Multiple Inheritance -> doesn't recognize as Module throws ValueError: parent must be None, Module or Scope

[Avital Oliver]

Discussed in #1390

^{Originally posted by SauravMaheshkar June 26, 2021}
I'm working on a Flax implementation for ProteinBERT: A universal deep-learning model of protein sequence and function. My work so far is in SauravMaheshkar/ProteinBERT.

I've made a simple test.py to check instantiation using the .init() function. My test script is as follows :

from proteinbert import ProteinBERT
import jax
from jax import random


def test():

    seq = jax.random.randint(
        key=random.PRNGKey(0), minval=0, maxval=21, shape=(2, 2048)
    )
    annotation = jax.random.randint(
        key=random.PRNGKey(0), minval=0, maxval=1, shape=(2, 8943)
    )

    init_rngs = {"params": random.PRNGKey(0), "layers": random.PRNGKey(1)}

    ProteinBERT().init(init_rngs, seq, annotation)


if __name__ == "__main__":
    test()

And I've been getting this error message

Error Message

Traceback (most recent call last):
  File "/Users/sauravmaheshkar/github/protein_bert/test.py", line 21, in <module>
    test()
  File "/Users/sauravmaheshkar/github/protein_bert/test.py", line 17, in test
    ProteinBERT().init(init_rngs, seq, annotation)
  File "/Users/sauravmaheshkar/opt/anaconda3/envs/proteinbert/lib/python3.7/site-packages/flax/linen/module.py", line 1000, in init
    method=method, mutable=mutable, **kwargs)
  File "/Users/sauravmaheshkar/opt/anaconda3/envs/proteinbert/lib/python3.7/site-packages/flax/linen/module.py", line 969, in init_with_output
    {}, *args, rngs=rngs, method=method, mutable=mutable, **kwargs)
  File "/Users/sauravmaheshkar/opt/anaconda3/envs/proteinbert/lib/python3.7/site-packages/flax/linen/module.py", line 939, in apply
    )(variables, *args, **kwargs, rngs=rngs)
  File "/Users/sauravmaheshkar/opt/anaconda3/envs/proteinbert/lib/python3.7/site-packages/flax/core/scope.py", line 687, in wrapper
    y = fn(root, *args, **kwargs)
  File "/Users/sauravmaheshkar/opt/anaconda3/envs/proteinbert/lib/python3.7/site-packages/flax/linen/module.py", line 1178, in scope_fn
    return fn(module.clone(parent=scope), *args, **kwargs)
  File "/Users/sauravmaheshkar/opt/anaconda3/envs/proteinbert/lib/python3.7/site-packages/flax/linen/module.py", line 266, in wrapped_module_method
    self._try_setup()
  File "/Users/sauravmaheshkar/opt/anaconda3/envs/proteinbert/lib/python3.7/site-packages/flax/linen/module.py", line 679, in _try_setup
    self.setup()
  File "/Users/sauravmaheshkar/opt/anaconda3/envs/proteinbert/lib/python3.7/site-packages/flax/linen/module.py", line 275, in wrapped_module_method
    y = fun(self, *args, **kwargs)
  File "/Users/sauravmaheshkar/github/protein_bert/proteinbert/model.py", line 82, in setup
    Reduce("b n d -> b d", "mean"),
  File "<string>", line 5, in __init__
  File "/Users/sauravmaheshkar/opt/anaconda3/envs/proteinbert/lib/python3.7/site-packages/flax/linen/module.py", line 599, in __post_init__
    raise ValueError("parent must be None, Module or Scope")
ValueError: parent must be None, Module or Scope

The problem lies in the Reduce defined in proteinbert/utils.py class which is defined as follows:

class Reduce(ReduceMixin, nn.Module):
    """
    Flax Module to act as a Reduce layer (from einops)
    """
    def __call__(self, input):
        return self._apply_recipe(input)

The idea is to create a Reduce layer/Module for flax which performs the reduce operation from einops. Although the module inherits from flax.linen.Module it still throws a ValueError.

Any help would be much appreciated 😊.

[Avital Oliver]

@marcvanzee will be investigating this.

A few guiding questions:

1. Why would einops be implemented as a Module instead of just a function?
2. Why is multiple inheritance needed here?
3. Regardless, this error shouldn't happen. So even if we answer (1) and (2) in a way that means there's a workaround, we should still fix this bug.

[Anselm Levskaya]

Just noticing this issue for the first time... I've seen similarly weird issues with Mixins and Flax resolved in the past by simply changing the order of the multiple inheritance - e.g. class Reduce(nn.Module, ReduceMixin): to put nn.Module first. I'm not 100% sure this is the same kind of issue that I've seen before w. mixins, but I'd certainly be curious if that would have fixed the issue...

[Cristian Garcia]

I was playing around with mixins to see how they interact with Module.

See experiments

Following case works:

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

class Mixin:
    def __call__(self, x):
        return self.dense(x)

class MyModule(nn.Module, Mixin):
    def setup(self):
        self.dense = nn.Dense(2)
    
module_a = MyModule()
variables = module_a.init(jax.random.PRNGKey(0), jnp.ones((1, 1)))

However, passing setup to Mixin fails:

class Mixin:
    def setup(self):
        self.dense = nn.Dense(2)
    
    def __call__(self, x):
        return self.dense(x)

class MyModule(nn.Module, Mixin):
    pass

# AttributeError: "MyModule" object has no attribute "dense"

This is again fixed if Mixin is set as the first parent:

class Mixin:
    def setup(self):
        self.dense = nn.Dense(2)
    
    def __call__(self, x):
        return self.dense(x)

class MyModule(Mixin, nn.Module):
    pass

Also, you cannot define compact methods on mixins (this is probably expected?):

class Mixin:
    @nn.compact
    def __call__(self, x):
        return nn.Dense(2)(x)

class MyModule(nn.Module, Mixin): # swapping doesn't help
    pass

Discussion

Based on these experiments the only insight I see is: don't define scope-dependent operations (compact, self.param/variable) in inside mixins as their methods will not be wrapped appropriately. Not sure if there is a way to properly wrap mixin methods in __init_subclass__, either _get_local_method_names is not detecting them or they are not available when __init_subclass__ is called.