🥷 Ninjax: Flexible Modules for JAX

Ninjax is a general module system for JAX. It gives the user complete and transparent control over updating the state of each module, bringing the flexibility of PyTorch and TensorFlow to JAX. Moreover, Ninjax makes it easy to mix and match modules from different libraries, such as Flax and Haiku.

Motivation

Existing deep learning libraries for JAX provide modules, but those modules only specify neural networks and cannot easily implement training logic. Orchestrating training all in one place, outside of the modules, is fine for simple code bases. But it becomes a problem when there are many 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. 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.

The main differences to existing deep learning libraries are:

• Ninjax does not need separate apply()/init() functions. Instead, the first function call creates variables automatically.
• Ninjax lets you access and update model parameters inside of impure functions, so modules can handle their own optimizers and update logic.
• Natural support for modules with multiple functions without need for Flax's setup() function or Haiku's hk.multi_transform().
• Ninjax' flexible state handling makes it trivial to mix and match modules from other deep learning libraries in your models.

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 flax
import jax
import jax.numpy as jnp
import ninjax as nj
import optax

# Supports all Flax and Haiku modules
Linear = nj.FromFlax(flax.linen.Dense)


class MyModel(nj.Module):

  lr: float = 0.01

  def __init__(self, size):
    self.size = size
    self.opt = optax.adam(self.lr)
    # Define submodules upfront
    self.h1 = Linear(128, name='h1')
    self.h2 = Linear(128, name='h2')

  def predict(self, x):
    x = jax.nn.relu(self.h1(x))
    x = jax.nn.relu(self.h2(x))
    # Define submodules inline
    x = self.get('h3', Linear, self.size, use_bias=False)(x)
    # Create state entries inline
    x += self.get('bias', jnp.zeros, self.size)
    return x

  def train(self, x, y):
    # Gradient with respect to submodules or state entries
    keys = [self.h1, self.h2, f'{self.path}/h3', f'{self.path}/bias']
    loss, params, grads = nj.grad(self.loss, keys)(x, y)
    # Update weights
    optstate = self.get('optstate', self.opt.init, params)
    updates, optstate = self.opt.update(grads, optstate)
    new_params = optax.apply_updates(params, updates)
    self.put(new_params)  # Store the new params
    return loss

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


# Create model and example data
model = MyModel(3, lr=0.01, name='model')
x = jnp.ones((64, 32), jnp.float32)
y = jnp.ones((64, 3), jnp.float32)

# Populate initial state from one or more functions
state = {}
state = nj.init(model.train)(state, x, y, seed=0)
print(state['model/bias'])  # [0., 0., 0.]

# Purify and jit one or more functions
train = nj.pure(model.train)
train = jax.jit(train)

# Training loop
for x, y in [(x, y)] * 10:
  state, loss = train(state, x, y)
  print('Loss:', float(loss))

# Look at the parameters
print(state['model/bias'])  # [-1.2e-09  1.8e-08 -2.5e-09]

Tutorial

How can I create state entries?

Ninjax gives modules full control over reading and updating their state entries. Use self.get(name, ctor, *args, **kwargs) to define state entries. The first call creates the entry as ctor(*args, **kwargs). Later calls return the current value:

class Module(nj.Module):

  def compute(self, x):
    init = jax.nn.initializers.variance_scaling(1, 'fan_avg', 'uniform')
    weights = self.get('weights', init, nj.rng(), (64, 32))
    bias = self.get('bias', jnp.zeros, (32,), jnp.float32)
    print(self.getm())  # {'/path/to/module/weights': ..., '/path/to/module/bias': ...}
    return x @ weights + bias

How can I update state entries?

To update the state entries of a module, use self.put(name, value) for individual entries of self.putm(mapping) to update multiple values:

class Module(nj.Module):

  def counting(self):
    counter = nj.get('counter', jnp.zeros, (), jnp.int32)
    self.put('counter', counter + 1)
    print(self.get('counter'))  # 1
    state = self.getm()
    state['counter'] += 1
    self.putm(state)
    print(self.getm()['counter'])  # 2
    print(self.get('counter'))  # 2

How can I use JIT compilation?

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

model = Model()
train = jax.jit(nj.pure(model.train))
params = {}
result, params = train(param, rng, ...)

Calling the pure function will create any parameters that are not yet in the params dictionary and return the new parameters alongside the function output.

You can speed up the first function call (where parameters are created) by using nj.jit instead of jax.jit. Internally, this avoids compiling two versions of the function.

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

The self.getm(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.getm('.*'), self.loss, x, y)
    return metrics  # {'loss': ..., 'grad_norm': ...}

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

Questions

If you have a question, please file an issue.