tfhub.dev hosts the following model formats: TF2 SavedModel, TF1 Hub format, TF.js and TFLite. This page provides an overview of each model format.
Content published to tfhub.dev can be automatically mirrored to other model hubs, provided it follows a specified format and is permitted by our Terms (https://tfhub.dev/terms). See our publishing documentation for more details, and our contribution documentation if you'd like to opt-out of mirroring.
tfhub.dev hosts TensorFlow models in the TF2 SavedModel format and TF1 Hub format. We recommend using models in the standardized TF2 SavedModel format instead of the deprecated TF1 Hub format when possible.
TF2 SavedModel is the recommended format for sharing TensorFlow models. You can learn more about the SavedModel format in the TensorFlow SavedModel guide.
You can use SavedModels from tfhub.dev without depending on the
library, since this format is a part of core TensorFlow.
Learn more about SavedModels on TF Hub:
TF1 Hub format
The TF1 Hub format is a custom serialization format used in by TF Hub library.
The TF1 Hub format is similar to the SavedModel format of TensorFlow 1 on a
syntactic level (same file names and protocol messages) but semantically
different to allow for module reuse, composition and re-training (e.g.,
different storage of resource initializers, different tagging conventions for
metagraphs). The easiest way to tell them apart on disk is the presence or
absence of the
Learn more about models in TF1 Hub format on TF Hub:
- Using TF1 Hub format models
- Exporting a model in the TF1 Hub format
- TF1/TF2 compatibility of TF1 Hub format
The TFLite format is used for on-device inference. You can learn more at the TFLite documentation.
The TF.js format is used for in-browser ML. You can learn more at the TF.js documentation.