The purpose of this repo is to collect activations when auditing CLIP models.

We focus on TinyCLIP although the code should be model agnostic.

Order of Operations

1. get_neuron_indices.py: Get random neuron indices from that model. DONE for.
2. test_model_accuracy.py: Test the accuracy of the model on your dataset. DONE for ImageNet on CLIP-B.
3. cache_activations.py: Cache activations and image indices for those neurons
4. plot_all_neurons_histogram.py: Define activation intervals for each layer, both raw values and SD values. Also get intervals. 4a. get_intervals.py: Get the percentiles per activation.
5. sample_images_from_interval.py: Get corresponding heatmaps for those n image indices per interval and save!