def faces_data(resize_factor = None):
print "Loading faces data ..."
t0 = time.time()
faces, labels = q.load_from_pkl("../data/faces.bzpkl")
if resize_factor is not None:
new_size = (int(64*resize_factor), int(64*resize_factor))
print new_size
for i in range(len(faces)):
print i
#t = cv2.resize(faces[i].reshape(64,64),new_size).ravel()
labels = np.asarray(labels, dtype=np.int32)
faces = np.asarray(faces, dtype=np.float32)
data = q.newObject()
data.train = q.newObject()
data.valid = q.newObject()
data.test = q.newObject()
mean = 128
stdev = 75
faces = (faces - mean) / stdev
np.random.seed(10)
np.random.shuffle(faces)
np.random.seed(10)
np.random.shuffle(labels)
data.train.x = np.vstack((faces[labels==1,:][ :1500], faces[labels==-1,:][ :1500]))
data.valid.x = np.vstack((faces[labels==1,:][1500:2000], faces[labels==-1,:][1500:2000]))
data.test.x = np.vstack((faces[labels==1,:][2000:2400], faces[labels==-1,:][2000:2400]))
data.train.y = np.hstack((np.ones((1500, ), dtype=np.int32),
np.zeros((1500, ),dtype=np.int32)))
data.valid.y = np.hstack((np.ones((500, ),dtype=np.int32),
np.zeros((500, ),dtype=np.int32)))
data.test.y = np.hstack((np.ones((400, ),dtype=np.int32),
np.zeros((400, ),dtype=np.int32)))
data.train.x = theano.shared(data.train.x, borrow=True)
data.train.y = theano.shared(data.train.y, borrow=True)
data.valid.x = theano.shared(data.valid.x, borrow=True)
data.valid.y = theano.shared(data.valid.y, borrow=True)
data.test.x = theano.shared(data.test.x , borrow=True)
data.test.y = theano.shared(data.test.y , borrow=True)
data.dim = (64, 64)
data_digits.dim_out = 2
def show(img):
q.show(img.reshape((64,64)))
cv2.destroyAllWindows()
for k in range(10):
cv2.waitKey(10)
data.show = show
t1 = time.time()
print "Data loaded in %0.2f seconds" % ( t1-t0)
return data
def compile_model_ops(self):
self.model_ops = q.newObject()
# Define the model training and testing functions for both
# minibatches and arbitrary data
index = self.model.index
x = self.model.x
y = self.model.y
X = self.model.X
Y = self.model.Y
"""
self.model_ops.train_model = theano.function(inputs=[X, Y],
outputs=self.model.errors(y),
updates=self.model.updates,
givens={
x: X,
y: Y})
"
self.model_ops.test_model2 = theano.function( inputs=[],
outputs=self.model.errors(y),
givens={
x: self.model.data.test.x,
y: self.model.data.test.y})
self.model_ops.validate_model2 = theano.function(inputs=[],
outputs=self.model.errors(y),
givens={
x: self.model.data.valid.x,
y: self.model.data.valid.y})
"""
self.params = self.model.params
self.acc_cost = self.model.acc_cost
self.L1 = np.sum([abs(param).sum() for param in self.params])
self.L2 = np.sum([(param**2).sum() for param in self.params])
self.reg_cost = self.L1_reg * self.L1 + self.L2_reg * self.L2
self.cost = self.acc_cost + self.reg_cost
self.grads = [T.grad(cost=self.cost, wrt=param) for param in self.params]
self.updates = [(param, param - self.lr * grad) for
param, grad in zip(self.params, self.grads)]
self.model_ops.minibatch = q.newObject()
self.model_ops.minibatch.test = theano.function( inputs=[index],
outputs=self.model.errors(y),
givens={
x: self.model.data.test.x[index * self.batch_size:(index + 1) * self.batch_size],
y: self.model.data.test.y[index * self.batch_size:(index + 1) * self.batch_size]})
self.model_ops.minibatch.validate = theano.function(inputs=[index],
outputs=self.model.errors(y),
givens={
x: self.model.data.valid.x[index * self.batch_size:(index + 1) * self.batch_size],
y: self.model.data.valid.y[index * self.batch_size:(index + 1) * self.batch_size]})
self.model_ops.minibatch.train = theano.function(inputs=[index],
outputs=self.cost,
updates=self.updates,
givens={
x: self.model.data.train.x[index * self.batch_size:(index + 1) * self.batch_size],
y: self.model.data.train.y[index * self.batch_size:(index + 1) * self.batch_size]})
def get_faces(resize_factor=None):
print "Loading faces data ..."
t0 = time.time()
faces, labels = q.load_from_pkl("../data/faces.bzpkl")
if resize_factor is not None:
new_size = (int(64 * resize_factor), int(64 * resize_factor))
print new_size
for i in range(len(faces)):
print i
#t = cv2.resize(faces[i].reshape(64,64),new_size).ravel()
labels = np.asarray(labels, dtype=np.int32)
faces = np.asarray(faces, dtype=np.float32)
data = q.newObject()
data.train = q.newObject()
data.valid = q.newObject()
data.test = q.newObject()
mean = 128
stdev = 75
faces = (faces - mean) / stdev
np.random.seed(10)
np.random.shuffle(faces)
np.random.seed(10)
np.random.shuffle(labels)
data.train.x = np.vstack(
(faces[labels == 1, :][:1500], faces[labels == -1, :][:1500]))
data.valid.x = np.vstack(
(faces[labels == 1, :][1500:2000], faces[labels == -1, :][1500:2000]))
data.test.x = np.vstack(
(faces[labels == 1, :][2000:2400], faces[labels == -1, :][2000:2400]))
data.train.y = np.hstack((np.ones(
(1500, ), dtype=np.int32), np.zeros((1500, ), dtype=np.int32)))
data.valid.y = np.hstack((np.ones(
(500, ), dtype=np.int32), np.zeros((500, ), dtype=np.int32)))
data.test.y = np.hstack((np.ones(
(400, ), dtype=np.int32), np.zeros((400, ), dtype=np.int32)))
data.train.x = theano.shared(data.train.x, borrow=True)
data.train.y = theano.shared(data.train.y, borrow=True)
data.valid.x = theano.shared(data.valid.x, borrow=True)
data.valid.y = theano.shared(data.valid.y, borrow=True)
data.test.x = theano.shared(data.test.x, borrow=True)
data.test.y = theano.shared(data.test.y, borrow=True)
data.dim = (64, 64)
data.dim_out = 2
def show(img):
q.show(img.reshape((64, 64)))
cv2.destroyAllWindows()
for k in range(10):
cv2.waitKey(10)
data.show = show
t1 = time.time()
print "Data loaded in %0.2f seconds" % (t1 - t0)
return data
def faces_data():
faces, labels = q.load_from_pkl("/home/tc/faces.bzpkl")
labels = np.asarray(labels, dtype=np.float64)
faces = np.asarray(faces, dtype=np.float64)
data = q.newObject()
data.train = q.newObject()
data.valid = q.newObject()
data.test = q.newObject()
data.train.x = np.vstack(
(faces[labels == 1, :][:1000], faces[labels == -1, :][:1000]))
data.valid.x = np.vstack(
(faces[labels == 1, :][1000:1500], faces[labels == -1, :][1000:1500]))
data.test.x = np.vstack(
(faces[labels == 1, :][1500:2000], faces[labels == -1, :][1500:2000]))
data.train.y = np.hstack((np.ones(
(1000, ), dtype=np.int32), np.zeros((1000, ), dtype=np.int32)))
data.valid.y = np.hstack((np.ones(
(500, ), dtype=np.int32), np.zeros((500, ), dtype=np.int32)))
data.test.y = np.hstack((np.ones(
(500, ), dtype=np.int32), np.zeros((500, ), dtype=np.int32)))
data.train.x = theano.shared(data.train.x, borrow=True)
data.train.y = theano.shared(data.train.y, borrow=True)
data.valid.x = theano.shared(data.valid.x, borrow=True)
data.valid.y = theano.shared(data.valid.y, borrow=True)
data.test.x = theano.shared(data.test.x, borrow=True)
data.test.y = theano.shared(data.test.y, borrow=True)
#np.random.seed(10)
#np.random.shuffle(data.train.x)
#np.random.seed(10)
#np.random.shuffle(data.train.y)
def show(img):
q.show(img.reshape((64, 64)))
cv2.destroyAllWindows()
for k in range(10):
cv2.waitKey(10)
data.show = show
return data
def get_mnist(dataset='mnist.pkl.gz'):
t0 = time.time()
print "Loading digits data ..."
dataset = "../data/" + dataset
f = gzip.open(dataset, 'rb')
train_set, valid_set, test_set = cPickle.load(f)
f.close()
def shared_dataset(data_xy, borrow=True):
data_x, data_y = data_xy
shared_x = theano.shared(np.asarray(data_x,
dtype=theano.config.floatX),
borrow=borrow)
shared_y = theano.shared(np.asarray(data_y, dtype=np.int32),
borrow=borrow)
return shared_x, shared_y
test_set_x, test_set_y = shared_dataset(test_set)
valid_set_x, valid_set_y = shared_dataset(valid_set)
train_set_x, train_set_y = shared_dataset(train_set)
data_digits = q.newObject()
data_digits.train = q.newObject()
data_digits.valid = q.newObject()
data_digits.test = q.newObject()
data_digits.train.x = train_set_x
data_digits.valid.x = valid_set_x
data_digits.test.x = test_set_x
data_digits.train.y = train_set_y
data_digits.valid.y = valid_set_y
data_digits.test.y = test_set_y
data_digits.dim = (28, 28)
data_digits.dim_out = 10
t1 = time.time()
print "Data loaded in %0.2f seconds" % (t1 - t0)
return data_digits
def mnist_data(dataset='mnist.pkl.gz'):
t0 = time.time()
print "Loading digits data ..."
dataset = "../data/"+dataset
f = gzip.open(dataset, 'rb')
train_set, valid_set, test_set = cPickle.load(f)
f.close()
def shared_dataset(data_xy, borrow=True):
data_x, data_y = data_xy
shared_x = theano.shared(np.asarray(data_x,
dtype=theano.config.floatX), borrow=borrow)
shared_y = theano.shared(np.asarray(data_y,
dtype=np.int32), borrow=borrow)
return shared_x, shared_y
test_set_x, test_set_y = shared_dataset(test_set)
valid_set_x, valid_set_y = shared_dataset(valid_set)
train_set_x, train_set_y = shared_dataset(train_set)
data_digits = q.newObject()
data_digits.train = q.newObject()
data_digits.valid = q.newObject()
data_digits.test = q.newObject()
data_digits.train.x = train_set_x
data_digits.valid.x = valid_set_x
data_digits.test.x = test_set_x
data_digits.train.y = train_set_y
data_digits.valid.y = valid_set_y
data_digits.test.y = test_set_y
data_digits.dim = (28, 28)
data_digits.dim_out = 10
t1 = time.time()
print "Data loaded in %0.2f seconds" % ( t1-t0)
return data_digits
def compile_model_ops(self):
self.model_ops = q.newObject()
# Define the model training and testing functions for both
# minibatches and arbitrary data
index = self.model.index
x = self.model.x
y = self.model.y
X = self.model.X
Y = self.model.Y
"""
self.model_ops.train_model = theano.function(inputs=[X, Y],
outputs=self.model.errors(y),
updates=self.model.updates,
givens={
x: X,
y: Y})
"
self.model_ops.test_model2 = theano.function( inputs=[],
outputs=self.model.errors(y),
givens={
x: self.model.data.test.x,
y: self.model.data.test.y})
self.model_ops.validate_model2 = theano.function(inputs=[],
outputs=self.model.errors(y),
givens={
x: self.model.data.valid.x,
y: self.model.data.valid.y})
"""
self.params = self.model.params
self.acc_cost = self.model.acc_cost
self.L1 = np.sum([abs(param).sum() for param in self.params])
self.L2 = np.sum([(param**2).sum() for param in self.params])
self.reg_cost = self.L1_reg * self.L1 + self.L2_reg * self.L2
self.cost = self.acc_cost + self.reg_cost
self.grads = [
T.grad(cost=self.cost, wrt=param) for param in self.params
]
self.updates = [(param, param - self.lr * grad)
for param, grad in zip(self.params, self.grads)]
self.model_ops.minibatch = q.newObject()
self.model_ops.minibatch.test = theano.function(
inputs=[index],
outputs=self.model.errors(y),
givens={
x:
self.model.data.test.x[index * self.batch_size:(index + 1) *
self.batch_size],
y:
self.model.data.test.y[index * self.batch_size:(index + 1) *
self.batch_size]
})
self.model_ops.minibatch.validate = theano.function(
inputs=[index],
outputs=self.model.errors(y),
givens={
x:
self.model.data.valid.x[index * self.batch_size:(index + 1) *
self.batch_size],
y:
self.model.data.valid.y[index * self.batch_size:(index + 1) *
self.batch_size]
})
self.model_ops.minibatch.train = theano.function(
inputs=[index],
outputs=self.cost,
updates=self.updates,
givens={
x:
self.model.data.train.x[index * self.batch_size:(index + 1) *
self.batch_size],
y:
self.model.data.train.y[index * self.batch_size:(index + 1) *
self.batch_size]
})
def __init__(self,
n_in,
n_out,
data,
layerSpecs,
batch_size,
rng,
learning_rate,
activation=T.tanh,
L1_reg=0.,
L2_reg=0.0001):
# Define classifier independent stuff
# preliminaries
#######################################################################
index = T.lscalar() # index to a [mini]batch
x = T.matrix('x')
y = T.ivector('y')
X = T.matrix('X')
Y = T.ivector('Y')
self.lr = theano.shared(learning_rate)
if (rng is None):
rng = np.random.RandomState(23455)
self.batch_size = batch_size
self.n_train_batches = data.train.x.get_value(
borrow=True).shape[0] / batch_size
self.n_valid_batches = data.valid.x.get_value(
borrow=True).shape[0] / batch_size
self.n_test_batches = data.test.x.get_value(
borrow=True).shape[0] / batch_size
self.data = data
# Define the model structure
#######################################################################
layerClasses = {
'conv': voo.layers.ConvLayer,
'hidden': voo.layers.FullyConnectedLayer,
'hidden_decomp': voo.layers.decomp.FullyConnectedDecompLayer,
'logistic': voo.layers.LogisticRegressionLayer,
'pooling': voo.layers.PoolingLayer,
'dropout': voo.layers.DropOutLayer
}
Layers = []
input_layer = voo.layers.InputLayer(x, n_in, batch_size)
Layers.append(input_layer)
prev_layer = input_layer
for layer_idx in range(len(layerSpecs)):
# First we get the layer class and specs
layerType, layerConfig = layerSpecs[layer_idx]
layerClass = layerClasses[layerType]
new_layer = voo.layers.GenLayer(layerClass, prev_layer, batch_size,
rng, layerConfig)
Layers.append(new_layer)
prev_layer = new_layer
self.Layers = Layers
#
# Define all the model specifics that relate to the parameters:
# vector of parameters, gradients, errors, updates, accuracy cost
# regularization cost, overall cost
######################################################################
self.params = []
for layer in Layers:
self.params.extend(layer.params)
self.acc_cost = self.Layers[-1].negative_log_likelihood(y)
self.L1 = np.sum([abs(param).sum() for param in self.params])
self.L2 = np.sum([(param**2).sum() for param in self.params])
self.reg_cost = L1_reg * self.L1 + L2_reg * self.L2
self.cost = self.acc_cost + self.reg_cost
self.grads = [
T.grad(cost=self.cost, wrt=param) for param in self.params
]
self.updates = [(param, param - self.lr * grad)
for param, grad in zip(self.params, self.grads)]
# Define the model training and testing functions for both
# minibatches and arbitrary data
self.train_model = theano.function(inputs=[X, Y],
outputs=self.errors(y),
updates=self.updates,
givens={
x: X,
y: Y
})
self.test_model2 = theano.function(inputs=[],
outputs=self.errors(y),
givens={
x: self.data.test.x,
y: self.data.test.y
})
self.validate_model2 = theano.function(inputs=[],
outputs=self.errors(y),
givens={
x: self.data.valid.x,
y: self.data.valid.y
})
self.minibatch = q.newObject()
self.minibatch.test = theano.function(
inputs=[index],
outputs=self.errors(y),
givens={
x:
self.data.test.x[index * self.batch_size:(index + 1) *
self.batch_size],
y:
self.data.test.y[index * self.batch_size:(index + 1) *
self.batch_size]
})
self.minibatch.validate = theano.function(
inputs=[index],
outputs=self.errors(y),
givens={
x:
self.data.valid.x[index * self.batch_size:(index + 1) *
self.batch_size],
y:
self.data.valid.y[index * self.batch_size:(index + 1) *
self.batch_size]
})
self.minibatch.train = theano.function(
inputs=[index],
outputs=self.cost,
updates=self.updates,
givens={
x:
self.data.train.x[index * self.batch_size:(index + 1) *
self.batch_size],
y:
self.data.train.y[index * self.batch_size:(index + 1) *
self.batch_size]
})
def __init__(self, n_in, n_out, data, layerSpecs, batch_size, rng, learning_rate, activation=T.tanh, L1_reg=0., L2_reg=0.0001):
# Define classifier independent stuff
# preliminaries
#######################################################################
index = T.lscalar() # index to a [mini]batch
x = T.matrix('x')
y = T.ivector('y')
X = T.matrix('X')
Y = T.ivector('Y')
self.lr = theano.shared(learning_rate)
if (rng is None):
rng = np.random.RandomState(23455)
self.batch_size = batch_size
self.n_train_batches = data.train.x.get_value(borrow=True).shape[0] / batch_size
self.n_valid_batches = data.valid.x.get_value(borrow=True).shape[0] / batch_size
self.n_test_batches = data.test.x.get_value (borrow=True).shape[0] / batch_size
self.data = data
# Define the model structure
#######################################################################
layerClasses = {
'conv': voo.layers.ConvLayer,
'hidden': voo.layers.FullyConnectedLayer,
'hidden_decomp': voo.layers.decomp.FullyConnectedDecompLayer,
'logistic': voo.layers.LogisticRegressionLayer,
'pooling': voo.layers.PoolingLayer,
'dropout': voo.layers.DropOutLayer
}
Layers = []
input_layer = voo.layers.InputLayer(x, n_in, batch_size)
Layers.append(input_layer)
prev_layer = input_layer
for layer_idx in range(len(layerSpecs)):
# First we get the layer class and specs
layerType, layerConfig = layerSpecs[layer_idx]
layerClass = layerClasses[layerType]
new_layer = voo.layers.GenLayer(layerClass, prev_layer, batch_size, rng, layerConfig)
Layers.append(new_layer)
prev_layer = new_layer
self.Layers=Layers
#
# Define all the model specifics that relate to the parameters:
# vector of parameters, gradients, errors, updates, accuracy cost
# regularization cost, overall cost
######################################################################
self.params = []
for layer in Layers:
self.params.extend(layer.params)
self.acc_cost = self.Layers[-1].negative_log_likelihood(y)
self.L1 = np.sum([abs(param).sum() for param in self.params])
self.L2 = np.sum([(param**2).sum() for param in self.params])
self.reg_cost = L1_reg * self.L1 + L2_reg * self.L2
self.cost = self.acc_cost + self.reg_cost
self.grads = [T.grad(cost=self.cost, wrt=param) for param in self.params]
self.updates = [(param, param - self.lr * grad) for
param, grad in zip(self.params, self.grads)]
# Define the model training and testing functions for both
# minibatches and arbitrary data
self.train_model = theano.function(inputs=[X, Y],
outputs=self.errors(y),
updates=self.updates,
givens={
x: X,
y: Y})
self.test_model2 = theano.function( inputs=[],
outputs=self.errors(y),
givens={
x: self.data.test.x,
y: self.data.test.y})
self.validate_model2 = theano.function(inputs=[],
outputs=self.errors(y),
givens={
x: self.data.valid.x,
y: self.data.valid.y})
self.minibatch = q.newObject()
self.minibatch.test = theano.function( inputs=[index],
outputs=self.errors(y),
givens={
x: self.data.test.x[index * self.batch_size:(index + 1) * self.batch_size],
y: self.data.test.y[index * self.batch_size:(index + 1) * self.batch_size]})
self.minibatch.validate = theano.function(inputs=[index],
outputs=self.errors(y),
givens={
x: self.data.valid.x[index * self.batch_size:(index + 1) * self.batch_size],
y: self.data.valid.y[index * self.batch_size:(index + 1) * self.batch_size]})
self.minibatch.train = theano.function(inputs=[index],
outputs=self.cost,
updates=self.updates,
givens={
x: self.data.train.x[index * self.batch_size:(index + 1) * self.batch_size],
y: self.data.train.y[index * self.batch_size:(index + 1) * self.batch_size]})
epoch = epoch + 1
if epoch == max_epochs or patience <= iteration:
done_looping = True
end_time = time.clock()
print(('Optimization complete with best validation score of %f %%,'
'with test performance %f %%') %
(best_validation_loss * 100., test_score * 100.))
print 'The code run for %d epochs, with %f epochs/sec' % (
epoch, 1. * epoch / (end_time - start_time))
print('The code ran for %.1fs' % ((end_time - start_time)))
data_faces = faces_data()
mnist = load_data('mnist.pkl.gz')
data_digits = q.newObject()
data_digits.train = q.newObject()
data_digits.valid = q.newObject()
data_digits.test = q.newObject()
data_digits.train.x = mnist[0][0]
data_digits.valid.x = mnist[1][0]
data_digits.test.x = mnist[2][0]
data_digits.train.y = mnist[0][1]
data_digits.valid.y = mnist[1][1]
data_digits.test.y = mnist[2][1]
print "Ready, Cap\'n!"
M = LogisticRegression_model(28 * 28,
