How to type hint `flax.linen.Module.apply`'s output correctly?

As of writing, this code does not pass the PyRight type checker:

import jax
import jax.numpy as jnp
import jax.typing as jt
import flax.linen as nn

class MLP(nn.Module):
    @nn.compact
    def __call__(self, inputs: jt.ArrayLike):
        inputs = jnp.array(inputs)
        return nn.Dense(4)(inputs)

if __name__ == "__main__":
    inputs = jnp.ones((2, 4))
    mlp = MLP()
    prng_key = jax.random.key(7)
    outputs: jax.Array = mlp.apply(mlp.init(prng_key, inputs), inputs)
    print(f"type of outputs is {type(outputs)}")

Its output is: type of outputs is <class 'jaxlib.xla_extension.ArrayImpl'>.

On like outputs: jax.Array = mlp.apply(mlp.init(prng_key, inputs), inputs), Pylance shows this error:

Type "Any | tuple[Any, FrozenVariableDict | dict[str, Any]]" is not assignable to declared type "Array"
  Type "Any | tuple[Any, FrozenVariableDict | dict[str, Any]]" is not assignable to type "Array"
    "tuple[Any, FrozenVariableDict | dict[str, Any]]" is not assignable to "Array"
PylancereportAssignmentType

Any other operations that treat outputs like a JAX array succeeds at runtime but also fails to type check.

Yet the output of that function is an array. This code type checks correctly:

import jax
import jax.numpy as jnp

x: jax.Array = jnp.zeros((2, 2))
print(f"x's type is {type(x)}")

and its output is: x's type is <class 'jaxlib.xla_extension.ArrayImpl'>, the same type.

How should I use this module to get the output array in a way that pyright can type check correctly?

Answer

The error you're seeing from PyRight is due to the fact that the apply method of a Flax module can return a tuple, which includes the output and some additional state information. This can lead to type mismatches when you try to assign it directly to a variable of type jax.Array

Try following changes:

import jax
import jax.numpy as jnp
import jax.typing as jt
import flax.linen as nn

class MLP(nn.Module):
    @nn.compact
    def __call__(self, inputs: jt.ArrayLike):
        inputs = jnp.array(inputs)
        return nn.Dense(4)(inputs)

if __name__ == "__main__":
    inputs = jnp.ones((2, 4))
    mlp = MLP()
    prng_key = jax.random.key(7)
    
    # unpack the output
    outputs, _ = mlp.apply(mlp.init(prng_key, inputs), inputs)
    
    # type hint outputs correctly
    outputs: jax.Array = outputs  # pass the type checker
    print(f"type of outputs is {type(outputs)}")