Deep learning：四十四(Pylearn2中的Quick-start例子)

　　前言：

　　听说Pylearn2是个蛮适合搞深度学习的库，它建立在Theano之上，支持GPU(估计得以后工作才玩这个，现在木有这个硬件条件)运算，由DL大牛Bengio小组弄出来的，再加上Pylearn2里面已经集成了一部分常见的DL算法，本着很想读读这些算法的源码和细节这一想法，打算学习下Pylearn2的使用. 网上这方面的中文资料简直是太少了，虽然本博文没什么实质内容，但也写贴出来，说不定可以帮到一些初学者。

　　从Bengio的一篇paper: Pylearn2: a machine learning research library可以看出，Pylearn2主要是针对机器学习开发者而设计的（说明使用该库的人需要有一定的机器学习背景知识），利用Pylearn2可以灵活设计自己的机器学习模型和算法，可扩展性较强（具体怎么弄？？）。而根据Pylearn2库的特征(官网)上的介绍可知，在pylearn2里，有一些常见的数据模块、模型模块、训练算法模块。数据模块中有常见的MNIST, CIFAR10, CIFAR100, STL10, NORB等。DL模型模块包含：RBM系列，AutoEncoder系列，LCC, maxout等。训练算法模块主要是SGD系列。

　　Pylearn2安装简单介绍：

　　好吧，进入正题。首先是库的安装，我是运行在64bit-ubuntu13.10上的。

　　1. 在此之前还需安装Theano(python下进行符号运算的库，类似Numpy，但在多维矩阵处理上功能更强)，安装Theano的方法请参考：Installing Theano（Bleeding-edge install instruction），里面有Ubuntu下安装的链接，按照里面的步骤一步步进行下去就行（期间遇到的各种问题多google吧！）。需要提一下的是，安装成功后，我们需要将Theano升级到开发版本Bleeding-edge下，因为后面的Pylearn2用到了开发版Theano的新特征。具体的升级方法参考网页中的Bleeding-edge install instructions小节。

　　2. Pylearn2的安装可以参考博文pylearn2安装及测试（lucktroy的csdn博客）。主要有3步：

　　a. 在想安装Pylearn2的目录下打开vim,输入命令：

　　git clone git://github.com/lisa-lab/pylearn2.git

　　b. 配置Pylearn2所用数据目录的环境变量（做一些标准实验时，可将数据放入该目录），即在vim里输入命令行:vim ~/.bashrc ,然后在打开的.bashrc文件最后一行加入语句:export PYLEARN2_DATA_PATH=YourPath/data 保存后退出。其中的YourPath为你想放入数据的目录全称。接着在vim里执行source ~/.vimrc命令

　　c. 进入pylearn2目录(刚用git下载后会有该文件的)，执行命令:python setup.py build.

　　运行Quick-start例子：

　　安装完Pylearn2后就想弄个sample爽一把，选的是GRBM算法例子，可参考官网的Quick-start example教程。这个例子中主要有3个步骤（如果实验过程中出现一些问题，可以参考下本博文的附录，看能否提供一些帮助）：

　　步骤一：创建数据。

　　在YourPath/pylearn2/scripts/tutorials/grbm_smd/ 目录下执行下列命令：python make_dataset.py

　　从make_dataset.py的源码中可以看出，这里使用的是CIFAR10图片库（http://www.cs.toronto.edu/~kriz/cifar.html（CIFAR10数据库）），为32*32大小的彩色图片，共5w个训练样本和1w个测试样本。训练grbm的patch大小为8*8的，有15w个patch。当然还对该图片库进行了一些预处理，比如ZCA白化等等。最后将预处理好的结果保存为pickle文件（pickle是python中用于序列处理的模块，保存数据为.pkl格式到硬盘，下次要使用该数据时可重新加载）：cifar10_preprocessed_train.pkl.

　　步骤二：GRBM模型参数的训练。

　　使用的命令（还是在原来的目录下）为：python ../../ train.py cifar_grbm_smd.yaml

　　其中的cifar_grbm_smd.yaml文件是该实验的配置文件，需要配置数据，模型，算法3个模块的一些参数，yaml文件是我们与pylearn2打交道的文件，如果是使用常见的深度学习模型和常见的优化算法来做实验的话，则只需把配置好这个.yaml文件就可以了。这可以简化不少工作。下面来看看这个cifar_grbm_smd.yaml的代码及一些注释，关于yaml语法的简单介绍可参考：YAML for Pylearn2. 另外，如果想了解GRBM，则可参考网友博文：DeepLearning（深度学习）原理与实现（四），写得很不错。

# pylearn2 tutorial example: cifar_grbm_smd.yaml by Ian Goodfellow
#
# Read the README file before reading this file
#
# This is an example of yaml file, which is the main way that an experimenter
# interacts with pylearn2.
#
# A yaml file is very similar to a python dictionary, with a bit of extra
# syntax.

# The !obj tag allows us to create a specific class of object. The text after
# the : indicates what class should be loaded. This is followed by a pair of
# braces containing the arguments to that class's __init__ method.
#
# Here, we allocate a Train object, which represents the main loop of the
# training script. The train script will run this loop repeatedly. Each time
# through the loop, the model is trained on data from a training dataset, then
# saved to file.

!obj:pylearn2.train.Train {
    # The !pkl tag is used to create an object from a pkl file. Here we retrieve
    # the dataset made by make_dataset.py and use it as our training dataset.
    dataset: !pkl: "cifar10_preprocessed_train.pkl",

    # Next we make the model to be trained. It is a Binary Gaussian RBM
    model: !obj:pylearn2.models.rbm.GaussianBinaryRBM {

        # The RBM needs 192 visible units (its inputs are 8x8 patches with 3
        # color channels)
        nvis : 192,

        # We'll use 400 hidden units for this RBM. That's a small number but we
        # want this example script to train quickly.
        nhid : 400,

        # The elements of the weight matrices of the RBM will be drawn
        # independently from U(-0.05, 0.05)
        irange : 0.05,

        # There are many ways to parameterize a GRBM. Here we use a
        # parameterization that makes the correspondence to denoising
        # autoencoders more clear.
        energy_function_class : !obj:pylearn2.energy_functions.rbm_energy.grbm_type_1 {},

        # Some learning algorithms are capable of estimating the standard
        # deviation of the visible units of a GRBM successfully, others are not
        # and just fix the standard deviation to 1.  We're going to show off
        # and learn the standard deviation.
        learn_sigma : True,

        # Learning works better if we provide a smart initialization for the
        # parameters.  Here we start sigma at .4 , which is about the same
        # standard deviation as the training data. We start the biases on the
        # hidden units at -2, which will make them have fairly sparse
        # activations.
        init_sigma : .4,
        init_bias_hid : -2.,

        # Some GRBM training algorithms can't handle the visible units being
        # noisy and just use their mean for all computations. We will show off
        # and not use that hack here.
        mean_vis : False,

        # One hack we will make is we will scale back the gradient steps on the
        # sigma parameter. This way we don't need to worry about sigma getting
        # too small prematurely (if it gets too small too fast the learning
        # signal gets weak).
        sigma_lr_scale : 1e-3

    },

    # Next we need to specify the training algorithm that will be used to train
    # the model.  Here we use stochastic gradient descent.

    algorithm: !obj:pylearn2.training_algorithms.sgd.SGD {
        # The learning rate determines how big of steps the learning algorithm
        # takes.  Here we use fairly big steps initially because we have a
        # learning rate adjustment scheme that will scale them down if
        # necessary.
        learning_rate : 1e-1,

        # Each gradient step will be based on this many examples
        batch_size : 5,

        # We'll monitor our progress by looking at the first 20 batches of the
        # training dataset. This is an estimate of the training error. To be
        # really exhaustive, we could use the entire training set instead,
        # or to avoid overfitting, we could use held out data instead.
        monitoring_batches : 20,

        monitoring_dataset : !pkl: "cifar10_preprocessed_train.pkl",

        # Here we specify the objective function that stochastic gradient
        # descent should minimize.  In this case we use denoising score
        # matching, which makes this RBM behave as a denoising autoencoder.
        # See
        # Pascal Vincent. "A Connection Between Score Matching and Denoising
        # Auutoencoders." Neural Computation, 2011
        # for details.

        cost : !obj:pylearn2.costs.ebm_estimation.SMD {

            # Denoising score matching uses a corruption process to transform
            # the raw data.  Here we use additive gaussian noise.

            corruptor : !obj:pylearn2.corruption.GaussianCorruptor {
                    stdev : 0.4
            },
        },

        # We'll use the monitoring dataset to figure out when to stop training.
        #
        # In this case, we stop if there is less than a 1% decrease in the
        # training error in the last epoch.  You'll notice that the learned
        # features are a bit noisy. If you'd like nice smooth features you can
        # make this criterion stricter so that the model will train for longer.
        # (setting N to 10 should make the weights prettier, but will make it
        # run a lot longer)

        termination_criterion : !obj:pylearn2.termination_criteria.MonitorBased {
            prop_decrease : 0.01,
            N : 1,
        },

        # Let's throw a learning rate adjuster into the training algorithm.
        # To do this we'll use an "extension," which is basically an event
        # handler that can be registered with the Train object.
        # This particular one is triggered on each epoch.
        # It will shrink the learning rate if the objective goes up and increase
        # the learning rate if the objective decreases too slowly. This makes
        # our learning rate hyperparameter less important to get right.
        # This is not a very mathematically principled approach, but it works
        # well in practice.
        },
    extensions : [!obj:pylearn2.training_algorithms.sgd.MonitorBasedLRAdjuster {}],
    #Finally, request that the model be saved after each epoch
    save_freq : 1
}

　　由上面的yaml文件可知，yaml中的内容有点类似python中的字典：一个关键字key对应一个值value。而这些key都是对应类的构造函数__init__()中的参数，也就是说将这些value传入到这些构造函数中，并由其对象接收。上面yaml代码中data来源于步骤一的cifar10_preprocessed_train.pkl文件。model来源于pylearn2库下的pylearn2.models.rbm.GaussianBinaryRBM类，而algorithm来源于pylearn2库下的pylearn2.training_algorithms.sgd.SGD类。

　　当.yaml文件都配置好后，我们就需要启动对应的程序来训练参数了，train.py就是执行的这个功能的，其代码为：

#!/usr/bin/env python
"""
Script implementing the logic for training pylearn2 models.

This is intended to be a "driver" for most training experiments. A user
specifies an object hierarchy in a configuration file using a dictionary-like
syntax and this script takes care of the rest.

For example configuration files that are consumable by this script, see

    pylearn2/scripts/train_example
    pylearn2/scripts/autoencoder_example
"""
__authors__ = "Ian Goodfellow"
__copyright__ = "Copyright 2010-2012, Universite de Montreal"
__credits__ = ["Ian Goodfellow", "David Warde-Farley"]
__license__ = "3-clause BSD"
__maintainer__ = "Ian Goodfellow"
__email__ = "goodfeli@iro"
# Standard library imports
import argparse
import gc
import logging
import os

# Third-party imports
import numpy as np

# Local imports
from pylearn2.utils import serial
from pylearn2.utils.logger import (
    CustomStreamHandler, CustomFormatter, restore_defaults
)

class FeatureDump(object):
    def __init__(self, encoder, dataset, path, batch_size=None, topo=False):
        self.encoder = encoder
        self.dataset = dataset
        self.path = path
        self.batch_size = batch_size
        self.topo = topo

    def main_loop(self):
        if self.batch_size is None:
            if self.topo:
                data = self.dataset.get_topological_view()
            else:
                data = self.dataset.get_design_matrix()
            output = self.encoder.perform(data)
        else:
            myiterator = self.dataset.iterator(mode='sequential',
                                               batch_size=self.batch_size,
                                               topo=self.topo)
            chunks = []
            for data in myiterator:
                chunks.append(self.encoder.perform(data))
            output = np.concatenate(chunks)
        np.save(self.path, output)

def make_argument_parser():
    parser = argparse.ArgumentParser(
        description="Launch an experiment from a YAML configuration file.",
        epilog='\n'.join(__doc__.strip().split('\n')[1:]).strip(),
        formatter_class=argparse.RawTextHelpFormatter
    ) #parser是用来接收参数的
    parser.add_argument('--level-name', '-L',
                        action='store_true',
                        help='Display the log level (e.g. DEBUG, INFO) '
                             'for each logged message')
    parser.add_argument('--timestamp', '-T',
                        action='store_true',
                        help='Display human-readable timestamps for '
                             'each logged message')
    parser.add_argument('--verbose-logging', '-V',
                        action='store_true',
                        help='Display timestamp, log level and source '
                             'logger for every logged message '
                             '(implies -T).')
    parser.add_argument('--debug', '-D',
                        action='store_true',
                        help='Display any DEBUG-level log messages, '
                             'suppressed by default.')
    parser.add_argument('config', action='store', #按照格式输入参数，比如这里的输入的参数会保存在config中
                        choices=None,
                        help='A YAML configuration file specifying the '
                             'training procedure')
    return parser

if __name__ == "__main__":
    parser = make_argument_parser()
    args = parser.parse_args() #读取传入进来的参数，这里是直接在命令行读取该文件,参数放入args.config中
    train_obj = serial.load_train_file(args.config) #serial.load_train_file()函数最后返回的是：
　　　　　　　　　　# return yaml_parse.load_path(args.config) 也就是说调用的是ymal_parse.load_path()函数。返回的是一个train类的对象。
　　　　　　　　　　# 其中的ymal_parse是pylearn2.config中的函数。

return yaml_parse.load_path(config_file_path)
    try:
        iter(train_obj) #iter()是个迭代器函数
        iterable = True
    except TypeError as e:
        iterable = False

    # Undo our custom logging setup.
    restore_defaults()
    # Set up the root logger with a custom handler that logs stdout for INFO
    # and DEBUG and stderr for WARNING, ERROR, CRITICAL.
    root_logger = logging.getLogger() #logging主要是python中用于处理日志的模块,这里是返回一个logger实例，由于没有指定name，所以是root logger
    if args.verbose_logging:
        formatter = logging.Formatter(fmt="%(asctime)s %(name)s %(levelname)s "
                                          "%(message)s")
        handler = CustomStreamHandler(formatter=formatter)
    else:
        if args.timestamp:
            prefix = '%(asctime)s '
        else:
            prefix = '' #这里为空
        formatter = CustomFormatter(prefix=prefix, only_from='pylearn2')
        handler = CustomStreamHandler(formatter=formatter)
    root_logger.addHandler(handler) #给root_lgger添加handler来帮助处理日志
    # Set the root logger level.
    if args.debug:
        root_logger.setLevel(logging.DEBUG)
    else:
        root_logger.setLevel(logging.INFO) #给root_logger设置级别，为INFO级别，因为每个日志消息都会关联一个级别

    if iterable: #enumerate()为对一个list或者array既要遍历索引又要遍历元素时使用
        for number, subobj in enumerate(iter(train_obj)):#train_obj里面装的是ymal文件内容，类似字典
            # Publish a variable indicating the training phase.
            phase_variable = 'PYLEARN2_TRAIN_PHASE'
            phase_value = 'phase%d' % (number + 1)
            os.environ[phase_variable] = phase_value
            os.putenv(phase_variable, phase_value)

            # Execute this training phase.
            subobj.main_loop()

            # Clean up, in case there's a lot of memory used that's
            # necessary for the next phase.
            del subobj
            gc.collect()
    else:
        train_obj.main_loop() #因为train_obj中已经包含了数据，模型，算法，所以调用main_loop()后表示采用对应算法用对应数据在对应的模型上训练
                              #直到满足迭代终止条件

　　其中最核心的就是main_loop()函数了，在调用main_loop()后，程序会自动用algorithm对象使用model对象在data上来训练参数了。至于具体该函数是怎样将data, model, algorithm联系起来的呢？我们可以试着去读一下源码：

　　首先是由train_obj.main_loop()函数将data, model, algorithm联系起来的。从名字train_obj可以看出它是一个某个类的对象，猜测应该是Pylearn2下的Train类对象，因为在库Pylearn2的子目录下有个model为train.py，该文件有个Train类，并且这个Train类有一个方法：main_loop()。看来一切符合猜测，那么是否真是的呢？

　　首先来看看train_obj从哪里来的(因为main_loop()是由train_obj来调用的)。由上面的程序可知：train_obj = serial.load_train_file(args.config), 需要跟踪serial, 找到serial.load_train_file()的源代码，最后一句为：return yaml_parse.load_path(args.config). 继续跟踪发现load_path()函数里面调用了load()函数，而里面最调用的是yaml.load()函数，由源码中的注释可知它是将.yaml配置文件转换成一个graph, 而这个graph应该就是一个Train对象...

　　好吧，到了该看main_loop()的内容了：

    def main_loop(self):
        """
        Repeatedly runs an epoch of the training algorithm, runs any
        epoch-level callbacks, and saves the model.
        """
        if self.algorithm is None:
            self.model.monitor = Monitor.get_monitor(self.model)
            self.setup_extensions()
            self.run_callbacks_and_monitoring()
            while True:
                rval = self.model.train_all(dataset=self.dataset)
                if rval is not None:
                    raise ValueError("Model.train_all should not return anything. Use Model.continue_learning to control whether learning continues.")
                self.model.monitor.report_epoch()
                if self.save_freq > 0 and self.model.monitor.epochs_seen % self.save_freq == 0:
                    self.save()
                continue_learning = self.model.continue_learning()
                assert continue_learning in [True, False, 0, 1]
                if not continue_learning:
                    break
        else:
            self.algorithm.setup(model=self.model, dataset=self.dataset) #这一句将data,model, dataset联系起来了
            self.setup_extensions() #和.yaml文件中的extensions项联系起来了
            if not hasattr(self.model, 'monitor'):
                # TODO: is this really necessary? I just put this error here
                # to prevent an AttributeError later, but I think we could
                # rewrite to avoid the AttributeError
                raise RuntimeError("The algorithm is responsible for setting"
                        " up the Monitor, but failed to.")
            if len(self.model.monitor._datasets)>0:
                # This monitoring channel keeps track of a shared variable,
                # which does not need inputs nor data.
                self.model.monitor.add_channel(name="monitor_seconds_per_epoch",
                                               ipt=None,
                                               val=self.monitor_time,
                                               data_specs=(NullSpace(), ''),
                                               dataset=self.model.monitor._datasets[0])
            self.run_callbacks_and_monitoring()
            while True: #循环中，直到满足终止条件
                with log_timing(log, None, final_msg='Time this epoch:',
                                callbacks=[self.monitor_time.set_value]):
                    rval = self.algorithm.train(dataset=self.dataset) #算法训练的核心函数
                if rval is not None:
                    raise ValueError("TrainingAlgorithm.train should not return anything. Use TrainingAlgorithm.continue_learning to control whether learning continues.")
                self.model.monitor.report_epoch()
                self.run_callbacks_and_monitoring()
                if self.save_freq > 0 and self.model.monitor._epochs_seen % self.save_freq == 0:
                    self.save()
                continue_learning =  self.algorithm.continue_learning(self.model) #终止条件测试
                assert continue_learning in [True, False, 0, 1]
                if not continue_learning:
                    break

        self.model.monitor.training_succeeded = True

        if self.save_freq > 0:
            self.save()

　　步骤三:

　　这部分就是看结果显示了，执行命令：python ../../show_weights.py cifar_grbm_smd.pkl 比如我这里执行后的结果显示如下：

　　

　　当然了你还可以使用plot_monitor.py来看一些对应结果。

　　总结：

　　当使用Pylearn2中已有的一些DL模型，采用其中已有的一些优化算法来做实验时，我们只需要配置好实验的.yaml文件即可，调参过程就是不断更改.ymal中的配置。但是如果需要采用自己新提出来的DL模型，或者采用自己新提出的目标函数及优化方法，则还需要自己写出对应的类，具体这部分该怎么做（比如说怎样去实现这个类，接口怎样设计，.ymal文件需要更改哪些地方等），本人暂时没任何经验。希望懂这些的可以大家可贡献贡献下想法，交流交流下。网上有个教程是把Pylearn2当做通常的python库来用，实现了一个异或网络，很不错，见：Neural network example using Pylearn2.

　　另外，分析Pylearn2的源码可知，每个algorithm中，必须有下面4个函数:__init(), setup(), train(),  continue_training(), 作用分别为构造函数, 根据model建立网络的结构，模型参数的训练，模型训练终止处理。model模块中，应该也有一些统一的函数。

　　附录：

　　我实验过程中可能出现的一些错误处理：

A:

　　如果执行 python make_dataset.py后出现错误：

　raise IOError("permission error creating %s" % filepath) IOError: permission error creating cifar10_preprocessed_train.pkl

　　看错误提示应该是权限问题，这时改为命令：

　　sudo python make_dataset.py

　　如果继续出现错误：

　　pylearn2.datasets.exc.NoDataPathError: You need to define your PYLEARN2_DATA_PATH environment variable. If you are using a computer at LISA, this should be set to /data/lisa/data.

　　说明PYLEARN2_DATA_PATH环境变量没有设置，但是前面却是设置了啊！为什么呢？有可能是你设置环境变量时用的是root权限，而执行该命令只是普通用户。如果切换到root下再执行 root#:python make_dataset.py成功！生成了cifar10_preprocessed_train.pkl

　　但是后面执行：../../train.py cifar_grbm_smd.yaml出现错误：ImportError: Could not import pylearn2.models but could import pylearn2. Original exception: No module named compat.python2x

　　到这里基本可以确定是权限问题，解决方法是：重新用普通用户安装了下pylearn2,设置好环境变量，放着好下载的数据后，执行（普通用户下）：

　　python make_dataset

　　则成功生成了cifar10_preprocessed_train.pkl 可恶的是后续的../../train.py cifar_grbm_smd.yaml还是会出现刚刚的错误。

　　当然了这个问题主要是因为Theano的版本不对，在使用pylearn2时，应该使用development版本的Theano，按照本文前面的方法更新下Theano即可。

B.

　　如果在显示权值阶段，当执行下面命令后：sudo python ../../show_weights.py cifar_grbm_smd.pkl.可能会出现下面提示：

　　You need to choose an image viewer program that pylearn2 should use. Then tell pylearn2 to usethat image viewer program by defining your PYLEARN2_VIEWER_COMMAND environment variable.You need to choose PYLEARN_VIEWER_COMMAND such that running ${PYLEARN2_VIEWER_COMMAND} image.png

in a command prompt on your machine will do the following:

-open an image viewer in a new process.

-not return until you have closed the image.

Acceptable commands include:

gwenview

eog --new-instance

This is assuming that you have gwenview or a version of eog that supports --new-instance

......

……

　　这说明pylearn2中没有指定图片显示的软件。首先安装gwenview软件：sudo apt-get Install gwenview.

　　然后设置一下PYLEARN2_VIEWER_COMMAND环境变量。vim ~/.bashrc 在最后一行加入gwenview的安装目录，比如我按照默认的安装目录加入的为：

　　export PYLEARN2_VIEWER_COMMAND=/usr/bin/gwenview

　　保存好后执行source ~/.bashrc

　  参考资料：

Pylearn2: a machine learning research library

Pylearn2库的特征(官网)

Installing Theano（Bleeding-edge install instruction）

pylearn2安装及测试（lucktroy的csdn博客）

Quick-start example

http://www.cs.toronto.edu/~kriz/cifar.html（CIFAR10数据库）

YAML for Pylearn2

　  DeepLearning（深度学习）原理与实现（四）

Neural network example using Pylearn2.