一、Win7 64 安装theano：
1、 下载Anaconda1.9.2，自带MniGw；
C:\Anaconda; C:\Anaconda\Scripts;C :\Anaconda\MinGW\超级的豌豆; C:\Anaconda\MinGW\x86_64-w64-mingw32\超级的豌豆; 加入到PATH，安装git,并将C:\git\超级的豌豆;加入到PATH
2、 Github下载theano然后解压，讲theano文件夹里面的theano文件夹拷贝到
C:\Anaconda\Lib\site-packages
3、 下载安装CUDA，版本随意
将
C:\Program Files x86)\NVIDIA Corporation\PhysX\Common;C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.0\超级的豌豆;C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.0\libnvvp;C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.0\lib\x64;C:\Program Files x86)\Microsoft SDKs\Windows\v7.0A\Include;
加入到path。
测试Cuda安装成功与否，使用
nvcc –V
看cuda版本
采用VS2010，将C:\Program Files x86)\Microsoft SDKs\Windows\v7.0A\Include;加入到PATH，则最终PATH有：
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v6.0\超级的豌豆;C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v6.0\libnvvp;c:\Program Files x86)\NVIDIA Corporation\PhysX\Common;C:\Anaconda;C:\Anaconda\Scripts;C:\Anaconda\MinGW\超级的豌豆;C:\Anaconda\MinGW\x86_64-w64-mingw32\超级的豌豆;C:\git\超级的豌豆;C:\Program Files x86)\Microsoft SDKs\Windows\v7.0A\Include;
4、 在C:\Users\Administrator下建立一个文件：.theanorc.txt
[global]openmp = Falsedevice = gpu0floatX = float32allow_input_downcast=True [blas]ldflags = [gcc]cxxflags = -IC:\Anaconda\MinGW [cuda]root=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.0\超级的豌豆 [nvcc]flags = -LC:\Anaconda\libs compiler_超级的豌豆dir = C:\Program Files x86)\Microsoft Visual Studio 10.0\VC\超级的豌豆fastmath = Trueflags = -arch=sm_30
注意错误：
1、c:\anaconda\include\pyconfig.h227) : fatal error C1083: Cannot open include fil
e: ‘basetsd.h’: No such file or directory；
ERROR theano.sandbox.cuda): Failed to compile cuda_ndarray.cu: ‘nvcc return st
atus’, 2, ‘for cmd’, ‘nvcc -shared -O3 -arch=sm_30 -use_fast_math –compiler-超级的豌豆
根据windows SDK查找这个文件，将这个文件对应的include加入到PATH，没有解决问题，后来compiler_超级的豌豆dir Vs2012更改为Vs2010解决问题。
compiler_超级的豌豆dir = C:\Program Files x86)\Microsoft Visual Studio 10.0\VC\超级的豌豆
二、安装Lasagne
1、目前还只能从源代码安装。
git clone https://github.com/Lasagne/Lasagne.git # 会建立一个Lasagne目录cd Lasagnepip install -r requirements.txt # 比较久python setup.py install # 这一步需要root权限
Daniel Nouri[Tutorial]上是这样安装的
pip install -r https://raw.githubusercontent.com/dnouri/kfkd-tutorial/master/requirements.txt
这样需要以root用户来执行，不推荐这样做。
尝试正确的安装方式：
git clone https://github.com/Lasagne/Lasagne.git # 会建立一个Lasagne目录
cd Lasagne
pip install -r https://raw.githubusercontent.com/dnouri/kfkd-tutorial/master/requirements.txt
问题：
1. 这是Python 2 mimetypes的bug
2. 需要将Python2.7/lib/mimetypes.py文件中如下片段注释或删除：
try: ctype = ctype.encodedefault_encoding) # omit in 3.x!except UnicodeEncodeError: pass
补充其它解决办法
解决办法：
在报错的页面添加代码：
import sys reloadsys) sys.setdefaultencoding’utf8′)
执行 Python ez_setup.py，报错：
UnicodeDecodeError: ‘utf8’ codec can’t decode byte 0xb0 in position 35: invalid
start byte
解决办法：
在报错的页面添加代码：
import sysreloadsys)sys.setdefaultencoding’utf-8′)
安装失败之后，重新安装，需要清理C:\Users\Administrator\AppData\Local\pip里面的内容
2、test on mnist
cd examplespython mnist.pyEpoch 103 of 500 took 11.717s training loss: 0.045202 validation loss: 0.059163 validation accuracy: 98.16 %%Epoch 104 of 500 took 11.702s training loss: 0.046228 validation loss: 0.058582 validation accuracy: 98.14 %%Epoch 105 of 500 took 11.704s training loss: 0.044530 validation loss: 0.058295 validation accuracy: 98.18 %%
三、Facial Keypoints Detection
1、从https://github.com/dylansun/Kaggle-Facial-Keypoint-Detection 上下载data数据进行训练：
Data Files：training .zip 60.10 mb)test .zip 15.99 mb)SampleSubmission .csv 201.08 kb)IdLookupTable .csv 842.51 kb)
Each predicted keypoint is specified by an x,y) real-valued pair in the space of pixel indices. There are 15 keypoints, which represent the following elements of the face:
left_eye_center, right_eye_center, left_eye_inner_corner, left_eye_outer_corner, right_eye_inner_corner, right_eye_outer_corner, left_eyebrow_inner_end, left_eyebrow_outer_end, right_eyebrow_inner_end, right_eyebrow_outer_end, nose_tip, mouth_left_corner, mouth_right_corner, mouth_center_top_lip, mouth_center_bottom_lip Left and right here refers to the point of view of the subject.
In some examples, some of the target keypoint positions are misssing encoded as missing entries in the csv, i.e., with nothing between two commas).
The input image is given in the last field of the data files, and consists of a list of pixels ordered by row), as integers in 0,255). The images are 96×96 pixels.
Data files
training.csv: list of training 7049 images. Each row contains the x,y) coordinates for 15 keypoints, and image data as row-ordered list of pixels.
test.csv: list of 1783 test images. Each row contains ImageId and image data as row-ordered list of pixels
submissionFileFormat.csv: list of 27124 keypoints to predict. Each row contains a RowId, ImageId, FeatureName, Location. FeatureName are “left_eye_center_x,” “right_eyebrow_outer_end_y,” etc. Location is what you need to predict.
2、采用kfkd.py进行训练：
“””
To use this script, first run this to fit your first model:
python kfkd.py fit
Then train a bunch of specialists that intiliaze their weights from
your first model:
python kfkd.py fit_specialists net.pickle
Plot their error curves:
python kfkd.py plot_learning_curves net-specialists.pickle
此处报错：
Traceback most recent call last):
File “C:\Program Files x86)\JetBrains\PyCharm 4.0.4\helpers\pydev\pydevd.py”, line 2222, in <module>
globals = debugger.runsetup[‘file’], None, None)
File “C:\Program Files x86)\JetBrains\PyCharm 4.0.4\helpers\pydev\pydevd.py”, line 1648, in run
pydev_imports.execfilefile, globals, locals) # execute the script
File “C:/Anaconda/src/lasagne/examples/Kaggle Facial Keypoints Detection/kfkd.py”, line 417, in <module>
func*sys.argv[2:])
File “C:/Anaconda/src/lasagne/examples/Kaggle Facial Keypoints Detection/kfkd.py”, line 371, in plot_learning_curves
models = pickle.loadf)
EOFError
解决方法：直接执行
plot_learning_curves
And finally make predictions to submit to Kaggle:
python kfkd.py predict net-specialists.pickle