v2.7.3

• enabled multi-layer feature extraction for both PyTorch and TensorFlow
• more unittests
• bug fixes
• deprecated Python versions 3.8 and 3.9 (Python 3.10+ needed)

Bug fixes with TF

• fixed TF preprocessing issues
• now, Harmonized models work as expected

Bug fixes

• Bug fixes regarding the SSL models and torch hub

New models: SAM and ALIGN

• bug fixes

New models: SAM and ALIGN

• Enabled feature extraction for new models: SAM and Kakaobrain's ALIGN implementation
• Fixed small bugs in rbf kernel of CKA class

PyTorch 2.0 support and CKA speedups + improvements

• added support for torch>=2.0.0 and torchvision=0.15.* (fixes issue #165)
• implemented CKA in both NumPy and PyTorch (PyTorch runs 2x faster than NumPy)
• added CUDA support for CKA in PyTorch (100x faster than CKA in NumPy)
• added new function called get_cka where the user can set the backend (torch or numpy); function returns the CKA object
• added an unbiased CKA variant (from here) in addition to vanilla CKA

Efficient batch-wise feature extraction + MAEs

• improved the efficiency of batch-wise feature extraction (with custom data pipeline)
• added pretrained MAEs to the thingsvision model collection

Efficient batch-wise feature extraction

• efficient batch-wise feature extraction (using a with statement for PyTorch models) for custom data pipeline

Batch-wise feature extraction and human alignment

• added a public batch-wise extraction method to the feature extractor (for custom dataset and data loader classes)
• updated README and docs with explanations about how and when to use the mini-batch extraction method
• small bug fixes and refactorings

v2.5.0

• added method to extractor class for aligning representations using gLocal
• small bug fixes and refactoring