class LastSelectorBlock(object):
"""
TODO(igor).
Parameters
----------
x : Matrix (GpuMatrix or CpuMatrix)
"""
def __init__(self, x):
device_id = x[0].device_id
learning = x[0].bpropagable
self.context = Context(device_id)
self.output = Matrix.empty_like(x[0])
self.output = Connector(self.output, device_id if learning else None)
if learning:
self.x, self.dL_dx = izip(*x.register_usage(device_id, device_id))
else:
self.x = x.register_usage(device_id)
self.last_idx = x.length - 1
def fprop(self):
self.output.assign(self.context, self.x[self.last_idx])
self.output.fprop()
def bprop(self):
self.dL_dx[self.last_idx].add(self.context,
self.output.backward_matrix)
class LastSelectorBlock(object):
"""
TODO(igor).
Parameters
----------
x : Matrix (GpuMatrix or CpuMatrix)
"""
def __init__(self, x):
device_id = x[0].device_id
learning = x[0].bpropagable
self.context = Context(device_id)
self.output = Matrix.empty_like(x[0])
self.output = Connector(self.output, device_id if learning else None)
if learning:
self.x, self.dL_dx = izip(*x.register_usage(device_id, device_id))
else:
self.x = x.register_usage(device_id)
self.last_idx = x.length - 1
def fprop(self):
self.output.assign(self.context, self.x[self.last_idx])
self.output.fprop()
def bprop(self):
self.dL_dx[self.last_idx].add(self.context, self.output.backward_matrix)
class GaussianNoiseBlock(object):
"""
Adds Gaussian noise to the block's input. Adding Gaussian noise can be
viewed as a regularization.
Parameters
----------
mean : float
Expected value of Gaussian noise
std : float
Standard deviation of added Gaussian noise
x : matrix
Block's input
seed : int
Seed for :func:`quagga.cuda.curand.create_generator`
device_id: int
Defines the device's id on which the computation will take place
"""
def __init__(self, mean, std, x, seed=42, device_id=None):
self.mean = mean
self.std = std
self.f_context = Context(device_id)
device_id = self.f_context.device_id
self.generator = Matrix.get_random_generator(seed)
if x.bpropagable:
self.b_context = Context(device_id)
self.x, self.dL_dx = x.register_usage(device_id, device_id)
else:
self.x = x.register_usage(device_id)
self.output = Matrix.empty_like(self.x)
self.output = Connector(self.output,
device_id if x.bpropagable else None)
self.training_mode = True
def fprop(self):
if self.training_mode:
self.x.add_gaussian_noise(self.f_context, self.generator,
self.mean, self.std, self.output)
else:
self.output.assign(self.f_context, self.x)
self.output.fprop()
def bprop(self):
self.dL_dx.add(self.b_context, self.output.backward_matrix)
def set_training_mode(self):
self.training_mode = True
def set_testing_mode(self):
self.training_mode = False
class GaussianNoiseBlock(object):
"""
Adds Gaussian noise to the block's input. Adding Gaussian noise can be
viewed as a regularization.
Parameters
----------
mean : float
Expected value of Gaussian noise
std : float
Standard deviation of added Gaussian noise
x : matrix
Block's input
seed : int
Seed for :func:`~quagga.cuda.curand.create_generator`
device_id: int
Defines the device's id on which the computation will take place
"""
def __init__(self, mean, std, x, seed=42, device_id=None):
self.mean = mean
self.std = std
self.f_context = Context(device_id)
device_id = self.f_context.device_id
self.generator = Matrix.get_random_generator(seed)
if x.bpropagable:
self.b_context = Context(device_id)
self.x, self.dL_dx = x.register_usage(device_id, device_id)
else:
self.x = x.register_usage(device_id)
self.output = Matrix.empty_like(self.x)
self.output = Connector(self.output, device_id if x.bpropagable else None)
self.training_mode = True
def fprop(self):
if self.training_mode:
self.x.add_gaussian_noise(self.f_context, self.generator, self.mean, self.std, self.output)
else:
self.output.assign(self.f_context, self.x)
self.output.fprop()
def bprop(self):
self.dL_dx.add(self.b_context, self.output.backward_matrix)
def set_training_mode(self):
self.training_mode = True
def set_testing_mode(self):
self.training_mode = False
class ScheduledSamplingBlock(object):
def __init__(self, probs, true_labels, schedule, seed, device_id=None):
self.schedule = schedule
self.rnd = np.random.RandomState(seed)
self.context = Context(device_id)
device_id = self.context.device_id
self.probs = probs.register_usage(device_id)
self.true_labels = true_labels.register_usage(device_id)
self.output = Connector(Matrix.empty_like(self.true_labels))
def fprop(self):
if self.rnd.binomial(1, self.schedule.value):
self.output.assign(self.context, self.true_labels)
else:
self.probs.argmax(self.context, self.output, axis=1)
self.output.fprop()
