def AutoEncoder_demo(learning_rate=0.1,
training_epochs=2,
dataset='mnist.pkl.gz',
batch_size=20,
output_folder='dA_plots'):
datasets = load_data(dataset)
train_set_x, train_set_y = datasets[0]
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
index = T.lscalar() # index to a [mini]batch
x = T.matrix('x')
if not os.path.isdir(output_folder):
os.makedirs(output_folder)
os.chdir(output_folder)
#####################################
# BUILDING THE MODEL CORRUPTION 0% #
#####################################
rng = numpy.random.RandomState(123)
theano_rng = RandomStreams(rng.randint(2**30))
da = AutoEncoder(np_rng=rng,
theano_rng=theano_rng,
input=x,
n_vis=28 * 28,
n_hid=500)
cost, updates = da.get_cost_updates(corruption_level=0.,
learning_rate=learning_rate)
train_da = theano.function(
inputs=[index],
outputs=[cost],
updates=updates,
givens={x: train_set_x[index * batch_size:(index + 1) * batch_size]})
start_time = time.clock()
for epoch in xrange(training_epochs):
# go through trainng set
c = []
for batch_index in xrange(n_train_batches):
c.append(train_da(batch_index))
print 'Training epoch %d, cost ' % epoch, numpy.mean(c)
end_time = time.clock()
training_time = (end_time - start_time)
print >> sys.stderr, ('The no corruption code ran for %.2fm' %
((training_time) / 60.))
image = PIL.Image.fromarray(
tile_raster_images(X=da.W.get_value(borrow=True).T,
img_shape=(28, 28),
tile_shape=(10, 10),
tile_spacing=(1, 1)))
image.save('filters_corruption_0.jpg')
#####################################
# BUILDING THE MODEL CORRUPTION 30% #
#####################################
rng = numpy.random.RandomState(123)
theano_rng = RandomStreams(rng.randint(2**30))
da = AutoEncoder(np_rng=rng,
theano_rng=theano_rng,
input=x,
n_vis=28 * 28,
n_hid=500)
cost, updates = da.get_cost_updates(corruption_level=0.3,
learning_rate=learning_rate)
train_da = theano.function(
inputs=[index],
outputs=[cost],
updates=updates,
givens={x: train_set_x[index * batch_size:(index + 1) * batch_size]})
start_time = time.clock()
for epoch in xrange(training_epochs):
c = []
for batch_index in xrange(n_train_batches):
c.append(train_da(batch_index))
print 'Training epoch %d, cost ' % epoch, numpy.mean(c)
end_time = time.clock()
training_time = (end_time - start_time)
print >> sys.stderr, ('The 30 percent corruption code ran for %.2fm' %
((training_time) / 60.))
image = PIL.Image.fromarray(
tile_raster_images(X=da.W.get_value(borrow=True).T,
img_shape=(28, 28),
tile_shape=(10, 10),
tile_spacing=(1, 1)))
image.save('filters_corruption_30.jpg')
os.chdir('../')
# -*- coding: utf-8 -*-
import theano
import theano.tensor as T
from loadDataset import load_data
from LogisticRegression import LogisticRegression
batch_size = 20
datasets = load_data(dataset)
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
print '... building the model'
index = T.lscalar()
x = T.matrix('x')
y = T.ivector('y')
classifier = LogisticRegression(x, y, n_in=784, n_out=10)
#cost = T.log(classifier.p_y_given_x)[T.arange(y.shape[0]), y]
cost = T.log(classifier.p_y_given_x)[T.arange(y.shape[0]), y]
train_model = theano.function(
# -*- coding: utf-8 -*-
import theano
import theano.tensor as T
from loadDataset import load_data
from LogisticRegression import LogisticRegression
batch_size = 20
datasets = load_data(dataset)
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
print '... building the model'
index = T.lscalar()
x = T.matrix('x')
y = T.ivector('y')
classifier = LogisticRegression(x, y, n_in=784, n_out=10)
#cost = T.log(classifier.p_y_given_x)[T.arange(y.shape[0]), y]
import numpy as np
from pylab import *
from loadDataset import load_data_2d as load_data
from random import random
import cv2
#For testing dataset for training. Randomly shows data and print it's value.
(X_train, y_train, X_test, y_test) = load_data()
labels = ['A','B','C','D','E','F','G','H','I','J','K','L','M','N','P','Q','R','S','T','U','V','W','X','Y','Z','0','1','2','3','4','5','6','7','8','9']
t =0
while True:
#t = int(random()*120000)
if labels[np.argmax(y_train[t])] != ' ':
im = np.reshape(X_train[t],(28,28))
cv2.imshow('a',im)
print(labels[np.argmax(y_train[t])])
cv2.waitKey(0)
t=t+1
cv2.destroyAllWindows()
def AutoEncoder_demo(learning_rate=0.1, training_epochs=2, dataset='mnist.pkl.gz', batch_size=20, output_folder='dA_plots'):
datasets = load_data(dataset)
train_set_x, train_set_y = datasets[0]
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
index = T.lscalar() # index to a [mini]batch
x = T.matrix('x')
if not os.path.isdir(output_folder):
os.makedirs(output_folder)
os.chdir(output_folder)
#####################################
# BUILDING THE MODEL CORRUPTION 0% #
#####################################
rng = numpy.random.RandomState(123)
theano_rng = RandomStreams(rng.randint(2 ** 30))
da = AutoEncoder(np_rng=rng, theano_rng=theano_rng, input=x, n_vis=28 * 28, n_hid=500)
cost, updates = da.get_cost_updates(corruption_level=0., learning_rate=learning_rate)
train_da = theano.function(inputs=[index],
outputs=[cost],
updates=updates,
givens={x: train_set_x[index * batch_size: (index + 1) * batch_size]})
start_time = time.clock()
for epoch in xrange(training_epochs):
# go through trainng set
c = []
for batch_index in xrange(n_train_batches):
c.append(train_da(batch_index))
print 'Training epoch %d, cost ' % epoch, numpy.mean(c)
end_time = time.clock()
training_time = (end_time - start_time)
print >> sys.stderr, ('The no corruption code ran for %.2fm' % ((training_time) / 60.))
image = PIL.Image.fromarray(tile_raster_images(X=da.W.get_value(borrow=True).T,
img_shape=(28, 28), tile_shape=(10, 10),
tile_spacing=(1, 1)))
image.save('filters_corruption_0.jpg')
#####################################
# BUILDING THE MODEL CORRUPTION 30% #
#####################################
rng = numpy.random.RandomState(123)
theano_rng = RandomStreams(rng.randint(2 ** 30))
da = AutoEncoder(np_rng=rng, theano_rng=theano_rng, input=x, n_vis=28 * 28, n_hid=500)
cost, updates = da.get_cost_updates(corruption_level=0.3, learning_rate=learning_rate)
train_da = theano.function(inputs=[index],
outputs=[cost],
updates=updates,
givens={x: train_set_x[index * batch_size:(index + 1) * batch_size]})
start_time = time.clock()
for epoch in xrange(training_epochs):
c = []
for batch_index in xrange(n_train_batches):
c.append(train_da(batch_index))
print 'Training epoch %d, cost ' % epoch, numpy.mean(c)
end_time = time.clock()
training_time = (end_time - start_time)
print >> sys.stderr, ('The 30 percent corruption code ran for %.2fm' % ((training_time) / 60.))
image = PIL.Image.fromarray(tile_raster_images(X=da.W.get_value(borrow=True).T,
img_shape=(28, 28), tile_shape=(10, 10),
tile_spacing=(1, 1)))
image.save('filters_corruption_30.jpg')
os.chdir('../')
