🥷 Ninjax

Ninjax is a lightweight module system for JAX that gives modules full control over their state (e.g. to have their own train() functions). It also makes it easy to use modules from different libraries together, such as Flax and Haiku.

Motivation

Existing neural network libraries for JAX provide modules, but their modules only specify neural graphs and cannot easily implement their own training logic. Orchestrating training logic all in one place, outside of the modules, is fine for simple code bases. But it becomes a problem when there are mnay modules with their own training logic and optimizers.

Ninjax solves this problem by giving each nj.Module full read and write access to its state, while remaining functional via nj.run(). This means modules can have train() functions to implement custom training logic, and call each other's train functions. Ninjax is intended to be used with one or more neural network libraries, such as Haiku and Flax.

Installation

Ninjax is a single file, so you can just copy it to your project directory. Or you can install the package:

pip install ninjax

Quickstart

import haiku as hk
import flax.linen as nn
import jax
import jax.numpy as jnp
import ninjax as nj


class Model(nj.Module):

  def __init__(self, size, act=jax.nn.relu):
    self.size = size
    self.act = act
    self.h1 = nj.HaikuModule(hk.Linear, 128)
    self.h2 = nj.HaikuModule(hk.Linear, 128)
    self.h3 = nj.FlaxModule(nn.Linear, size)

  def __call__(self, x):
    x = self.act(self.h1(x))
    x = self.act(self.h2(x))
    x = self.h3(x)
    return x

  def train(self, x, y):
    self(x)  # Ensure parameters are created.
    state = self.get_state()
    loss, grad = nj.grad(self.loss, state)(x, y)
    state = jax.tree_map(lambda p, g: p - 0.01 * g, state, grad)
    self.set_state(state)
    return loss

  def loss(self, x, y):
    return ((self(x) - y) ** 2).mean()


model = Model(8)
main = jax.random.PRNGKey(0)
state = {}
for x, y in dataset:
  rng, main = jax.random.split(main)
  state, loss = nj.run(model.train, state, rng, x, y)
  print('Loss:', float(loss))

API

Ninjax uses a simple API that provides flexible access to state. The nj.run() lets you call your modules from the outside. All other functions are to be used in your modules or other code that is within the nj.run() call.

# Run a function or method that uses Ninjax state.
state, out = nj.run(state, rng, fn, *args, **kwargs)

# Inherit your modules from this class for automatic name scoping and helper
# functions for accesing the state that belongs to this module.
class nj.Module:
  @path                                 # Unique scope string for this module.
  def get(name, ctor, *args, **kwargs)  # Get or create state entry.
  def set(name, value)                  # Update state entry.
  def get_state(filter='.*')            # Get multiple state entries.
  def set_state(entries)                # Update multiple state entries.

# Return the mutable global state dictionary.
state = nj.state()

# Get a unique random number generator key.
rng = nj.rng()

# Compute the gradient with respect to global state entries, specified by key.
grad = nj.grad(fn, keys)(*args, **kwargs)

# Ninjax provides convenience wrappers for popular JAX libraries. These
# automate the manual initialization and state passing that these libraries
# require for you:

mlp = nj.HaikuModule(hk.nets.MLP, [128, 128, 32])
outputs = mlp(inputs)

opt = nj.OptaxModule(optax.adam, 1e-3)
opt(mlp.get_state(), loss, data)  # Train the MLP with a loss function.

Tutorial

How can I use JIT compilation?

The nj.run() function makes the state your JAX code uses explicit, so it can be jitted and transformed freely:

model = Model()
train = jax.jit(functools.partial(nj.run, model.train))
train(state, rng, ...)

How can I compute gradients?

You can use jax.grad as normal for computing gradients with respect to explicit inputs of your function. To compute gradients with respect to Ninjax state, use nj.grad(fn, keys):

class Module(nj.Module):

  def train(self, x, y):
    params = self.get_state('.*')
    loss, grads = nj.grad(self.loss, params.keys())(x, y)
    params = jax.tree_map(lambda p, g: p - 0.01 * g, params, grads)
    self.set_state(params)

The self.get_state(filter='.*') method optionally accepts a regex pattern to select only a subset of the state dictionary. It also returns only state entries of the current module. To access the global state, use nj.state().

How can I define modules compactly?

You can use self.get(name, ctor, *args, **kwargs) inside methods of your modules. When called for the first time, it creates a new state entry from the constructor ctor(*args, **kwargs). Later calls return the existing entry:

class Module(nj.Module):

  def __call__(self, x):
    x = jax.nn.relu(self.get('h1', Linear, 128)(x))
    x = jax.nn.relu(self.get('h2', Linear, 128)(x))
    x = self.get('h3', Linear, 32)(x)
    return x

How can I use Haiku modules?

There is nothing special about using external libraries with Ninjax. Haiku requires its modules to be passed through hk.transform and the initialized via transformed.init(rng, batch). For convenience, Ninjax provides nj.HaikuModule to do this for you:

class Module(nj.Module):

  def __init__(self):
    self.mlp = nj.HaikuModule(hk.nets.MLP, [128, 128, 32])

  def __call__(self, x):
    return self.mlp(x)

You can also predefine a list of aliases for Haiku modules that you want to use frequently:

Linear = functools.partial(nj.HaikuModule, hk.Linear)
Conv2D = functools.partial(nj.HaikuModule, hk.Conv2D)
MLP = functools.partial(nj.HaikuModule, hk.nets.MLP)
# ...

How can I use Flax modules?

There is nothing special about using external libraries with Ninjax. Flax requires its modules to be initialized via params = model.init(rng, batch) and used via model.apply(params, data). For convenience, Ninjax provides nj.FlaxModule to do this for you:

class Module(nj.Module):

  def __init__(self):
    self.linear = nj.FlaxModule(nn.Dense, 128)

  def __call__(self, x):
    return self.linear(x)

You can also predefine a list of aliases for Flax modules that you want to use frequently:

Dense = functools.partial(nj.FlaxModule, nn.Dense)
Conv = functools.partial(nj.FlaxModule, nn.Conv)
# ...

How can I use Optax optimizers?

There is nothing special about using external libraries like Optax with Ninjax. Optax requires its optimizers to be initialized, their state to be passed through the optimizer call, and the resulting updates to be applied. For convenience, Ninjax provides nj.OptaxModule to do this for you:

class Module(nj.Module):

  def __init__(self):
    self.mlp = MLP()
    self.opt = nj.OptaxModule(optax.adam, 1e-3)

  def train(self, x, y):
    self.mlp(x)  # Ensure paramters are created.
    metrics = self.opt(self.mlp.get_state('.*'), self.loss, x, y)
    return metrics  # {'loss': ..., 'grad_norm': ...}

  def loss(self, x, y):
    return ((self.mlp(x) - y) ** 2).mean()

Limitations

Ninjax is still a young library. One current limitation is that LAX symbolic control flow and computing gradients of gradients has not been tested and might not work correctly. If you are interested in this functionality or encounter any other issues, let me know.

Questions

If you have a question, please file an issue.