def main():
workers = [
Worker(id=0, params=[0.9, 0.9], hyperparams=[1.0, 0.0], performance=0),
Worker(id=1, params=[0.9, 0.9], hyperparams=[0.0, 1.0], performance=0)
]
train(train_worker, workers, 'Exploit only')
def main():
workers = [
Worker(id=0, params=[0.9, 0.9], hyperparams=[1.0, 0.3], performance=0),
Worker(id=1, params=[0.9, 0.9], hyperparams=[0.3, 1.0], performance=0)
]
train(train_worker, workers, 'Population Based Training')
def main():
workers = [
Worker(id=0, params=[0.9, 0.9], hyperparams=[1.0, 0.0], performance=0),
Worker(id=1, params=[0.9, 0.9], hyperparams=[0.0, 1.0], performance=0)
]
train(train_worker, workers, 'Grid Search')
def train_worker(id, population, child):
model = Model()
params_store = []
params_store.append(population[id].params)
for i in range(0, 64):
params = population[id].params
hyperparams = population[id].hyperparams
params = step(model, params, hyperparams)
performance = eval(model, params, hyperparams)
params_store.append(params)
if i % 8 == 0:
_, best_params = exploit(hyperparams, params, performance,
population)
if not np.array_equal(best_params, params):
"""
best_params != params means that we have found a better model.
In this example we don't pass the hyperparameters of the best model to the explore function because we
want to analyse the effect on both model separately.
"""
hyperparams, params = explore(hyperparams, best_params,
performance, population)
performance = eval(model, params, hyperparams)
population[id] = Worker(id=id,
params=params,
hyperparams=hyperparams,
performance=performance)
# Send parameters to the other end of the connection
child.send(params_store)
def train_worker(id, population, child):
model = Model()
params_store = []
params_store.append(population[id].params)
for i in range(0, 64):
params = population[id].params
hyperparams = population[id].hyperparams
params = step(model, params, hyperparams)
performance = eval(model, params, hyperparams)
params_store.append(params)
population[id] = Worker(id=id,
params=params,
hyperparams=hyperparams,
performance=performance)
# Send parameters to the other end of the connection
child.send(params_store)