def load_data():
svhn = SVHN()
X_train, y_train = svhn.load_preprocessed_data('train')
X_test, y_test = svhn.load_preprocessed_data('test')
# svhn.visualize(X_train, y_train)
return X_train, y_train, X_test, y_test
momentum = 0.9
weightdecay = 0.01
finetune_lr = 1e-2
finetune_epc = 400
print " "
print "batchsize =", batchsize
print "momentum =", momentum
print "finetune: lr = %f, epc = %d" % (finetune_lr, finetune_epc)
#############
# LOAD DATA #
#############
svhn_data = SVHN()
train_x, train_y = svhn_data.get_train_set(include_extra=False)
test_x, test_y = svhn_data.get_test_set()
train_smp = train_x.shape[0]
test_smp = test_x.shape[0]
print "\n... pre-processing"
preprocess_model = SubtractMeanAndNormalizeH(train_x.shape[1])
map_fun = theano.function([preprocess_model.varin], preprocess_model.output())
train_x = map_fun(train_x.astype(theano.config.floatX)).reshape((train_smp, 32, 32, 3)).swapaxes(1, 3)
test_x = map_fun(test_x.astype(theano.config.floatX)).reshape((test_smp, 32, 32, 3)).swapaxes(1, 3)
train_x = theano.shared(value=train_x, name='train_x', borrow=True)
train_y = theano.shared(value=train_y.astype('int64') - 1, name='train_y', borrow=True)
test_x = theano.shared(value=test_x, name='test_x', borrow=True)
test_y = theano.shared(value=test_y.astype('int64') - 1, name='test_y', borrow=True)
weightdecay = 0.01
finetune_lr = 1e-4
finetune_epc = 400
part_size = 100000
print " "
print "batchsize =", batchsize
print "momentum =", momentum
print "finetune: lr = %f, epc = %d" % (finetune_lr, finetune_epc)
#############
# LOAD DATA #
#############
svhn_data = SVHN()
train_x_np, train_y_np = svhn_data.get_train_set(include_extra=True)
test_x, test_y = svhn_data.get_test_set()
train_smp = train_x_np.shape[0]
test_smp = test_x.shape[0]
print "\n... pre-processing"
preprocess_model = SubtractMeanAndNormalizeH(train_x_np.shape[1])
map_fun = theano.function([preprocess_model.varin], preprocess_model.output())
train_x_accu = []
for i in range(train_smp/part_size):
train_x_accu.append(map_fun(
train_x_np[i * part_size : (i + 1) * part_size].astype(theano.config.floatX)
).reshape((part_size, 32, 32, 3)).swapaxes(1, 3))
train_x_accu.append(map_fun(