def __init__(self, window_size=1, m=1, r=2, batch_size=1):
self.r = r
self.window_size = window_size
self.n_batches = (self.window_size / batch_size)
self.n_input = m * (r + 1)
self.n_output = m
self.data = [[0 for j in xrange(m)]
for i in xrange(window_size + (r + 1))]
self.x_value_pred = numpy.zeros((window_size, self.n_input),
dtype=theano.config.floatX)
self.x_value = numpy.zeros((window_size, self.n_input),
dtype=theano.config.floatX)
self.x = theano.shared(self.x_value, borrow=True)
self.y_value = numpy.zeros((window_size, self.n_output),
dtype=theano.config.floatX)
self.y = theano.shared(self.y_value, borrow=True)
numpy_rng = numpy.random.RandomState(89677)
print '... building the model'
# construct the stacked denoising autoencoder class
self.sda = SdA(numpy_rng=numpy_rng,
n_ins=self.n_input,
hidden_layers_sizes=[10],
n_outs=self.n_output)
# retrieving functions
self.pretraining_fns = self.sda.pretraining_functions(
train_set_x=self.x, batch_size=batch_size)
self.train_fn = self.sda.build_finetune_function(
train_set_x=self.x,
train_set_y=self.y,
batch_size=batch_size,
)
self.predict_fn = self.sda.build_predict_function()
def __init__(self, window_size=1, m=1, r=2, batch_size=1, hidden_layers_sizes=[10]):
self.r = r
self.window_size = window_size
self.n_batches = (self.window_size / batch_size)
self.n_input = m*(r+1)
self.n_output = m
self.data = [[0 for j in xrange(m)] for i in xrange(window_size + (r+1))]
self.x_value_pred = numpy.zeros((window_size, self.n_input), dtype=theano.config.floatX)
self.x_value = numpy.zeros((window_size, self.n_input), dtype=theano.config.floatX)
self.x = theano.shared(self.x_value, borrow=True)
self.y_value = numpy.zeros((window_size, self.n_output), dtype=theano.config.floatX)
self.y = theano.shared(self.y_value, borrow=True)
numpy_rng = numpy.random.RandomState(89677)
print '... building the model'
# construct the stacked denoising autoencoder class
self.sda = SdA(
numpy_rng=numpy_rng,
n_ins=self.n_input,
hidden_layers_sizes=hidden_layers_sizes,
n_outs=self.n_output
)
# retrieving functions
self.pretraining_fns = self.sda.pretraining_functions(
train_set_x=self.x,
batch_size=batch_size
)
self.train_fn = self.sda.build_finetune_function(
train_set_x=self.x,
train_set_y=self.y,
batch_size=batch_size,
)
self.predict_fn = self.sda.build_prediction_function()
class TestBed:
def __init__(self, window_size=1, m=1, r=2, batch_size=1, hidden_layers_sizes=[10]):
self.r = r
self.window_size = window_size
self.n_batches = (self.window_size / batch_size)
self.n_input = m*(r+1)
self.n_output = m
self.data = [[0 for j in xrange(m)] for i in xrange(window_size + (r+1))]
self.x_value_pred = numpy.zeros((window_size, self.n_input), dtype=theano.config.floatX)
self.x_value = numpy.zeros((window_size, self.n_input), dtype=theano.config.floatX)
self.x = theano.shared(self.x_value, borrow=True)
self.y_value = numpy.zeros((window_size, self.n_output), dtype=theano.config.floatX)
self.y = theano.shared(self.y_value, borrow=True)
numpy_rng = numpy.random.RandomState(89677)
print '... building the model'
# construct the stacked denoising autoencoder class
self.sda = SdA(
numpy_rng=numpy_rng,
n_ins=self.n_input,
hidden_layers_sizes=hidden_layers_sizes,
n_outs=self.n_output
)
# retrieving functions
self.pretraining_fns = self.sda.pretraining_functions(
train_set_x=self.x,
batch_size=batch_size
)
self.train_fn = self.sda.build_finetune_function(
train_set_x=self.x,
train_set_y=self.y,
batch_size=batch_size,
)
self.predict_fn = self.sda.build_prediction_function()
def supply(self, y):
self.data.append(y)
while self.window_size + (self.r+1) < len(self.data):
self.data.pop(0)
for i in xrange(self.x_value.shape[0]-1):
self.x_value_pred[i] = self.x_value_pred[i+1]
self.x_value[i] = self.x_value[i+1]
self.y_value[i] = self.y_value[i+1]
self.x_value_pred[-1] = [self.data[-1-i%(self.r+1)][int(i/(self.r+1))] for i in xrange(self.x_value.shape[1])]
self.x_value[-1] = self.x_value_pred[-2]
self.y_value[-1] = y
def pretrain(self, pretraining_epochs, pretraining_lr):
total_epochs = 0
avg_cost = 0
for i in xrange(self.sda.n_layers):
corruption_level = i / 10.0
# go through pretraining epochs
for epoch in xrange(pretraining_epochs):
total_epochs = total_epochs + 1
# go through the training set
c = []
for batch_index in xrange(self.n_batches):
c.append(self.pretraining_fns[i](index=batch_index, corruption=corruption_level, lr=pretraining_lr))
minibatch_avg_cost = numpy.mean(c)
avg_cost = avg_cost + minibatch_avg_cost
avg_cost = avg_cost / total_epochs if 0 < total_epochs else 0
return avg_cost
def finetune(self, finetunning_epochs, finetunning_lr=0.1):
done_looping = False
epoch = 0
avg_cost = 0
while (epoch < finetunning_epochs) and (not done_looping):
epoch = epoch + 1
batch_idx = range(self.n_batches)
numpy.random.shuffle(batch_idx)
for minibatch_index in batch_idx:
minibatch_avg_cost = self.train_fn(minibatch_index, lr=finetunning_lr)
avg_cost = avg_cost + minibatch_avg_cost
avg_cost = avg_cost / epoch if 0 < epoch else 0
return avg_cost
def predict(self):
return self.predict_fn(self.x_value_pred)[-1]
class TestBed:
def __init__(self, window_size=1, m=1, r=2, batch_size=1):
self.r = r
self.window_size = window_size
self.n_batches = (self.window_size / batch_size)
self.n_input = m * (r + 1)
self.n_output = m
self.data = [[0 for j in xrange(m)]
for i in xrange(window_size + (r + 1))]
self.x_value_pred = numpy.zeros((window_size, self.n_input),
dtype=theano.config.floatX)
self.x_value = numpy.zeros((window_size, self.n_input),
dtype=theano.config.floatX)
self.x = theano.shared(self.x_value, borrow=True)
self.y_value = numpy.zeros((window_size, self.n_output),
dtype=theano.config.floatX)
self.y = theano.shared(self.y_value, borrow=True)
numpy_rng = numpy.random.RandomState(89677)
print '... building the model'
# construct the stacked denoising autoencoder class
self.sda = SdA(numpy_rng=numpy_rng,
n_ins=self.n_input,
hidden_layers_sizes=[10],
n_outs=self.n_output)
# retrieving functions
self.pretraining_fns = self.sda.pretraining_functions(
train_set_x=self.x, batch_size=batch_size)
self.train_fn = self.sda.build_finetune_function(
train_set_x=self.x,
train_set_y=self.y,
batch_size=batch_size,
)
self.predict_fn = self.sda.build_predict_function()
def supply(self, y):
self.data.append(y)
while self.window_size + (self.r + 1) < len(self.data):
self.data.pop(0)
for i in xrange(self.x_value.shape[0] - 1):
self.x_value_pred[i] = self.x_value_pred[i + 1]
self.x_value[i] = self.x_value[i + 1]
self.y_value[i] = self.y_value[i + 1]
self.x_value_pred[-1] = [
self.data[-1 - i % (self.r + 1)][int(i / (self.r + 1))]
for i in xrange(self.x_value.shape[1])
]
self.x_value[-1] = self.x_value_pred[-2]
self.y_value[-1] = y
def pretrain(self, pretraining_epochs, pretraining_lr):
corruption_levels = [.1, .2, .3]
total_epochs = 0
avg_cost = 0
for i in xrange(self.sda.n_layers):
# go through pretraining epochs
for epoch in xrange(pretraining_epochs):
total_epochs = total_epochs + 1
# go through the training set
c = []
for batch_index in xrange(self.n_batches):
c.append(self.pretraining_fns[i](
index=batch_index,
corruption=corruption_levels[i],
lr=pretraining_lr))
minibatch_avg_cost = numpy.mean(c)
avg_cost = avg_cost + minibatch_avg_cost
avg_cost /= total_epochs
return avg_cost
def finetune(self, finetunning_epochs, finetunning_lr=0.1):
done_looping = False
epoch = 0
avg_cost = 0
while (epoch < finetunning_epochs) and (not done_looping):
epoch = epoch + 1
for minibatch_index in xrange(self.n_batches):
minibatch_avg_cost = self.train_fn(minibatch_index,
lr=finetunning_lr)
avg_cost = avg_cost + minibatch_avg_cost
avg_cost /= epoch
return avg_cost
def predict(self):
return self.predict_fn(self.x_value_pred)[0]
