def validate(self, params, server_stamp, unblock):
"""
Optionally run validation policy.
This may include all sorts of logging etc.
"""
if unblock and (server_stamp) % self.validate_frequency == 0:
# Make sure that we run validation with the server parameters
param_copy(params, self.classifier.params)
validation_losses = [
self.validate_model(i) for i in xrange(self.n_valid_batches)
]
this_validation_loss = np.mean(validation_losses)
logging.info('server_stamp %i, validation error %f %%' %
(server_stamp, this_validation_loss * 100.))
self.validation_timestamps.append(server_stamp)
self.validation_results.append(this_validation_loss)
def validate(self, params, server_stamp, unblock):
"""
Optionally run validation policy.
This may include all sorts of logging etc.
"""
if unblock and (server_stamp) % self.validate_frequency == 0:
# Make sure that we run validation with the server parameters
param_copy(params, self.classifier.params)
validation_losses = [self.validate_model(i)
for i in xrange(self.n_valid_batches)]
this_validation_loss = np.mean(validation_losses)
logging.info('server_stamp %i, validation error %f %%' %
( server_stamp, this_validation_loss * 100.)
)
self.validation_timestamps.append(server_stamp)
self.validation_results.append(this_validation_loss)
def fetch_update(self, iteration):
"""
Decide where the next grad update will come from.
Also decides whether to drop the grad
"""
client_id = sample(self.client_priorities, self.blocking)
self.update_priorities()
# overwrite the models params with client params
param_copy(self.client_params[client_id], self.classifier.params)
minibatch_index = iteration % self.n_train_batches
grads, parameter_timestamp = self.pusher(self,
minibatch_index,
client_id,
j=self.j)
return (grads, parameter_timestamp, client_id)
def fetch_update(self, iteration):
"""
Decide where the next grad update will come from.
Also decides whether to drop the grad
"""
client_id = sample(self.client_priorities, self.blocking)
self.update_priorities()
# overwrite the models params with client params
param_copy(self.client_params[client_id], self.classifier.params)
minibatch_index = iteration % self.n_train_batches
grads, parameter_timestamp = self.pusher(self,
minibatch_index,
client_id,
j=self.j)
return (grads, parameter_timestamp, client_id)
def update_parameters(self, params, client, server_stamp, unblock, v):
"""
Give new parameters to a client.
"""
# update grad magnitude mean
self.v = v
# add client to blocking set no matter what
if client not in self.blocking:
self.blocking.append(client)
if unblock: # update params of all blocking clients
for client_id in self.blocking:
if self.fetcher(self, self.k, client):
# we've decided to pull updates
param_copy(params, self.client_params[client_id])
else:
# we're going to ignore that (we could have not copied it)
pass
self.parameter_stamps[client_id] = server_stamp
self.blocking = [] # nothing is blocking now
def update_parameters(self, params, client, server_stamp, unblock, v):
"""
Give new parameters to a client.
"""
# update grad magnitude mean
self.v = v
# add client to blocking set no matter what
if client not in self.blocking:
self.blocking.append(client)
if unblock: # update params of all blocking clients
for client_id in self.blocking:
if self.fetcher(self, self.k, client):
# we've decided to pull updates
param_copy(params, self.client_params[client_id])
else:
# we're going to ignore that (we could have not copied it)
pass
self.parameter_stamps[client_id] = server_stamp
self.blocking = [] # nothing is blocking now
