################################################################################
# Copyright (c) 2021 ContinualAI. #
# Copyrights licensed under the MIT License. #
# See the accompanying LICENSE file for terms. #
# #
# Date: 30-12-2020 #
# Author(s): Lorenzo Pellegrini #
# E-mail: contact@continualai.org #
# Website: www.continualai.org #
################################################################################
from typing import List, Union, Dict
import torch
from torch import Tensor
from avalanche.evaluation import Metric, PluginMetric, GenericPluginMetric
from avalanche.evaluation.metrics.mean import Mean
from avalanche.evaluation.metric_utils import phase_and_task
from collections import defaultdict
[docs]class Accuracy(Metric[float]):
"""Accuracy metric. This is a standalone metric.
The update method computes the accuracy incrementally
by keeping a running average of the <prediction, target> pairs
of Tensors provided over time.
The "prediction" and "target" tensors may contain plain labels or
one-hot/logit vectors.
Each time `result` is called, this metric emits the average accuracy
across all predictions made since the last `reset`.
The reset method will bring the metric to its initial state. By default
this metric in its initial state will return an accuracy value of 0.
"""
[docs] def __init__(self):
"""Creates an instance of the standalone Accuracy metric.
By default this metric in its initial state will return an accuracy
value of 0. The metric can be updated by using the `update` method
while the running accuracy can be retrieved using the `result` method.
"""
self._mean_accuracy = Mean()
"""The mean utility that will be used to store the running accuracy."""
@torch.no_grad()
def update(
self,
predicted_y: Tensor,
true_y: Tensor,
) -> None:
"""Update the running accuracy given the true and predicted labels.
:param predicted_y: The model prediction. Both labels and logit vectors
are supported.
:param true_y: The ground truth. Both labels and one-hot vectors
are supported.
:return: None.
"""
true_y = torch.as_tensor(true_y)
predicted_y = torch.as_tensor(predicted_y)
if len(true_y) != len(predicted_y):
raise ValueError("Size mismatch for true_y and predicted_y tensors")
# Check if logits or labels
if len(predicted_y.shape) > 1:
# Logits -> transform to labels
predicted_y = torch.max(predicted_y, 1)[1]
if len(true_y.shape) > 1:
# Logits -> transform to labels
true_y = torch.max(true_y, 1)[1]
true_positives = float(torch.sum(torch.eq(predicted_y, true_y)))
total_patterns = len(true_y)
self._mean_accuracy.update(
true_positives / total_patterns, total_patterns
)
def result(self) -> float:
"""Retrieves the running accuracy.
Calling this method will not change the internal state of the metric.
:return: The current running accuracy, which is a float value
between 0 and 1.
"""
return self._mean_accuracy.result()
def reset(self) -> None:
"""Resets the metric.
:return: None.
"""
self._mean_accuracy.reset()
[docs]class TaskAwareAccuracy(Metric[float]):
"""The task-aware Accuracy metric.
The metric computes a dictionary of <task_label, accuracy value> pairs.
update/result/reset methods are all task-aware.
See :class:`avalanche.evaluation.Accuracy` for the documentation about
the `Accuracy` metric.
"""
[docs] def __init__(self):
"""Creates an instance of the task-aware Accuracy metric."""
self._mean_accuracy = defaultdict(Accuracy)
"""
The mean utility that will be used to store the running accuracy
for each task label.
"""
@torch.no_grad()
def update(
self,
predicted_y: Tensor,
true_y: Tensor,
task_labels: Union[float, Tensor],
) -> None:
"""Update the running accuracy given the true and predicted labels.
Parameter `task_labels` is used to decide how to update the inner
dictionary: if Float, only the dictionary value related to that task
is updated. If Tensor, all the dictionary elements belonging to the
task labels will be updated.
:param predicted_y: The model prediction. Both labels and logit vectors
are supported.
:param true_y: The ground truth. Both labels and one-hot vectors
are supported.
:param task_labels: the int task label associated to the current
experience or the task labels vector showing the task label
for each pattern.
:return: None.
"""
if len(true_y) != len(predicted_y):
raise ValueError("Size mismatch for true_y and predicted_y tensors")
if isinstance(task_labels, Tensor) and len(task_labels) != len(true_y):
raise ValueError("Size mismatch for true_y and task_labels tensors")
if isinstance(task_labels, int):
self._mean_accuracy[task_labels].update(predicted_y, true_y)
elif isinstance(task_labels, Tensor):
for pred, true, t in zip(predicted_y, true_y, task_labels):
if isinstance(t, Tensor):
t = t.item()
self._mean_accuracy[t].update(
pred.unsqueeze(0), true.unsqueeze(0)
)
else:
raise ValueError(
f"Task label type: {type(task_labels)}, "
f"expected int/float or Tensor"
)
def result(self, task_label=None) -> Dict[int, float]:
"""
Retrieves the running accuracy.
Calling this method will not change the internal state of the metric.
task label is ignored if `self.split_by_task=False`.
:param task_label: if None, return the entire dictionary of accuracies
for each task. Otherwise return the dictionary
`{task_label: accuracy}`.
:return: A dict of running accuracies for each task label,
where each value is a float value between 0 and 1.
"""
assert task_label is None or isinstance(task_label, int)
if task_label is None:
return {k: v.result() for k, v in self._mean_accuracy.items()}
else:
return {task_label: self._mean_accuracy[task_label].result()}
def reset(self, task_label=None) -> None:
"""
Resets the metric.
task label is ignored if `self.split_by_task=False`.
:param task_label: if None, reset the entire dictionary.
Otherwise, reset the value associated to `task_label`.
:return: None.
"""
assert task_label is None or isinstance(task_label, int)
if task_label is None:
self._mean_accuracy = defaultdict(Accuracy)
else:
self._mean_accuracy[task_label].reset()
class AccuracyPluginMetric(GenericPluginMetric[float]):
"""
Base class for all accuracies plugin metrics
"""
def __init__(self, reset_at, emit_at, mode, split_by_task=False):
"""Creates the Accuracy plugin
:param reset_at:
:param emit_at:
:param mode:
:param split_by_task: whether to compute task-aware accuracy or not.
"""
self.split_by_task = split_by_task
if self.split_by_task:
self._accuracy = TaskAwareAccuracy()
else:
self._accuracy = Accuracy()
super(AccuracyPluginMetric, self).__init__(
self._accuracy, reset_at=reset_at, emit_at=emit_at, mode=mode
)
def reset(self, strategy=None) -> None:
self._metric.reset()
def result(self, strategy=None) -> float:
return self._metric.result()
def update(self, strategy):
if isinstance(self._accuracy, Accuracy):
self._accuracy.update(strategy.mb_output, strategy.mb_y)
elif isinstance(self._accuracy, TaskAwareAccuracy):
self._accuracy.update(
strategy.mb_output, strategy.mb_y, strategy.mb_task_id
)
else:
assert False, "should never get here."
[docs]class MinibatchAccuracy(AccuracyPluginMetric):
"""
The minibatch plugin accuracy metric.
This metric only works at training time.
This metric computes the average accuracy over patterns
from a single minibatch.
It reports the result after each iteration.
If a more coarse-grained logging is needed, consider using
:class:`EpochAccuracy` instead.
"""
[docs] def __init__(self):
"""
Creates an instance of the MinibatchAccuracy metric.
"""
super(MinibatchAccuracy, self).__init__(
reset_at="iteration", emit_at="iteration", mode="train"
)
def __str__(self):
return "Top1_Acc_MB"
[docs]class EpochAccuracy(AccuracyPluginMetric):
"""
The average accuracy over a single training epoch.
This plugin metric only works at training time.
The accuracy will be logged after each training epoch by computing
the number of correctly predicted patterns during the epoch divided by
the overall number of patterns encountered in that epoch.
"""
[docs] def __init__(self):
"""
Creates an instance of the EpochAccuracy metric.
"""
super(EpochAccuracy, self).__init__(
reset_at="epoch", emit_at="epoch", mode="train"
)
def __str__(self):
return "Top1_Acc_Epoch"
[docs]class RunningEpochAccuracy(AccuracyPluginMetric):
"""
The average accuracy across all minibatches up to the current
epoch iteration.
This plugin metric only works at training time.
At each iteration, this metric logs the accuracy averaged over all patterns
seen so far in the current epoch.
The metric resets its state after each training epoch.
"""
[docs] def __init__(self):
"""
Creates an instance of the RunningEpochAccuracy metric.
"""
super(RunningEpochAccuracy, self).__init__(
reset_at="epoch", emit_at="iteration", mode="train"
)
def __str__(self):
return "Top1_RunningAcc_Epoch"
[docs]class ExperienceAccuracy(AccuracyPluginMetric):
"""
At the end of each experience, this plugin metric reports
the average accuracy over all patterns seen in that experience.
This metric only works at eval time.
"""
[docs] def __init__(self):
"""
Creates an instance of ExperienceAccuracy metric
"""
super(ExperienceAccuracy, self).__init__(
reset_at="experience", emit_at="experience", mode="eval"
)
def __str__(self):
return "Top1_Acc_Exp"
[docs]class StreamAccuracy(AccuracyPluginMetric):
"""
At the end of the entire stream of experiences, this plugin metric
reports the average accuracy over all patterns seen in all experiences.
This metric only works at eval time.
"""
[docs] def __init__(self):
"""
Creates an instance of StreamAccuracy metric
"""
super(StreamAccuracy, self).__init__(
reset_at="stream", emit_at="stream", mode="eval"
)
def __str__(self):
return "Top1_Acc_Stream"
[docs]class TrainedExperienceAccuracy(AccuracyPluginMetric):
"""
At the end of each experience, this plugin metric reports the average
accuracy for only the experiences that the model has been trained on so far.
This metric only works at eval time.
"""
[docs] def __init__(self):
"""
Creates an instance of TrainedExperienceAccuracy metric by first
constructing AccuracyPluginMetric
"""
super(TrainedExperienceAccuracy, self).__init__(
reset_at="stream", emit_at="stream", mode="eval"
)
self._current_experience = 0
def after_training_exp(self, strategy) -> None:
self._current_experience = strategy.experience.current_experience
# Reset average after learning from a new experience
AccuracyPluginMetric.reset(self, strategy)
return AccuracyPluginMetric.after_training_exp(self, strategy)
def update(self, strategy):
"""
Only update the accuracy with results from experiences that have been
trained on
"""
if strategy.experience.current_experience <= self._current_experience:
AccuracyPluginMetric.update(self, strategy)
def __str__(self):
return "Accuracy_On_Trained_Experiences"
[docs]def accuracy_metrics(
*,
minibatch=False,
epoch=False,
epoch_running=False,
experience=False,
stream=False,
trained_experience=False,
) -> List[PluginMetric]:
"""
Helper method that can be used to obtain the desired set of
plugin metrics.
:param minibatch: If True, will return a metric able to log
the minibatch accuracy at training time.
:param epoch: If True, will return a metric able to log
the epoch accuracy at training time.
:param epoch_running: If True, will return a metric able to log
the running epoch accuracy at training time.
:param experience: If True, will return a metric able to log
the accuracy on each evaluation experience.
:param stream: If True, will return a metric able to log
the accuracy averaged over the entire evaluation stream of experiences.
:param trained_experience: If True, will return a metric able to log
the average evaluation accuracy only for experiences that the
model has been trained on
:return: A list of plugin metrics.
"""
metrics = []
if minibatch:
metrics.append(MinibatchAccuracy())
if epoch:
metrics.append(EpochAccuracy())
if epoch_running:
metrics.append(RunningEpochAccuracy())
if experience:
metrics.append(ExperienceAccuracy())
if stream:
metrics.append(StreamAccuracy())
if trained_experience:
metrics.append(TrainedExperienceAccuracy())
return metrics
__all__ = [
"Accuracy",
"TaskAwareAccuracy",
"MinibatchAccuracy",
"EpochAccuracy",
"RunningEpochAccuracy",
"ExperienceAccuracy",
"StreamAccuracy",
"TrainedExperienceAccuracy",
"accuracy_metrics",
]