TensorFlow Hub

Warnings

• breaking Versions 0.16.0 and 0.16.1 enforce the use of Keras 2 by requiring `tf-keras` to be installed. Older versions might have used a bundled Keras or a different Keras installation strategy.
  Fix: Ensure `tf-keras>=2.14.1` is installed in your environment: `pip install tf-keras`.
• breaking Deprecated TensorFlow 1 (TF1) APIs, such as `LatestModuleExporter` and `register_module_for_export`, were removed in v0.16.0. Code relying on these TF1 APIs will break.
  Fix: Migrate your code to use TensorFlow 2.x APIs, primarily `hub.KerasLayer` and SavedModels for model export/import.
• breaking Starting with version 0.15.0, TensorFlow Hub requires Python 3.9 or higher. Previous versions supported older Python versions (e.g., 3.7+ from v0.13.0).
  Fix: Upgrade your Python environment to Python 3.9 or a later compatible version.
• breaking The `make_image_classifier` and `make_nearest_neighbour_index` tools were removed in version 0.14.0.
  Fix: Refactor your code to implement these functionalities using `hub.KerasLayer` and standard TensorFlow/Keras operations, or look for alternative tools/models on tfhub.dev (Kaggle Models).
• gotcha As of November 15, 2023, `tfhub.dev` model links redirect to Kaggle Models. Crucially, unmigrated models (a specific list, including some iNaturalist and NVIDIA models) were permanently deleted on March 18, 2024, and are no longer retrievable via their old `tfhub.dev` handles.
  Fix: Use Kaggle Models counterparts for `tfhub.dev` links. If you were using one of the unmigrated models, you must find an alternative or retrain your own model. Check official TensorFlow Hub documentation for the list of deleted models and migration FAQs.
• gotcha The `hub.Module` class, part of the TensorFlow 1 API, is deprecated and has known compatibility issues with newer TensorFlow 2.x versions, especially concerning `tf.compat.v1.ragged.placeholder()` with TF2.3.
  Fix: Always prefer `hub.KerasLayer` for loading models in TensorFlow 2.x environments. Avoid `hub.Module` for new development and migrate existing code where possible.
• gotcha The TensorFlow Hub API is under active development and is not yet guaranteed to have a stable API or model format, meaning future updates might introduce breaking changes even outside of major version bumps.
  Fix: Pin `tensorflow-hub` to a specific version in your `requirements.txt` or `pyproject.toml` to ensure consistent behavior across deployments. Regularly check release notes for upcoming changes.

Install

• pip install tensorflow-hub PyPI
• conda install -c conda-forge tensorflow-hub Conda

Imports

• tensorflow_hub

  import tensorflow_hub as hub

  The standard and recommended import alias for brevity and convention.

Quickstart

This quickstart demonstrates how to load a pre-trained text embedding model from TensorFlow Hub using `hub.KerasLayer` and generate embeddings for sample text. It also shows how to integrate this `hub.KerasLayer` into a larger Keras Sequential model for transfer learning tasks like text classification.

import tensorflow as tf
import tensorflow_hub as hub

# Load a pre-trained text embedding model from TensorFlow Hub
# This model generates 128-dimensional embeddings for text inputs.
model_url = "https://tfhub.dev/google/nnlm-en-dim128/2"
embed = hub.KerasLayer(model_url, input_shape=[], dtype=tf.string, trainable=False)

# Example usage: Generate embeddings for a list of sentences
sentences = [
    "Hello TensorFlow Hub!",
    "This is a test sentence for text embedding.",
    "Machine learning is fun."
]

embeddings = embed(tf.constant(sentences))

print(f"Input sentences: {sentences}")
print(f"Generated embeddings shape: {embeddings.shape}")
print(f"First embedding (first 5 values): {embeddings[0, :5].numpy()}")

# You can then use these embeddings as input to another Keras layer
# for tasks like classification.
classifier_model = tf.keras.Sequential([
    embed,
    tf.keras.layers.Dense(16, activation='relu'),
    tf.keras.layers.Dense(1, activation='sigmoid')
])
classifier_model.summary()