Ray Tune ; combine population based training schedule with Hyperopt

Question

Are Population Based Training (PBT) and HyperOpt Search combinable ?

The AsyncHyperBandScheduler is used in the Hyperopt Example of ray.tune

Here config set some parameters for the run() function

config = {
    "num_samples": 10 if args.smoke_test else 1000,
    "config": {
        "iterations": 100,
    },
    "stop": {
        "timesteps_total": 100
    },
}

and space is the argument for the hyperparameter Space inside the Tune Search Algorithm with hyperopt functions:

space = {
    "width": hp.uniform("width", 0, 20),
    "height": hp.uniform("height", -100, 100),
    "activation": hp.choice("activation", ["relu", "tanh"])
}

algo = HyperOptSearch()
scheduler = AsyncHyperBandScheduler()
run(easy_objective, search_alg=algo, scheduler=scheduler, **config)

yet in the population-based example with Keras the hyperparameter space is loosely given by hyperparam_mutations inside the Tune Trials Scheduler with numpy functions

pbt = PopulationBasedTraining(
    hyperparam_mutations={
        "dropout": lambda: np.random.uniform(0, 1),
        "lr": lambda: 10**np.random.randint(-10, 0),
        "rho": lambda: np.random.uniform(0, 1)
    })

and config is used in a different way:

To set starting parameters for individuals of the population

run(MemNNModel,scheduler=pbt,
     config={
        "batch_size": 32,
        "epochs": 1,
        "dropout": 0.3,
        "lr": 0.01,
        "rho": 0.9
    })

To summarize:

Hyperopt takes its the Hyperparameter Space inside the Search Algorithm with hyperopt functions Population-Based Training takes its the Hyperparameter Space inside the Trials Scheduler with numpy functions Both use Config differently

From this answer I take that, the Hyperopt Search will only take completed Trials into account.

Is there a conflict with sampled hyperparameters from Hyperopt Search and "changed while running" hyperparameters from Population-Based Training? quote: "a single trial may see many different hyperparameters over its lifetime"

Ray Tune ; combine population based training schedule with Hyperopt

0 Answers0