jax-dataclasses 1.4.4

Dataclasses + JAX

jax_dataclasses

• Overview
• Installation
• Core interface
• Static fields
• Mutations
• Shape and data-type annotations
• Alternatives
• Misc

Overview

jax_dataclasses provides a wrapper around dataclasses.dataclass for use in JAX, which enables automatic support for:

• Pytree registration. This allows dataclasses to be used at API boundaries in JAX. (necessary for function transformations, JIT, etc)
• Serialization via flax.serialization.
• Static analysis with tools like mypy, jedi, pyright, etc. (including for constructors)
• Optional shape and data-type annotations, which are checked at runtime.

Heavily influenced by some great existing work (the obvious one being flax.struct.dataclass); see Alternatives for comparisons.

Installation

In Python >=3.7:

pip install jax_dataclasses

We can then import:

import jax_dataclasses as jdc

Core interface

jax_dataclasses is meant to provide a drop-in replacement for dataclasses.dataclass: jdc.pytree_dataclass has the same interface as dataclasses.dataclass, but also registers the target class as a pytree node.

We also provide several aliases: jdc.[field, asdict, astuples, is_dataclass, replace] are all identical to their counterparts in the standard dataclasses library.

Static fields

To mark a field as static (in this context: constant at compile-time), we can wrap its type with jdc.Static[]:

@jdc.pytree_dataclass
class A:
    a: jnp.ndarray
    b: jdc.Static[bool]

In a pytree node, static fields will be treated as part of the treedef instead of as a child of the node; all fields that are not explicitly marked static should contain arrays or child nodes.

Mutations

All dataclasses are automatically marked as frozen and thus immutable (even when no frozen= parameter is passed in). To make changes to nested structures easier, jdc.copy_and_mutate (a) makes a copy of a pytree and (b) returns a context in which any of that copy's contained dataclasses are temporarily mutable:

from jax import numpy as jnp
import jax_dataclasses as jdc

@jdc.pytree_dataclass
class Node:
  child: jnp.ndarray

obj = Node(child=jnp.zeros(3))

with jdc.copy_and_mutate(obj) as obj_updated:
  # Make mutations to the dataclass. This is primarily useful for nested
  # dataclasses.
  #
  # Also does input validation: if the treedef, leaf shapes, or dtypes of `obj`
  # and `obj_updated` don't match, an AssertionError will be raised.
  # This can be disabled with a `validate=False` argument.
  obj_updated.child = jnp.ones(3)

print(obj)
print(obj_updated)

Shape and data-type annotations

Subclassing from jdc.EnforcedAnnotationsMixin enables automatic shape and data-type validation. Arrays contained within dataclasses are validated on instantiation and a .get_batch_axes() method is exposed for grabbing any common batch axes to the shapes of contained arrays.

We can start by importing the standard Annotated type:

# Python >=3.9
from typing import Annotated

# Backport
from typing_extensions import Annotated

We can then add shape annotations:

@jdc.pytree_dataclass
class MnistStruct(jdc.EnforcedAnnotationsMixin):
    image: Annotated[
        jnp.ndarray,
        # Note that we can move the expected location of the batch axes by
        # shifting the ellipsis around.
        #
        # If the ellipsis is excluded, we assume batch axes at the start of the
        # shape.
        (..., 28, 28),
    ]
    label: Annotated[
        jnp.ndarray,
        (..., 10),
    ]

Or data-type annotations:

    image: Annotated[
        jnp.ndarray,
        jnp.float32,
    ]
    label: Annotated[
        jnp.ndarray,
        jnp.integer,
    ]

Or both (note that annotations are order-invariant):

    image: Annotated[
        jnp.ndarray,
        (..., 28, 28),
        jnp.float32,
    ]
    label: Annotated[
        jnp.ndarray,
        (..., 10),
        jnp.integer,
    ]

Then, assuming we've constrained both the shape and data-type:

# OK
struct = MnistStruct(
  image=onp.zeros((28, 28), dtype=onp.float32),
  label=onp.zeros((10,), dtype=onp.uint8),
)
print(struct.get_batch_axes()) # Prints ()

# OK
struct = MnistStruct(
  image=onp.zeros((32, 28, 28), dtype=onp.float32),
  label=onp.zeros((32, 10), dtype=onp.uint8),
)
print(struct.get_batch_axes()) # Prints (32,)

# AssertionError on instantiation because of type mismatch
MnistStruct(
  image=onp.zeros((28, 28), dtype=onp.float32),
  label=onp.zeros((10,), dtype=onp.float32), # Not an integer type!
)

# AssertionError on instantiation because of shape mismatch
MnistStruct(
  image=onp.zeros((28, 28), dtype=onp.float32),
  label=onp.zeros((5,), dtype=onp.uint8),
)

# AssertionError on instantiation because of batch axis mismatch
struct = MnistStruct(
  image=onp.zeros((64, 28, 28), dtype=onp.float32),
  label=onp.zeros((32, 10), dtype=onp.uint8),
)

Alternatives

A few other solutions exist for automatically integrating dataclass-style objects into pytree structures. Great ones include: chex.dataclass, flax.struct, and tjax.dataclass. These all influenced this library.

The main differentiators of jax_dataclasses are:

• Static analysis support. tjax has a custom mypy plugin to enable type checking, but isn't supported by other tools. flax.struct implements the dataclass_transform spec proposed by pyright, but isn't supported by other tools. Because @jdc.pytree_dataclass has the same API as @dataclasses.dataclass, it can include pytree registration behavior at runtime while being treated as the standard decorator during static analysis. This means that all static checkers, language servers, and autocomplete engines that support the standard dataclasses library should work out of the box with jax_dataclasses.

• Nested dataclasses. Making replacements/modifications in deeply nested dataclasses can be really frustrating. The three alternatives all introduce a .replace(self, ...) method to dataclasses that's a bit more convenient than the traditional dataclasses.replace(obj, ...) API for shallow changes, but still becomes really cumbersome to use when dataclasses are nested. jdc.copy_and_mutate() is introduced to address this.

• Static field support. Parameters that should not be traced in JAX should be marked as static. This is supported in flax, tjax, and jax_dataclasses, but not chex.

• Serialization. When working with flax, being able to serialize dataclasses is really handy. This is supported in flax.struct (naturally) and jax_dataclasses, but not chex or tjax.

• Shape and type annotations. See above.

You can also eschew the dataclass-style interface entirely; see how brax registers pytrees. This is a reasonable thing to prefer: it requires some floating strings and breaks things that I care about but you may not (like immutability and __post_init__), but gives more flexibility with custom __init__ methods.

Misc

This code was originally written for and factored out of jaxfg, where Nick Heppert provided valuable feedback.