def online_accuracy(self):
if self._get_current_predictions("train").size == 0:
raise Exception(
"You need to call add([prediction, label, task_id]) in order to compute an online accuracy "
"(add([prediction, label, None]) also works here, task_id is not needed)."
)
predictions = self._get_current_predictions("train")
targets = self._get_current_targets("train")
return accuracy(predictions, targets)
def average_incremental_accuracy(self):
"""Computes the average of the accuracies computed after each task.
Reference:
* iCaRL: Incremental Classifier and Representation Learning
Rebuffi et al. CVPR 2017
"""
return statistics.mean([
accuracy(self._predictions["test"][t], self._targets["test"][t])
for t in range(len(self._predictions["test"]))
])
def online_cumulative_performance(self):
"""Computes the accuracy of last task on the train set.
Reference:
* Online Fast Adaptation and Knowledge Accumulation: a New Approach to Continual Learning
Caccia et al. NeurIPS 2020
"""
return accuracy(
self._predictions["train"][-1],
self._targets["train"][-1]
)
def average_incremental_accuracy(self):
"""Computes the average of the accuracies computed after each task.
Reference:
* iCaRL: Incremental Classifier and Representation Learning
Rebuffi et al. CVPR 2017
"""
all_preds, all_targets, _ = self._get_best_epochs(subset="test")
return statistics.mean([
accuracy(all_preds[t], all_targets[t])
for t in range(len(all_preds))
])
def online_accuracy(self):
if len(self._batch_predictions) == 0:
raise Exception(
"You need to call <add_batch(preds, targets)> in order to get the online accuracy."
)
if len(self._batch_predictions) > 1:
p, t = np.concatenate(self._batch_predictions), np.concatenate(self._batch_targets)
else:
p, t = self._batch_predictions[0], self._batch_targets[0]
return accuracy(p, t)
def online_cumulative_performance(self):
"""Computes the accuracy of last task on the train set.
Reference:
* Online Fast Adaptation and Knowledge Accumulation: a New Approach to Continual Learning
Caccia et al. NeurIPS 2020
"""
preds = np.concatenate(
[dict_epoch['predictions'] for dict_epoch in self.logger_dict["train"]["performance"][self.current_task]]
)
targets = np.concatenate(
[dict_epoch['targets'] for dict_epoch in self.logger_dict["train"]["performance"][self.current_task]]
)
return accuracy(preds, targets)
def accuracy(self):
return accuracy(self._get_current_predictions("test"),
self._get_current_targets("test"))
def accuracy(self):
return accuracy(
self._predictions["test"][-1],
self._targets["test"][-1]
)
