generax 0.0.1

Generative Models using Jax

Generax

Generax provides implementations of different kinds of generative models. The library is built on top of Equinox which removes the need to worry about keeping track of model parameters. For example, the following code snippet shows how to create a neural spline flow and sample from it.

key = random.PRNGKey(0) # JAX random key
x = ... # some data

# Data dependent initialization
model = NeuralSpline(x=x,
                     key=key,
                     n_layers=3,
                     n_res_blocks=4,
                     hidden_size=32,
                     working_size=16)

# Sample from the model
samples = model.sample(key, n_samples=1000)

# Compute the log probability of data
log_prob = model.log_prob(x)

Installation

Roadmap

Implemented

• Normalizing flows
• Continuous normalizing flows
• Diffusion models

And these models can be trained using a variety of methods including:

• Maximum likelihood
• Score matching
• Flow matching
• Variational inference

Training

Generax provides an easy interface to train these models:

trainer = Trainer(checkpoint_path='tmp/RealNVP')

model = trainer.train(model=model,              # Generax model
                      objective=max_likelihood, # Objective function
                      evaluate_model=tester,    # Testing function
                      optimizer=optimizer,      # Optax optimizer
                      num_steps=10000,          # Number of training steps
                      data_iterator=train_ds,   # Training data iterator
                      double_batch=1000,        # Train these many batches in a scan loop
                      checkpoint_every=1000,    # Checkpoint interval
                      test_every=1000,          # Test interval
                      retrain=True)             # Retrain from checkpoint

See the tutorial for an example.