A package for tuning machine learning models (bayesian or genetic tuning), with or without a config file.
- Create your model and config files (if desired)
- Create your tuning config in the format as follows (json):
{
"attributes": {
"generic_param_example: {
"type": <"int"/"float"/"bool">,
"min": <min_value if int or float>,
"max": <max_value if int or float>,
<this is an optional param for configtune but NOT tuningbayes, default=1: "step": <step size value>,
<this is an optional param to enforce step limits (for configtune but NOT tuningbayes): "strict": <True/False>>
},
"int_you_want_to_tune_example": {
"type": "int",
"min": 1,
"max": 10,
"step": 2
},
"float_you_want_to_tune_example": {
"type": "int",
"min": 0,
"max": 1,
"step": 0.1
},
"bool_you_want_to_tune_example": {
"type": "bool"
},
"categorical_values_you_want_to_tune_example": {
"type": "categorical",
"values": ["a", "b", "c"]
}
}
Boolean values don't need any bounds. The parameter names should match those found in your model config file, if you have one. Categorical values will be randomly selected for initialization.
- Create your evaluation function. This function needs to take in a config file or a list of values being tuned if you're not using a config. It should output a scalar value.
Example overall usage of TuningDeap:
from configtune import TuningDeap
def eval_function(config_file):
return your_eval_function(config_file)
tune = TuningDeap(eval_function, tuning_config, model_config, n_generation=5, population_size=10,
minimize=True, output_dir="/tmp", verbose=False)
best_config, best_score = tune.run()
Example overall usage of TuningBayes:
from configtune import TuningBayes
def eval_function(config_file):
return your_eval_function(config_file)
tune = TuningBayes(eval_function, tuning_config, model_config, n_calls=10, n_random_starts=2,
output_dir="/tmp", verbose=True)
best_config, best_score = tune.run()