def before_training_exp(self, strategy, **kwargs):
if self.freeze_remaining_model and strategy.training_exp_counter > 0:
self.freeze_other_layers()
# Count current classes and number of samples for each of them.
data = strategy.experience.dataset
self.model.cur_j = examples_per_class(data.targets)
self.cur_class = [cls for cls in set(self.model.cur_j.keys()) if
self.model.cur_j[cls] > 0]
self.reset_weights(self.cur_class)
def _before_training_exp(self, **kwargs):
self.model.eval()
self.model.end_features.train()
self.model.output.train()
if self.clock.train_exp_counter > 0:
# In AR1 batch 0 is treated differently as the feature extractor is
# left more free to learn.
# This if is executed for batch > 0, in which we freeze layers
# below "self.freeze_below_layer" (which usually is the latent
# replay layer!) and we also change the parameters of BatchReNorm
# layers to a more conservative configuration.
# "freeze_up_to" will freeze layers below "freeze_below_layer"
# Beware that Batch ReNorm layers are not frozen!
freeze_up_to(
self.model,
freeze_until_layer=self.freeze_below_layer,
layer_filter=AR1.filter_bn_and_brn,
)
# Adapt the parameters of BatchReNorm layers
change_brn_pars(
self.model,
momentum=self.inc_update_rate,
r_d_max_inc_step=0,
r_max=self.max_r_max,
d_max=self.max_d_max,
)
# Adapt the model and optimizer
self.model = self.model.to(self.device)
self.optimizer = SGD(
self.model.parameters(),
lr=self.lr,
momentum=self.momentum,
weight_decay=self.l2,
)
# super()... will run S.I. and CWR* plugin callbacks
super()._before_training_exp(**kwargs)
# Update cur_j of CWR* to consider latent patterns
if self.clock.train_exp_counter > 0:
for class_id, count in examples_per_class(self.rm[1]).items():
self.model.cur_j[class_id] += count
self.cwr_plugin.cur_class = [
cls
for cls in set(self.model.cur_j.keys())
if self.model.cur_j[cls] > 0
]
self.cwr_plugin.reset_weights(self.cwr_plugin.cur_class)