In [ ]:

    

import numpy as np
import copy
import bayes_logistic as bl
from sklearn import datasets


    

In [ ]:

    

# Digit we want to classify
DIGIT = 1


    

In [ ]:

    

# Prep the MNIST image classification data. The data is divided into training, 
# cross validation (cv) and test sets

TRAINING_PERCENTAGE = 0.6
CV_PERCENTAGE = 0.2
TEST_PERCENTAGE  = 0.2

digits = datasets.load_digits()
n_samples, n_features = digits.data.shape

# Add bias column to the feature matrix
X = np.ones((n_samples, n_features + 1))
X[:, 1:] = digits.data

# Change the vector to binary data
y = digits.target
for idx in range(0, y.shape[0]):
    y[idx] = 1.0 if y[idx] == DIGIT else 0.0

training_cnt = int(n_samples*TRAINING_PERCENTAGE)
training_X = X[:training_cnt,:]
training_y = y[:training_cnt]

cv_cnt = int(n_samples*CV_PERCENTAGE)
cv_X = X[training_cnt:(training_cnt + cv_cnt),:]
cv_y = y[training_cnt:(training_cnt + cv_cnt)]

test_cnt = int(n_samples*TEST_PERCENTAGE)
test_X = X[(training_cnt + cv_cnt):,:]
test_y = y[(training_cnt + cv_cnt):]


    

In [ ]:

    

# Train the model 
GD_BATCH_SIZE = training_cnt
ITERATION_CNT = 1
w = training_X.shape[1]
w_prior = np.zeros(w)
H_prior = np.diag(np.ones(w))*0.001

for i in range(0, ITERATION_CNT):
    for idx in range(0, training_cnt, GD_BATCH_SIZE):
        batch_size = GD_BATCH_SIZE if (idx + GD_BATCH_SIZE) < training_cnt else training_cnt - idx
        w_posterior, H_posterior = bl.fit_bayes_logistic(training_y[idx:batch_size], 
                                                     training_X[idx:batch_size,:], 
                                                     w_prior, H_prior, solver = 'BFGS')
        w_prior = copy.copy(w_posterior)
        H_prior = copy.copy(H_posterior)
    
    # Perform Cross validation
    y_cnt = 0
    cv_probs = bl.bayes_logistic_prob(cv_X, w_prior, H_prior)
    prediction_cnt = 0
    for idx in range(0, cv_cnt):
        if cv_probs[idx] > 0.9 and cv_y[idx] == 1:
            prediction_cnt += 1
        y_cnt += 1
        
    prediction_accuracy = (100.0*prediction_cnt)/y_cnt
    print "Prediction Accuracy for cross validation set %.02f" % prediction_accuracy


    

In [ ]:

    

# Perform Test
y_cnt = 0
test_probs = bl.bayes_logistic_prob(test_X, w_prior, H_prior)
prediction_cnt = 0
for idx in range(0, test_cnt):
    if cv_probs[idx] > 0.5 and cv_y[idx] == 1:
        prediction_cnt += 1
    y_cnt += 1
        
prediction_accuracy = (100.0*prediction_cnt)/y_cnt
print "Prediction Accuracy for test set %.02f" % prediction_accuracy