def evaluate_lenet5(learning_rate=0.1,
n_epochs=200,
dataset='emotion',
nkerns=[20, 50],
batch_size=500):
""" Demonstrates lenet on MNIST dataset
:type learning_rate: float
:param learning_rate: learning rate used (factor for the stochastic
gradient)
:type n_epochs: int
:param n_epochs: maximal number of epochs to run the optimizer
:type dataset: string
:param dataset: path to the dataset used for training /testing (MNIST here)
:type nkerns: list of ints
:param nkerns: number of kernels on each layer
"""
rng = numpy.random.RandomState(23455)
datasets = Ld.load_share(dataset)
if dataset == 'mnist':
ishape = (28, 28) # this is the size of MNIST images
num_label = 10
elif dataset == 'emotion':
ishape = (48, 48) # this is the size of MNIST images
num_label = 7
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
# compute number of minibatches for training, validation and testing
n_train_batches = train_set_x.get_value(borrow=True).shape[0]
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0]
n_test_batches = test_set_x.get_value(borrow=True).shape[0]
n_train_batches /= batch_size
n_valid_batches /= batch_size
n_test_batches /= batch_size
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
x = T.matrix('x') # the data is presented as rasterized images
y = T.ivector('y') # the labels are presented as 1D vector of
# [int] labels
######################
# BUILD ACTUAL MODEL #
######################
print '... building the model'
# Reshape matrix of rasterized images of shape (batch_size,28*28)
# to a 4D tensor, compatible with our LeNetConvPoolLayer
layer0_input = x.reshape((batch_size, 1, ishape[0], ishape[1]))
# Construct the first convolutional pooling layer:
# filtering reduces the image size to (28-5+1,28-5+1)=(24,24)
# maxpooling reduces this further to (24/2,24/2) = (12,12)
# 4D output tensor is thus of shape (batch_size,nkerns[0],12,12)
layer0 = LeNetConvPoolLayer(rng,
input=layer0_input,
image_shape=(batch_size, 1, ishape[0],
ishape[1]),
filter_shape=(nkerns[0], 1, 5, 5),
poolsize=(2, 2))
# Construct the second convolutional pooling layer
# filtering reduces the image size to (12-5+1,12-5+1)=(8,8)
# maxpooling reduces this further to (8/2,8/2) = (4,4)
# 4D output tensor is thus of shape (nkerns[0],nkerns[1],4,4)
if dataset == 'emotion':
layer05 = LeNetConvPoolLayer(rng,
input=layer0.output,
image_shape=(batch_size, nkerns[0], 22,
22),
filter_shape=(nkerns[0], nkerns[0], 3, 3),
poolsize=(2, 2))
layer1 = LeNetConvPoolLayer(rng,
input=layer05.output,
image_shape=(batch_size, nkerns[0], 10,
10),
filter_shape=(nkerns[1], nkerns[0], 3, 3),
poolsize=(2, 2))
elif dataset == 'mnist':
layer1 = LeNetConvPoolLayer(rng,
input=layer0.output,
image_shape=(batch_size, nkerns[0], 12,
12),
filter_shape=(nkerns[1], nkerns[0], 5, 5),
poolsize=(2, 2))
# the TanhLayer being fully-connected, it operates on 2D matrices of
# shape (batch_size,num_pixels) (i.e matrix of rasterized images).
# This will generate a matrix of shape (20,32*4*4) = (20,512)
layer2_input = layer1.output.flatten(2)
# construct a fully-connected sigmoidal layer
layer2 = HiddenLayer(rng,
input=layer2_input,
n_in=nkerns[1] * 4 * 4,
n_out=500,
activation=T.tanh)
# classify the values of the fully-connected sigmoidal layer
layer3 = LogisticRegression(input=layer2.output, n_in=500, n_out=num_label)
# the cost we minimize during training is the NLL of the model
cost = layer3.negative_log_likelihood(y)
# create a function to compute the mistakes that are made by the model
test_model = theano.function(
[index],
layer3.errors(y),
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
y: test_set_y[index * batch_size:(index + 1) * batch_size]
})
validate_model = theano.function(
[index],
layer3.errors(y),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size]
})
# create a list of all model parameters to be fit by gradient descent
params = layer3.params + layer2.params + layer1.params + layer0.params
# create a list of gradients for all model parameters
grads = T.grad(cost, params)
# train_model is a function that updates the model parameters by
# SGD Since this model has many parameters, it would be tedious to
# manually create an update rule for each model parameter. We thus
# create the updates list by automatically looping over all
# (params[i],grads[i]) pairs.
updates = []
for param_i, grad_i in zip(params, grads):
updates.append((param_i, param_i - learning_rate * grad_i))
train_model = theano.function(
[index],
cost,
updates=updates,
givens={
x: train_set_x[index * batch_size:(index + 1) * batch_size],
y: train_set_y[index * batch_size:(index + 1) * batch_size]
})
###############
# TRAIN MODEL #
###############
print '... training'
# early-stopping parameters
patience = 10000 # look as this many examples regardless
patience_increase = 2 # wait this much longer when a new best is
# found
improvement_threshold = 0.995 # a relative improvement of this much is
# considered significant
validation_frequency = min(n_train_batches, patience / 2)
# go through this many
# minibatche before checking the network
# on the validation set; in this case we
# check every epoch
best_params = None
best_validation_loss = numpy.inf
best_iter = 0
test_score = 0.
start_time = time.clock()
epoch = 0
done_looping = False
while (epoch < n_epochs) and (not done_looping):
epoch = epoch + 1
for minibatch_index in xrange(n_train_batches):
iter = (epoch - 1) * n_train_batches + minibatch_index
if iter % 100 == 0:
print 'training @ iter = ', iter
cost_ij = train_model(minibatch_index)
if (iter + 1) % validation_frequency == 0:
# compute zero-one loss on validation set
validation_losses = [
validate_model(i) for i in xrange(n_valid_batches)
]
this_validation_loss = numpy.mean(validation_losses)
print('epoch %i, minibatch %i/%i, validation error %f %%' % \
(epoch, minibatch_index + 1, n_train_batches, \
this_validation_loss * 100.))
# if we got the best validation score until now
if this_validation_loss < best_validation_loss:
#improve patience if loss improvement is good enough
if this_validation_loss < best_validation_loss * \
improvement_threshold:
patience = max(patience, iter * patience_increase)
# save best validation score and iteration number
best_validation_loss = this_validation_loss
best_iter = iter
# test it on the test set
test_losses = [
test_model(i) for i in xrange(n_test_batches)
]
test_score = numpy.mean(test_losses)
print(
(' epoch %i, minibatch %i/%i, test error of best '
'model %f %%') % (epoch, minibatch_index + 1,
n_train_batches, test_score * 100.))
if patience <= iter:
done_looping = True
break
end_time = time.clock()
print('Optimization complete.')
print('Best validation score of %f %% obtained at iteration %i,'\
'with test performance %f %%' %
(best_validation_loss * 100., best_iter + 1, test_score * 100.))
print >> sys.stderr, ('The code for file ' + os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.))
class CDBN:
def __init__(self,
input=None,
label=None,
n_ins=2,
hidden_layer_sizes=[3, 3],
n_outs=2,
rng=None):
self.x = input
self.y = label
self.sigmoid_layers = []
self.rbm_layers = []
self.n_layers = len(hidden_layer_sizes)
if rng is None:
rng = numpy.random.RandomState(1234)
assert self.n_layers > 0
# construct multi-layer
for i in xrange(self.n_layers):
# layer_size
if i == 0:
input_size = n_ins
else:
input_size = hidden_layer_sizes[i - 1]
# layer_input
if i == 0:
layer_input = self.x
else:
layer_input = self.sigmoid_layers[-1].sample_h_given_v()
# construct sigmoid_layer
sigmoid_layer = HiddenLayer(input=layer_input,
n_in=input_size,
n_out=hidden_layer_sizes[i],
rng=rng,
activation=sigmoid)
self.sigmoid_layers.append(sigmoid_layer)
# construct rbm_layer
if i == 0:
rbm_layer = CRBM(
input=layer_input, # continuous-valued inputs
n_visible=input_size,
n_hidden=hidden_layer_sizes[i],
W=sigmoid_layer.W, # W, b are shared
hbias=sigmoid_layer.b)
else:
rbm_layer = RBM(
input=layer_input,
n_visible=input_size,
n_hidden=hidden_layer_sizes[i],
W=sigmoid_layer.W, # W, b are shared
hbias=sigmoid_layer.b)
self.rbm_layers.append(rbm_layer)
# layer for output using Logistic Regression
self.log_layer = LogisticRegression(
input=self.sigmoid_layers[-1].sample_h_given_v(),
label=self.y,
n_in=hidden_layer_sizes[-1],
n_out=n_outs)
# finetune cost: the negative log likelihood of the logistic regression layer
self.finetune_cost = self.log_layer.negative_log_likelihood()
class SdA:
def __init__(self,
input=None,
label=None,
n_ins=2,
hidden_layer_sizes=[3, 3],
n_outs=2,
rng=None):
self.x = input
self.y = label
self.sigmoid_layers = []
self.dA_layers = []
self.n_layers = len(hidden_layer_sizes) # = len(self.rbm_layers)
if rng is None:
rng = numpy.random.RandomState(1234)
assert self.n_layers > 0
# construct multi-layer
for i in xrange(self.n_layers):
# layer_size
if i == 0:
input_size = n_ins
else:
input_size = hidden_layer_sizes[i - 1]
# layer_input
if i == 0:
layer_input = self.x
else:
layer_input = self.sigmoid_layers[-1].sample_h_given_v()
# construct sigmoid_layer
sigmoid_layer = HiddenLayer(input=layer_input,
n_in=input_size,
n_out=hidden_layer_sizes[i],
rng=rng,
activation=sigmoid)
self.sigmoid_layers.append(sigmoid_layer)
# construct dA_layers
dA_layer = dA(input=layer_input,
n_visible=input_size,
n_hidden=hidden_layer_sizes[i],
W=sigmoid_layer.W,
hbias=sigmoid_layer.b)
self.dA_layers.append(dA_layer)
# layer for output using Logistic Regression
self.log_layer = LogisticRegression(
input=self.sigmoid_layers[-1].sample_h_given_v(),
label=self.y,
n_in=hidden_layer_sizes[-1],
n_out=n_outs)
# finetune cost: the negative log likelihood of the logistic regression layer
self.finetune_cost = self.log_layer.negative_log_likelihood()
def pretrain(self, lr=0.1, corruption_level=0.3, epochs=100):
for i in xrange(self.n_layers):
if i == 0:
layer_input = self.x
else:
layer_input = self.sigmoid_layers[i - 1].sample_h_given_v(
layer_input)
da = self.dA_layers[i]
for epoch in xrange(epochs):
da.train(lr=lr,
corruption_level=corruption_level,
input=layer_input)
def finetune(self, lr=0.1, epochs=100):
layer_input = self.sigmoid_layers[-1].sample_h_given_v()
# train log_layer
epoch = 0
while epoch < epochs:
self.log_layer.train(lr=lr, input=layer_input)
lr *= 0.95
epoch += 1
def predict(self, x):
layer_input = x
for i in xrange(self.n_layers):
sigmoid_layer = self.sigmoid_layers[i]
layer_input = sigmoid_layer.output(input=layer_input)
return self.log_layer.predict(layer_input)
def __init__(self, n_conv_layers=2, filter_shapes=((20, 1, 5, 5), (50, 20, 5, 5)),
image_shape=(500, 1, 28, 28), poolsize=(2, 2), n_hidden_neurons = 500,
learning_rate=0.1, dataset_name='mnist.pkl.gz'):
"""
Инициализирует сеть, в соответствии с переданными параметрами.
:type n_conv_layers: int > 0
:param n_conv_layers: количество свёрточных слоёв
:type filter_shapes: tuple, длины n_conv_layers, состоящий из описаний фильтров(tuple длины 4)
:param filter_shapes: каждый filter_shape имеет следующий формат:
(количество фильтров, количество входных каналов, высота фильтра, ширина фильтра)
количество фильтров соотвествует количеству выходных каналов
:type image_shape: tuple или list длины 4
:param image_shape: (размер одного пакета, количество входных каналов(карт признаков),
высота изображения, ширина изображения)
:param poolsize: tuple длины 2
:type n_hidden_neurons: int > 0
:param n_hidden_neurons: количество нейронов в полносвязном скрытом слое
:type learning_rate: double
:param learning_rate: параметр, отвественный за скорость обучения методом градиентного спуска
"""
# Проверка корректности входных параметров
assert len(image_shape) == 4
assert len(filter_shapes) == n_conv_layers
assert len(poolsize) == 2
self.n_conv_layers = n_conv_layers
self.batch_size = image_shape[0]
self.rng = numpy.random.RandomState(23455)
# резервируем символические переменные для данных
######################
# BUILD ACTUAL MODEL #
######################
print '... building the model'
x = T.matrix('x') # the data is presented as rasterized images
y = T.ivector('y') # the labels are presented as 1D vector of [int] labels
layer_input = x.reshape(image_shape)
# лист из параметров скрытых слоёв
params = []
# Инициализирую сверточные слоя, в соответствии с архитектурой, указанной в параметрах
for i in xrange(n_conv_layers):
batch_size, n_filter_maps, image_height, image_weight = image_shape
n_filters, n_input_filter_maps, filer_heigth, filter_weight = filter_shapes[i]
pool_height, pool_weight = poolsize
# Построение свёртного слоя + пулинг
conv_layer = LeNetConvPoolLayer(
self.rng,
input=layer_input,
image_shape=image_shape,
filter_shape=filter_shapes[i],
poolsize=poolsize
)
layer_input = conv_layer.output
# Сохраняю параметры сети
params.append(conv_layer.params)
# Фильтр сокращает размер изображение, новый размер: (28-5+1 , 28-5+1) = (24, 24)
image_height = image_height - filer_heigth + 1
image_weight = image_weight - filter_weight + 1
# maxpooling также сокращает размер, новый размер: (24/2, 24/2) = (12, 12)
image_height /= pool_height
image_weight /= pool_weight
# Таким образом новый размер тензора изображения: (batch_size, 20, 12, 12)
image_shape = (batch_size, n_filters, image_height, image_weight)
# Результат прохождения изображений через свёрточные слои записан в layer_input
# Размер потока данный теперь соответсвует image_shape
# Если в сети два слоя с дефолтными параметрами, то выходное изображением имеет формат:
# filtering reduces the image size to (12-5+1, 12-5+1) = (8, 8)
# maxpooling reduces this further to (8/2, 8/2) = (4, 4)
# 4D output tensor is thus of shape (batch_size, 50, 4, 4)
batch_size, n_filter_maps, image_height, image_weight = image_shape
# Полносвязный скытый слой принимает на вход 2D матрицу, но у нас есть 4D тензор
# Поэтому превращаем тензор в матрицу размера (batch_size, n_filters * image_height * image_weight)
fully_connected_layer_input = layer_input.flatten(2)
n_filter_maps = image_shape[1]
# n_in: размерность входа
# n_out: количество нейронов в скрытом слое
# Построение полносвязного слоя
# TODO: тут по-хорошему можно задавать активационную функцию в качестве параметра
fully_connected_layer = HiddenLayer(
self.rng,
input=fully_connected_layer_input,
n_in=n_filter_maps * image_height * image_weight,
n_out=n_hidden_neurons,
activation=T.tanh
)
# classify the values of the fully-connected sigmoidal layer
logistic_regression_layer = LogisticRegression(input=fully_connected_layer.output,
n_in=n_hidden_neurons, n_out=10)
index = T.lscalar() # индекс пакета
x_set = T.matrix('x_set')
# Функция, распознающая изображения (используется уже после обучения)
self.prediction_model = theano.function(
[index, x_set],
outputs=logistic_regression_layer.y_pred,
givens={
x: x_set[index * self.batch_size: (index + 1) * self.batch_size]
}
)
# the cost we minimize during training is the NLL of the model
cost = logistic_regression_layer.negative_log_likelihood(y)
self.load(dataset_name=dataset_name)
# Создаём функцию, подсчитывающую ошибку модели
self.test_model = theano.function(
[index],
logistic_regression_layer.errors(y),
givens={
x: self.test_set_x[index * self.batch_size: (index + 1) * self.batch_size],
y: self.test_set_y[index * self.batch_size: (index + 1) * self.batch_size]
}
)
self.validate_model = theano.function(
[index],
logistic_regression_layer.errors(y),
givens={
x: self.valid_set_x[index * self.batch_size: (index + 1) * self.batch_size],
y: self.valid_set_y[index * self.batch_size: (index + 1) * self.batch_size]
}
)
self.inverted_params = logistic_regression_layer.params + fully_connected_layer.params
for i in xrange(n_conv_layers - 1, -1, -1):
self.inverted_params += params[i]
# Создаём список градиентов для всех параметров модели
grads = T.grad(cost, self.inverted_params)
# train_model это функция, которая обновляет параметры модели с помощью SGD
# Так как модель имеет много парамметров, было бы утомтельным вручную создавать правила обновления
# для каждой модели, поэтому мы создали updates list для автоматического прохождения по парам
# (params[i], grads[i])
updates = [
(param_i, param_i - learning_rate * grad_i)
for param_i, grad_i in zip(self.inverted_params, grads)
]
self.train_model = theano.function(
[index],
cost,
updates=updates,
givens={
x: self.train_set_x[index * self.batch_size: (index + 1) * self.batch_size],
y: self.train_set_y[index * self.batch_size: (index + 1) * self.batch_size]
}
)
set_x = T.matrix("set_x")
self.predict = theano.function(
[set_x],
logistic_regression_layer.y_pred,
givens={
x: set_x
}
)
class DBN:
def __init__(self, input=None, label=None, n_ins=2, hidden_layer_sizes=[3, 3], n_outs=2, rng=None):
self.x = input
self.y = label
self.sigmoid_layers = []
self.rbm_layers = []
self.n_layers = len(hidden_layer_sizes) # = len(self.rbm_layers)
if rng is None:
rng = numpy.random.RandomState(1234)
assert self.n_layers > 0
# construct multi-layer
for i in xrange(self.n_layers):
# layer_size
if i == 0:
input_size = n_ins
else:
input_size = hidden_layer_sizes[i - 1]
# layer_input
if i == 0:
layer_input = self.x
else:
layer_input = self.sigmoid_layers[-1].sample_h_given_v()
# construct sigmoid_layer
sigmoid_layer = HiddenLayer(input=layer_input,
n_in=input_size,
n_out=hidden_layer_sizes[i],
rng=rng,
activation=sigmoid)
self.sigmoid_layers.append(sigmoid_layer)
# construct rbm_layer
rbm_layer = RBM(input=layer_input,
n_visible=input_size,
n_hidden=hidden_layer_sizes[i],
W=sigmoid_layer.W, # W, b are shared
hbias=sigmoid_layer.b)
self.rbm_layers.append(rbm_layer)
# layer for output using Logistic Regression
self.log_layer = LogisticRegression(input=self.sigmoid_layers[-1].sample_h_given_v(),
label=self.y,
n_in=hidden_layer_sizes[-1],
n_out=n_outs)
# finetune cost: the negative log likelihood of the logistic regression layer
self.finetune_cost = self.log_layer.negative_log_likelihood()
def pretrain(self, lr=0.1, k=1, epochs=100):
# pre-train layer-wise
for i in xrange(self.n_layers):
if i == 0:
layer_input = self.x
else:
layer_input = self.sigmoid_layers[i-1].sample_h_given_v(layer_input)
rbm = self.rbm_layers[i]
for epoch in xrange(epochs):
rbm.contrastive_divergence(lr=lr, k=k, input=layer_input)
def finetune(self, lr=0.1, epochs=100):
layer_input = self.sigmoid_layers[-1].sample_h_given_v()
# train log_layer
epoch = 0
done_looping = False
while (epoch < epochs) and (not done_looping):
self.log_layer.train(lr=lr, input=layer_input)
lr *= 0.95
epoch += 1
def predict(self, x):
layer_input = x
for i in xrange(self.n_layers):
sigmoid_layer = self.sigmoid_layers[i]
layer_input = sigmoid_layer.output(input=layer_input)
out = self.log_layer.predict(layer_input)
return out
def evaluate_lenet5(learning_rate=0.1, n_epochs=200, dataset="emotion", nkerns=[20, 50], batch_size=500):
""" Demonstrates lenet on MNIST dataset
:type learning_rate: float
:param learning_rate: learning rate used (factor for the stochastic
gradient)
:type n_epochs: int
:param n_epochs: maximal number of epochs to run the optimizer
:type dataset: string
:param dataset: path to the dataset used for training /testing (MNIST here)
:type nkerns: list of ints
:param nkerns: number of kernels on each layer
"""
rng = numpy.random.RandomState(23455)
datasets = Ld.load_share(dataset)
if dataset == "mnist":
ishape = (28, 28) # this is the size of MNIST images
num_label = 10
elif dataset == "emotion":
ishape = (48, 48) # this is the size of MNIST images
num_label = 7
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
# compute number of minibatches for training, validation and testing
n_train_batches = train_set_x.get_value(borrow=True).shape[0]
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0]
n_test_batches = test_set_x.get_value(borrow=True).shape[0]
n_train_batches /= batch_size
n_valid_batches /= batch_size
n_test_batches /= batch_size
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
x = T.matrix("x") # the data is presented as rasterized images
y = T.ivector("y") # the labels are presented as 1D vector of
# [int] labels
######################
# BUILD ACTUAL MODEL #
######################
print "... building the model"
# Reshape matrix of rasterized images of shape (batch_size,28*28)
# to a 4D tensor, compatible with our LeNetConvPoolLayer
layer0_input = x.reshape((batch_size, 1, ishape[0], ishape[1]))
# Construct the first convolutional pooling layer:
# filtering reduces the image size to (28-5+1,28-5+1)=(24,24)
# maxpooling reduces this further to (24/2,24/2) = (12,12)
# 4D output tensor is thus of shape (batch_size,nkerns[0],12,12)
layer0 = LeNetConvPoolLayer(
rng,
input=layer0_input,
image_shape=(batch_size, 1, ishape[0], ishape[1]),
filter_shape=(nkerns[0], 1, 5, 5),
poolsize=(2, 2),
)
# Construct the second convolutional pooling layer
# filtering reduces the image size to (12-5+1,12-5+1)=(8,8)
# maxpooling reduces this further to (8/2,8/2) = (4,4)
# 4D output tensor is thus of shape (nkerns[0],nkerns[1],4,4)
if dataset == "emotion":
layer05 = LeNetConvPoolLayer(
rng,
input=layer0.output,
image_shape=(batch_size, nkerns[0], 22, 22),
filter_shape=(nkerns[0], nkerns[0], 3, 3),
poolsize=(2, 2),
)
layer1 = LeNetConvPoolLayer(
rng,
input=layer05.output,
image_shape=(batch_size, nkerns[0], 10, 10),
filter_shape=(nkerns[1], nkerns[0], 3, 3),
poolsize=(2, 2),
)
elif dataset == "mnist":
layer1 = LeNetConvPoolLayer(
rng,
input=layer0.output,
image_shape=(batch_size, nkerns[0], 12, 12),
filter_shape=(nkerns[1], nkerns[0], 5, 5),
poolsize=(2, 2),
)
# the TanhLayer being fully-connected, it operates on 2D matrices of
# shape (batch_size,num_pixels) (i.e matrix of rasterized images).
# This will generate a matrix of shape (20,32*4*4) = (20,512)
layer2_input = layer1.output.flatten(2)
# construct a fully-connected sigmoidal layer
layer2 = HiddenLayer(rng, input=layer2_input, n_in=nkerns[1] * 4 * 4, n_out=500, activation=T.tanh)
# classify the values of the fully-connected sigmoidal layer
layer3 = LogisticRegression(input=layer2.output, n_in=500, n_out=num_label)
# the cost we minimize during training is the NLL of the model
cost = layer3.negative_log_likelihood(y)
# create a function to compute the mistakes that are made by the model
test_model = theano.function(
[index],
layer3.errors(y),
givens={
x: test_set_x[index * batch_size : (index + 1) * batch_size],
y: test_set_y[index * batch_size : (index + 1) * batch_size],
},
)
validate_model = theano.function(
[index],
layer3.errors(y),
givens={
x: valid_set_x[index * batch_size : (index + 1) * batch_size],
y: valid_set_y[index * batch_size : (index + 1) * batch_size],
},
)
# create a list of all model parameters to be fit by gradient descent
params = layer3.params + layer2.params + layer1.params + layer0.params
# create a list of gradients for all model parameters
grads = T.grad(cost, params)
# train_model is a function that updates the model parameters by
# SGD Since this model has many parameters, it would be tedious to
# manually create an update rule for each model parameter. We thus
# create the updates list by automatically looping over all
# (params[i],grads[i]) pairs.
updates = []
for param_i, grad_i in zip(params, grads):
updates.append((param_i, param_i - learning_rate * grad_i))
train_model = theano.function(
[index],
cost,
updates=updates,
givens={
x: train_set_x[index * batch_size : (index + 1) * batch_size],
y: train_set_y[index * batch_size : (index + 1) * batch_size],
},
)
###############
# TRAIN MODEL #
###############
print "... training"
# early-stopping parameters
patience = 10000 # look as this many examples regardless
patience_increase = 2 # wait this much longer when a new best is
# found
improvement_threshold = 0.995 # a relative improvement of this much is
# considered significant
validation_frequency = min(n_train_batches, patience / 2)
# go through this many
# minibatche before checking the network
# on the validation set; in this case we
# check every epoch
best_params = None
best_validation_loss = numpy.inf
best_iter = 0
test_score = 0.0
start_time = time.clock()
epoch = 0
done_looping = False
while (epoch < n_epochs) and (not done_looping):
epoch = epoch + 1
for minibatch_index in xrange(n_train_batches):
iter = (epoch - 1) * n_train_batches + minibatch_index
if iter % 100 == 0:
print "training @ iter = ", iter
cost_ij = train_model(minibatch_index)
if (iter + 1) % validation_frequency == 0:
# compute zero-one loss on validation set
validation_losses = [validate_model(i) for i in xrange(n_valid_batches)]
this_validation_loss = numpy.mean(validation_losses)
print (
"epoch %i, minibatch %i/%i, validation error %f %%"
% (epoch, minibatch_index + 1, n_train_batches, this_validation_loss * 100.0)
)
# if we got the best validation score until now
if this_validation_loss < best_validation_loss:
# improve patience if loss improvement is good enough
if this_validation_loss < best_validation_loss * improvement_threshold:
patience = max(patience, iter * patience_increase)
# save best validation score and iteration number
best_validation_loss = this_validation_loss
best_iter = iter
# test it on the test set
test_losses = [test_model(i) for i in xrange(n_test_batches)]
test_score = numpy.mean(test_losses)
print (
(" epoch %i, minibatch %i/%i, test error of best " "model %f %%")
% (epoch, minibatch_index + 1, n_train_batches, test_score * 100.0)
)
if patience <= iter:
done_looping = True
break
end_time = time.clock()
print ("Optimization complete.")
print (
"Best validation score of %f %% obtained at iteration %i,"
"with test performance %f %%" % (best_validation_loss * 100.0, best_iter + 1, test_score * 100.0)
)
print >> sys.stderr, (
"The code for file " + os.path.split(__file__)[1] + " ran for %.2fm" % ((end_time - start_time) / 60.0)
)