TensorFlow Probability Nightly

Common errors

• ModuleNotFoundError: No module named 'tensorflow_probability.python.internal.backend.keras'

  cause TensorFlow 2.16+ uses Keras 3, but TFP expects Keras 2. The Keras 2 backend module cannot be found at the expected path.
  fix Install `tf-keras` (or `tf-keras-nightly`) using `pip install tf-keras`. You may also need to explicitly configure TFP to use this Keras version if issues persist.

• ValueError: Python 3.8 is no longer supported by TensorFlow Probability.

  cause Attempting to install or run `tfp-nightly` on an unsupported Python 3.8 environment.
  fix Upgrade your Python environment to 3.9, 3.10, or 3.11.

• AttributeError: module 'tensorflow.python.framework.ops' has no attribute 'uid'

  cause This error typically indicates an incompatibility between the installed `tfp-nightly` and `tensorflow-nightly` versions, where one expects an API that the other no longer provides or has changed.
  fix Ensure both `tensorflow-nightly` and `tfp-nightly` are updated to their very latest versions: `pip install --upgrade tensorflow-nightly tfp-nightly`.

• TypeError: 'int' object is not iterable

  cause This can occur due to subtle API changes in nightly builds, especially concerning how shapes, dtypes, or specific arguments are handled in distributions or bijectors. It might also relate to NumPy deprecation warnings and changes in array casting.
  fix Check the latest `tfp-nightly` release notes or GitHub issues. Ensure all input tensors have explicit shapes and correct dtypes. For instance, ensure single-element tensors are not implicitly cast to Python integers when iterating is expected.

Warnings

• breaking TensorFlow Probability currently relies on Keras 2 for its Keras backend integration. With TensorFlow 2.16+ (including `tensorflow-nightly`), `tf.keras` now points to Keras 3. This leads to incompatibility issues.
  Fix: You must explicitly install `tf-keras` (or `tf-keras-nightly`) and ensure your `tfp-nightly` setup utilizes it instead of `tf.keras`. Some examples might involve `import tf_keras as keras` and configuring TFP to use this module.
• breaking Support for Python 3.8 has been officially removed from TensorFlow Probability.
  Fix: Upgrade your Python environment to version 3.9 or higher (e.g., 3.9, 3.10, 3.11).
• gotcha `tfp-nightly` is built and tested against `tensorflow-nightly` and specific JAX versions. Using `tfp-nightly` with stable TensorFlow releases or significantly older/newer `jax` versions can lead to runtime errors or unexpected behavior due to API mismatches.
  Fix: Always install `tensorflow-nightly` alongside `tfp-nightly` to maintain compatibility, e.g., `pip install --upgrade tensorflow-nightly tfp-nightly`. Consult TFP release notes for recommended JAX versions.
• gotcha As a nightly build, `tfp-nightly` may introduce API changes, experimental features, or deprecations without strict backward compatibility guarantees. Code written for stable `tensorflow-probability` might break with `tfp-nightly`.
  Fix: Be prepared for frequent updates and potential API churn. Review GitHub pull requests, issues, and release notes for changes. Use `tfp-nightly` primarily for testing new features or contributing, and `tensorflow-probability` for production workloads.

Install

• pip install tfp-nightly Install latest nightly build

Imports

• tfp

  import tensorflow_probability as tfp

• tfd

  import tensorflow_probability.distributions as tfd

• tfb

  import tensorflow_probability.bijectors as tfb

• tf

  import tensorflow as tf

Quickstart

This quickstart demonstrates how to create a basic Normal distribution, sample from it, calculate log probabilities, and apply a bijector to transform it into a LogNormal distribution. It showcases the core `distributions` and `bijectors` modules of TensorFlow Probability.

import tensorflow as tf
import tensorflow_probability as tfp
tfd = tfp.distributions
tfb = tfp.bijectors

# Create a Normal distribution
mean = tf.constant(0.0, dtype=tf.float32)
stddev = tf.constant(1.0, dtype=tf.float32)
normal_dist = tfd.Normal(loc=mean, scale=stddev)

# Sample from the distribution
samples = normal_dist.sample(10)
print("Samples from Normal distribution:", samples)

# Calculate log probability
log_prob = normal_dist.log_prob(0.5)
print("Log probability of 0.5:", log_prob)

# Demonstrate a bijector (e.g., Exp) to create a LogNormal distribution
exp_bijector = tfb.Exp()
log_normal_dist = exp_bijector(normal_dist)

# Sample from the transformed distribution
log_normal_samples = log_normal_dist.sample(5)
print("Samples from LogNormal distribution (via Exp bijector):", log_normal_samples)