import numpy as np
from sklearn.metrics import mean_squared_error
[docs]def evaluate(reference, prediction, metric="cosine"):
"""Function to evaluate the vector field related reference quantities vs. that from vector field prediction.
Parameters
----------
reference: `numpy.ndarray`
The reference quantity of the vector field (for example, simulated velocity vectors at each point or trajectory,
or estimated RNA velocity vector)
prediction: `numpy.ndarray`
The predicted quantity of the vector field (for example, velocity vectors calculated based on reconstructed vector
field function at each point or trajectory, or reconstructed RNA velocity vector)
metric: `str`
The metric for benchmarking the vector field quantities after reconstruction.
Returns
-------
res: `float`
The score between the reference vs. reconstructed quantities based on the metric.
"""
if metric == "cosine":
true_normalized = reference / np.linalg.norm(reference, axis=1).reshape(-1, 1)
predict_normalized = prediction / np.linalg.norm(prediction, axis=1).reshape(
-1, 1
)
res = np.mean(true_normalized * predict_normalized) * prediction.shape[1]
elif metric == "rmse":
res = mean_squared_error(y_true=reference, y_pred=prediction)
return res