fovi.utils.knnprobe

fovi.utils.knnprobe.knn_probe(trainer, k=20, temperature=0.07)

Perform k-nearest neighbors classification using pretrained backbone features.

Extracts features from both training and validation sets using the trainer’s model, then performs weighted KNN classification to evaluate representation quality.

Parameters:

• trainer – Trainer object with model, val_loader, and cfg attributes.

• k (int, optional) – Number of nearest neighbors to use for classification. Defaults to 20.

• temperature (float, optional) – Temperature for softmax weighting of neighbor distances. Lower values make voting sharper. Defaults to 0.07.

Returns:

Dictionary containing:

• knn_top1 (float): Top-1 accuracy percentage.

• knn_top5 (float): Top-5 accuracy percentage.

• train_features (torch.Tensor): Features from training set.

• val_features (torch.Tensor): Features from validation set.

• train_labels (torch.Tensor): Labels from training set.

• val_labels (torch.Tensor): Labels from validation set.

Return type:

dict

fovi.utils.knnprobe.knn_classifier(train_features, train_labels, test_features, test_labels, k, T, num_classes=1000)

Perform weighted KNN classification and compute accuracy.

Uses cosine similarity (via dot product on L2-normalized features) to find k nearest neighbors, then performs temperature-scaled weighted voting.

Parameters:

• train_features (torch.Tensor) – Training set feature vectors of shape (N, D).

• train_labels (torch.Tensor) – Training set labels of shape (N,).

• test_features (torch.Tensor) – Test set feature vectors of shape (M, D).

• test_labels (torch.Tensor) – Test set labels of shape (M,).

• k (int) – Number of nearest neighbors to use.

• T (float) – Temperature for softmax weighting. If None, uses unweighted voting.

• num_classes (int, optional) – Number of classes. Defaults to 1000.

Returns:

(top1, top5) accuracy percentages.

Return type:

tuple