TensorFlow Hub’s model store moved to Kaggle; the loader stayed

What it does The tensorflow_hub Python library fetches pre-trained TensorFlow SavedModels from the web and slots them into your program with minimal boilerplate. It handles download caching and version resolution, letting you treat published models—image embedders, text classifiers, and the like—as reusable layers in a larger graph. For years it was the official client for Google’s tfhub.dev marketplace.

The interesting bit The marketplace itself has moved: tfhub.dev now redirects to Kaggle Models, and a slate of unmigrated assets was permanently deleted in March 2024. The library soldiers on, still resolving legacy tfhub.dev URLs while Google recommends swapping them for Kaggle Models counterparts.

Key highlights

• Downloads and caches TF2 SavedModels (plus legacy TF1 Hub modules) for local reuse
• Implements the Reusable SavedModel interface so pre-trained weights compose like standard layers
• Continues to support legacy tfhub.dev URLs after the site’s migration to Kaggle Models
• Ships with tutorials for TF2 image retraining and text classification
• Licensed under Apache 2.0

Caveats

• A specific list of unmigrated model assets—including NVIDIA industrial classifiers and Silero speech-to-text models—were deleted in March 2024 and are irretrievable
• TF1 Hub format is deprecated; the docs steer users toward TF2 SavedModels and the Reusable SavedModel interface
• Google recommends replacing tfhub.dev links with Kaggle Models equivalents, though no explicit code migration is required

Verdict Worth keeping if you maintain a TensorFlow pipeline that relies on programmatically fetching SavedModels. New projects that don’t need the legacy URL layer may prefer browsing Kaggle Models directly.