强的部落格 – 第37页 – 量化自我和极简主义的窝藏点

使用octave运行DeepLearnToolbox

环境： Ubuntu 12.04, octave 3.8.2

DeepLearnToolbox是matlab/octave的深度学习工具箱，包括Deep Belief Nets, Stacked Autoencoders, Convolutional Neural Nets, Convolutional Autoencoders and vanilla Neural Nets。

下面运行下里面的测试程序：

1. 下载DeepLearnToolbox
git clone https://github.com/rasmusbergpalm/DeepLearnToolbox.git

2. 启动Octave

3. 安装DeepLearnToolbox
octave:1> addpath(genpath(‘DeepLearnToolbox’));

4. 将目录./DeepLearnToolbox/tests/test_example_SAE.m的内容复制到octave的命令行内，最后一行增加”end;”

5. 运行刚才定义的函数
octave:8> test_example_SAE;

6. 运行结果出现绘制的图片和执行结果
octave:11> test_example_SAE
warning: load: file found in load path
Training AE 1/1
epoch 1/1. Took 21.469 seconds. Mini-batch mean squared error on training set is 10.61; Full-batch train err = 11.049844
epoch 1/1. Took 8.4329 seconds. Mini-batch mean squared error on training set is 0.2197; Full-batch train err = 0.094171

cuda-convnet2使用

环境: Ubuntu 12.04, cuda-convnet2, CUDA 6

安装步骤：

1. 预安装需求库

sudo apt-get install python-dev python-numpy python-scipy python-magic python-matplotlib libatlas-base-dev libjpeg-dev libopencv-dev git

2. 安装CUDA 6.0
从http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1204/x86_64/cuda-repo-ubuntu1204_6.5-14_amd64.deb下载 cuda-repo-ubuntu1204_6.5-14_amd64.deb
$wget http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1204/x86_64/cuda-repo-ubuntu1204_6.5-14_amd64.deb
$ sudo dpkg -i cuda-repo-ubuntu1204_6.0-37_amd64.deb
$ sudo apt-get update
$ sudo apt-get install cuda

3. 配置CUDA环境变量

vi ~/.bashrc

export CUDA_HOME=/usr/local/cuda-6.0
export LD_LIBRARY_PATH=${CUDA_HOME}/lib64

PATH=${CUDA_HOME}/bin:${PATH}
export PATH

4. 下载cuda-convnet2源码

git clone https://code.google.com/p/cuda-convnet2/

5. 编译源码

jerry@hq:/u01/cuda-convnet2$ sh build.sh
mkdir -p ./bin//src
g++ -O3 -c -fPIC -DNUMPY_INTERFACE -I./include -I/usr/include/python2.7 -I/usr/lib/python2.7/dist-packages/numpy/core/include/numpy/ src/matrix.cpp -o ./bin//src/matrix.o
In file included from /usr/include/python2.7/Python.h:8:0,
from src/../include/matrix.h:22,
from src/matrix.cpp:17:
/usr/include/python2.7/pyconfig.h:1161:0: warning: “_POSIX_C_SOURCE” redefined [enabled by default]
#define _POSIX_C_SOURCE 200112L
^
In file included from /usr/include/stdlib.h:25:0,
from src/../include/matrix_funcs.h:20,
from src/../include/matrix.h:20,
from src/matrix.cpp:17:
/usr/include/features.h:164:0: note: this is the location of the previous definition
# define _POSIX_C_SOURCE 200809L
^
In file included from /usr/include/python2.7/Python.h:8:0,
from src/../include/matrix.h:22,
from src/matrix.cpp:17:
/usr/include/python2.7/pyconfig.h:1183:0: warning: “_XOPEN_SOURCE” redefined [enabled by default]
#define _XOPEN_SOURCE 600
^
In file included from /usr/include/stdlib.h:25:0,
from src/../include/matrix_funcs.h:20,
from src/../include/matrix.h:20,
from src/matrix.cpp:17:
/usr/include/features.h:166:0: note: this is the location of the previous definition
# define _XOPEN_SOURCE 700
^
cd ./bin/ && g++ -O3 -DNUMPY_INTERFACE -shared -Wl,-no-undefined -o libutilpy.so src/matrix.o -L/usr/lib/atlas-base -latlas -lcblas -lpython2.7
ln -sf ./bin//libutilpy.so .
mkdir -p ./bin/release
mkdir -p ./obj/release/src
mkdir -p ./bin/release
mkdir -p ./obj/release/src
mkdir -p ./bin/release
mkdir -p ./obj/release/src
mkdir -p ./bin//src
g++ -O3 -c -fPIC -I./include -I/usr/include/python2.7 src/pyext.cpp -o ./bin//src/pyext.o
In file included from /usr/include/python2.7/Python.h:8:0,
from src/../include/pyext.h:23,
from src/pyext.cpp:17:
/usr/include/python2.7/pyconfig.h:1161:0: warning: “_POSIX_C_SOURCE” redefined [enabled by default]
#define _POSIX_C_SOURCE 200112L
^
In file included from /usr/include/stdio.h:28:0,
from src/../include/pyext.h:20,
from src/pyext.cpp:17:
/usr/include/features.h:164:0: note: this is the location of the previous definition
# define _POSIX_C_SOURCE 200809L
^
In file included from /usr/include/python2.7/Python.h:8:0,
from src/../include/pyext.h:23,
from src/pyext.cpp:17:
/usr/include/python2.7/pyconfig.h:1183:0: warning: “_XOPEN_SOURCE” redefined [enabled by default]
#define _XOPEN_SOURCE 600
^
In file included from /usr/include/stdio.h:28:0,
from src/../include/pyext.h:20,
from src/pyext.cpp:17:
/usr/include/features.h:166:0: note: this is the location of the previous definition
# define _XOPEN_SOURCE 700
^
cd ./bin/ && g++ -O3 -shared -Wl,-no-undefined -o _MakeDataPyExt.so src/pyext.o -L/usr/local/cuda/lib64 `pkg-config –libs python` `pkg-config –libs opencv` -lpthread
ln -sf ./bin//_MakeDataPyExt.so .

6. 运行脚本
jerry@hq:/u01/cuda-convnet2$ python convnet.py –data-path=/u01/lisa/data/cifar10/cifar-10-batches-py –save-path=/u01/jerry/tmp –test-range=5 –train-range=1-4 –layer-def=./layers/layers-cifar10-11pct.cfg –layer-params=./layers/layer-params-cifar10-11pct.cfg –data-provider=cifar-cropped –test-freq=13 –epochs=100
Option –gpu (GPU override) not supplied
convnet.py usage:
Option Description Default
[–check-grads <0/1> ] – Check gradients and quit? [0]
[–color-noise <float> ] – Add PCA noise to color channels with given scale [0]
[–conserve-mem <0/1> ] – Conserve GPU memory (slower)? [0]
[–conv-to-local <string,…> ] – Convert given conv layers to unshared local []
[–epochs <int> ] – Number of epochs [50000]
[–feature-path <string> ] – Write test data features to this path (to be used with –write-features) []
[–force-save <0/1> ] – Force save before quitting [0]
[–inner-size <int> ] – Cropped DP: crop size (0 = don’t crop) [0]
[–layer-path <string> ] – Layer file path prefix []
[–load-file <string> ] – Load file []
[–logreg-name <string> ] – Logreg cost layer name (for –test-out) []
[–mini <int> ] – Minibatch size [128]
[–multiview-test <0/1> ] – Cropped DP: test on multiple patches? [0]
[–scalar-mean <float> ] – Subtract this scalar from image (-1 = don’t) [-1]
[–test-freq <int> ] – Testing frequency [57]
[–test-one <0/1> ] – Test on one batch at a time? [1]
[–test-only <0/1> ] – Test and quit? [0]
[–test-out <string> ] – Output test case predictions to given path []
[–unshare-weights <string,…>] – Unshare weight matrices in given layers []
[–write-features <string> ] – Write test data features from given layer []
–data-path <string> – Data path
–data-provider <string> – Data provider
–gpu <int,…> – GPU override
–layer-def <string> – Layer definition file
–layer-params <string> – Layer parameter file
–save-file <string> – Save file override
–save-path <string> – Save path
–test-range <int[-int]> – Data batch range: testing
–train-range <int[-int]> – Data batch range: training

由于ubuntu是在windows里的虚拟机，无法使用本机或外置的gpu显卡，故无法运行程序。有点遗憾

OpenCV训练自己的衣服探测分类器

环境: Ubuntu 12.4, OpenCV, 光影魔术手4, ObjectMarker

步骤如下：

1. 首先收集要探测的物体图片(正样本)和背景图片(负样本)，使用“光影魔术手”批处理图片为固定大小和bmp格式

2. 使用ObjectMarker来抽取物体图片(正样本)生成info.txt，内容如下：
rawdata/136e2b8aef176609829e23e54081db6d.bmp 1 149 36 127 193
rawdata/203d730e146b11af60eef38b5ee280a1.bmp 1 52 92 175 201
rawdata/21b7e18952a3af2975fb407510af943e.bmp 1 82 102 135 182
rawdata/2f9ed3b7e3c87b0890bcdd94fae54415.bmp 1 45 28 190 230

3. 将info.txt文件和正负物体图片导入到Ubuntu环境内。

4. 在背景图片中生成bg.txt文件，内容为相应的文件名：
jerry@hq:~$ more neg_pic/bg3.txt
c15108780ef9f873650f3cbd0259fa6f.bmp
e4f1dce1ed15262848f0b9e0efb0ad56.bmp
6c7fc84c405201ba2a92c38e5f828966.bmp
a00f7dc3c16e8b31d16bb9122f34aa0c.bmp
cad04ea931db04381bb5fed9cb8dea61.bmp
056f0cd1d04d62dfa4ab9fdb0b6c052b.bmp
95f5c539316d9261a83b3a98d90eac58.bmp
08178909a71d88750fc35d5b87e838fd.bmp

4. 准备训练数据
jerry@hq:~$ opencv_createsamples -info info.txt -vec pic.vec -w 30 -h 30 -b neg_pic/bg.txt -num 26

5. 训练haar分类器
jerry@hq:~$ mkdir cascade_data
jerry@hq:~$ opencv_traincascade -data cascade_data -vec pic.vec -bg neg_pic/bg3.txt -w 30 -h 30 -numPos 26 -numNeg 50 -numStages 10

Pylearn2的Stacked Autoencoders示例

环境：Ubuntu 12.4

1. 首先下载训练数据

cd /u01/lisa/data/mnist
wget http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz

gunzip train-images-idx3-ubyte.gz

wget http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz

gunzip train-labels-idx1-ubyte.gz wget http://yann.lecun.com/exdb/mnist/t10k-images-idx3-ubyte.gz gunzip t10k-images-idx3-ubyte.gz

wget http://yann.lecun.com/exdb/mnist/t10k-labels-idx1-ubyte.gz

gunzip t10k-labels-idx1-ubyte.gz

2. 修改文件dae_l1.yaml

进入python命令行模式

layer1_yaml = open('dae_l1.yaml', 'r').read()

hyper_params_l1 = {'train_stop' : 50000, 'batch_size' : 100, 'monitoring_batches' : 5, 'nhid' : 500, 'max_epochs' : 10, 'save_path' : '.'}

layer1_yaml = layer1_yaml % (hyper_params_l1)

print layer1_yaml

将输出的内容全部覆盖掉dae_l1.yaml文件的内容

3. 进入示例脚本目录
cd ~/pylearn2/pylearn2/scripts/tutorials/stacked_autoencoders
执行脚本
python ~/pylearn2/pylearn2/scripts/train.py dae_l1.yaml
输入日志如下：
/home/jerry/pylearn2/pylearn2/utils/call_check.py:98: UserWarning: the `one_hot` parameter is deprecated. To get one-hot e ncoded targets, request that they live in `VectorSpace` through the `data_specs` parameter of MNIST’s iterator method. `on e_hot` will be removed on or after September 20, 2014.
return to_call(**kwargs)
/home/jerry/.local/lib/python2.7/site-packages/theano/sandbox/rng_mrg.py:1183: UserWarning: MRG_RandomStreams Can’t determ ine #streams from size (Shape.0), guessing 60*256
nstreams = self.n_streams(size)
Parameter and initial learning rate summary:
vb: 0.001
hb: 0.001
W: 0.001
Wprime: 0.001
/home/jerry/pylearn2/pylearn2/models/model.py:71: UserWarning: The <class ‘pylearn2.models.autoencoder.DenoisingAutoencode r’> Model subclass seems not to call the Model constructor. This behavior may be considered an error on or after 2014-11-0 1.
warnings.warn(“The ” + str(type(self)) + ” Model subclass ”
Compiling sgd_update…
Compiling sgd_update done. Time elapsed: 7.379370 seconds
compiling begin_record_entry…
compiling begin_record_entry done. Time elapsed: 0.103046 seconds
Monitored channels:
learning_rate
objective
total_seconds_last_epoch
training_seconds_this_epoch
Compiling accum…
graph size: 19
Compiling accum done. Time elapsed: 0.876798 seconds
Monitoring step:
Epochs seen: 0
Batches seen: 0
Examples seen: 0
learning_rate: 0.001
objective: 89.1907964264
total_seconds_last_epoch: 0.0
training_seconds_this_epoch: 0.0
Time this epoch: 19.928861 seconds
……
Monitoring step:
Epochs seen: 10
Batches seen: 5000
Examples seen: 500000
learning_rate: 0.001
objective: 11.9511445315
total_seconds_last_epoch: 35.828732
training_seconds_this_epoch: 22.296131
Saving to ./dae_l1.pkl…
Saving to ./dae_l1.pkl done. Time elapsed: 0.936124 seconds
Saving to ./dae_l1.pkl…
Saving to ./dae_l1.pkl done. Time elapsed: 0.886536 seconds

4. 查看参数
>>> from pylearn2.utils import serial
>>> serial.load(‘dae_l1.pkl’)
<pylearn2.models.autoencoder.DenoisingAutoencoder object at 0x46855d0>
>>>
>>> model = serial.load(‘dae_l1.pkl’)
>>>
>>> dir(model)
[‘__call__’, ‘__class__’, ‘__delattr__’, ‘__dict__’, ‘__doc__’, ‘__format__’, ‘__getattribute__’, ‘__getstate__’, ‘__hash__’, ‘__init__’, ‘__metaclass__’, ‘__module__’, ‘__new__’, ‘__reduce__’, ‘__reduce_ex__’, ‘__repr__’, ‘__setattr__’, ‘__setstate__’, ‘__sizeof__’, ‘__str__’, ‘__subclasshook__’, ‘__weakref__’, ‘_disallow_censor_updates’, ‘_ensure_extensions’, ‘_hidden_activation’, ‘_hidden_input’, ‘_initialize_hidbias’, ‘_initialize_visbias’, ‘_initialize_w_prime’, ‘_initialize_weights’, ‘_modify_updates’, ‘_overrides_censor_updates’, ‘_params’, ‘_test_batch_size’, ‘act_dec’, ‘act_enc’, ‘censor_updates’, ‘continue_learning’, ‘corruptor’, ‘cpu_only’, ‘dataset_yaml_src’, ‘decode’, ‘encode’, ‘enforce_constraints’, ‘extensions’, ‘fn’, ‘free_energy’, ‘function’, ‘get_default_cost’, ‘get_input_dim’, ‘get_input_source’, ‘get_input_space’, ‘get_lr_scalers’, ‘get_monitoring_channels’, ‘get_monitoring_data_specs’, ‘get_output_dim’, ‘get_output_space’, ‘get_param_values’, ‘get_param_vector’, ‘get_params’, ‘get_target_source’, ‘get_target_space’, ‘get_test_batch_size’, ‘get_weights’, ‘get_weights_format’, ‘get_weights_topo’, ‘get_weights_view_shape’, ‘hidbias’, ‘input_space’, ‘inverse’, ‘irange’, ‘libv’, ‘modify_updates’, ‘monitor’, ‘nhid’, ‘output_space’, ‘perform’, ‘print_versions’, ‘reconstruct’, ‘redo_theano’, ‘register_names_to_del’, ‘rng’, ‘s_rng’, ‘score’, ‘set_batch_size’, ‘set_input_space’, ‘set_param_values’, ‘set_param_vector’, ‘set_visible_size’, ‘tag’, ‘tied_weights’, ‘train_all’, ‘train_batch’, ‘upward_pass’, ‘visbias’, ‘w_prime’, ‘weights’, ‘yaml_src’]
>>>

5. 类似步骤2，修改dae_l2.yaml文件

layer2_yaml = open('dae_l2.yaml', 'r').read()

hyper_params_l2 = {'train_stop' : 50000, 'batch_size' : 100, 'monitoring_batches' : 5, 'nvis' : 500, 'nhid' : 500, 'max_epochs' : 10, 'save_path' : '.'}

layer2_yaml = layer2_yaml % (hyper_params_l2)

print layer2_yaml

6. 执行dae_l2.yaml ，第二层模型训练
python ~/pylearn2/pylearn2/scripts/train.py dae_l2.yaml
/home/jerry/pylearn2/pylearn2/utils/call_check.py:98: UserWarning: the `one_hot` parameter is deprecated. To get one-hot encoded targets, request that they live in `VectorSpace` through the `data_specs` parameter of MNIST’s iterator method. `one_hot` will be removed on or after September 20, 2014.
return to_call(**kwargs)
/home/jerry/.local/lib/python2.7/site-packages/theano/sandbox/rng_mrg.py:1183: UserWarning: MRG_RandomStreams Can’t determine #streams from size (Shape.0), guessing 60*256
nstreams = self.n_streams(size)
Parameter and initial learning rate summary:
vb: 0.001
hb: 0.001
W: 0.001
Wprime: 0.001
/home/jerry/pylearn2/pylearn2/models/model.py:71: UserWarning: The <class ‘pylearn2.models.autoencoder.DenoisingAutoencoder’> Model subclass seems not to call the Model constructor. This behavior may be considered an error on or after 2014-11-01.
warnings.warn(“The ” + str(type(self)) + ” Model subclass ”
Compiling sgd_update…
Compiling sgd_update done. Time elapsed: 0.339660 seconds
compiling begin_record_entry…
compiling begin_record_entry done. Time elapsed: 0.023657 seconds
Monitored channels:
learning_rate
objective
total_seconds_last_epoch
training_seconds_this_epoch
Compiling accum…
graph size: 19
Compiling accum done. Time elapsed: 0.189965 seconds
Monitoring step:
Epochs seen: 0
Batches seen: 0
Examples seen: 0
learning_rate: 0.001
objective: 52.2956323286
total_seconds_last_epoch: 0.0
training_seconds_this_epoch: 0.0
Time this epoch: 17.452593 seconds
……
Monitoring step:
Epochs seen: 10
Batches seen: 5000
Examples seen: 500000
learning_rate: 0.001
objective: 4.33433924602
total_seconds_last_epoch: 30.433518
training_seconds_this_epoch: 19.303109
Saving to ./dae_l2.pkl…
Saving to ./dae_l2.pkl done. Time elapsed: 0.607150 seconds
Saving to ./dae_l2.pkl…
Saving to ./dae_l2.pkl done. Time elapsed: 0.588375 seconds

7. 类似步骤2修改dae_mlp.yaml文件

mlp_yaml = open('dae_mlp.yaml', 'r').read()

hyper_params_mlp = {'train_stop' : 50000, 'valid_stop' : 60000, 'batch_size' : 100, 'max_epochs' : 50, 'save_path' : '.'}

mlp_yaml = mlp_yaml % (hyper_params_mlp)

print mlp_yaml

（注：在原dae_mlp.yaml文件内没有save_path, save_freq这两项，造成参数数据没有保存，因而需要加入这两项，如下：
save_path : ‘./dae_mlp.pkl’,
save_freq : 1
)

8. 运行监督优化–Supervised fine-tuning
python ~/pylearn2/pylearn2/scripts/train.py dae_mlp.yaml
/home/jerry/pylearn2/pylearn2/utils/call_check.py:98: UserWarning: the `one_hot` parameter is deprecated. To get one-hot encoded targets, request that they live in `VectorSpace` through the `data_specs` parameter of MNIST’s iterator method. `one_hot` will be removed on or after September 20, 2014.
return to_call(**kwargs)
Parameter and initial learning rate summary:
vb: 0.05
hb: 0.05
W: 0.05
Wprime: 0.05
vb: 0.05
hb: 0.05
W: 0.05
Wprime: 0.05
softmax_b: 0.05
softmax_W: 0.05
Compiling sgd_update…
Compiling sgd_update done. Time elapsed: 17.156073 seconds
compiling begin_record_entry…
compiling begin_record_entry done. Time elapsed: 0.056943 seconds
Monitored channels:
learning_rate
momentum
total_seconds_last_epoch
training_seconds_this_epoch
valid_objective
valid_y_col_norms_max
valid_y_col_norms_mean
valid_y_col_norms_min
valid_y_max_max_class
valid_y_mean_max_class
valid_y_min_max_class
valid_y_misclass
valid_y_nll
valid_y_row_norms_max
valid_y_row_norms_mean
valid_y_row_norms_min
Compiling accum…
graph size: 63
Compiling accum done. Time elapsed: 8.821601 seconds
Monitoring step:
Epochs seen: 0
Batches seen: 0
Examples seen: 0
learning_rate: 0.05
momentum: 0.5
total_seconds_last_epoch: 0.0
training_seconds_this_epoch: 0.0
valid_objective: 2.30245763578
valid_y_col_norms_max: 0.0650026130651
valid_y_col_norms_mean: 0.0641744853852
valid_y_col_norms_min: 0.0624679393698
valid_y_max_max_class: 0.105532125739
valid_y_mean_max_class: 0.102753872501
valid_y_min_max_class: 0.101059172742
valid_y_misclass: 0.9031
valid_y_nll: 2.30245763578
valid_y_row_norms_max: 0.0125483545665
valid_y_row_norms_mean: 0.00897718040255
valid_y_row_norms_min: 0.00411555936503
Time this epoch: 18.159817 seconds
……
Monitoring step:
Epochs seen: 50
Batches seen: 25000
Examples seen: 2500000
learning_rate: 0.0183943399319
momentum: 0.539357429719
total_seconds_last_epoch: 21.789649
training_seconds_this_epoch: 19.881821
valid_objective: 0.0667691463031
valid_y_col_norms_max: 1.93649990002
valid_y_col_norms_mean: 1.93614117524
valid_y_col_norms_min: 1.93520053981
valid_y_max_max_class: 0.999997756761
valid_y_mean_max_class: 0.980073621031
valid_y_min_max_class: 0.548149309784
valid_y_misclass: 0.02
valid_y_nll: 0.0667691463031
valid_y_row_norms_max: 0.546499525611
valid_y_row_norms_mean: 0.264354016013
valid_y_row_norms_min: 0.101427414171

9. 至此整个训练过程结束
想调参数可以在yaml文件内调整，另外参数数据在三个文件内 dae_l1.pkl， dae_l2.pkl， dae_mlp.pkl

Pylearn2的使用简介

环境: ubuntu 12.4

Pylearn2是基于theano上封装的深度学习包。它实现一些常见的模型，具体请参考： http://deeplearning.net/software/pylearn2/library/index.html#libdoc，比theano在做实际的项目节约时间，只需要配置一些参数来实现模型的训练。
下面来讲解实际的安装和使用：

1. 安装Theano（Bleeding-edge install instruction）

jerry@hq:~$sudo pip install –upgrade –no-deps git+git://github.com/Theano/Theano.git –user

2. 下载Pylearn2
jerry@hq:~$git clone git://github.com/lisa-lab/pylearn2.git

3. 安装pylearn2
jerry@hq:~$cd pylearn2
jerry@hq:~$sudo python setup.py develop –user

4. 测试安装成功
jerry@hq:~$python
import pylearn2
能加载包即安装ok

5. 设置PYTHON2_DATA_PATH, PYLEARN2_VIEWR_COMMAND
vi ~/.bashrc
添加
export PYLEARN2_DATA_PATH=/u01/lisa/data
export PYLEARN2_VIEWER_COMMAND=/usr/bin/eog

如何运行一个示例

1. 下载数据
cd /u01/lisa/data/cifar10
wget http://www.cs.utoronto.ca/~kriz/cifar-10-python.tar.gz
tar xvf cifar-10-python.tar.gz

2. 修改make_dataset.py文件，指定路径/u01/lisa/data/ (由于本机上/空间不足，只能把数据放在其它路径上)
jerry@hq:~$vi /home/jerry/pylearn2/pylearn2/scripts/tutorials/grbm_smd/make_dataset.py
修改成这样：
“””
path = pylearn2.__path__[0]
train_example_path = os.path.join(path, ‘scripts’, ‘tutorials’, ‘grbm_smd’)
train.use_design_loc(os.path.join(train_example_path, ‘cifar10_preprocessed_train_design.npy’))
train_pkl_path = os.path.join(train_example_path, ‘cifar10_preprocessed_train.pkl’)
“””
train_pkl_path = os.path.join(‘/u01/lisa/data/’, ‘cifar10_preprocessed_train.pkl’)
serial.save(train_pkl_path, train)

3. 对下载数据进行数据预处理
python /home/jerry/pylearn2/pylearn2/scripts/tutorials/grbm_smd/make_dataset.py
处理完后在目录/u01/lisa/data下有一个文件 cifar10_preprocessed_train.pkl，大概652M左右

4. 对数据进行训练
cd /u01/lisa/data
python ~/pylearn2/pylearn2/scripts/train.py ~/pylearn2/pylearn2/scripts/tutorials/grbm_smd/cifar_grbm_smd.yaml

5. 查看结果
python ~/pylearn2/pylearn2/scripts/show_weights.py ~/pylearn2/pylearn2/scripts/tutorials/grbm_smd/cifar_grbm_smd.pkl

python ~/pylearn2/pylearn2/scripts/plot_monitor.py ~/pylearn2/pylearn2/scripts/tutorials/grbm_smd/cifar_grbm_smd.pkl

python ~/pylearn2/pylearn2/scripts/print_monitor.py ~/pylearn2/pylearn2/scripts/tutorials/grbm_smd/cifar_grbm_smd.pkl

python ~/pylearn2/pylearn2/scripts/summarize_model.py ~/pylearn2/pylearn2/scripts/tutorials/grbm_smd/cifar_grbm_smd.pkl

6. 直接查看生成参数的文件cifar_grbm_smd.pkl

加载模型文件
>>> from pylearn2.utils import serial
>>> model = serial.load(‘/home/jerry/pylearn2/pylearn2/scripts/tutorials/grbm_smd/cifar_grbm_smd.pkl’)
查下文件结构
>>> dir(model)
获取权重参数
>>> model.get_weights()
获取参数名
>>> model.get_params()
获取参数值
>>> model.get_param_values()

python 邮件发送

python 2.7代码如下：

#coding: utf-8
import smtplib
from email.mime.text import MIMEText

#connect smtp server
msg = MIMEText(‘Hello’,’plain’,’utf-8′)
msg[‘Subject’] = ‘ Load data sucess!’
#msg[‘Date’] = formatdate(localtime=True)
smtp = smtplib.SMTP()
smtp.connect(‘proxy-in.xxx.com’)
smtp.sendmail(‘bidev@xxx.com’, ‘swer@xxx.com’, msg.as_string())

python 2.4.3

#coding: utf-8
import smtplib
from email.MIMEText import MIMEText

#connect smtp server
msg = MIMEText(‘Hello’,’plain’,’utf-8′)
msg[‘Subject’] = ‘ Load data sucess!’
smtp = smtplib.SMTP()
smtp.connect(‘proxy-in.xxx.com’)
smtp.sendmail(‘bidev@xxx.com’, ‘hsdf@xxx.com’, msg.as_string())

http://m.baidu.com/news?fr=mohome&ssid=0&uid=&pu=sz%401320_2001%2Cta%40iphone_1_7.1_3_537&bd_page_type=1#page/info%3A互联网/http%3A%2F%2Fwww.huxiu.com%2Farticle%2F114327%2F1.html/深挖BAT内部级别和薪资待遇，你敢看？%20/虎嗅网/1430894729000/12489789468531375681

个人感言

越是看起来与我们背景接近的人取得成就越是会对身边的人产生一种不平衡感。

CXXNET安装

环境：ubuntu 14.04, cuda 6.5

先安装cuda-toolkit, cuda-cublas, cudart, cuda-curand这四个安装包

cuda_6.5.14_linux_64.run

cuda-cublas-6-5_6.5-14_amd64.deb
cuda-cudart-6-5_6.5-14_amd64.deb
cuda-curand-6-5_6.5-14_amd64.deb

下载路径：http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1404/x86_64/

安装 OpenCV

sudo apt-get install libopencv-2.4

配置环境变量

vi ~/.bashrc

export CUDA_HOME=/usr/local/cuda-6.5
export PATH=$CUDA_HOME/bin:$PATH
export LD_LIBRARY_PATH=$CUDA_HOME/lib64:/usr/local/lib:$LD_LIBRARY_PATH
export CPLUS_INCLUDE_PATH=/usr/local/cuda/include

下载一份cxxnet

git clone https://github.com/dmlc/cxxnet.git

切换至目录 cd cxxnet

拷贝一份配置到当前目录 cp make/config.mk .

修改 vi config.mk

USE_CUDA = 1

USE_BLAS = blas

USE_DIST_PS = 1
USE_OPENMP_ITER = 1

编辑 vi Makefile，修改如下：

CFLAGS += -g -O3 -I./mshadow/ -fPIC $(MSHADOW_CFLAGS) -fopenmp -I/usr/local/cuda/include
LDFLAGS = -pthread $(MSHADOW_LDFLAGS) -L/usr/local/cuda/lib64

最后编译文件

./build.sh

unable to correct problems you have held broken packages

OS: Ubuntu 14.04

When installing a ubuntu desktop KDE, some error like “unable to correct problems you have held broken packages” happen. So I finally found the problm is the package python3-software-properties is too new and can’t be compatible with the kde package. The soluation is follling:

sudo apt-get remove python3-software-properties

sudo apt-get install python3-software-properties=0.92.36