gismo07 / minka Goto Github PK

Minka is ment to be a middleware between your optimization problem and a modular set of supporting tools to:

• automatically perform hyperparameter optimization
• store the results, parameter sets and custom data
• compare different training runs
• visualize the results and training processes

The aim of the project is to provide a simple as possible interface without the need to change your existing code.

Could it be easier than just adding one line?

minka(yourOptimizationTask, 'configuration.json').opt(numberOfRuns)

!pip install --upgrade minka-johann-haselberger

• In order to store the data, a mongoDB is required. You can host your own free one here: https://cloud.mongodb.com
• If you want to use the w&b interface, an account is needed. Create one here: https://app.wandb.ai

Minka uses a very simple combination of a configuration json file and the actual optimization task, represented as a single class.

optimization template:

class yourOptimizationTask:
    def __init__(self):
        pass
    
    def prepare(self):
        pass
    
    def run(self, config):
        x = config['x']
        result = (x - 2) ** 2
        
        evalMetrics = {
            'error': result
        }
        
        logArrays = {
            'someValues': [1,2,3,4]
        }
        
        return result, evalMetrics, logArrays
    
    def cleanup(self):
        pass

content of the configuration.json file:

{
    "comment": "parameter types: fix, categorical, discrete_uniform, int, loguniform, uniform",
    "parameters": {
        "batchSize": {"type": "fix","values": 256},
        "epochs": {"type": "fix","values": 75},
        "x": {"type": "categorical","values": [11,22,26]}
    }
}