hyperparameters 0.4.0

Define your Hyperparameters once and use in argparse, hypertunning libraries, and as strongly-typed attributes in code!

Hyperparameters

Are you tired of repeating hyperparameters in code, argparse definitions, and hyperparameter tunning libraries?

Hyperparameters lets you define your hyperparameters once and use everywhere! Moreover, you get type linting and spelling checks for free!

Quickstart

Install Hyperparameters:

pip install hyperparameters

You need to install the ray.tune package separately if you want to use the ray.tune hypertunning integration:

pip install -U "ray[tune]"

Define your parameters once using the Hyperparams class:

from argparse import ArgumentParser
from typing import Optional

from hyperparameters import HP, Hyperparams


class MyHyperparams(Hyperparams):
    epochs: int = HP(
        "Number of epochs to train for",
        default=5,
    )
    lr: float = HP(
        "Learning rate",
        default=1e-3,
    )
    tokenizer: str = HP(
        "HF tokenizer to use",
        default="BPE",
        choices=["BPE", "WordPiece"],
    )
    train_data_path: str = HP(
        "Path to the training dataset",
    )
    use_dropout: bool = HP(
        "Whether the dropout layers should be activated",
        default=True,
    )
    pretrained_weights: Optional[str] = HP(
        "Path to the pretrained model weights, if any",
        default=None,
    )


parser = ArgumentParser()
MyHyperparams.add_arguments(parser)
params = MyHyperparams.from_arguments(parser.parse_args())

# access parameters as typed attributes
params.train_data_path

# convert to a dictionary
print(params.dict())

Let's see the registered arguments by running the code above with --help:

usage: example.py [-h] [--epochs EPOCHS] [--lr LR] [--tokenizer {BPE,WordPiece}] --train-data-path TRAIN_DATA_PATH [--use-dropout | --no-use-dropout]
                  [--pretrained-weights PRETRAINED_WEIGHTS]

options:
  -h, --help            show this help message and exit
  --epochs EPOCHS       Number of epochs to train for
  --lr LR               Learning rate
  --tokenizer {BPE,WordPiece}
                        HF tokenizer to use
  --train-data-path TRAIN_DATA_PATH
                        Path to the training dataset
  --use-dropout         Whether the dropout layers should be activated
  --no-use-dropout      Disable: Whether the dropout layers should be activated
  --pretrained-weights PRETRAINED_WEIGHTS
                        Path to the pretrained model weights, if any

As can be seen from the above, Hyperparameters takes care of low level details for you:

1. The --train-data-path parameter is required because we didn't provide a default value for it. All other parameters are optional.
2. Even though the default value for --pretrained-weights is None, this parameter is optional. However, we had to use the Optional[str] type hint.
3. You can provide the --use-dropout or the --no-use-dropout flag, but not both at the same time. Neither flag is required, as the use_dropout parameter has a default value.
4. The --tokenizer parameter can only be BPE or WordPiece. Providing any other value results in an error.
5. The data types of the parameters are parsed from strings and validated according to the type hints in the MyHyperparams class.
6. The default values are used whenever an argument is ommitted. Let's check that by running with --train-data-path mydata/: the script prints: {'epochs': 5, 'lr': 0.001, 'tokenizer': 'BPE', 'train_data_path': 'mydata/', 'use_dropout': True, 'pretrained_weights': None}.

Default values and required parameters

1. If a default value is omitted completely, the parameter will be required.
2. Otherwise, if any default value is specified, even None, the parameter will be optional.
3. You must use the Optional type hint if the default value is None.
4. You can't have bool parameters with None as default. Flags that can be None just don't make any sense.
5. You can't have choice parameters with None as default. If you need this, just add a "null" value to the list of choices.

Hypertunning

Different hypertunning libraries provide different APIs for defining search spaces. Hyperparameters can be easily extended to support any hypertunning library. You can do it yourself following the steps discussed below - it's easy! The ray.tune library is supported out of the box.

When defining hypertunable parameters with Hyperparameters, make sure to inherit from both the Hyperparams class and the mixin class specific to your hypertunning library. Here is an example for ray.tune:

from hyperparameters import HP, Hyperparams
from hyperparameters.ray_tune_hyperparams import RayTuneHyperparamsMixin


class MyHyperparams(Hyperparams, RayTuneHyperparamsMixin):
    ...

Next, you can specify the search spaces:

1. use the hypertunning library API to describe the search space and provide this object to the HP(..., search_space=) parameter;
2. if you already specify the HP(..., choices=[]) parameter you can treat the possible choices as a search space by setting the HP(..., tunable=True) parameter.
3. for bool fields just set the HP(..., tunable=True) parameter, which is equivallent to HP(..., tunable=True, choices=[False, True]).

Note that when using the search_space parameter, you don't need to use the tunable parameter.

Finally, you are ready to use the mixin-specific methods to extract information about your search spaces. For ray.tune these are:

1. .ray_tune_param_space() method that returns a dictionary describing the search space.
2. .ray_tune_best_values() method that returns a dictionary of the best values to start hypertunning from.

Here is a complete example of defining parameters hypertunable with ray.tune:

from ray import tune

from hyperparameters import HP, Hyperparams
from hyperparameters.ray_tune_hyperparams import RayTuneHyperparamsMixin


class MyHyperparams(Hyperparams, RayTuneHyperparamsMixin):
    lr: float = HP(
        "Learning rate",
        default=1e-3,
        search_space=tune.loguniform(1e-5, 1e-2),
    )
    layers_num: int = HP(
        "Number of model layers",
        default=8,
        tunable=True,
        choices=[4, 8, 16, 24],
    )
    use_dropout: bool = HP(
        "Whether the dropout layers should be activated",
        default=True,
        tunable=True,
    )


params = MyHyperparams()

# Search space config that ray.tune understands
params.ray_tune_param_space()

# Best values to start the hypertunning from
params.ray_tune_best_values()

Supporting other hypertunning libraries

In Hyperparameters, the logic specific to hypertunning libraries is implemented with mixin classes. This means that you can add many mixins to your parameters and support several hypertunning libraries at once.

The best place to start is to review the implementation of the RayTuneHyperparamsMixin class in hyperparameters/ray_tune_hyperparams.py.

The class implementing a new hypertunning library must:

1. Inherit from hyperparameters.hyperparams.HyperparamsProtocol.
2. Provide class methods for converting the params data stored in hyperparameters.hyperparams.HyperparamInfo structure into a format that the hypertunning library understands.
3. Use the cls._tunable_params() to get the info about the parameters.
4. Wrap info.choices in a proper type for the hypertunning library.
5. Provide a method for returning the best parameter values to start the hypertunning from.