print __doc__
with open(sys.argv[1], 'r') as f:
W, _, bh = pic.load(f)
try:
with open('logreg_backup_' + str(W.shape[0]), + '.pkl', 'r') as f:
W2, b2 = pic.load(f)
pretrain = False
except:
W2 = None
b2 = None
pretrain = True
labels, train, test = pp.load_from_csv(sys.argv[2], sys.argv[3])
train, test, labels, test_labels = cv.train_test_split(train,
labels,
test_size=0.2)
mlp = ae.TwoLayerPerceptron(784,
W.shape[1],
10,
W_1_init=W,
b_1_init=bh,
W_2_init=W2,
b_2_init=b2)
if pretrain:
#greedy pretraining
"""
import preprocess as pp
import sys
import autoencoder as ae
import csv
import cPickle as pic
import sklearn.decomposition as dec
if __name__ == '__main__':
print __doc__
labels, train_features, test_features = pp.load_from_csv(
sys.argv[1], sys.argv[2])
try:
if sys.argv[6] == 'whiten':
pca = dec.PCA(whiten=True)
pca.fit(train_features)
train_features = pca.transform(train_features)
test_features = pca.transform(test_features)
else:
import preprocess as pp
import train_classifier as clf
from autoencoder import LogisticRegression
import numpy as np
if __name__ == "__main__":
labels, train, test = pp.load_from_csv( 'train.csv', 'test.csv' )
train_batches = train.reshape((420,100,784))
label_batches = np.array(labels).reshape((420,100))
logreg = LogisticRegression( 784, 10 )
logreg.fit( train_batches , label_batches, nbatches = 420 )
import preprocess as pp
import train_classifier as clf
from autoencoder import LogisticRegression
import numpy as np
if __name__ == "__main__":
labels, train, test = pp.load_from_csv('train.csv', 'test.csv')
train_batches = train.reshape((420, 100, 784))
label_batches = np.array(labels).reshape((420, 100))
logreg = LogisticRegression(784, 10)
logreg.fit(train_batches, label_batches, nbatches=420)
"""
import preprocess as pp
import sys
import autoencoder as ae
import csv
import cPickle as pic
import sklearn.decomposition as dec
if __name__=='__main__':
print __doc__
labels, train_features, test_features = pp.load_from_csv(sys.argv[1], sys.argv[2])
try:
if sys.argv[6]=='whiten':
pca = dec.PCA( whiten=True )
pca.fit( train_features )
train_features = pca.transform( train_features )
test_features = pca.transform( test_features )
else:
#otherwise, if the costs are diminishing too slowly,
#increase the learning rate by 1.1
if np.abs(rel_cost_change) < tol_low and rel_cost_change < 0.0:
learning_rate = 1.1*learning_rate
#if costs is changing less than the tolerance, stop, learning is done
if np.abs(rel_cost_change) < tolerance:
break
if __name__ == '__main__':
print __doc__%(sys.argv[1],sys.argv[2],sys.argv[3])
labels, train, test = pp.load_from_csv( sys.argv[1], sys.argv[2] )
#set values to something useful, batch_size and number of epochs
# doesn't seem to make much of a difference
training_epochs = 300
training_batches = 100
#patch_size, for 28x28 images, 10x10 patches seemed reasonable
patch_size = 15
batch_size = int(train.shape[0] / training_batches)
n_filters = 500
output_file = ('ae_' + str(patch_size) + 'x' + str(patch_size) + '_'
+ str(n_filters) + '_filters_backup.pikle')