flarejax 0.4.10

Flexible Neural Networks in JAX

FlareJax

FlareJax is a Python library for building neural networks and optimizers in Jax. It is designed to minimize the time between a new research idea and its implementation.

Features

• Mutable modules for quick and dirty modifications via Module
• Serialization of modules via @saveable, save, and load
• Systematically modifying modules by using flatten and unflatten
• Safely handling shared/cyclical references and static arguments through filter_jit
• Commonly used NN layers and optimizers are included
• As a small codebase, it is relatively easy to understand and extend

Quick Example

Define new Modules by subclassing Module. All methods are callable PyTrees.

@flr.saveable("Example:Linear")  # optional, make saveable
class Linear(flr.Module):
    def __init__(self, key: PRNGKeyArray, dim: int):
        self.dim = dim
        self.w = None

    def __call__(self, x):
        # lazy initialization dependent on the input shape
        if self.w is None:
            self.w = jrn.normal(key, (x.shape[-1], self.dim))

        return x @ self.w

layer = Linear(jrn.PRNGKey(0), 3)
x = jnp.zeros((1, 4))

# the model is initialized after the first call
y = layer(x)
assert layer.w.shape == (4, 3)

For optimization, define a loss function, which takes the module as the first argument.

def loss_fn(module, x, y):
    return jnp.mean((module(x) - y) ** 2)

opt = flr.opt.Adam(3e-4)

# automatically just-in-time compiled
opt, model, loss = flr.train(opt, model, loss_fn, x, y)

Models can be saved and loaded.

flr.save(layer, "model.npz")

# load the model
layer = flr.load("model.npz")
assert isinstance(layer, Linear)

Installation

FlareJax can be installed via pip. It requires Python 3.10 or higher and Jax 0.4.33 or higher.

pip install flarejax

See Also

• Jax Docs: Jax is a library for numerical computing that is designed to be composable and fast.
• Equinox library: FlareJax is heavily inspired by this awesome library.
• torch.nn.Module: Many of the principles of mutability are inspired by PyTorch's torch.nn.Module.
• NNX Docs: NNX is a library for neural networks in flax that also supports mutability.
• Always feel free to reach out to me via email.