def training_updates(visible_batch, model, sampler, optimizer):
"""
Combine together updates from various sources for RBM training.
Parameters
----------
visible_batch : tensor_like
Theano symbolic representing a minibatch on the visible units,
with the first dimension indexing training examples and the second
indexing data dimensions.
rbm : object
An instance of `RBM` or a derived class, or one implementing
the RBM interface.
sampler : object
An instance of `Sampler` or a derived class, or one implementing
the sampler interface.
optimizer : object
An instance of `Optimizer` or a derived class, or one implementing
the optimizer interface (typically an `SGDOptimizer`).
"""
pos_v = visible_batch
neg_v = sampler.particles
grads = model.ml_gradients(pos_v, neg_v)
ups = optimizer.updates(gradients=grads)
# Add the sampler's updates (negative phase particles, etc.).
safe_update(ups, sampler.updates())
return ups
def updates(self, gradients):
"""
Return symbolic updates to apply given a set of gradients
on the parameters being optimized.
Parameters
----------
gradients : list of tensor_likes
List of symbolic gradients for the parameters contained
in self.params, in the same order as in self.params.
Returns
-------
updates : dict
A dictionary with the shared variables in self.params as keys
and a symbolic expression of how they are to be updated each
SGD step as values.
Notes
-----
`cost_updates` is a convenient helper function that takes all
necessary gradients with respect to a given symbolic cost.
"""
ups = {}
# Add the learning rate/iteration updates
l_ups, learn_rates = self.learning_rate_updates(gradients)
safe_update(ups, l_ups)
# Get the updates from sgd_updates, a PyLearn library function.
p_up = dict(self.sgd_updates(self.params, gradients, learn_rates))
# Add the things in p_up to ups
safe_update(ups, p_up)
# Clip the values if needed.
# We do not want the clipping values to force an upcast
# of the update: updates should have the same type as params
for param, (p_min, p_max) in self.clipping_values.iteritems():
p_min = tensor.as_tensor(p_min)
p_max = tensor.as_tensor(p_max)
dtype = param.dtype
if p_min.dtype != dtype:
p_min = tensor.cast(p_min, dtype)
if p_max.dtype != dtype:
p_max = tensor.cast(p_max, dtype)
ups[param] = tensor.clip(ups[param], p_min, p_max)
# Return the updates dictionary.
return ups
def updates(self, gradients):
"""
Return symbolic updates to apply given a set of gradients
on the parameters being optimized.
Parameters
----------
gradients : list of tensor_likes
List of symbolic gradients for the parameters contained
in self.params, in the same order as in self.params.
Returns
-------
updates : dict
A dictionary with the shared variables in self.params as keys
and a symbolic expression of how they are to be updated each
SGD step as values.
Notes
-----
`cost_updates` is a convenient helper function that takes all
necessary gradients with respect to a given symbolic cost.
"""
ups = {}
# Add the learning rate/iteration updates
l_ups, learn_rates = self.learning_rate_updates()
safe_update(ups, l_ups)
# Get the updates from sgd_updates, a PyLearn library function.
p_up = dict(sgd_updates(self.params, gradients, learn_rates))
# Add the things in p_up to ups
safe_update(ups, p_up)
# Clip the values if needed.
# We do not want the clipping values to force an upcast
# of the update: updates should have the same type as params
for param, (p_min, p_max) in self.clipping_values.iteritems():
p_min = tensor.as_tensor(p_min)
p_max = tensor.as_tensor(p_max)
dtype = param.dtype
if p_min.dtype != dtype:
p_min = tensor.cast(p_min, dtype)
if p_max.dtype != dtype:
p_max = tensor.cast(p_max, dtype)
ups[param] = tensor.clip(ups[param], p_min, p_max)
# Return the updates dictionary.
return ups
def training_updates(visible_batch, model, sampler, optimizer):
"""
Combine together updates from various sources for RBM training.
Parameters
----------
visible_batch : tensor_like
Theano symbolic representing a minibatch on the visible units, \
with the first dimension indexing training examples and the second \
indexing data dimensions.
rbm : object
An instance of `RBM` or a derived class, or one implementing \
the RBM interface.
sampler : object
An instance of `Sampler` or a derived class, or one implementing \
the sampler interface.
optimizer : object
An instance of `_Optimizer` or a derived class, or one implementing \
the optimizer interface (typically an `_SGDOptimizer`).
Returns
-------
WRITEME
"""
# TODO: the Optimizer object got deprecated, and this is the only
# functionality that requires it. We moved the Optimizer
# here with an _ before its name.
# We should figure out how best to refactor the code.
# Optimizer was problematic because people kept using SGDOptimizer
# instead of training_algorithms.sgd.
# Compute negative phase updates.
sampler_updates = sampler.updates()
# Compute SML gradients.
pos_v = visible_batch
#neg_v = sampler_updates[sampler.particles]
neg_v = sampler.particles
grads = model.ml_gradients(pos_v, neg_v)
# Build updates dictionary combining (gradient, sampler) updates.
ups = optimizer.updates(gradients=grads)
safe_update(ups, sampler_updates)
return ups
def training_updates(visible_batch, model, sampler, optimizer):
"""
Combine together updates from various sources for RBM training.
Parameters
----------
visible_batch : tensor_like
Theano symbolic representing a minibatch on the visible units, \
with the first dimension indexing training examples and the second \
indexing data dimensions.
rbm : object
An instance of `RBM` or a derived class, or one implementing \
the RBM interface.
sampler : object
An instance of `Sampler` or a derived class, or one implementing \
the sampler interface.
optimizer : object
An instance of `_Optimizer` or a derived class, or one implementing \
the optimizer interface (typically an `_SGDOptimizer`).
Returns
-------
WRITEME
"""
# TODO: the Optimizer object got deprecated, and this is the only
# functionality that requires it. We moved the Optimizer
# here with an _ before its name.
# We should figure out how best to refactor the code.
# Optimizer was problematic because people kept using SGDOptimizer
# instead of training_algorithms.sgd.
# Compute negative phase updates.
sampler_updates = sampler.updates()
# Compute SML gradients.
pos_v = visible_batch
#neg_v = sampler_updates[sampler.particles]
neg_v = sampler.particles
grads = model.ml_gradients(pos_v, neg_v)
# Build updates dictionary combining (gradient, sampler) updates.
ups = optimizer.updates(gradients=grads)
safe_update(ups, sampler_updates)
return ups
def get_monitoring_channels(self, data):
"""
Notes
-----
Monitors quantities related to the approximate posterior parameters phi
and the conditional and prior parameters theta.
"""
space, source = self.get_monitoring_data_specs()
space.validate(data)
rval = OrderedDict()
X = data
epsilon_shape = (X.shape[0], self.nhid)
epsilon = self.sample_from_epsilon(shape=epsilon_shape)
phi = self.encode_phi(X)
z = self.sample_from_q_z_given_x(epsilon=epsilon, phi=phi)
theta = self.decode_theta(z)
X_r = self.means_from_theta(theta)
rval["reconstruction_mse"] = T.sqr(X - X_r).mean()
posterior_channels = \
self.posterior.monitoring_channels_from_conditional_params(phi)
safe_update(rval, posterior_channels)
conditional_channels = \
self.conditional.monitoring_channels_from_conditional_params(theta)
safe_update(rval, conditional_channels)
prior_channels = self.prior.monitoring_channels_from_prior_params()
safe_update(rval, prior_channels)
return rval
def get_lr_scalers(self):
rval = OrderedDict()
if self.encoder is not None:
safe_update(rval, self.encoder.get_lr_scalers())
return rval
def get_lr_scalers(self):
rval = self.prior.get_lr_scalers()
safe_update(rval, self.conditional.get_lr_scalers())
safe_update(rval, self.posterior.get_lr_scalers())
return rval
def get_lr_scalers(self):
rval = OrderedDict()
if self.encoder is not None:
safe_update(rval, self.encoder.get_lr_scalers())
return rval
