Entrypoint loader

Class Resolver

Lookup and instantiate classes with style.

💪 Getting Started

from class_resolver import ClassResolver
from dataclasses import dataclass

class Base: pass

@dataclass
class A(Base):
   name: str

@dataclass
class B(Base):
   name: str

# Index
resolver = ClassResolver([A, B], base=Base)

# Lookup
assert A == resolver.lookup('A')

# Instantiate with a dictionary
assert A(name='hi') == resolver.make('A', {'name': 'hi'})

# Instantiate with kwargs
assert A(name='hi') == resolver.make('A', name='hi')

# A pre-instantiated class will simply be passed through
assert A(name='hi') == resolver.make(A(name='hi'))

🤖 Writing Extensible Machine Learning Models with `class-resolver`

Assume you've implemented a simple multi-layer perceptron in PyTorch:

from itertools import chain

from more_itertools import pairwise
from torch import nn

class MLP(nn.Sequential):
    def __init__(self, dims: list[int]):
        super().__init__(chain.from_iterable(
            (
                nn.Linear(in_features, out_features),
                nn.ReLU(),
            )
            for in_features, out_features in pairwise(dims)
        ))

This MLP uses a hard-coded rectified linear unit as the non-linear activation function between layers. We can generalize this MLP to use a variety of non-linear activation functions by adding an argument to its __init__() function like in:

from itertools import chain

from more_itertools import pairwise
from torch import nn

class MLP(nn.Sequential):
    def __init__(self, dims: list[int], activation: str = "relu"):
        if activation == "relu":
            activation = nn.ReLU()
        elif activation == "tanh":
            activation = nn.Tanh()
        elif activation == "hardtanh":
            activation = nn.Hardtanh()
        else:
            raise KeyError(f"Unsupported activation: {activation}")
        super().__init__(chain.from_iterable(
            (
                nn.Linear(in_features, out_features),
                activation,
            )
            for in_features, out_features in pairwise(dims)
        ))

The first issue with this implementation is it relies on a hard-coded set of conditional statements and is therefore hard to extend. It can be improved by using a dictionary lookup:

from itertools import chain

from more_itertools import pairwise
from torch import nn

activation_lookup: dict[str, nn.Module] = {
   "relu": nn.ReLU(),
   "tanh": nn.Tanh(),
   "hardtanh": nn.Hardtanh(),
}

class MLP(nn.Sequential):
    def __init__(self, dims: list[int], activation: str = "relu"):
        activation = activation_lookup[activation]
        super().__init__(chain.from_iterable(
            (
                nn.Linear(in_features, out_features),
                activation,
            )
            for in_features, out_features in pairwise(dims)
        ))

This approach is rigid because it requires pre-instantiation of the activations. If we needed to vary the arguments to the nn.HardTanh class, the previous approach wouldn't work. We can change the implementation to lookup on the class before instantiation then optionally pass some arguments:

from itertools import chain

from more_itertools import pairwise
from torch import nn

activation_lookup: dict[str, type[nn.Module]] = {
   "relu": nn.ReLU,
   "tanh": nn.Tanh,
   "hardtanh": nn.Hardtanh,
}

class MLP(nn.Sequential):
    def __init__(
        self, 
        dims: list[int], 
        activation: str = "relu", 
        activation_kwargs: None | dict[str, any] = None,
    ):
        activation_cls = activation_lookup[activation]
        activation = activation_cls(**(activation_kwargs or {}))
        super().__init__(chain.from_iterable(
            (
                nn.Linear(in_features, out_features),
                activation,
            )
            for in_features, out_features in pairwise(dims)
        ))

This is pretty good, but it still has a few issues:

you have to manually maintain the activation_lookup dictionary,
you can't pass an instance or class through the activation keyword
you have to get the casing just right
the default is hard-coded as a string, which means this has to get copied (error-prone) in any place that creates an MLP
you have to re-write this logic for all of your classes

Enter the class_resolver package, which takes care of all of these things using the following:

from itertools import chain

from class_resolver import ClassResolver, Hint
from more_itertools import pairwise
from torch import nn

activation_resolver = ClassResolver(
    [nn.ReLU, nn.Tanh, nn.Hardtanh],
    base=nn.Module,
    default=nn.ReLU,
)

class MLP(nn.Sequential):
    def __init__(
        self, 
        dims: list[int], 
        activation: Hint[nn.Module] = None,  # Hint = Union[None, str, nn.Module, type[nn.Module]]
        activation_kwargs: None | dict[str, any] = None,
    ):
        super().__init__(chain.from_iterable(
            (
                nn.Linear(in_features, out_features),
                activation_resolver.make(activation, activation_kwargs),
            )
            for in_features, out_features in pairwise(dims)
        ))

Because this is such a common pattern, we've made it available through contrib module in class_resolver.contrib.torch:

from itertools import chain

from class_resolver import Hint
from class_resolver.contrib.torch import activation_resolver
from more_itertools import pairwise
from torch import nn

class MLP(nn.Sequential):
    def __init__(
        self, 
        dims: list[int], 
        activation: Hint[nn.Module] = None,
        activation_kwargs: None | dict[str, any] = None,
    ):
        super().__init__(chain.from_iterable(
            (
                nn.Linear(in_features, out_features),
                activation_resolver.make(activation, activation_kwargs),
            )
            for in_features, out_features in pairwise(dims)
        ))

Now, you can instantiate the MLP with any of the following:

MLP(dims=[10, 200, 40])  # uses default, which is ReLU
MLP(dims=[10, 200, 40], activation="relu")  # uses lowercase
MLP(dims=[10, 200, 40], activation="ReLU")  # uses stylized
MLP(dims=[10, 200, 40], activation=nn.ReLU)  # uses class
MLP(dims=[10, 200, 40], activation=nn.ReLU())  # uses instance

MLP(dims=[10, 200, 40], activation="hardtanh", activation_kwargs={"min_val": 0.0, "max_value": 6.0})  # uses kwargs
MLP(dims=[10, 200, 40], activation=nn.HardTanh, activation_kwargs={"min_val": 0.0, "max_value": 6.0})  # uses kwargs
MLP(dims=[10, 200, 40], activation=nn.HardTanh(0.0, 6.0))  # uses instance

In practice, it makes sense to stick to using the strings in combination with hyper-parameter optimization libraries like Optuna.

⬇️ Installation

The most recent release can be installed from PyPI with:

$ pip install class_resolver

The most recent code and data can be installed directly from GitHub with:

$ pip install git+https://github.com/cthoyt/class-resolver.git

To install in development mode, use the following:

$ git clone git+https://github.com/cthoyt/class-resolver.git
$ cd class-resolver
$ pip install -e .

🙏 Contributing

Contributions, whether filing an issue, making a pull request, or forking, are appreciated. See CONTRIBUTING.rst for more information on getting involved.

👋 Attribution

⚖️ License

The code in this package is licensed under the MIT License.

🍪 Cookiecutter

This package was created with @audreyfeldroy's cookiecutter package using @cthoyt's cookiecutter-snekpack template.

🛠️ For Developers

See developer instructions

The final section of the README is for if you want to get involved by making a code contribution.

❓ Testing

After cloning the repository and installing tox with pip install tox, the unit tests in the tests/ folder can be run reproducibly with:

$ tox

Additionally, these tests are automatically re-run with each commit in a GitHub Action.

📦 Making a Release

After installing the package in development mode and installing tox with pip install tox, the commands for making a new release are contained within the finish environment in tox.ini. Run the following from the shell:

$ tox -e finish

This script does the following:

Uses BumpVersion to switch the version number in the setup.cfg and src/{{cookiecutter.package_name}}/version.py to not have the -dev suffix
Packages the code in both a tar archive and a wheel
Uploads to PyPI using twine. Be sure to have a .pypirc file configured to avoid the need for manual input at this step
Push to GitHub. You'll need to make a release going with the commit where the version was bumped.
Bump the version to the next patch. If you made big changes and want to bump the version by minor, you can use tox -e bumpversion minor after.

	def train(optimizer: Hint[Optimizer] = "adam", optimizer_kwargs: OptionalKwargs = None):
	model = [Parameter(torch.randn(2, 2, requires_grad=True))]
	optimizer = optimizer_resolver.make(optimizer, optimizer_kwargs, model=model)

cthoyt / class-resolver Goto Github PK

class-resolver's Introduction

Class Resolver

💪 Getting Started

🤖 Writing Extensible Machine Learning Models with class-resolver

⬇️ Installation

🙏 Contributing

👋 Attribution

⚖️ License

🍪 Cookiecutter

🛠️ For Developers

❓ Testing

📦 Making a Release

class-resolver's People

Contributors

Stargazers

Watchers

Forkers

class-resolver's Issues

Recommend Projects

Recommend Topics

Recommend Org

🤖 Writing Extensible Machine Learning Models with `class-resolver`