seli 0.1.0

A Python package called seli

Seli

Minimizing the time from idea to implementation with flexible neural networks in seli.

Features

• Mutable modules for quick and dirty modifications via Module
• Serialization of modules via @saveable, save, and load
• Systematically modifying modules by traversing nested modules
• Safely handling shared/cyclical references and static arguments through seli.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 layers by subclassing seli.Module. All modules are PyTrees.

import seli

# add a name to make the module saveable
class Linear(sl.Module, name="example:Linear");
    def __init__(self, dim: int)
        self.dim = dim

        # parameters can be directly initialized
        # or an initialization method can be passed
        self.weight = seli.Param(init=seli.init.Normal("Xavier"))

    def __call__(self, x):
        # the weight gets initialized on the first call
        # by providing the shape, the value is stored
        return x @ self.weight((x.shape[-1], self.dim))


# set the rngs for all submodules at once
# no code for passing rngs around is needed
model = Linear(10).set_rngs(42)
y = model(jnp.ones(8))

A training step can be written as follows, it requires python 3.11 or newer.

optimizer = seli.opt.Adam(1e-3)
loss = seli.opt.MeanSquaredError()

x = jnp.ones(32, 8)
y = jnp.ones(32, 10)

optimizer, model, loss_value = optimizer.minimize(loss, model, y, x)

Models can be serialized and loaded. This process is safe and does not use pickling.

seli.save(model, "model.npz")

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

Installation

You can install from PiPy using pip

pip install seli

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.