README.rst

sedna

pip install sedna

Simple interface to Orion hyperparameter search, no storage, no setup required, sedna gives you back the full control over the optimization process.

Leverage Orion optimizers

• Random Search
• Grid Search
• Hyperband
• ASHA
• BOHB
• DEHB
• Population Based Training (PBT)
• Population Based Bandits (PB2)
• TPE
• Ax
• Evolution-ES
• MOFA
• Nevergrad
• HEBO

Simplified space definition

As decorated function

@hyperparameter(a=uniform(0, 1), b=uniform(1, 2))
def objective(a, b):
    return a + b

As annotated function

def objective(a: uniform(0, 1), b: uniform(1, 2)) -> float:
    return a + b

As dataclass

@dataclass
class MySpace:
    a: uniform(0, 1) = 0
    b: uniform(1, 2) = 1

Gives you back the power over the optimization process

Setup your own workflow

from sedna.core.space import fidelity, get_space, hyperparameter, uniform
from sedna.core.hunt import Optimize


@hyperparameter(epoch=fidelity(2, 10, base=2), a=uniform(0, 1), b=uniform(1, 2))
def fun(epoch, a, b):
   return (a + b) / epoch


def main():
   space = get_space(fun)

   opt = Optimize("hyperband", space, max_trials=10)

   while not opt.is_done():
      samples = opt.suggest(2)

      for sample in samples:
         result = fun(**sample.params)

         opt.observe(sample, result)

Integrate it with your current workflow

from sedna.core.space import fidelity, get_space, hyperparameter, uniform
from sedna.core.hunt import Optimize


@hyperparameter(epoch=fidelity(2, 10, base=2), a=uniform(0, 1), b=uniform(1, 2))
def fun(epoch, a, b):
   return (a + b) / epoch


def main(njob):

   import submitit

   executor = submitit.AutoExecutor(folder="log_test")
   executor.update_parameters(timeout_min=1, slurm_partition="dev")

   opt = Optimize("hyperband", space, max_trials=10)

   while not opt.is_done():
      samples = opt.suggest(njob)
      futures = []

      for sample in samples:
         job = executor.submit(fun, **sample.params)
         futures.append((sample, job)

      for sample, future in futures:
         result = job.result()
         opt.observe(sample, result)