Category Archives: python

Heat map subplots sharing same color bar pandas with seaborn

Make sure you have pandas and seaborn installed

plt.cla()
plt.close()
fig, (ax0,ax1) = plt.subplots(1, 2, sharex=True, sharey=True)
cbar_ax = fig.add_axes([.91,.3,.03,.4])
sns.heatmap(pd1.corr(),ax=ax0,cbar=True,vmin=-1,vmax=1,cbar_ax = cbar_ax)
ax0.set_title('title1')
sns.heatmap(pd2.corr(),ax=ax1,cbar=True,vmin=-1,vmax=1,cbar_ax = cbar_ax)
ax1.set_title('title 2')
fig.suptitle('big title',fontsize=20)
#saving figure for publication if needed
plt.savefig('save.tif', dpi=300)
plt.show()

thats it! I guess the most important thing here is cbar_ax = fig.add_axes([.91,.3,.03,.4]) and make sure you have a fixed vmin and vmax.

source: http://stackoverflow.com/questions/24653986/saving-matplotlib-figure-with-add-axes

Advertisements

How to make Custom estimator class and custom scorer to do cross validation using sklearn api on your custom model

I made a combined weak classifier model, needed a custom estimator and custom scorer. I went through a few stack overflow articles however none actually targeted specifically for cross validation in sklearn.

Then I figured I would try to implement baseestimator class, and make my own scorer. It WORKED. :>

Therefore, I am posting instructions here on how to use it, hopefully its gonna be useful to you.

Steps:

  1. Write your own estimator class, just make sure to implement base estimator (or extend I am not sure how this works in python but its similar. base estimator is like an interface or abstract class provides basic functionalities for estimator)
  2. Write your loss function or gain function, and then make your own scorer
  3. Use the sklearn api to do cross validation. Using whatever you have created in 1 and 2.

Code: Please read comments. Important.

#create a custom estimator class

#Keep in mind. This is just a simplified version. You can treat it as any other class, just make sure the signitures should stay same, or you should add default value to other parameters

from sklearn.base import BaseEstimator
class custom_classifier(BaseEstimator):
  from sklearn import tree
  from sklearn.cluster import KMeans
  import numpy as np
  from sklearn.cluster import KMeans
  #Kmeans clustering model
  __clusters = None
  #decision tree model
  __tree = None
  #x library.
  __X = None
  #y library.
  __y = None
  #columns selected from pandas dataframe.
  __columns = None

  def fit(self, X, y, **kwargs):
    self.fit_kmeans(self.__X,self.__y)
    self.fit_decisiontree(self.__X,self.__y)

  def predict(self,X):
    result_kmeans = self.__clusters.predict(X)
    result_tree = self.__tree.predict(X)
    result = result_tree
    return np.array(result)

  def fit_kmeans(self,X,y):
    clusters = KMeans(n_clusters=4, random_state=0).fit(X)
    #the error center should have the lowest number of labels.(implementation not shown here)
    self.__clusters = clusters

  def fit_decisiontree(self,X,y):
    temp_tree = tree.DecisionTreeClassifier(criterion='entropy',max_depth=3)
    temp_tree.fit(X,y)
    self.__tree = temp_tree

Now we have our class. We need to build hit/loss function:

#again, feel free to change any thing in the hit function. As long as the function signature remain the same.

def seg_tree_hit_func(ground_truth, predictions):
  total_hit = 0
  total_number = 0
  for i in xrange(len(predictions)):
    if predictions[i]==2:
      continue
    else:
      total_hit += (1-abs(ground_truth[i]-predictions[i]))
      total_number+=1.0
    print 'skipped: ',len(predictions)- total_number,'/',len(predictions),'instances'
  return total_hit/total_number if total_number!=0 else 0

Now we still need to build scorer.

from sklearn.metrics.scorer import make_scorer

#make our own scorer
score = make_scorer(seg_tree_hit_func, greater_is_better=True)

We have our scorer, our estimator, and so we can start doing cross-validation task:

#change the 7 to whatever fold validation you are running.

scores = cross_val_score(custom_classifier(), X, Y, cv=7, scoring=score)

There it is! You have your own scorer and estimator, and you can use sklearn api to plug it in anything from sklearn easily.

 

Hope this helps.

Install new version of python and set up virtualenvwrapper under your user account (e.g. you don’t have admin rights)

I ran into a problem today trying to install python libraries on server which I don’t have admin rights. Not happy but got a solution around. This way its easier to maintain the environment myself. Hopefully someone can make a software that manages software environment in user space on linux.

Install python 2.7.10 under your user/.localpython
Replace the USER_NAME with your own username.

mkdir ~/src
mkdir ~/.localpython
cd ~/src
wget https://www.python.org/ftp/python/2.7.10/Python-2.7.10.tgz
tar -zxvf Python-2.7.10.tgz
cd Python-2.7.10

./configure -prefix=/home/USER_NAME/.localpython
make
make install

Install virtualenv

cd ~/src
wget https://pypi.python.org/packages/source/v/virtualenv/virtualenv-13.1.2.tar.gz#md5=b989598f068d64b32dead530eb25589a
tar -zxvf virtualenv-13.1.2.tar.gz
cd virtualenv-13.1.2
~/.localpython/bin/python setup.py install

install per

cd ~/src
wget https://pypi.python.org/packages/source/p/pbr/pbr-1.8.1.tar.gz#md5=c8f9285e1a4ca6f9654c529b158baa3a
tar -zxvf pbr-1.8.1.tar.gz
cd pbr-1.8.1
~/.localpython/bin/python setup.py install

 

install virtualenvwrapper

cd ~/src
wget https://pypi.python.org/packages/source/v/virtualenvwrapper/virtualenvwrapper-4.7.1.tar.gz#md5=3789e0998818d9a8a4ec01cfe2a339b2
tar -zxvf virtualenvwrapper-4.7.1.tar.gz
cd virtualenvwrapper-4.7.1
~/.localpython/bin/python setup.py install

install stevedore (dependency for virtualenvwrapper

cd ~/src
wget https://pypi.python.org/packages/source/s/stevedore/stevedore-1.9.0.tar.gz#md5=53e2bc3b49dd9c920cfce7f63822b1a5
tar -zxvf stevedore-1.9.0.tar.gz
cd stevedore-1.9.0
~/.localpython/bin/python setup.py install

Now last step:
Edit your ~/.bashrc file so your python distribution is the one when you type which python command. Add the following lines at the top.


export PATH="/home/USER_NAME/.localpython/bin:$PATH"
export WORKON_HOME=~/.virtualenvs
export PROJECT_HOME=~/Devel
source /home/USER_NAME/.localpython/bin/virtualenvwrapper.sh

exit your server and log in again. You should be good to go to make mkvirtualenv and do whatever you want.

The above guide is tested on my University cluster where python2.6 was installed however i don’t have control over custom libraries. I could use virtualenv though.

Installing scipy on redhat with error “no lapack/blas resources found”

Update Feb21 2016:
For centos: 
wget http://mirror.centos.org/centos/6/os/x86_64/Packages/lapack-devel-3.2.1-4.el6.x86_64.rpm
wget http://mirror.centos.org/centos/6/os/x86_64/Packages/blas-devel-3.2.1-4.el6.x86_64.rpm
wget http://mirror.centos.org/centos/6/os/x86_64/Packages/texinfo-tex-4.13a-8.el6.x86_64.rpm
wget http://mirror.centos.org/centos/6/os/x86_64/Packages/libicu-devel-4.2.1-9.1.el6_2.x86_64.rpm
sudo yum localinstall *.rpm

sudo pip install scipy
This is copied from http://stackoverflow.com/questions/24708213/install-r-on-redhat-errors-on-dependencies-that-dont-exist

For Ubuntu 14.04:
sudo apt-get install gfortran
sudo apt-get install libblas-dev liblapack-dev