Tunafish is an autotuner for plain ol' Python functions. Under the hood, it uses genetic algorithms to determine which arguments maximize a function's return value (i.e. it's fitness); however, Tunafish eliminates the need to think about genetic algorithms. Instead, you can focus more on just writing code.
To use Tunafish, your objective function must satisfy the following two conditions:
- Arguments must be annotated as primitives types, like
float,int,str, andbool,typing.Text, etc. - The return value must be expressed as a single
float(i.e. a fitness value).
Consider a function whose objective is to place orders to buy and sell stocks. The input arguments are aggression, which regulates the a minimum amount of time between orders, and window, which determines how far back the trading algorithm should looks when deciding when take action. The return value is simply the net gain or loss generated while trading, which is used as a proxy for fitness.
from example_project import create_trader, load_historical_trading_data
trader = create_trader()
training_data = load_historical_trading_data(start, stop, interval)
def trade(aggression: float, window: int) -> float:
gains = trader.trade(training_data, aggression, window)
return gains # AKA fitnessfrom tunafish import FunctionTuner
tuner = FunctionTuner()
arguments = tuner.tune(trade, options={
'aggression': {'min': 0.01, 'max': 1.0},
'window': {'min': 5, 'max': 20}
})Working examples can be found in tunafish.examples. The "basic" and "early_stopping" examples differ only in that "early_stopping" shows you how to control a bit more of the internals of the genetic algorithm. In particular, we tell it to exit the training loop early if we reach a fitness goal before all 500 epochs have run. Running these examples should generate a plot, showing convergence of fitness versus time.
Just do python -m tunafish.examples.basic!