def on_epoch_end(self, epoch, logs: dict = None):
self.loss.append(logs.get(METRIC_LOSS))
self.accuracy.append(logs.get(METRIC_ACCURACY))
self.val_loss.append(logs.get(METRIC_VAL_LOSS, 0))
self.val_accuracy.append(logs.get(METRIC_VAL_ACCURACY, 0))
if self.config.dp:
# Account for tf-privacy library writing to stdout
with redirect_stdout(io.StringIO()):
eps, _ = compute_epsilon(self.total_token_count, self.config,
epoch)
logs[METRIC_EPSILON] = eps
# NOTE: this is just a list of the same value, but
# is simpler for creating the history csv
delta = 1 / float(self.total_token_count)
logs[METRIC_DELTA] = delta
self.epsilon.append(logs.get(METRIC_EPSILON, 0))
self.delta.append(logs.get(METRIC_DELTA, 0))
# NOTE: When we do the final history saving, one of these items
# will flip to 1 to denote the actual best model that is saved
# This just seeds the column for now
self.best.append(0)
def on_epoch_end(self, epoch, logs: dict = None):
self.loss.append(logs.get(METRIC_LOSS))
self.accuracy.append(logs.get(METRIC_ACCURACY))
self.val_loss.append(logs.get(METRIC_VAL_LOSS, 0))
self.val_accuracy.append(logs.get(METRIC_VAL_ACCURACY, 0))
if self.config.dp:
# Account for tf-privacy library writing to stdout
with redirect_stdout(io.StringIO()):
eps, _ = compute_epsilon(self.total_token_count, self.config,
epoch)
self.epsilon.append(eps)
# NOTE: this is just a list of the same value, but
# is simpler for creating the history csv
self.delta.append(1 / float(self.total_token_count))
else:
# Append 0,0 for epsilon and delta. These will be dropped afterwards anyway in
# non-DP mode.
self.epsilon.append(0)
self.delta.append(0)
# NOTE: When we do the final history saving, one of these items
# will flip to 1 to denote the actual best model that is saved
# This just seeds the column for now
self.best.append(0)