import cPickle
import numpy as np
from cifar import load_cifar
from cifar import load_cifar_file
import matplotlib.pyplot as plt
from logreg import sigmoid
# Try to do classification based on logistic regression for some CIFAR-10 images.
#IT will be difficult to get convergence using all 3072 features values and 10000-images in bach-1
# We will the images to grayscale to reduce the nof. features

trainimages, trainlab = load_cifar_file('data_batch_1')

#print trainlab.shape
#print trainimages.shape
#Reduce to gray level image by averaging the 3 bands
grayimages = (trainimages[:,:,:,0]+trainimages[:,:,:,1]+trainimages[:,:,:,2])/3
print grayimages.shape
nsamp,nx,ny,nb=trainimages.shape
trainvec =grayimages.reshape(nsamp,nx*ny)



#print grayimages.shape
#nsamp,nx,ny,nb=grayimages.shape
#trainvec =grayimages.reshape(nsamp,nx*ny)
print trainvec.shape
nofsamp, nfeat = trainvec.shape
trainvecAppend = np.ones((nofsamp,nfeat+1))
trainvecAppend[:,1:]= trainvec


#Reshape images to vectors
#trainimgvector = trainimages.reshape
noftrain = 500;
subtrain = trainvecAppend[1:noftrain,:]
sublab = trainlab[1:noftrain]
testvec = trainvecAppend[noftrain:noftrain+noftrain,:]
testlab = trainlab[noftrain:noftrain+noftrain]
from sklearn.linear_model import LogisticRegression
clf = LogisticRegression(penalty='l2',C=1,solver='sag',max_iter=2000, random_state=1,multi_class='multinomial').fit(subtrain,sublab)
#print training score
print("training score : %3.3f (%s)") % (clf.score(subtrain,sublab),'multinomial')
print("test score : %3.3f (%s)") % (clf.score(testvec,testlab),'multinomial')