🔬 unsupervised_evaluator.py

Runs feature-based evaluations for self-supervised checkpoints. It can execute multiple modes in one pass:

• eval_knn – k‑NN classification on extracted features.
• eval_linear – logistic regression (linear probing).
• eval_umap – dimensionality reduction and clustering metrics with UMAP.

Call it directly to analyse a saved experiment:

python evaluators/unsupervised_evaluator.py eval.mode='eval_knn,eval_linear,eval_umap' eval.experiment_path=path/to/exp

The script loads the model, extracts features once, then applies the selected evaluations. Results are saved to the experiment folder.

Automatic integration

During DINO training, if eval.interval is set in the config, the trainer will automatically call run_evaluation every N epochs. Each evaluation run is stored under the training output directory with the epoch number.