OCLMetrics

class capymoa.evaluation.ocl.OCLMetrics

Bases: object

A collection of metrics evaluating an online continual learner.

We define some metrics in terms of a matrix \(R\in\mathbb{R}^{T \times T}\) (accuracy_matrix) where each element \(R_{i,j}\) contains the the test accuracy on task \(j\) after sequentially training on tasks \(1\) through \(i\).

Online learning make predictions continuously during training, so we also provide “anytime” versions of the metrics. These metrics are collected periodically during training. Specifically, \(H\) times per task. The results of this evaluation are stored in a matrix \(A\in\mathbb{R}^{T \times H \times T}\) (anytime_accuracy_matrix) where each element \(A_{i,h,j}\) contains the test accuracy on task \(j\) after sequentially training on tasks \(1\) through \(i-1\) and step \(h\) of task \(i\).

anytime_accuracy_all: ndarray

The accuracy on all tasks after training on each step in each task.

\[a_\text{any all}(t, h) = \frac{1}{T}\sum^T_{i=1} A_{t,h,i}\]

We flatten the $t,h$ dimensions to a 1D array. Use anytime_task_index to get the corresponding task index for plotting.

anytime_accuracy_all_avg: float

The average of anytime_accuracy_all over all tasks.

\[\bar{a}_\text{any all} = \frac{1}{T}\sum_{t=1}^T \frac{1}{H}\sum_{h=1}^H a_\text{any all}(t, h)\]

anytime_accuracy_seen: ndarray

The accuracy on seen tasks after training on each step in each task.

\[a_\text{any seen}(t, h) = \frac{1}{t}\sum^t_{i=1} A_{t,h,i}\]

We flatten the $t,h$ dimensions to a 1D array. Use anytime_task_index to get the corresponding task index for plotting.

anytime_accuracy_seen_avg: float

The average of anytime_accuracy_seen over all tasks.

\[\bar{a}_\text{any seen} = \frac{1}{T}\sum_{t=1}^T \frac{1}{H}\sum_{h=1}^H a_\text{any seen}(t, h)\]

anytime_task_index: ndarray

The position in each task where the anytime accuracy was measured.

accuracy_all: ndarray

The accuracy on all tasks after training on each task.

\[a_\text{all}(t) = \frac{1}{T} \sum_{i=1}^{T} R_{t,i}\]

Use task_index to get the corresponding task index for plotting.

accuracy_all_avg: float

The average of accuracy_all over all tasks.

\[\bar{a}_\text{all} = \frac{1}{T}\sum_{t=1}^T a_\text{all}(t)\]

accuracy_seen: ndarray

The accuracy on seen tasks after training on each task.

\[a_\text{seen}(t) = \frac{1}{t}\sum^t_{i=1} R_{t,i}\]

Use task_index to get the corresponding task index for plotting.

accuracy_seen_avg: float

The average of accuracy_seen over all tasks.

\[\bar{a}_\text{seen} = \frac{1}{T}\sum_{t=1}^T a_\text{seen}(t)\]

accuracy_final: float

The accuracy on the final task after training on all tasks.

\[a_\text{final} = a_\text{all}(T)\]

task_index: ndarray

The position of each task in the metrics.

forward_transfer: float

A scalar measuring the impact learning had on future tasks.

\[r_\text{FWT} = \frac{2}{T(T-1)}\sum_{i=1}^{T} \sum_{j=i+1}^{T} R_{i,j}\]

backward_transfer: float

A scalar measuring the impact learning had on past tasks.

\[r_\text{BWT} = \frac{2}{T(T-1)} \sum_{i=2}^{T} \sum_{j=1}^{i-1} (R_{i,j} - R_{j,j})\]

accuracy_matrix: ndarray

A matrix measuring the accuracy on each task after training on each task.

R[i, j] is the accuracy on task \(j\) after training on tasks \(1\) through \(i\).

anytime_accuracy_matrix: ndarray

A matrix measuring the accuracy on each task after training on each task and step.

This matrix is \(A\) with the first two dimensions flattened to a 2D array.

__init__(

anytime_accuracy_all: ndarray,

anytime_accuracy_all_avg: float,

anytime_accuracy_seen: ndarray,

anytime_accuracy_seen_avg: float,

anytime_task_index: ndarray,

accuracy_all: ndarray,

accuracy_all_avg: float,

accuracy_seen: ndarray,

accuracy_seen_avg: float,

accuracy_final: float,

task_index: ndarray,

forward_transfer: float,

backward_transfer: float,

accuracy_matrix: ndarray,

anytime_accuracy_matrix: ndarray,

prequential_cumulative_accuracy: float,

) → None

prequential_cumulative_accuracy: float

The average test-then-train accuracy over the entire data-stream.