Combinations API

The Combinations API in Hypster allows you to generate and query all possible combinations of hyperparameters defined using hp.select() in your configuration. This is particularly useful for hyperparameter tuning and exploring different model configurations.

Generating Combinations

Using get_combinations()

The get_combinations() method generates all possible combinations of hyperparameters for a given configuration, but only for parameters defined with hp.select().

Example:

from hypster import HP, config


@config
def model_config(hp: HP):
    model_type = hp.select(["cnn", "rnn"], default="cnn")
    if model_type == "cnn":
        num_layers = hp.select([3, 5], default=3)
    else:
        cell_type = hp.select(["lstm", "gru"], default="lstm")

    learning_rate = hp.number(default=0.001)
    model_name = hp.text(default="my_model")


# Generate combinations
combinations = model_config.get_combinations()
combinations

---------------------------------------------------------------------------
ModuleNotFoundError                       Traceback (most recent call last)
Cell In[1], line 1
----> 1 from hypster import HP, config
      4 @config
      5 def model_config(hp: HP):
      6     model_type = hp.select(["cnn", "rnn"], default="cnn")

ModuleNotFoundError: No module named 'hypster'

Note that learning_rate and model_name are not included in the combinations as they are not defined using hp.select().

Nested Configurations

For nested configurations using hp.propagate(), the combinations will include the nested structure:

from hypster import HP, config, save

save(model_config, "model_config.py")


@config
def training_config(hp: HP):
    import hypster

    batch_size = hp.select([32, 64], default=32)
    num_epochs = hp.number(default=100)
    model_config = hypster.load("model_config.py")
    model_params = hp.propagate(model_config, name="model")


# Generate combinations
combinations = training_config.get_combinations()
combinations

[{'batch_size': 32, 'model.model_type': 'cnn', 'model.num_layers': 3},
 {'batch_size': 32, 'model.model_type': 'cnn', 'model.num_layers': 5},
 {'batch_size': 32, 'model.model_type': 'rnn', 'model.cell_type': 'lstm'},
 {'batch_size': 32, 'model.model_type': 'rnn', 'model.cell_type': 'gru'},
 {'batch_size': 64, 'model.model_type': 'cnn', 'model.num_layers': 3},
 {'batch_size': 64, 'model.model_type': 'cnn', 'model.num_layers': 5},
 {'batch_size': 64, 'model.model_type': 'rnn', 'model.cell_type': 'lstm'},
 {'batch_size': 64, 'model.model_type': 'rnn', 'model.cell_type': 'gru'}]

Querying Combinations

The query_combinations() function allows you to filter the generated combinations based on specific criteria.

Using query_combinations()

from hypster import query_combinations

filtered_combinations = query_combinations(combinations, {"model.model_type": "cnn", "batch_size": 64})
filtered_combinations

[{'batch_size': 64, 'model.model_type': 'cnn', 'model.num_layers': 3},
 {'batch_size': 64, 'model.model_type': 'cnn', 'model.num_layers': 5}]

Warning

get_combinations() runs your config code multiple times to dynamically explore the options space. Make sure your configuration code runs fast and doesn’t incur any charges.