v0.4.7

What's Changed

Backbone weights migrated to HuggingFace Hub

• Pretrained backbone weights are now hosted at magcil/deepaudiox-backbones on HuggingFace Hub, replacing the previous GitHub raw file hosting.
• Downloads are handled via hf_hub_download, which manages caching automatically in ~/.cache/huggingface/hub/. Files are only downloaded once per machine.
• A download message is printed only on first download — subsequent calls load silently from cache.
• Unauthenticated HuggingFace Hub warnings are suppressed; actual errors still surface. Users can set HF_TOKEN in their environment for higher rate limits.

from_checkpoint no longer re-downloads pretrained weights

• Backbone.from_checkpoint() and AudioClassifier.from_checkpoint() now reconstruct the model with pretrained=False, regardless of what was saved in the checkpoint config. Previously, if the config contained pretrained=True, the pretrained weights were downloaded from the Hub and then immediately overwritten by the checkpoint weights — a redundant and slow operation.

load_state_dict uses strict=False

• Backbone weight loading now uses strict=False to accommodate minor mismatches between pretrained checkpoint keys and the model definition.

Full Changelog: v0.4.6...v0.4.7