Defining Metrics
Azimuth builds upon the HuggingFace Metric API. Please refer to HuggingFace Metric Hub that details all available metrics that can be added to Azimuth.
Metric definition
A metric definition is a simple Custom Object that loads a HuggingFace metric.
The metric custom object has an extra-field additional_kwargs
that will be pass to the compute()
function of the HuggingFace metric. This will be required for some metrics, as shown in the example
with precision, recall and F1 below.
By default, Azimuth loads Precision, Recall and F1 with average=weighted
:
{
"metrics": {
"Precision": {
"class_name": "datasets.load_metric",
"kwargs": {
"path": "precision"
},
"additional_kwargs": {
"average": "weighted"
}
},
"Recall": {
"class_name": "datasets.load_metric",
"kwargs": {
"path": "recall"
},
"additional_kwargs": {
"average": "weighted"
}
},
"F1": {
"class_name":"datasets.load_metric",
"kwargs": {
"path": "f1"
},
"additional_kwargs": {
"average": "weighted"
}
}
}
}
Custom Metrics
For more experienced users, HuggingFace presents how to create custom metrics in this tutorial.
For custom metrics, Azimuth supplies probabilities when possible. The compute
method must follow
this signature:
from typing import Any, Dict, List
import numpy as np
def _compute(
self,
predictions: List[int], # (1)
references: List[int], # (2)
probabilities: np.ndarray, # (3)
) -> Dict[str, Any]: # (4)
...
- Predicted labels (1 class per sample).
- Ground truth labels.
- Probabilities computed by the pipeline.
- A dictionary with the computed values.