters, - stride=stride, - padding=padding, - act=LinearActivation(), - bias_attr=False) - return batch_norm_layer( - name=name + "_bn", input=tmp, act=active_type, use_global_stats=is_test) - - -def bottleneck_block(name, input, num_filters1, num_filters2): - """ - A wrapper for bottlenect building block in ResNet. - Last conv_bn_layer has no activation. - Addto layer has activation of relu. - """ - last_name = conv_bn_layer( - name=name + '_branch2a', - input=input, - filter_size=1, - num_filters=num_filters1, - stride=1, - padding=0) - last_name = conv_bn_layer( - name=name + '_branch2b', - input=last_name, - filter_size=3, - num_filters=num_filters1, - stride=1, - padding=1) - last_name = conv_bn_layer( - name=name + '_branch2c', - input=last_name, - filter_size=1, - num_filters=num_filters2, - stride=1, - padding=0, - active_type=LinearActivation()) - - return addto_layer( - name=name + "_addto", input=[input, last_name], act=ReluActivation()) - - -def mid_projection(name, input, num_filters1, num_filters2, stride=2): - """ - A wrapper for middile projection in ResNet. - projection shortcuts are used for increasing dimensions, - and other shortcuts are identity - branch1: projection shortcuts are used for increasing - dimensions, has no activation. - branch2x: bottleneck building block, shortcuts are identity. - """ - # stride = 2 - branch1 = conv_bn_layer( - name=name + '_branch1', - input=input, - filter_size=1, - num_filters=num_filters2, - stride=stride, - padding=0, - active_type=LinearActivation()) - - last_name = conv_bn_layer( - name=name + '_branch2a', - input=input, - filter_size=1, - num_filters=num_filters1, - stride=stride, - padding=0) - last_name = conv_bn_layer( - name=name + '_branch2b', - input=last_name, - filter_size=3, - num_filters=num_filters1, - stride=1, - padding=1) - - last_name = conv_bn_layer( - name=name + '_branch2c', - input=last_name, - filter_size=1, - num_filters=num_filters2, - stride=1, - padding=0, - active_type=LinearActivation()) - - return addto_layer( - name=name + "_addto", input=[branch1, last_name], act=ReluActivation()) - - -def deep_res_net(res2_num=3, res3_num=4, res4_num=6, res5_num=3): - """ - A wrapper for 50,101,152 layers of ResNet. - res2_num: number of blocks stacked in conv2_x - res3_num: number of blocks stacked in conv3_x - res4_num: number of blocks stacked in conv4_x - res5_num: number of blocks stacked in conv5_x - """ - # For ImageNet - # conv1: 112x112 - img = data_layer(name='input', size=224 * 224 * 3) - tmp = conv_bn_layer( - "conv1", - img, - filter_size=7, - channels=3, - num_filters=64, - stride=2, - padding=3) - tmp = img_pool_layer(name="pool1", input=tmp, pool_size=3, stride=2) - - # conv2_x: 56x56 - tmp = mid_projection( - name="res2_1", input=tmp, num_filters1=64, num_filters2=256, stride=1) - for i in xrange(2, res2_num + 1, 1): - tmp = bottleneck_block( - name="res2_" + str(i), input=tmp, num_filters1=64, num_filters2=256) - - # conv3_x: 28x28 - tmp = mid_projection( - name="res3_1", input=tmp, num_filters1=128, num_filters2=512) - for i in xrange(2, res3_num + 1, 1): - tmp = bottleneck_block( - name="res3_" + str(i), - input=tmp, - num_filters1=128, - num_filters2=512) - - # conv4_x: 14x14 - tmp = mid_projection( - name="res4_1", input=tmp, num_filters1=256, num_filters2=1024) - for i in xrange(2, res4_num + 1, 1): - tmp = bottleneck_block( - name="res4_" + str(i), - input=tmp, - num_filters1=256, - num_filters2=1024) - - # conv5_x: 7x7 - tmp = mid_projection( - name="res5_1", input=tmp, num_filters1=512, num_filters2=2048) - for i in xrange(2, res5_num + 1, 1): - tmp = bottleneck_block( - name="res5_" + str(i), - input=tmp, - num_filters1=512, - num_filters2=2048) - - tmp = img_pool_layer( - name='avgpool', - input=tmp, - pool_size=7, - stride=1, - pool_type=AvgPooling()) - - output = fc_layer( - name='output', input=tmp, size=1000, act=SoftmaxActivation()) - - if not is_predict: - classification_cost( - input=output, label=data_layer( - name='label', size=1)) - - -def res_net_50(): - deep_res_net(3, 4, 6, 3) - - -def res_net_101(): - deep_res_net(3, 4, 23, 3) - - -def res_net_152(): - deep_res_net(3, 8, 36, 3) - - -if not is_predict: - Inputs("input", "label") -else: - Inputs("input") -# Outputs("cost-softmax" if not is_predict else "output") -Outputs("res5_3_branch2c_conv", "res5_3_branch2c_bn") - -if layer_num == 50: - res_net_50() -elif layer_num == 101: - res_net_101() -elif layer_num == 152: - res_net_152() -else: - print("Wrong layer number.") diff --git a/v1_api_demo/quick_start/.gitignore b/v1_api_demo/quick_start/.gitignore deleted file mode 100644 index f71662563ff96d6227dd568d9951a90b0d09456e..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/.gitignore +++ /dev/null @@ -1,15 +0,0 @@ -*.pyc -data/dict.txt -data/dict_all.txt -data/labels.list -data/mosesdecoder-master/ -data/reviews_Electronics_5.json.gz -data/test.list -data/test.txt -data/train.list -data/train.txt -data/pred.list -data/pred.txt -dataprovider_copy_1.py -train.log -output diff --git a/v1_api_demo/quick_start/api_predict.py b/v1_api_demo/quick_start/api_predict.py deleted file mode 100755 index 9bdffe1006281c58a595e2771561ba62e4c2d6bd..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/api_predict.py +++ /dev/null @@ -1,147 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import os, sys -import numpy as np -from optparse import OptionParser -from py_paddle import swig_paddle, DataProviderConverter -from paddle.trainer.PyDataProvider2 import sparse_binary_vector -from paddle.trainer.config_parser import parse_config -""" -Usage: run following command to show help message. - python api_predict.py -h -""" - - -class QuickStartPrediction(): - def __init__(self, train_conf, dict_file, model_dir=None, label_file=None): - """ - train_conf: trainer configure. - dict_file: word dictionary file name. - model_dir: directory of model. - """ - self.train_conf = train_conf - self.dict_file = dict_file - self.word_dict = {} - self.dict_dim = self.load_dict() - self.model_dir = model_dir - if model_dir is None: - self.model_dir = os.path.dirname(train_conf) - - self.label = None - if label_file is not None: - self.load_label(label_file) - - conf = parse_config(train_conf, "is_predict=1") - self.network = swig_paddle.GradientMachine.createFromConfigProto( - conf.model_config) - self.network.loadParameters(self.model_dir) - input_types = [sparse_binary_vector(self.dict_dim)] - self.converter = DataProviderConverter(input_types) - - def load_dict(self): - """ - Load dictionary from self.dict_file. - """ - for line_count, line in enumerate(open(self.dict_file, 'r')): - self.word_dict[line.strip().split('\t')[0]] = line_count - return len(self.word_dict) - - def load_label(self, label_file): - """ - Load label. - """ - self.label = {} - for v in open(label_file, 'r'): - self.label[int(v.split('\t')[1])] = v.split('\t')[0] - - def get_index(self, data): - """ - transform word into integer index according to the dictionary. - """ - words = data.strip().split() - word_slot = [self.word_dict[w] for w in words if w in self.word_dict] - return word_slot - - def batch_predict(self, data_batch): - input = self.converter(data_batch) - output = self.network.forwardTest(input) - prob = output[0]["id"].tolist() - print("predicting labels is:") - print prob - - -def option_parser(): - usage = "python predict.py -n config -w model_dir -d dictionary -i input_file " - parser = OptionParser(usage="usage: %s [options]" % usage) - parser.add_option( - "-n", - "--tconf", - action="store", - dest="train_conf", - help="network config") - parser.add_option( - "-d", - "--dict", - action="store", - dest="dict_file", - help="dictionary file") - parser.add_option( - "-b", - "--label", - action="store", - dest="label", - default=None, - help="dictionary file") - parser.add_option( - "-c", - "--batch_size", - type="int", - action="store", - dest="batch_size", - default=1, - help="the batch size for prediction") - parser.add_option( - "-w", - "--model", - action="store", - dest="model_path", - default=None, - help="model path") - return parser.parse_args() - - -def main(): - options, args = option_parser() - train_conf = options.train_conf - batch_size = options.batch_size - dict_file = options.dict_file - model_path = options.model_path - label = options.label - swig_paddle.initPaddle("--use_gpu=0") - predict = QuickStartPrediction(train_conf, dict_file, model_path, label) - - batch = [] - labels = [] - for line in sys.stdin: - [label, text] = line.split("\t") - labels.append(int(label)) - batch.append([predict.get_index(text)]) - print("labels is:") - print labels - predict.batch_predict(batch) - - -if __name__ == '__main__': - main() diff --git a/v1_api_demo/quick_start/api_predict.sh b/v1_api_demo/quick_start/api_predict.sh deleted file mode 100755 index 4d9aa9e8854ed79446a47dbc593f419cdda077b4..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/api_predict.sh +++ /dev/null @@ -1,30 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -set -e - -#Note the default model is pass-00002, you shold make sure the model path -#exists or change the mode path. -#only test on trainer_config.lr.py -model=output/model/pass-00001/ -config=trainer_config.lr.py -label=data/labels.list -dict=data/dict.txt -batch_size=20 -head -n$batch_size data/test.txt | python api_predict.py \ - --tconf=$config\ - --model=$model \ - --label=$label \ - --dict=$dict \ - --batch_size=$batch_size diff --git a/v1_api_demo/quick_start/api_train.py b/v1_api_demo/quick_start/api_train.py deleted file mode 100644 index 5699789daa4051661b0a72c69f4668f2d8bb9cb2..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/api_train.py +++ /dev/null @@ -1,122 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import argparse -import itertools -import random - -from paddle.trainer.config_parser import parse_config -from py_paddle import swig_paddle as api -from py_paddle import DataProviderConverter -from paddle.trainer.PyDataProvider2 \ - import integer_value, integer_value_sequence, sparse_binary_vector - - -def parse_arguments(): - parser = argparse.ArgumentParser() - parser.add_argument( - "--train_data", type=str, required=False, help="train data file") - parser.add_argument("--test_data", type=str, help="test data file") - parser.add_argument( - "--config", type=str, required=True, help="config file name") - parser.add_argument("--dict_file", required=True, help="dictionary file") - parser.add_argument( - "--seq", default=1, type=int, help="whether use sequence training") - parser.add_argument( - "--use_gpu", default=0, type=int, help="whether use GPU for training") - parser.add_argument( - "--trainer_count", - default=1, - type=int, - help="Number of threads for training") - parser.add_argument( - "--num_passes", default=5, type=int, help="Number of training passes") - return parser.parse_args() - - -UNK_IDX = 0 - - -def load_data(file_name, word_dict): - with open(file_name, 'r') as f: - for line in f: - label, comment = line.strip().split('\t') - words = comment.split() - word_slot = [word_dict.get(w, UNK_IDX) for w in words] - yield word_slot, int(label) - - -def load_dict(dict_file): - word_dict = dict() - with open(dict_file, 'r') as f: - for i, line in enumerate(f): - w = line.strip().split()[0] - word_dict[w] = i - return word_dict - - -def main(): - options = parse_arguments() - api.initPaddle("--use_gpu=%s" % options.use_gpu, - "--trainer_count=%s" % options.trainer_count) - - word_dict = load_dict(options.dict_file) - train_dataset = list(load_data(options.train_data, word_dict)) - if options.test_data: - test_dataset = list(load_data(options.test_data, word_dict)) - else: - test_dataset = None - - trainer_config = parse_config(options.config, - "dict_file=%s" % options.dict_file) - # No need to have data provider for trainer - trainer_config.ClearField('data_config') - trainer_config.ClearField('test_data_config') - - # create a GradientMachine from the model configuratin - model = api.GradientMachine.createFromConfigProto( - trainer_config.model_config) - # create a trainer for the gradient machine - trainer = api.Trainer.create(trainer_config, model) - - # create a data converter which converts data to PaddlePaddle - # internal format - input_types = [ - integer_value_sequence(len(word_dict)) if options.seq else - sparse_binary_vector(len(word_dict)), integer_value(2) - ] - converter = DataProviderConverter(input_types) - - batch_size = trainer_config.opt_config.batch_size - trainer.startTrain() - for train_pass in xrange(options.num_passes): - trainer.startTrainPass() - random.shuffle(train_dataset) - for pos in xrange(0, len(train_dataset), batch_size): - batch = itertools.islice(train_dataset, pos, pos + batch_size) - size = min(batch_size, len(train_dataset) - pos) - trainer.trainOneDataBatch(size, converter(batch)) - trainer.finishTrainPass() - if test_dataset: - trainer.startTestPeriod() - for pos in xrange(0, len(test_dataset), batch_size): - batch = itertools.islice(test_dataset, pos, pos + batch_size) - size = min(batch_size, len(test_dataset) - pos) - trainer.testOneDataBatch(size, converter(batch)) - trainer.finishTestPeriod() - trainer.finishTrain() - - -if __name__ == '__main__': - main() diff --git a/v1_api_demo/quick_start/api_train.sh b/v1_api_demo/quick_start/api_train.sh deleted file mode 100755 index 9b2a4e2f224b1677c458ede66a6a3bac09d8ad61..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/api_train.sh +++ /dev/null @@ -1,29 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -set -e - -# Note: if using trainer_config.emb.py, trainer_config.cnn.py -# or trainer_config.lstm.py, you need to change --seq to --seq=1 -# because they are sequence models. -python api_train.py \ - --config=trainer_config.lr.py \ - --trainer_count=2 \ - --num_passes=15 \ - --use_gpu=0 \ - --seq=0 \ - --train_data=data/train.txt \ - --test_data=data/test.txt \ - --dict_file=data/dict.txt \ - 2>&1 | tee 'train.log' diff --git a/v1_api_demo/quick_start/cluster/cluster_train.sh b/v1_api_demo/quick_start/cluster/cluster_train.sh deleted file mode 100755 index a7b1f01064b29cf6abc4cd6b706ee466a6d6da36..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/cluster/cluster_train.sh +++ /dev/null @@ -1,45 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -set -e - -# Should run pserver.sh before run this script. -bin_dir=$(cd `dirname $0`; pwd) -home_dir=$(cd "${bin_dir}/.."; pwd) -source "$bin_dir/env.sh" - -model_dir="$bin_dir/output" -log_file="$bin_dir/train.log" - -pushd "$home_dir" -cfg=trainer_config.lr.py -paddle train \ - --start_pserver=false \ - --config=$cfg \ - --save_dir=${model_dir} \ - --trainer_count=4 \ - --local=0 \ - --log_period=100 \ - --num_passes=15 \ - --use_gpu=false \ - --show_parameter_stats_period=100 \ - --test_all_data_in_one_period=1 \ - --num_gradient_servers=1 \ - --nics=`get_nics` \ - --port=7164 \ - --ports_num=1 \ - --pservers="127.0.0.1" \ - --comment="paddle_trainer" \ - 2>&1 | tee "$log_file" -popd diff --git a/v1_api_demo/quick_start/cluster/env.sh b/v1_api_demo/quick_start/cluster/env.sh deleted file mode 100644 index a404993835d0e479f65c89c5561855293b7b66f0..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/cluster/env.sh +++ /dev/null @@ -1,28 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -set -e - -function get_nics() { - machine=`uname -s` - local nics="" - if [ "$machine" == "Linux" ]; then - nics="lo" - elif [ "$machine" == "Darwin" ]; then - nics="lo0" - else - nics="unsupport" - fi - echo $nics -} diff --git a/v1_api_demo/quick_start/cluster/pserver.sh b/v1_api_demo/quick_start/cluster/pserver.sh deleted file mode 100755 index b187c1d9b9108a607ed310253d54ecc096f0e792..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/cluster/pserver.sh +++ /dev/null @@ -1,26 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -set -e -bin_dir=$(cd `dirname $0`; pwd) -source "$bin_dir/env.sh" - -paddle pserver \ - --nics=`get_nics` \ - --port=7164 \ - --ports_num=1 \ - --ports_num_for_sparse=1 \ - --num_gradient_servers=1 \ - --comment="paddle_pserver" \ - 2>&1 | tee 'pserver.log' diff --git a/v1_api_demo/quick_start/data/README.md b/v1_api_demo/quick_start/data/README.md deleted file mode 100644 index 63abcf7ebf31903213e44cf492b93e09f61db14e..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/data/README.md +++ /dev/null @@ -1,9 +0,0 @@ -This dataset consists of electronics product reviews associated with -binary labels (positive/negative) for sentiment classification. - -The preprocessed data can be downloaded by script `get_data.sh`. -The data was derived from reviews_Electronics_5.json.gz at - -http://snap.stanford.edu/data/amazon/productGraph/categoryFiles/reviews_Electronics_5.json.gz - -If you want to process the raw data, you can use the script `proc_from_raw_data/get_data.sh`. diff --git a/v1_api_demo/quick_start/data/get_data.sh b/v1_api_demo/quick_start/data/get_data.sh deleted file mode 100755 index a09a18f919e5a84f1f7c889a43f0a5fbf4a60a77..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/data/get_data.sh +++ /dev/null @@ -1,27 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -set -e - -DIR="$( cd "$(dirname "$0")" ; pwd -P )" -cd $DIR - -# Download the preprocessed data -wget http://paddlepaddle.bj.bcebos.com/demo/quick_start_preprocessed_data/preprocessed_data.tar.gz - -# Extract package -tar zxvf preprocessed_data.tar.gz - -# Remove compressed package -rm preprocessed_data.tar.gz diff --git a/v1_api_demo/quick_start/data/proc_from_raw_data/get_data.sh b/v1_api_demo/quick_start/data/proc_from_raw_data/get_data.sh deleted file mode 100755 index d976eaebfaa600778e0ab6bb0adbd7159f1cce2f..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/data/proc_from_raw_data/get_data.sh +++ /dev/null @@ -1,79 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -# 1. size of pos : neg = 1:1. -# 2. size of testing set = min(25k, len(all_data) * 0.1), others is traning set. -# 3. distinct train set and test set. - -set -e - -DIR="$( cd "$(dirname "$0")" ; pwd -P )" -cd $DIR - -# Download data -echo "Downloading Amazon Electronics reviews data..." -# http://jmcauley.ucsd.edu/data/amazon/ -wget http://snap.stanford.edu/data/amazon/productGraph/categoryFiles/reviews_Electronics_5.json.gz -echo "Downloading mosesdecoder..." -# https://github.com/moses-smt/mosesdecoder -wget https://github.com/moses-smt/mosesdecoder/archive/master.zip - -unzip master.zip -rm master.zip - -################## -# Preprocess data -echo "Preprocess data..." -export LC_ALL=C -UNAME_STR=`uname` - -if [ ${UNAME_STR} == 'Linux' ]; then - SHUF_PROG='shuf' -else - SHUF_PROG='gshuf' -fi - -mkdir -p tmp -python preprocess.py -i reviews_Electronics_5.json.gz -# uniq and shuffle -cd tmp -echo 'Uniq and shuffle...' -cat pos_*|sort|uniq|${SHUF_PROG}> pos.shuffed -cat neg_*|sort|uniq|${SHUF_PROG}> neg.shuffed - -min_len=`sed -n '$=' neg.shuffed` -test_num=$((min_len/10)) -if [ $test_num -gt 12500 ];then - test_num=12500 -fi -train_num=$((min_len-test_num)) - -head -n$train_num pos.shuffed >train.pos -head -n$train_num neg.shuffed >train.neg -tail -n$test_num pos.shuffed >test.pos -tail -n$test_num neg.shuffed >test.neg - -cat train.pos train.neg | ${SHUF_PROG} >../train.txt -cat test.pos test.neg | ${SHUF_PROG} >../test.txt - -cd - -echo 'train.txt' > train.list -echo 'test.txt' > test.list - -# use 30k dict -rm -rf tmp -mv dict.txt dict_all.txt -cat dict_all.txt | head -n 30001 > dict.txt -echo 'Done.' diff --git a/v1_api_demo/quick_start/data/proc_from_raw_data/preprocess.py b/v1_api_demo/quick_start/data/proc_from_raw_data/preprocess.py deleted file mode 100755 index 5706351a21fbd15d9bbf197156bb0fdabcb07295..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/data/proc_from_raw_data/preprocess.py +++ /dev/null @@ -1,236 +0,0 @@ -# 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. -# -*- coding: UTF-8 -*- - -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -""" -1. Tokenize the words and punctuation -2. pos sample : rating score 5; neg sample: rating score 1-2. - -Usage: - python preprocess.py -i data_file [random seed] -""" - -import sys -import os -import operator -import gzip -from subprocess import Popen, PIPE -from optparse import OptionParser -import json -from multiprocessing import Queue -from multiprocessing import Pool -import multiprocessing - -batch_size = 5000 -word_count = {} -num_tokenize = max(1, - multiprocessing.cpu_count() - 2) # parse + tokenize + save -max_queue_size = 8 -parse_queue = Queue(maxsize=max_queue_size + num_tokenize) -tokenize_queue = Queue(maxsize=max_queue_size + num_tokenize) - - -def create_dict(data): - """ - Create dictionary based on data, and saved in data_dir/dict.txt. - The first line is unk \t -1. - data: list, input data by batch. - """ - for seq in data: - try: - for w in seq.lower().split(): - if w not in word_count: - word_count[w] = 1 - else: - word_count[w] += 1 - except: - sys.stderr.write(seq + "\tERROR\n") - - -def parse(path): - """ - Open .gz file. - """ - sys.stderr.write(path) - g = gzip.open(path, 'r') - for l in g: - yield json.loads(l) - g.close() - - -def tokenize(sentences): - """ - Use tokenizer.perl to tokenize input sentences. - tokenizer.perl is tool of Moses. - sentences : a list of input sentences. - return: a list of processed text. - """ - dir = './mosesdecoder-master/scripts/tokenizer/tokenizer.perl' - if not os.path.exists(dir): - sys.exit( - "The ./mosesdecoder-master/scripts/tokenizer/tokenizer.perl does not exists." - ) - tokenizer_cmd = [dir, '-l', 'en', '-q', '-'] - assert isinstance(sentences, list) - text = "\n".join(sentences) - tokenizer = Popen(tokenizer_cmd, stdin=PIPE, stdout=PIPE) - tok_text, _ = tokenizer.communicate(text) - toks = tok_text.split('\n')[:-1] - return toks - - -def save_data(instance, data_dir, pre_fix, batch_num): - """ - save data by batch - """ - label = ['1' if pre_fix == 'pos' else '0' for i in range(len(instance))] - lines = ['%s\t%s' % (label[i], instance[i]) for i in range(len(label))] - file_name = os.path.join(data_dir, "%s_%s.txt" % (pre_fix, batch_num)) - file(file_name, 'w').write('\n'.join(lines) + '\n') - - -def tokenize_batch(id): - """ - tokenize data by batch - """ - while True: - num_batch, instance, pre_fix = parse_queue.get() - if num_batch == -1: ### parse_queue finished - tokenize_queue.put((-1, None, None)) - sys.stderr.write("Thread %s finish\n" % (id)) - break - tokenize_instance = tokenize(instance) - tokenize_queue.put((num_batch, tokenize_instance, pre_fix)) - sys.stderr.write('.') - - -def save_batch(data_dir, num_tokenize, data_dir_dict): - """ - save data by batch - build dict.txt - """ - token_count = 0 - while True: - num_batch, instance, pre_fix = tokenize_queue.get() - if num_batch == -1: - token_count += 1 - if token_count == num_tokenize: #### tokenize finished. - break - else: - continue - save_data(instance, data_dir, pre_fix, num_batch) - create_dict(instance) ## update dict - - sys.stderr.write("save file finish\n") - f = open(data_dir_dict, 'w') - f.write('%s\t%s\n' % ('unk', '-1')) - for k, v in sorted(word_count.items(), key=operator.itemgetter(1), \ - reverse=True): - f.write('%s\t%s\n' % (k, v)) - f.close() - sys.stderr.write("build dict finish\n") - - -def parse_batch(data, num_tokenize): - """ - parse data by batch - parse -> tokenize -> save - """ - raw_txt = parse(data) - neg, pos = [], [] - count = 0 - sys.stderr.write("extract raw data\n") - for l in raw_txt: - rating = l["overall"] - text = l["reviewText"].lower() # # convert words to lower case - if rating == 5.0 and text: - pos.append(text) - if rating < 3.0 and text: - neg.append(text) - if len(pos) == batch_size or len(neg) == batch_size: - if len(pos) == batch_size: - batch = pos - pre_fix = 'pos' - else: - batch = neg - pre_fix = 'neg' - - parse_queue.put((count, batch, pre_fix)) - count += 1 - if pre_fix == 'pos': - pos = [] - else: - neg = [] - - if len(pos) > 0: - parse_queue.put((count, pos, 'pos')) - count += 1 - if len(neg) > 0: - parse_queue.put((count, neg, 'neg')) - count += 1 - for i in range(num_tokenize): - parse_queue.put((-1, None, None)) #### for tokenize's input finished - sys.stderr.write("parsing finish\n") - - -def option_parser(): - parser = OptionParser(usage="usage: python preprcoess.py "\ - "-i data_path [options]") - parser.add_option( - "-i", "--data", action="store", dest="input", help="Input data path.") - parser.add_option( - "-s", - "--seed", - action="store", - dest="seed", - default=1024, - help="Set random seed.") - return parser.parse_args() - - -def main(): - reload(sys) - sys.setdefaultencoding('utf-8') - options, args = option_parser() - data = options.input - seed = options.seed - data_dir_dict = os.path.join(os.path.dirname(data), 'dict.txt') - data_dir = os.path.join(os.path.dirname(data), 'tmp') - pool = Pool(processes=num_tokenize + 2) - pool.apply_async(parse_batch, args=(data, num_tokenize)) - for i in range(num_tokenize): - pool.apply_async(tokenize_batch, args=(str(i), )) - pool.apply_async(save_batch, args=(data_dir, num_tokenize, data_dir_dict)) - pool.close() - pool.join() - - file(os.path.join(os.path.dirname(data), 'labels.list'), - 'w').write('neg\t0\npos\t1\n') - - -if __name__ == '__main__': - main() diff --git a/v1_api_demo/quick_start/dataprovider_bow.py b/v1_api_demo/quick_start/dataprovider_bow.py deleted file mode 100644 index 2745495586449b5d1eb64ae570f73eb6b14dbdfe..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/dataprovider_bow.py +++ /dev/null @@ -1,86 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer.PyDataProvider2 import * - -# id of the word not in dictionary -UNK_IDX = 0 - - -# initializer is called by the framework during initialization. -# It allows the user to describe the data types and setup the -# necessary data structure for later use. -# `settings` is an object. initializer need to properly fill settings.input_types. -# initializer can also store other data structures needed to be used at process(). -# In this example, dictionary is stored in settings. -# `dictionay` and `kwargs` are arguments passed from trainer_config.lr.py -def initializer(settings, dictionary, **kwargs): - # Put the word dictionary into settings - settings.word_dict = dictionary - - # setting.input_types specifies what the data types the data provider - # generates. - settings.input_types = { - # The first input is a sparse_binary_vector, - # which means each dimension of the vector is either 0 or 1. It is the - # bag-of-words (BOW) representation of the texts. - 'word': sparse_binary_vector(len(dictionary)), - # The second input is an integer. It represents the category id of the - # sample. 2 means there are two labels in the dataset. - # (1 for positive and 0 for negative) - 'label': integer_value(2) - } - - -# Delaring a data provider. It has an initializer 'data_initialzer'. -# It will cache the generated data of the first pass in memory, so that -# during later pass, no on-the-fly data generation will be needed. -# `setting` is the same object used by initializer() -# `file_name` is the name of a file listed train_list or test_list file given -# to define_py_data_sources2(). See trainer_config.lr.py. -@provider(init_hook=initializer, cache=CacheType.CACHE_PASS_IN_MEM) -def process(settings, file_name): - # Open the input data file. - with open(file_name, 'r') as f: - # Read each line. - for line in f: - # Each line contains the label and text of the comment, separated by \t. - label, comment = line.strip().split('\t') - - # Split the words into a list. - words = comment.split() - - # convert the words into a list of ids by looking them up in word_dict. - word_vector = [settings.word_dict.get(w, UNK_IDX) for w in words] - - # Return the features for the current comment. The first is a list - # of ids representing a 0-1 binary sparse vector of the text, - # the second is the integer id of the label. - yield {'word': word_vector, 'label': int(label)} - - -def predict_initializer(settings, dictionary, **kwargs): - settings.word_dict = dictionary - settings.input_types = {'word': sparse_binary_vector(len(dictionary))} - - -# Declaring a data provider for prediction. The difference with process -# is that label is not generated. -@provider(init_hook=predict_initializer, should_shuffle=False) -def process_predict(settings, file_name): - with open(file_name, 'r') as f: - for line in f: - comment = line.strip().split() - word_vector = [settings.word_dict.get(w, UNK_IDX) for w in comment] - yield {'word': word_vector} diff --git a/v1_api_demo/quick_start/dataprovider_emb.py b/v1_api_demo/quick_start/dataprovider_emb.py deleted file mode 100755 index ddfa3ce9b73555cb3b7f5a44314ca35b12d41ede..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/dataprovider_emb.py +++ /dev/null @@ -1,52 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer.PyDataProvider2 import * - -UNK_IDX = 0 - - -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 - 'word': integer_value_sequence(len(dictionary)), - # Define the second input for label id - 'label': integer_value(2) - } - - -@provider(init_hook=initializer, cache=CacheType.CACHE_PASS_IN_MEM) -def process(settings, file_name): - with open(file_name, 'r') as f: - for line in f: - label, comment = line.strip().split('\t') - words = comment.split() - word_slot = [settings.word_dict.get(w, UNK_IDX) for w in words] - yield {'word': word_slot, 'label': int(label)} - - -def predict_initializer(settings, dictionary, **kwargs): - settings.word_dict = dictionary - settings.input_types = {'word': integer_value_sequence(len(dictionary))} - - -@provider(init_hook=predict_initializer, should_shuffle=False) -def process_predict(settings, file_name): - with open(file_name, 'r') as f: - for line in f: - comment = line.strip().split() - word_slot = [settings.word_dict.get(w, UNK_IDX) for w in comment] - yield {'word': word_slot} diff --git a/v1_api_demo/quick_start/predict.sh b/v1_api_demo/quick_start/predict.sh deleted file mode 100755 index e47c2dd01fb5c919203964e298018e6dc2bd366e..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/predict.sh +++ /dev/null @@ -1,32 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -set -e - -cfg=trainer_config.lr.py -#cfg=trainer_config.emb.py -#cfg=trainer_config.cnn.py -#cfg=trainer_config.lstm.py -model="output/pass-00003" -paddle train \ - --config=$cfg \ - --use_gpu=false \ - --job=test \ - --init_model_path=$model \ - --config_args=is_predict=1 \ - --predict_output_dir=. \ -2>&1 | tee 'predict.log' -paddle usage -l 'predict.log' -e $? -n "quick_start_predict_${cfg}" >/dev/null 2>&1 - -mv rank-00000 result.txt diff --git a/v1_api_demo/quick_start/train.sh b/v1_api_demo/quick_start/train.sh deleted file mode 100755 index 01697fed48054be8ad98a01d4cbb5029e6a1ead0..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/train.sh +++ /dev/null @@ -1,34 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -set -e - -cfg=trainer_config.lr.py -#cfg=trainer_config.emb.py -#cfg=trainer_config.cnn.py -#cfg=trainer_config.lstm.py -#cfg=trainer_config.bidi-lstm.py -#cfg=trainer_config.db-lstm.py -#cfg=trainer_config.resnet-lstm.py -paddle train \ - --config=$cfg \ - --save_dir=./output \ - --trainer_count=4 \ - --log_period=100 \ - --num_passes=15 \ - --use_gpu=false \ - --show_parameter_stats_period=100 \ - --test_all_data_in_one_period=1 \ - 2>&1 | tee 'train.log' -paddle usage -l "train.log" -e $? -n "quick_start_${cfg}" >/dev/null 2>&1 diff --git a/v1_api_demo/quick_start/trainer_config.bidi-lstm.py b/v1_api_demo/quick_start/trainer_config.bidi-lstm.py deleted file mode 100644 index 3deff4aa00b1ea5d66097514867d1a392393a523..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/trainer_config.bidi-lstm.py +++ /dev/null @@ -1,74 +0,0 @@ -# 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. -# edit-mode: -*- python -*- - -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -dict_file = "./data/dict.txt" -word_dict = dict() -with open(dict_file, 'r') as f: - for i, line in enumerate(f): - w = line.strip().split()[0] - word_dict[w] = i - -is_predict = get_config_arg('is_predict', bool, False) -trn = 'data/train.list' if not is_predict else None -tst = 'data/test.list' if not is_predict else 'data/pred.list' -process = 'process' if not is_predict else 'process_predict' -define_py_data_sources2( - train_list=trn, - test_list=tst, - module="dataprovider_emb", - obj=process, - args={"dictionary": word_dict}) - -batch_size = 128 if not is_predict else 1 -settings( - batch_size=batch_size, - learning_rate=2e-3, - learning_method=AdamOptimizer(), - regularization=L2Regularization(8e-4), - gradient_clipping_threshold=25) - -bias_attr = ParamAttr(initial_std=0., l2_rate=0.) -data = data_layer(name="word", size=len(word_dict)) -emb = embedding_layer(input=data, size=128) - -bi_lstm = bidirectional_lstm(input=emb, size=128) -dropout = dropout_layer(input=bi_lstm, dropout_rate=0.5) - -output = fc_layer( - input=dropout, size=2, bias_attr=bias_attr, act=SoftmaxActivation()) - -if is_predict: - maxid = maxid_layer(output) - outputs([maxid, output]) -else: - label = data_layer(name="label", size=2) - cls = classification_cost(input=output, label=label) - outputs(cls) diff --git a/v1_api_demo/quick_start/trainer_config.cnn.py b/v1_api_demo/quick_start/trainer_config.cnn.py deleted file mode 100644 index e09e41484d30db385a1d276b7f346b444fe79d3d..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/trainer_config.cnn.py +++ /dev/null @@ -1,68 +0,0 @@ -# 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. -# edit-mode: -*- python -*- - -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -dict_file = "./data/dict.txt" -word_dict = dict() -with open(dict_file, 'r') as f: - for i, line in enumerate(f): - w = line.strip().split()[0] - word_dict[w] = i - -is_predict = get_config_arg('is_predict', bool, False) -trn = 'data/train.list' if not is_predict else None -tst = 'data/test.list' if not is_predict else 'data/pred.list' -process = 'process' if not is_predict else 'process_predict' -define_py_data_sources2( - train_list=trn, - test_list=tst, - module="dataprovider_emb", - obj=process, - args={"dictionary": word_dict}) - -batch_size = 128 if not is_predict else 1 -settings( - batch_size=batch_size, - learning_rate=2e-3, - learning_method=AdamOptimizer(), - regularization=L2Regularization(8e-4), - gradient_clipping_threshold=25) - -data = data_layer(name="word", size=len(word_dict)) -embedding = embedding_layer(input=data, size=128) -conv = sequence_conv_pool(input=embedding, context_len=3, hidden_size=512) -output = fc_layer(input=conv, size=2, act=SoftmaxActivation()) -if is_predict: - maxid = maxid_layer(output) - outputs([maxid, output]) -else: - label = data_layer(name="label", size=2) - cls = classification_cost(input=output, label=label) - outputs(cls) diff --git a/v1_api_demo/quick_start/trainer_config.db-lstm.py b/v1_api_demo/quick_start/trainer_config.db-lstm.py deleted file mode 100644 index fba802b4600b33cfbfd0820cce1f47e4d0f948ae..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/trainer_config.db-lstm.py +++ /dev/null @@ -1,74 +0,0 @@ -# edit-mode: -*- python -*- - -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -dict_file = "./data/dict.txt" -word_dict = dict() -with open(dict_file, 'r') as f: - for i, line in enumerate(f): - w = line.strip().split()[0] - word_dict[w] = i - -is_predict = get_config_arg('is_predict', bool, False) -trn = 'data/train.list' if not is_predict else None -tst = 'data/test.list' if not is_predict else 'data/pred.list' -process = 'process' if not is_predict else 'process_predict' -define_py_data_sources2( - train_list=trn, - test_list=tst, - module="dataprovider_emb", - obj=process, - args={"dictionary": word_dict}) - -batch_size = 128 if not is_predict else 1 -settings( - batch_size=batch_size, - learning_rate=2e-3, - learning_method=AdamOptimizer(), - regularization=L2Regularization(8e-4), - gradient_clipping_threshold=25) - -bias_attr = ParamAttr(initial_std=0., l2_rate=0.) - -data = data_layer(name="word", size=len(word_dict)) -emb = embedding_layer(input=data, size=128) - -hidden_0 = mixed_layer(size=128, input=[full_matrix_projection(input=emb)]) -lstm_0 = lstmemory(input=hidden_0, layer_attr=ExtraAttr(drop_rate=0.1)) - -input_layers = [hidden_0, lstm_0] - -for i in range(1, 8): - fc = fc_layer(input=input_layers, size=128) - lstm = lstmemory( - input=fc, - layer_attr=ExtraAttr(drop_rate=0.1), - reverse=(i % 2) == 1, ) - input_layers = [fc, lstm] - -lstm_last = pooling_layer(input=lstm, pooling_type=MaxPooling()) - -output = fc_layer( - input=lstm_last, size=2, bias_attr=bias_attr, act=SoftmaxActivation()) - -if is_predict: - maxid = maxid_layer(output) - outputs([maxid, output]) -else: - label = data_layer(name="label", size=2) - cls = classification_cost(input=output, label=label) - outputs(cls) diff --git a/v1_api_demo/quick_start/trainer_config.emb.py b/v1_api_demo/quick_start/trainer_config.emb.py deleted file mode 100644 index f69f98ff7fc885d3fe16d3aaf66967389b3b3240..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/trainer_config.emb.py +++ /dev/null @@ -1,64 +0,0 @@ -# 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. -# edit-mode: -*- python -*- - -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -dict_file = "./data/dict.txt" -word_dict = dict() -with open(dict_file, 'r') as f: - for i, line in enumerate(f): - w = line.strip().split()[0] - word_dict[w] = i - -is_predict = get_config_arg('is_predict', bool, False) -trn = 'data/train.list' if not is_predict else None -tst = 'data/test.list' if not is_predict else 'data/pred.list' -process = 'process' if not is_predict else 'process_predict' -define_py_data_sources2( - train_list=trn, - test_list=tst, - module="dataprovider_emb", - obj=process, - args={"dictionary": word_dict}) - -batch_size = 128 if not is_predict else 1 -settings( - batch_size=batch_size, learning_rate=2e-3, learning_method=AdamOptimizer()) - -data = data_layer(name="word", size=len(word_dict)) -embedding = embedding_layer(input=data, size=128) -avg = pooling_layer(input=embedding, pooling_type=AvgPooling()) -output = fc_layer(input=avg, size=2, act=SoftmaxActivation()) -if is_predict: - maxid = maxid_layer(output) - outputs([maxid, output]) -else: - label = data_layer(name="label", size=2) - cls = classification_cost(input=output, label=label) - outputs(cls) diff --git a/v1_api_demo/quick_start/trainer_config.lr.py b/v1_api_demo/quick_start/trainer_config.lr.py deleted file mode 100644 index b7b694940e338acbc40ffd3e5597f209bf07488f..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/trainer_config.lr.py +++ /dev/null @@ -1,85 +0,0 @@ -# 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. -# edit-mode: -*- python -*- - -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -dict_file = get_config_arg('dict_file', str, "./data/dict.txt") -word_dict = dict() -with open(dict_file, 'r') as f: - for i, line in enumerate(f): - w = line.strip().split()[0] - word_dict[w] = i - -is_predict = get_config_arg('is_predict', bool, False) -trn = 'data/train.list' if not is_predict else None -tst = 'data/test.list' if not is_predict else 'data/pred.list' -process = 'process' if not is_predict else 'process_predict' - -# define the data sources for the model. -# We need to use different process for training and prediction. -# For training, the input data includes both word IDs and labels. -# For prediction, the input data only includs word Ids. -define_py_data_sources2( - train_list=trn, - test_list=tst, - module="dataprovider_bow", - obj=process, - args={"dictionary": word_dict}) - -batch_size = 128 if not is_predict else 1 -settings( - batch_size=batch_size, - learning_rate=2e-3, - learning_method=AdamOptimizer(), - regularization=L2Regularization(8e-4), - gradient_clipping_threshold=25) - -# Define the data for text features. The size of the data layer is the number -# of words in the dictionary. -data = data_layer(name="word", size=len(word_dict)) - -# Define a fully connected layer with logistic activation. -# (also called softmax activation). -output = fc_layer(input=data, size=2, act=SoftmaxActivation()) - -if not is_predict: - # For training, we need label and cost - - # define the category id for each example. - # The size of the data layer is the number of labels. - label = data_layer(name="label", size=2) - - # Define cross-entropy classification loss and error. - cls = classification_cost(input=output, label=label) - outputs(cls) -else: - # For prediction, no label is needed. We need to output - # We need to output classification result, and class probabilities. - maxid = maxid_layer(output) - outputs([maxid, output]) diff --git a/v1_api_demo/quick_start/trainer_config.lstm.py b/v1_api_demo/quick_start/trainer_config.lstm.py deleted file mode 100644 index 8967d78807b9bbf990f5dd36240c18199b86954e..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/trainer_config.lstm.py +++ /dev/null @@ -1,70 +0,0 @@ -# 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. -# edit-mode: -*- python -*- - -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -dict_file = "./data/dict.txt" -word_dict = dict() -with open(dict_file, 'r') as f: - for i, line in enumerate(f): - w = line.strip().split()[0] - word_dict[w] = i - -is_predict = get_config_arg('is_predict', bool, False) -trn = 'data/train.list' if not is_predict else None -tst = 'data/test.list' if not is_predict else 'data/pred.list' -process = 'process' if not is_predict else 'process_predict' -define_py_data_sources2( - train_list=trn, - test_list=tst, - module="dataprovider_emb", - obj=process, - args={"dictionary": word_dict}) - -batch_size = 128 if not is_predict else 1 -settings( - batch_size=batch_size, - learning_rate=2e-3, - learning_method=AdamOptimizer(), - regularization=L2Regularization(8e-4), - gradient_clipping_threshold=25) - -data = data_layer(name="word", size=len(word_dict)) -emb = embedding_layer(input=data, size=128) -lstm = simple_lstm( - input=emb, size=128, lstm_cell_attr=ExtraAttr(drop_rate=0.25)) -lstm_max = pooling_layer(input=lstm, pooling_type=MaxPooling()) -output = fc_layer(input=lstm_max, size=2, act=SoftmaxActivation()) -if is_predict: - maxid = maxid_layer(output) - outputs([maxid, output]) -else: - label = data_layer(name="label", size=2) - cls = classification_cost(input=output, label=label) - outputs(cls) diff --git a/v1_api_demo/quick_start/trainer_config.resnet-lstm.py b/v1_api_demo/quick_start/trainer_config.resnet-lstm.py deleted file mode 100644 index 32d0596f250c0f0c5a4004d3af7adb794b3f0f1b..0000000000000000000000000000000000000000 --- a/v1_api_demo/quick_start/trainer_config.resnet-lstm.py +++ /dev/null @@ -1,104 +0,0 @@ -# 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. -# edit-mode: -*- python -*- - -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -""" -This configuration is a demonstration of how to implement the stacked LSTM -with residual connections, i.e. an LSTM layer takes the sum of the hidden states -and inputs of the previous LSTM layer instead of only the hidden states. -This architecture is from: -Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi, -Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, -Jeff Klingner, Apurva Shah, Melvin Johnson, Xiaobing Liu, Lukasz Kaiser, -Stephan Gouws, Yoshikiyo Kato, Taku Kudo, Hideto Kazawa, Keith Stevens, -George Kurian, Nishant Patil, Wei Wang, Cliff Young, Jason Smith, Jason Riesa, -Alex Rudnick, Oriol Vinyals, Greg Corrado, Macduff Hughes, Jeffrey Dean. 2016. -Google's Neural Machine Translation System: Bridging the Gap between Human and -Machine Translation. In arXiv https://arxiv.org/pdf/1609.08144v2.pdf -Different from the architecture described in the paper, we use a stack single -direction LSTM layers as the first layer instead of bi-directional LSTM. Also, -since this is a demo code, to reduce computation time, we stacked 4 layers -instead of 8 layers. -""" - -from paddle.trainer_config_helpers import * - -dict_file = "./data/dict.txt" -word_dict = dict() -with open(dict_file, 'r') as f: - for i, line in enumerate(f): - w = line.strip().split()[0] - word_dict[w] = i - -is_predict = get_config_arg('is_predict', bool, False) -trn = 'data/train.list' if not is_predict else None -tst = 'data/test.list' if not is_predict else 'data/pred.list' -process = 'process' if not is_predict else 'process_predict' -define_py_data_sources2( - train_list=trn, - test_list=tst, - module="dataprovider_emb", - obj=process, - args={"dictionary": word_dict}) - -batch_size = 128 if not is_predict else 1 -settings( - batch_size=batch_size, - learning_rate=2e-3, - learning_method=AdamOptimizer(), - regularization=L2Regularization(8e-4), - gradient_clipping_threshold=25) - -bias_attr = ParamAttr(initial_std=0., l2_rate=0.) - -data = data_layer(name="word", size=len(word_dict)) -emb = embedding_layer(input=data, size=128) -lstm = simple_lstm(input=emb, size=128, lstm_cell_attr=ExtraAttr(drop_rate=0.1)) - -previous_input, previous_hidden_state = emb, lstm - -for i in range(3): - # The input to the current layer is the sum of the hidden state - # and input of the previous layer. - current_input = addto_layer(input=[previous_input, previous_hidden_state]) - hidden_state = simple_lstm( - input=current_input, size=128, lstm_cell_attr=ExtraAttr(drop_rate=0.1)) - previous_input, previous_hidden_state = current_input, hidden_state - -lstm = previous_hidden_state - -lstm_last = pooling_layer(input=lstm, pooling_type=MaxPooling()) -output = fc_layer( - input=lstm_last, size=2, bias_attr=bias_attr, act=SoftmaxActivation()) - -if is_predict: - maxid = maxid_layer(output) - outputs([maxid, output]) -else: - label = data_layer(name="label", size=2) - cls = classification_cost(input=output, label=label) - outputs(cls) diff --git a/v1_api_demo/sequence_tagging/data/get_data.sh b/v1_api_demo/sequence_tagging/data/get_data.sh deleted file mode 100755 index 0cdb394035e782b3a647f7f13e79d55b5d3dff48..0000000000000000000000000000000000000000 --- a/v1_api_demo/sequence_tagging/data/get_data.sh +++ /dev/null @@ -1,21 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -set -e - -DIR="$( cd "$(dirname "$0")" ; pwd -P )" -cd $DIR - -wget http://www.cnts.ua.ac.be/conll2000/chunking/train.txt.gz -wget http://www.cnts.ua.ac.be/conll2000/chunking/test.txt.gz diff --git a/v1_api_demo/sequence_tagging/data/test.list b/v1_api_demo/sequence_tagging/data/test.list deleted file mode 100644 index 073c0a0c9063ac55f762ac261746aa73057d70e8..0000000000000000000000000000000000000000 --- a/v1_api_demo/sequence_tagging/data/test.list +++ /dev/null @@ -1 +0,0 @@ -data/test.txt.gz diff --git a/v1_api_demo/sequence_tagging/data/train.list b/v1_api_demo/sequence_tagging/data/train.list deleted file mode 100644 index 43c24d5f6484a90fe883ad5516fe100d27c9ce47..0000000000000000000000000000000000000000 --- a/v1_api_demo/sequence_tagging/data/train.list +++ /dev/null @@ -1 +0,0 @@ -data/train.txt.gz diff --git a/v1_api_demo/sequence_tagging/dataprovider.py b/v1_api_demo/sequence_tagging/dataprovider.py deleted file mode 100644 index bb4b4465bc7e032c50c1d21263651e2578af67be..0000000000000000000000000000000000000000 --- a/v1_api_demo/sequence_tagging/dataprovider.py +++ /dev/null @@ -1,260 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer.PyDataProvider2 import * -import gzip -import logging - -logging.basicConfig( - format='[%(levelname)s %(asctime)s %(filename)s:%(lineno)s] %(message)s', ) -logger = logging.getLogger('paddle') -logger.setLevel(logging.INFO) - -OOV_POLICY_IGNORE = 0 -OOV_POLICY_USE = 1 -OOV_POLICY_ERROR = 2 - -num_original_columns = 3 - -# Feature combination patterns. -# [[-1,0], [0,0]] means previous token at column 0 and current token at -# column 0 are combined as one feature. -patterns = [ - [[-2, 0]], - [[-1, 0]], - [[0, 0]], - [[1, 0]], - [[2, 0]], - [[-1, 0], [0, 0]], - [[0, 0], [1, 0]], - [[-2, 1]], - [[-1, 1]], - [[0, 1]], - [[1, 1]], - [[2, 1]], - [[-2, 1], [-1, 1]], - [[-1, 1], [0, 1]], - [[0, 1], [1, 1]], - [[1, 1], [2, 1]], - [[-2, 1], [-1, 1], [0, 1]], - [[-1, 1], [0, 1], [1, 1]], - [[0, 1], [1, 1], [2, 1]], -] - -dict_label = { - 'B-ADJP': 0, - 'I-ADJP': 1, - 'B-ADVP': 2, - 'I-ADVP': 3, - 'B-CONJP': 4, - 'I-CONJP': 5, - 'B-INTJ': 6, - 'I-INTJ': 7, - 'B-LST': 8, - 'I-LST': 9, - 'B-NP': 10, - 'I-NP': 11, - 'B-PP': 12, - 'I-PP': 13, - 'B-PRT': 14, - 'I-PRT': 15, - 'B-SBAR': 16, - 'I-SBAR': 17, - 'B-UCP': 18, - 'I-UCP': 19, - 'B-VP': 20, - 'I-VP': 21, - 'O': 22 -} - - -def make_features(sequence): - length = len(sequence) - num_features = len(sequence[0]) - - def get_features(pos): - if pos < 0: - return ['#B%s' % -pos] * num_features - if pos >= length: - return ['#E%s' % (pos - length + 1)] * num_features - return sequence[pos] - - for i in xrange(length): - for pattern in patterns: - fname = '/'.join([get_features(i + pos)[f] for pos, f in pattern]) - sequence[i].append(fname) - - -''' -Source file format: -Each line is for one timestep. The features are separated by space. -An empty line indicates end of a sequence. - -cutoff: a list of numbers. If count of a feature is smaller than this, - it will be ignored. -if oov_policy[i] is OOV_POLICY_USE, id 0 is reserved for OOV features of -i-th column. - -return a list of dict for each column -''' - - -def create_dictionaries(filename, cutoff, oov_policy): - def add_to_dict(sequence, dicts): - num_features = len(dicts) - for features in sequence: - l = len(features) - assert l == num_features, "Wrong number of features " + line - for i in xrange(l): - if features[i] in dicts[i]: - dicts[i][features[i]] += 1 - else: - dicts[i][features[i]] = 1 - - num_features = len(cutoff) - dicts = [] - for i in xrange(num_features): - dicts.append(dict()) - - f = gzip.open(filename, 'rb') - - sequence = [] - - for line in f: - line = line.strip() - if not line: - make_features(sequence) - add_to_dict(sequence, dicts) - sequence = [] - continue - features = line.split(' ') - sequence.append(features) - - for i in xrange(num_features): - dct = dicts[i] - n = 1 if oov_policy[i] == OOV_POLICY_USE else 0 - todo = [] - for k, v in dct.iteritems(): - if v < cutoff[i]: - todo.append(k) - else: - dct[k] = n - n += 1 - - if oov_policy[i] == OOV_POLICY_USE: - # placeholder so that len(dct) will be the number of features - # including OOV - dct['#OOV#'] = 0 - - logger.info('column %d dict size=%d, ignored %d' % (i, n, len(todo))) - for k in todo: - del dct[k] - - f.close() - return dicts - - -def initializer(settings, **xargs): - cutoff = [3, 1, 0] - cutoff += [3] * len(patterns) - oov_policy = [OOV_POLICY_IGNORE, OOV_POLICY_ERROR, OOV_POLICY_ERROR] - oov_policy += [OOV_POLICY_IGNORE] * len(patterns) - dicts = create_dictionaries('data/train.txt.gz', cutoff, oov_policy) - dicts[2] = dict_label - settings.dicts = dicts - settings.oov_policy = oov_policy - input_types = [] - num_features = len(dicts) - for i in xrange(num_original_columns): - input_types.append(integer_sequence(len(dicts[i]))) - logger.info("slot %s size=%s" % (i, len(dicts[i]))) - if patterns: - dim = 0 - for i in xrange(num_original_columns, num_features): - dim += len(dicts[i]) - input_types.append(sparse_binary_vector_sequence(dim)) - logger.info("feature size=%s" % dim) - settings.input_types = input_types - - -''' -if oov_policy[i] == OOV_POLICY_USE, features in i-th column which are not -existed in dicts[i] will be assigned to id 0. -if oov_policy[i] == OOV_POLICY_ERROR, all features in i-th column MUST exist -in dicts[i]. -''' - - -@provider(init_hook=initializer, cache=CacheType.CACHE_PASS_IN_MEM) -def process(settings, filename): - input_file = filename - dicts = settings.dicts - oov_policy = settings.oov_policy - - def gen_sample(sequence): - num_features = len(dicts) - sample = [list() for i in xrange(num_original_columns)] - if patterns: - sample.append([]) - for features in sequence: - assert len(features) == num_features, \ - "Wrong number of features: " + line - for i in xrange(num_original_columns): - id = dicts[i].get(features[i], -1) - if id != -1: - sample[i].append(id) - elif oov_policy[i] == OOV_POLICY_IGNORE: - sample[i].append(0xffffffff) - elif oov_policy[i] == OOV_POLICY_ERROR: - logger.fatal("Unknown token: %s" % features[i]) - else: - sample[i].append(0) - - if patterns: - dim = 0 - vec = [] - for i in xrange(num_original_columns, num_features): - id = dicts[i].get(features[i], -1) - if id != -1: - vec.append(dim + id) - elif oov_policy[i] == OOV_POLICY_IGNORE: - pass - elif oov_policy[i] == OOV_POLICY_ERROR: - logger.fatal("Unknown token: %s" % features[i]) - else: - vec.ids.append(dim + 0) - - dim += len(dicts[i]) - sample[-1].append(vec) - return sample - - num_features = len(dicts) - f = gzip.open(input_file, 'rb') - - num_sequences = 0 - sequence = [] - for line in f: - line = line.strip() - if not line: - make_features(sequence) - yield gen_sample(sequence) - sequence = [] - num_sequences += 1 - continue - features = line.split(' ') - sequence.append(features) - - f.close() - - logger.info("num_sequences=%s" % num_sequences) diff --git a/v1_api_demo/sequence_tagging/linear_crf.py b/v1_api_demo/sequence_tagging/linear_crf.py deleted file mode 100644 index ea012ba1ae9c790ccefd3dd5f066aa92202128a2..0000000000000000000000000000000000000000 --- a/v1_api_demo/sequence_tagging/linear_crf.py +++ /dev/null @@ -1,83 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -import math - -define_py_data_sources2( - train_list="data/train.list", - test_list="data/test.list", - module="dataprovider", - obj="process") - -batch_size = 1 -settings( - learning_method=MomentumOptimizer(), - batch_size=batch_size, - regularization=L2Regularization(batch_size * 1e-4), - model_average=ModelAverage(0.5), - learning_rate=1e-1, - learning_rate_decay_a=1e-5, - learning_rate_decay_b=0.25, ) - -num_label_types = 23 - - -def get_simd_size(size): - return int(math.ceil(float(size) / 8)) * 8 - - -# Currently, in order to use sparse_update=True, -# the size has to be aligned. -num_label_types = get_simd_size(num_label_types) - -features = data_layer(name="features", size=76328) -word = data_layer(name="word", size=6778) -pos = data_layer(name="pos", size=44) -chunk = data_layer(name="chunk", size=num_label_types) - -crf_input = fc_layer( - input=features, - size=num_label_types, - act=LinearActivation(), - bias_attr=False, - param_attr=ParamAttr( - initial_std=0, sparse_update=True)) - -crf = crf_layer( - input=crf_input, - label=chunk, - param_attr=ParamAttr( - name="crfw", initial_std=0), ) - -crf_decoding = crf_decoding_layer( - size=num_label_types, - input=crf_input, - label=chunk, - param_attr=ParamAttr(name="crfw"), ) - -sum_evaluator( - name="error", - input=crf_decoding, ) - -chunk_evaluator( - name="chunk_f1", - input=crf_decoding, - label=chunk, - chunk_scheme="IOB", - num_chunk_types=11, ) - -inputs(word, pos, chunk, features) -outputs(crf) diff --git a/v1_api_demo/sequence_tagging/readme.md b/v1_api_demo/sequence_tagging/readme.md deleted file mode 100644 index 2e17fffb83c532f5e5fec1227f169c97c1f20e22..0000000000000000000000000000000000000000 --- a/v1_api_demo/sequence_tagging/readme.md +++ /dev/null @@ -1,45 +0,0 @@ -# Sequence Tagging - -This demo is a sequence model for assigning tags to each token in a sentence. The task is described at CONLL2000 Text Chunking task. - -## Download data -```bash -cd demo/sequence_tagging -./data/get_data.sh -``` - -## Train model -```bash -cd demo/sequence_tagging -./train.sh -``` - -## Model description - -We provide two models. One is a linear CRF model (linear_crf.py) with is equivalent to the one at leon.bottou.org/projects/sgd. The second one is a stacked bidirectional RNN and CRF model (rnn_crf.py). -
- - - - - - - - - - - - - - - - - - - - - - -
Model nameNumber of parametersF1 score
linear_crf 1.8M 0.937
rnn_crf 960K 0.941
-
-
diff --git a/v1_api_demo/sequence_tagging/rnn_crf.py b/v1_api_demo/sequence_tagging/rnn_crf.py deleted file mode 100644 index 937a34df103663ecf0f0827bbfb9d82823c9b902..0000000000000000000000000000000000000000 --- a/v1_api_demo/sequence_tagging/rnn_crf.py +++ /dev/null @@ -1,121 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * - -import math - -define_py_data_sources2( - train_list="data/train.list", - test_list="data/test.list", - module="dataprovider", - obj="process") - -batch_size = 16 -settings( - learning_method=MomentumOptimizer(), - batch_size=batch_size, - regularization=L2Regularization(batch_size * 1e-5), - model_average=ModelAverage(0.5), - learning_rate=2e-3, - learning_rate_decay_a=5e-7, - learning_rate_decay_b=0.5, ) - -word_dim = 128 -hidden_dim = 128 -with_rnn = True - -initial_std = 1 / math.sqrt(hidden_dim) -param_attr = ParamAttr(initial_std=initial_std) -cpu_layer_attr = ExtraLayerAttribute(device=-1) - -default_device(0) - -num_label_types = 23 - -features = data_layer(name="features", size=76328) -word = data_layer(name="word", size=6778) -pos = data_layer(name="pos", size=44) -chunk = data_layer( - name="chunk", size=num_label_types, layer_attr=cpu_layer_attr) - -emb = embedding_layer( - input=word, size=word_dim, param_attr=ParamAttr(initial_std=0)) - -hidden1 = mixed_layer( - size=hidden_dim, - act=STanhActivation(), - bias_attr=True, - input=[ - full_matrix_projection(emb), table_projection( - pos, param_attr=param_attr) - ]) - -if with_rnn: - rnn1 = recurrent_layer( - act=ReluActivation(), - bias_attr=True, - input=hidden1, - param_attr=ParamAttr(initial_std=0), ) - -hidden2 = mixed_layer( - size=hidden_dim, - act=STanhActivation(), - bias_attr=True, - input=[full_matrix_projection(hidden1)] + - ([full_matrix_projection( - rnn1, param_attr=ParamAttr(initial_std=0))] if with_rnn else []), ) - -if with_rnn: - rnn2 = recurrent_layer( - reverse=True, - act=ReluActivation(), - bias_attr=True, - input=hidden2, - param_attr=ParamAttr(initial_std=0), ) - -crf_input = mixed_layer( - size=num_label_types, - bias_attr=False, - input=[full_matrix_projection(hidden2), ] + - ([full_matrix_projection( - rnn2, param_attr=ParamAttr(initial_std=0))] if with_rnn else []), ) - -crf = crf_layer( - input=crf_input, - label=chunk, - param_attr=ParamAttr( - name="crfw", initial_std=0), - layer_attr=cpu_layer_attr, ) - -crf_decoding = crf_decoding_layer( - size=num_label_types, - input=crf_input, - label=chunk, - param_attr=ParamAttr(name="crfw"), - layer_attr=cpu_layer_attr, ) - -sum_evaluator( - name="error", - input=crf_decoding, ) - -chunk_evaluator( - name="chunk_f1", - input=crf_decoding, - label=chunk, - chunk_scheme="IOB", - num_chunk_types=11, ) - -inputs(word, pos, chunk, features) -outputs(crf) diff --git a/v1_api_demo/sequence_tagging/train.sh b/v1_api_demo/sequence_tagging/train.sh deleted file mode 100755 index 37e196c84200dc26ccb523076a81dbc393b1280f..0000000000000000000000000000000000000000 --- a/v1_api_demo/sequence_tagging/train.sh +++ /dev/null @@ -1,12 +0,0 @@ -#!/bin/bash - -paddle train \ - --config rnn_crf.py \ - --parallel_nn=1 \ - --use_gpu=1 \ - --dot_period=10 \ - --log_period=1000 \ - --test_period=0 \ - --num_passes=10 \ -2>&1 | tee 'train.log' -paddle usage -l 'train.log' -e $? -n "sequence_tagging_train" >/dev/null 2>&1 diff --git a/v1_api_demo/sequence_tagging/train_linear.sh b/v1_api_demo/sequence_tagging/train_linear.sh deleted file mode 100755 index ad6e2d8ee7f813c69f9dd250c6f7bbb4403a0ed5..0000000000000000000000000000000000000000 --- a/v1_api_demo/sequence_tagging/train_linear.sh +++ /dev/null @@ -1,11 +0,0 @@ -#!/bin/bash - -paddle train \ - --config linear_crf.py \ - --use_gpu=0 \ - --dot_period=100 \ - --log_period=10000 \ - --test_period=0 \ - --num_passes=10 -2>&1 | tee 'train_linear.log' -paddle usage -l 'train_linear.log' -e $? -n "sequence_tagging_train_linear" >/dev/null 2>&1 diff --git a/v1_api_demo/traffic_prediction/README b/v1_api_demo/traffic_prediction/README deleted file mode 100644 index 4c95188583513c332b7d7cb0a32d59336208e1aa..0000000000000000000000000000000000000000 --- a/v1_api_demo/traffic_prediction/README +++ /dev/null @@ -1,7 +0,0 @@ -run by: -cd ./data -sh get_data.sh -cd .. -sh train.sh -sh predict.sh - diff --git a/v1_api_demo/traffic_prediction/data/get_data.sh b/v1_api_demo/traffic_prediction/data/get_data.sh deleted file mode 100755 index f2fa548d4709c0361334f117bfb49e18d83c32f4..0000000000000000000000000000000000000000 --- a/v1_api_demo/traffic_prediction/data/get_data.sh +++ /dev/null @@ -1,34 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors, Inc. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -set -e -set -x - -DIR="$( cd "$(dirname "$0")" ; pwd -P )" -cd $DIR - -#download the dataset -echo "Downloading traffic data..." -wget http://paddlepaddle.cdn.bcebos.com/demo/traffic/traffic_data.tar.gz - -#extract package -echo "Unzipping..." -tar -zxvf traffic_data.tar.gz - -echo "data/speeds.csv" > train.list -echo "data/speeds.csv" > test.list -echo "data/speeds.csv" > pred.list - -echo "Done." diff --git a/v1_api_demo/traffic_prediction/dataprovider.py b/v1_api_demo/traffic_prediction/dataprovider.py deleted file mode 100644 index c7883b6950c369ee67c39b80ce1cefbbf9350459..0000000000000000000000000000000000000000 --- a/v1_api_demo/traffic_prediction/dataprovider.py +++ /dev/null @@ -1,82 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors, Inc. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer.PyDataProvider2 import * -import sys -import numpy as np -TERM_NUM = 24 -FORECASTING_NUM = 24 -LABEL_VALUE_NUM = 4 - - -def initHook(settings, file_list, **kwargs): - """ - Init hook is invoked before process data. It will set obj.slots and store data meta. - - :param settings: global object. It will passed to process routine. - :type obj: object - :param file_list: the meta file object, which passed from trainer_config.py,but unused in this function. - :param kwargs: unused other arguments. - """ - del kwargs #unused - - settings.pool_size = sys.maxint - #Use a time seires of the past as feature. - #Dense_vector's expression form is [float,float,...,float] - settings.input_types = [dense_vector(TERM_NUM)] - #There are next FORECASTING_NUM fragments you need predict. - #Every predicted condition at time point has four states. - for i in range(FORECASTING_NUM): - settings.input_types.append(integer_value(LABEL_VALUE_NUM)) - - -@provider( - init_hook=initHook, cache=CacheType.CACHE_PASS_IN_MEM, should_shuffle=True) -def process(settings, file_name): - with open(file_name) as f: - #abandon fields name - f.next() - for row_num, line in enumerate(f): - speeds = map(int, line.rstrip('\r\n').split(",")[1:]) - # Get the max index. - end_time = len(speeds) - # Scanning and generating samples - for i in range(TERM_NUM, end_time - FORECASTING_NUM): - # For dense slot - pre_spd = map(float, speeds[i - TERM_NUM:i]) - - # Integer value need predicting, values start from 0, so every one minus 1. - fol_spd = [j - 1 for j in speeds[i:i + FORECASTING_NUM]] - - # Predicting label is missing, abandon the sample. - if -1 in fol_spd: - continue - yield [pre_spd] + fol_spd - - -def predict_initHook(settings, file_list, **kwargs): - settings.pool_size = sys.maxint - settings.input_types = [dense_vector(TERM_NUM)] - - -@provider(init_hook=predict_initHook, should_shuffle=False) -def process_predict(settings, file_name): - with open(file_name) as f: - #abandon fields name - f.next() - for row_num, line in enumerate(f): - speeds = map(int, line.rstrip('\r\n').split(",")) - end_time = len(speeds) - pre_spd = map(float, speeds[end_time - TERM_NUM:end_time]) - yield pre_spd diff --git a/v1_api_demo/traffic_prediction/gen_result.py b/v1_api_demo/traffic_prediction/gen_result.py deleted file mode 100644 index 3da70b30315f863fd3582583e9a29540a09c1e7f..0000000000000000000000000000000000000000 --- a/v1_api_demo/traffic_prediction/gen_result.py +++ /dev/null @@ -1,61 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors, Inc. 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. - -res = [] -with open('./rank-00000') as f: - for line in f: - pred = map(int, line.strip('\r\n;').split(";")) - #raw prediction range from 0 to 3 - res.append([i + 1 for i in pred]) - -file_name = open('./data/pred.list').read().strip('\r\n') - -FORECASTING_NUM = 24 -header = [ - 'id', - '201604200805', - '201604200810', - '201604200815', - '201604200820', - '201604200825', - '201604200830', - '201604200835', - '201604200840', - '201604200845', - '201604200850', - '201604200855', - '201604200900', - '201604200905', - '201604200910', - '201604200915', - '201604200920', - '201604200925', - '201604200930', - '201604200935', - '201604200940', - '201604200945', - '201604200950', - '201604200955', - '201604201000', -] -################### -## To CSV format ## -################### -with open(file_name) as f: - f.next() - print ','.join(header) - for row_num, line in enumerate(f): - fields = line.rstrip('\r\n').split(',') - linkid = fields[0] - print linkid + ',' + ','.join(map(str, res[row_num])) diff --git a/v1_api_demo/traffic_prediction/predict.sh b/v1_api_demo/traffic_prediction/predict.sh deleted file mode 100755 index 2dbd5e8805dd97d35c7d58917f8ec6b5033bda03..0000000000000000000000000000000000000000 --- a/v1_api_demo/traffic_prediction/predict.sh +++ /dev/null @@ -1,30 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors, Inc. 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. -set -e - -cfg=trainer_config.py -# pass choice -model="output/pass-00000" -paddle train \ - --config=$cfg \ - --use_gpu=false \ - --job=test \ - --init_model_path=$model \ - --config_args=is_predict=1 \ - --predict_output_dir=. - -python gen_result.py > result.csv - -rm -rf rank-00000 diff --git a/v1_api_demo/traffic_prediction/train.sh b/v1_api_demo/traffic_prediction/train.sh deleted file mode 100755 index 48dfc5604f80042598c5c779bd450a5808fdfb64..0000000000000000000000000000000000000000 --- a/v1_api_demo/traffic_prediction/train.sh +++ /dev/null @@ -1,27 +0,0 @@ -#!/bin/bash -# Copyright (c) 2016 PaddlePaddle Authors, Inc. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -set -e - -cfg=trainer_config.py -paddle train \ - --config=$cfg \ - --save_dir=./output \ - --trainer_count=4 \ - --log_period=1000 \ - --dot_period=10 \ - --num_passes=10 \ - --use_gpu=false \ - --show_parameter_stats_period=3000 \ - 2>&1 | tee 'train.log' diff --git a/v1_api_demo/traffic_prediction/trainer_config.py b/v1_api_demo/traffic_prediction/trainer_config.py deleted file mode 100755 index 52d678624aff7ca2264c3c20e320004217d14397..0000000000000000000000000000000000000000 --- a/v1_api_demo/traffic_prediction/trainer_config.py +++ /dev/null @@ -1,52 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors, Inc. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -from paddle.trainer_config_helpers import * - -################################### DATA Configuration ############################################# -is_predict = get_config_arg('is_predict', bool, False) -trn = './data/train.list' if not is_predict else None -tst = './data/test.list' if not is_predict else './data/pred.list' -process = 'process' if not is_predict else 'process_predict' -define_py_data_sources2( - train_list=trn, test_list=tst, module="dataprovider", obj=process) -################################### Parameter Configuaration ####################################### -TERM_NUM = 24 -FORECASTING_NUM = 24 -emb_size = 16 -batch_size = 128 if not is_predict else 1 -settings( - batch_size=batch_size, - learning_rate=1e-3, - learning_method=RMSPropOptimizer()) -################################### Algorithm Configuration ######################################## - -output_label = [] - -link_encode = data_layer(name='link_encode', size=TERM_NUM) -for i in xrange(FORECASTING_NUM): - # Each task share same weight. - link_param = ParamAttr( - name='_link_vec.w', initial_max=1.0, initial_min=-1.0) - link_vec = fc_layer(input=link_encode, size=emb_size, param_attr=link_param) - score = fc_layer(input=link_vec, size=4, act=SoftmaxActivation()) - if is_predict: - maxid = maxid_layer(score) - output_label.append(maxid) - else: - # Multi-task training. - label = data_layer(name='label_%dmin' % ((i + 1) * 5), size=4) - cls = classification_cost( - input=score, name="cost_%dmin" % ((i + 1) * 5), label=label) - output_label.append(cls) -outputs(output_label) diff --git a/v1_api_demo/vae/README.md b/v1_api_demo/vae/README.md deleted file mode 100644 index e55d483b023773900729622a6cac44116fc79c76..0000000000000000000000000000000000000000 --- a/v1_api_demo/vae/README.md +++ /dev/null @@ -1,13 +0,0 @@ -#Variational Autoencoder (VAE) - -This demo implements VAE training described in the original paper (https://arxiv.org/abs/1312.6114). - - -In order to run the model, first download the MNIST dataset by running the shell script in ./data. - -Then you can run the command below. The flag --useGpu specifies whether to use gpu for training (0 is cpu, 1 is gpu). - -$python vae_train.py [--use_gpu 1] - -The generated images will be stored in ./samples/ -The corresponding models will be stored in ./params/ diff --git a/v1_api_demo/vae/data/get_mnist_data.sh b/v1_api_demo/vae/data/get_mnist_data.sh deleted file mode 100755 index a77c81bf5af9ddb6634ff89460797ca543c5e517..0000000000000000000000000000000000000000 --- a/v1_api_demo/vae/data/get_mnist_data.sh +++ /dev/null @@ -1,17 +0,0 @@ -#!/usr/bin/env sh -# This script downloads the mnist data and unzips it. -set -e -DIR="$( cd "$(dirname "$0")" ; pwd -P )" -rm -rf "$DIR/mnist_data" -mkdir "$DIR/mnist_data" -cd "$DIR/mnist_data" - -echo "Downloading..." - -for fname in train-images-idx3-ubyte train-labels-idx1-ubyte t10k-images-idx3-ubyte t10k-labels-idx1-ubyte -do - if [ ! -e $fname ]; then - wget --no-check-certificate http://yann.lecun.com/exdb/mnist/${fname}.gz - gunzip ${fname}.gz - fi -done diff --git a/v1_api_demo/vae/dataloader.py b/v1_api_demo/vae/dataloader.py deleted file mode 100644 index e9ff95d44f825cd941b5687f754618e66d491e7f..0000000000000000000000000000000000000000 --- a/v1_api_demo/vae/dataloader.py +++ /dev/null @@ -1,60 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import numpy as np - - -class MNISTloader(): - def __init__(self, - data_path="./data/mnist_data/", - batch_size=60, - process='train'): - self.batch_size = batch_size - self.data_path = data_path - self._pointer = 0 - self.image_batches = np.array([]) - self.process = process - - def _extract_images(self, filename, n): - f = open(filename, 'rb') - f.read(16) - data = np.fromfile(f, 'ubyte', count=n * 28 * 28).reshape((n, 28 * 28)) - #Mapping data into [-1, 1] - data = data / 255. * 2. - 1 - data_batches = np.split(data, 60000 / self.batch_size, 0) - - f.close() - - return data_batches - - @property - def pointer(self): - return self._pointer - - def load_data(self): - TRAIN_IMAGES = '%s/train-images-idx3-ubyte' % self.data_path - TEST_IMAGES = '%s/t10k-images-idx3-ubyte' % self.data_path - - if self.process == 'train': - self.image_batches = self._extract_images(TRAIN_IMAGES, 60000) - else: - self.image_batches = self._extract_images(TEST_IMAGES, 10000) - - def next_batch(self): - batch = self.image_batches[self._pointer] - self._pointer = (self._pointer + 1) % (60000 / self.batch_size) - return np.array(batch) - - def reset_pointer(self): - self._pointer = 0 diff --git a/v1_api_demo/vae/vae_conf.py b/v1_api_demo/vae/vae_conf.py deleted file mode 100644 index 301dd23793d19ec5946cc7bb07e32c53c04a972b..0000000000000000000000000000000000000000 --- a/v1_api_demo/vae/vae_conf.py +++ /dev/null @@ -1,116 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from paddle.trainer_config_helpers import * -import numpy as np - -is_generating = get_config_arg("is_generating", bool, False) - -settings(batch_size=32, learning_rate=1e-3, learning_method=AdamOptimizer()) - -X_dim = 28 * 28 -h_dim = 128 -z_dim = 100 - - -def reparameterization(mu, logvar): - eps = ParamAttr(initial_mean=0., initial_std=1) - with mixed_layer() as sigma: - sigma += dotmul_projection(layer_math.exp(logvar) * 0.5, param_attr=eps) - return mu + sigma - - -def q_func(X): - """ - xavier initialization - """ - param_attr = ParamAttr( - name='share.w', initial_mean=0., initial_std=1. / np.sqrt(X_dim / 2.)) - mu_param = ParamAttr( - name='mu.w', initial_mean=0., initial_std=1. / np.sqrt(h_dim / 2.)) - logvar_param = ParamAttr( - name='logvar.w', initial_mean=0., initial_std=1. / np.sqrt(h_dim / 2.)) - - bias_attr = ParamAttr(name='share.bias', initial_mean=0., initial_std=0.) - mu_bias = ParamAttr(name='mu.bias', initial_mean=0., initial_std=0.) - logvar_bias = ParamAttr(name='logvar.bias', initial_mean=0., initial_std=0.) - - share_layer = fc_layer( - X, - size=h_dim, - param_attr=param_attr, - bias_attr=bias_attr, - act=ReluActivation()) - - return (fc_layer( - share_layer, - size=z_dim, - param_attr=mu_param, - bias_attr=mu_bias, - act=LinearActivation()), fc_layer( - share_layer, - size=z_dim, - param_attr=logvar_param, - bias_attr=logvar_bias, - act=LinearActivation())) - - -def generator(z): - - hidden_param = ParamAttr( - name='hidden.w', initial_mean=0., initial_std=1. / np.sqrt(z_dim / 2.)) - hidden_bias = ParamAttr(name='hidden.bias', initial_mean=0., initial_std=0.) - prob_param = ParamAttr( - name='prob.w', initial_mean=0., initial_std=1. / np.sqrt(h_dim / 2.)) - prob_bias = ParamAttr(name='prob.bias', initial_mean=0., initial_std=0.) - - hidden_layer = fc_layer( - z, - size=h_dim, - act=ReluActivation(), - param_attr=hidden_param, - bias_attr=hidden_bias) - prob = fc_layer( - hidden_layer, - size=X_dim, - act=SigmoidActivation(), - param_attr=prob_param, - bias_attr=prob_bias) - - return prob - - -def reconstruct_error(prob, X): - cost = multi_binary_label_cross_entropy(input=prob, label=X) - return cost - - -def KL_loss(mu, logvar): - with mixed_layer() as mu_square: - mu_square += dotmul_operator(mu, mu, scale=1.) - - cost = 0.5 * sum_cost(layer_math.exp(logvar) + mu_square - 1. - logvar) - - return cost - - -if not is_generating: - x_batch = data_layer(name='x_batch', size=X_dim) - mu, logvar = q_func(x_batch) - z_samples = reparameterization(mu, logvar) - prob = generator(z_samples) - outputs(reconstruct_error(prob, x_batch) + KL_loss(mu, logvar)) -else: - z_samples = data_layer(name='noise', size=z_dim) - outputs(generator(z_samples)) diff --git a/v1_api_demo/vae/vae_train.py b/v1_api_demo/vae/vae_train.py deleted file mode 100644 index 1babb011c77b92861cc680a2e1aaa8c9ae5d97b5..0000000000000000000000000000000000000000 --- a/v1_api_demo/vae/vae_train.py +++ /dev/null @@ -1,175 +0,0 @@ -# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import argparse -import random -import numpy as np -import cPickle -import sys, os -from PIL import Image - -from paddle.trainer.config_parser import parse_config -from paddle.trainer.config_parser import logger -import py_paddle.swig_paddle as api -import dataloader -import matplotlib.pyplot as plt - - -def plot_samples(samples): - fig = plt.figure(figsize=(4, 4)) - gs = gridspec.GridSpec(4, 4) - gs.update(wspace=0.05, hspace=0.05) - for i, sample in enumerate(samples): - plt.subplot(gs[i]) - plt.axis('off') - plt.imshow(sample.reshape(28, 28), cmap='Greys_r') - - return fig - - -def CHECK_EQ(a, b): - assert a == b, "a=%s, b=%s" % (a, b) - - -def get_fake_samples(generator_machine, batch_size, noise): - gen_inputs = api.Arguments.createArguments(1) - gen_inputs.setSlotValue(0, api.Matrix.createDenseFromNumpy(noise)) - gen_outputs = api.Arguments.createArguments(0) - generator_machine.forward(gen_inputs, gen_outputs, api.PASS_TEST) - fake_samples = gen_outputs.getSlotValue(0).copyToNumpyMat() - return fake_samples - - -def copy_shared_parameters(src, dst): - ''' - copy the parameters from src to dst - :param src: the source of the parameters - :type src: GradientMachine - :param dst: the destination of the parameters - :type dst: GradientMachine - ''' - src_params = [src.getParameter(i) for i in xrange(src.getParameterSize())] - src_params = dict([(p.getName(), p) for p in src_params]) - - for i in xrange(dst.getParameterSize()): - dst_param = dst.getParameter(i) - src_param = src_params.get(dst_param.getName(), None) - if src_param is None: - continue - src_value = src_param.getBuf(api.PARAMETER_VALUE) - dst_value = dst_param.getBuf(api.PARAMETER_VALUE) - CHECK_EQ(len(src_value), len(dst_value)) - dst_value.copyFrom(src_value) - dst_param.setValueUpdated() - - -def find(iterable, cond): - for item in iterable: - if cond(item): - return item - return None - - -def get_layer_size(model_conf, layer_name): - layer_conf = find(model_conf.layers, lambda x: x.name == layer_name) - assert layer_conf is not None, "Cannot find '%s' layer" % layer_name - return layer_conf.size - - -def main(): - parser = argparse.ArgumentParser() - parser.add_argument( - "--use_gpu", default="1", help="1 means use gpu for training") - parser.add_argument("--gpu_id", default="0", help="the gpu_id parameter") - args = parser.parse_args() - use_gpu = args.use_gpu - assert use_gpu in ["0", "1"] - - if not os.path.exists("./samples/"): - os.makedirs("./samples/") - - if not os.path.exists("./params/"): - os.makedirs("./params/") - - api.initPaddle('--use_gpu=' + use_gpu, '--dot_period=10', - '--log_period=1000', '--gpu_id=' + args.gpu_id, - '--save_dir=' + "./params/") - - conf = "vae_conf.py" - - trainer_conf = parse_config(conf, "is_generating=False") - gener_conf = parse_config(conf, "is_generating=True") - - batch_size = trainer_conf.opt_config.batch_size - - noise_dim = get_layer_size(gener_conf.model_config, "noise") - - mnist = dataloader.MNISTloader(batch_size=batch_size) - mnist.load_data() - - training_machine = api.GradientMachine.createFromConfigProto( - trainer_conf.model_config) - - generator_machine = api.GradientMachine.createFromConfigProto( - gener_conf.model_config) - - trainer = api.Trainer.create(trainer_conf, training_machine) - - trainer.startTrain() - - for train_pass in xrange(100): - trainer.startTrainPass() - mnist.reset_pointer() - i = 0 - it = 0 - while mnist.pointer != 0 or i == 0: - X = mnist.next_batch().astype('float32') - - inputs = api.Arguments.createArguments(1) - inputs.setSlotValue(0, api.Matrix.createDenseFromNumpy(X)) - - trainer.trainOneDataBatch(batch_size, inputs) - - if it % 1000 == 0: - - outputs = api.Arguments.createArguments(0) - training_machine.forward(inputs, outputs, api.PASS_TEST) - loss = np.mean(outputs.getSlotValue(0).copyToNumpyMat()) - print "\niter: {}".format(str(it).zfill(3)) - print "VAE loss: {}".format(str(loss).zfill(3)) - - #Sync parameters between networks (GradientMachine) at the beginning - copy_shared_parameters(training_machine, generator_machine) - - z_samples = np.random.randn(batch_size, - noise_dim).astype('float32') - samples = get_fake_samples(generator_machine, batch_size, - z_samples) - - #Generating the first 16 images for a picture. - figure = plot_samples(samples[:16]) - plt.savefig( - "./samples/{}_{}.png".format( - str(train_pass).zfill(3), str(i).zfill(3)), - bbox_inches='tight') - plt.close(figure) - i += 1 - it += 1 - - trainer.finishTrainPass() - trainer.finishTrain() - - -if __name__ == '__main__': - main()