class NoiseExtension(SimpleExtension, RNGMixin):
def __init__(self, noise_parameters=None, **kwargs):
kwargs.setdefault("before_training", True)
kwargs.setdefault("after_training", True)
self.noise_parameters = noise_parameters
std = 1.0
self.noise_init = IsotropicGaussian(std=std)
theano_seed = self.rng.randint(np.iinfo(np.int32).max)
self.theano_generator = UnitNoiseGenerator(std=std,
theano_seed=theano_seed)
self.noise_updates = OrderedDict([(param,
self.theano_generator.apply(param))
for param in self.noise_parameters])
super(NoiseExtension, self).__init__(**kwargs)
def do(self, callback_name, *args):
self.parse_args(callback_name, args)
if callback_name == 'before_training':
# Before training, intiaizlize noise
for p in self.noise_parameters:
self.noise_init.initialize(p, self.rng)
# And set up update to change noise on every update
self.main_loop.algorithm.add_updates(self.noise_updates)
if callback_name == 'after_training':
# After training, zero noise again.
for p in self.noise_parameters:
v = p.get_value()
p.set_value(np.zeros(v.shape, dtype=v.dtype))
class NoiseExtension(SimpleExtension, RNGMixin):
def __init__(self, noise_parameters=None, **kwargs):
kwargs.setdefault("before_training", True)
kwargs.setdefault("after_training", True)
self.noise_parameters = noise_parameters
std = 1.0
self.noise_init = IsotropicGaussian(std=std)
theano_seed = self.rng.randint(np.iinfo(np.int32).max)
self.theano_generator = UnitNoiseGenerator(
std=std, theano_seed=theano_seed)
self.noise_updates = OrderedDict(
[(param, self.theano_generator.apply(param))
for param in self.noise_parameters])
super(NoiseExtension, self).__init__(**kwargs)
def do(self, callback_name, *args):
self.parse_args(callback_name, args)
if callback_name == 'before_training':
# Before training, intiaizlize noise
for p in self.noise_parameters:
self.noise_init.initialize(p, self.rng)
# And set up update to change noise on every update
self.main_loop.algorithm.add_updates(self.noise_updates)
if callback_name == 'after_training':
# After training, zero noise again.
for p in self.noise_parameters:
v = p.get_value()
p.set_value(np.zeros(v.shape, dtype=v.dtype))
if __name__ == "__main__":
theano.config.optimizer='None'
import numpy as np
from blocks.initialization import IsotropicGaussian, Constant
x = tensor.tensor3("inp_variable")
#x = tensor.tensor3("inp_variable")
n_hid = 512
n_in = 512
np.random.seed(1)
rng = np.random
init = IsotropicGaussian(0.02)
#init = Constant(0.00)
inp_to_state = shared_floatx_zeros((n_in, n_hid))
init.initialize(inp_to_state, rng)
inp_to_update = shared_floatx_zeros((n_in, n_hid))
init.initialize(inp_to_update, rng)
inp_to_reset = shared_floatx_zeros((n_in, n_hid))
init.initialize(inp_to_reset, rng)
inp_to_state_b = shared_floatx_zeros((n_hid,))
init.initialize(inp_to_state_b, rng)
inp_to_update_b = shared_floatx_zeros((n_hid,))
init.initialize(inp_to_update_b, rng)
inp_to_reset_b = shared_floatx_zeros((n_hid,))
init.initialize(inp_to_reset_b, rng)
state_to_state = shared_floatx_zeros((n_hid, n_hid))
init.initialize(state_to_state, rng)
state_to_update = shared_floatx_zeros((n_hid, n_hid))
