def run(uri,
database,
namespace,
function,
num_experiments,
num_repro,
objective,
variables,
defaults,
params,
resumable,
sleep_time=60,
save_dir='.'):
client = new_client(uri, database)
defaults.update(dict(list(variables.items()) + list(params.items())))
configs = generate(num_experiments, num_repro, objective,
list(sorted(variables)), defaults, resumable)
register(client, function, namespace, configs)
wait(client, namespace, sleep=sleep_time)
data = fetch_results(client, namespace, configs, list(sorted(variables)),
params, defaults)
save_results(namespace, data, save_dir)
test(data, num_experiments, num_repro, objective, variables, resumable)
def test_register_uniques(client):
defaults = {'a': 1000, 'b': 1001, 'c': 1002, 'd': 3}
namespace = 'test'
configs = generate(range(3), 'abc', defaults=defaults)
variables = list('abc') + ['reference']
namespaces = [env(namespace, v) for v in variables]
assert fetch_registered(client, namespaces) == set()
register(client, foo, namespace, configs)
assert len(fetch_registered(client, namespaces)) == 4 * 3
def test_fetch_results_non_completed(client):
defaults = {'a': 0, 'b': 1}
params = {'c': 2, 'd': 3}
defaults = {'epoch': 0}
medians = ['a']
configs = generate(range(2), 'ab', params)
namespace = 'test'
register(client, foo, namespace, configs)
with pytest.raises(RuntimeError) as exc:
fetch_results(client, namespace, configs, medians, params, defaults)
assert exc.match('Not all trials are completed')
def test_register_uniques(client):
num_replicates = 10
num_experiments = 5
replicates = generate_mocked_replicates(num_replicates, num_experiments)
def count_registered():
status = fetch_vars_stats(client, NAMESPACE)
return sum(status[key]['count'] for key in status.keys())
assert count_registered() == 0
register(client, foo, NAMESPACE, replicates)
assert count_registered(
) == num_experiments * num_replicates * 3 # (biased+simfree+simfixed)
def test_register_resume(client):
defaults = {'a': 0, 'b': 1, 'c': 2, 'd': 3}
namespace = 'test'
configs = generate(range(3), 'abc', defaults=defaults)
variables = list('abc') + ['reference']
namespaces = [env(namespace, v) for v in variables]
assert fetch_registered(client, namespaces) == set()
new_registered = register(client, foo, namespace, configs)
assert len(fetch_registered(client, namespaces)) == 4 * 3
assert fetch_registered(client, namespaces) == new_registered
# Resume with 10 seeds per configs this time.
configs = generate(range(10), 'abc', defaults=defaults)
new_registered = register(client, foo, namespace, configs)
assert len(fetch_registered(client, namespaces)) == 4 * 3 + 4 * 7
assert fetch_registered(client, namespaces) != new_registered
assert len(new_registered) == 4 * 7
def test_fetch_results_corrupt_completed(client):
defaults = {'a': 0, 'b': 1}
params = {'c': 2, 'd': 3}
defaults = {'epoch': 0}
medians = ['a']
num_items = 2
configs = generate(range(num_items), 'ab', defaults=defaults)
namespace = 'test'
register(client, foo, namespace, configs)
for variable in configs.keys():
for i in range(num_items):
workitem = client.dequeue(WORK_QUEUE, env(namespace, variable))
client.mark_actioned(WORK_QUEUE, workitem)
with pytest.raises(RuntimeError) as exc:
fetch_results(client, namespace, configs, medians, params, defaults)
assert exc.match('Nothing found in result queue for trial')
def test_register_resume(client):
num_replicates = 10
num_experiments = 5
replicates = generate_mocked_replicates(num_replicates, num_experiments)
def count_registered():
status = fetch_vars_stats(client, NAMESPACE)
return sum(status[key]['count'] for key in status.keys())
assert count_registered() == 0
register(client, foo, NAMESPACE, replicates)
assert count_registered(
) == num_experiments * num_replicates * 3 # (biased+simfree+simfixed)
# Resume with 10 more replicates per configs this time.
replicates = generate_mocked_replicates(num_replicates + 10,
num_experiments)
register(client, foo, NAMESPACE, replicates)
assert count_registered() == num_experiments * (num_replicates + 10) * 3
def test_fetch_results_all_completed(client):
defaults = {'a': 1000, 'b': 1001}
params = {'c': 2, 'd': 3, 'epoch': 5}
defaults.update(params)
medians = ['a']
num_items = 2
configs = generate(range(num_items), 'ab', defaults=defaults)
namespace = 'test'
register(client, foo, namespace, configs)
print(configs)
worker = TrialWorker(URI, DATABASE, 0, None)
worker.max_retry = 0
worker.timeout = 1
worker.run()
print(fetch_vars_stats(client, namespace))
data = fetch_results(client, namespace, configs, medians, params, defaults)
assert data.medians == ['a']
assert data.noise.values.tolist() == ['a', 'b']
assert data.params.values.tolist() == ['c', 'd']
assert data.order.values.tolist() == [0, 1]
assert data.epoch.values.tolist() == list(range(params['epoch'] + 1))
assert data.uid.shape == (3, 2)
assert data.seed.values.tolist(
) == data.noise.values.tolist() + ['reference']
assert data.a.values.tolist() == [[0, 1000, 1000], [1, 1000, 1000]]
assert data.b.values.tolist() == [[1001, 0, 1001], [1001, 1, 1001]]
assert data.c.values.tolist() == [[2, 2, 2], [2, 2, 2]]
assert data.d.values.tolist() == [[3, 3, 3], [3, 3, 3]]
assert (data.obj.loc[dict(order=0, seed='a')].values.tolist() == list(
range(2002, 2002 + params['epoch'] + 1)))
def test_remaining(client):
defaults = {'a': 0, 'b': 1, 'c': 2, 'd': 3}
configs = generate(range(1), 'ab', defaults=defaults)
namespace = 'test'
register(client, foo, namespace, configs)
def get_stats(variables):
trial_stats = fetch_vars_stats(client, namespace)
print(trial_stats)
return {v: trial_stats[f'test-{v}'] for v in variables}
assert remaining(get_stats('ab'))
assert remaining(get_stats('a'))
workitem = client.dequeue(WORK_QUEUE, env(namespace, 'a'))
assert remaining(get_stats('a'))
client.mark_actioned(WORK_QUEUE, workitem)
assert not remaining(get_stats('a'))
assert remaining(get_stats('ab'))
workitem = client.dequeue(WORK_QUEUE, env(namespace, 'b'))
assert remaining(get_stats('ab'))
client.mark_actioned(WORK_QUEUE, workitem)
assert not remaining(get_stats('ab'))
def test_consolidate_results(client):
num_experiments = 5
num_replicates = 10
space = {
'a': 'uniform(lower=-1, upper=1)',
'b': 'uniform(lower=-1, upper=1)',
'c': 'uniform(lower=-1, upper=1)'
}
variables = {'d': 5}
defaults = {'e': 2, 'epoch': 5}
seed = 2
hpo = 'random_search'
objective = 'obj'
fidelity = Fidelity(5, 5, name='epoch').to_dict()
surrogate_budget = 10
hpo_budget = 5
configs = generate_hpos(list(range(num_experiments)), [hpo],
budget=surrogate_budget,
fidelity=fidelity,
search_space=space,
namespace=NAMESPACE,
defaults=defaults)
to_replicate = get_configs_to_replicate(configs, num_experiments)
reset_pool_size(configs['random_search'])
randomize_seeds(configs['random_search'], variables, seed)
variable_names = list(sorted(variables.keys()))
hpo_stats = fetch_all_hpo_stats(client, NAMESPACE)
namespaces = register_hpos(client, NAMESPACE, foo, configs, defaults,
hpo_stats)
worker = TrialWorker(URI, DATABASE, 0, None)
worker.max_retry = 0
worker.timeout = 0.02
worker.run()
data = defaultdict(dict)
hpos_ready, remainings = fetch_hpos_valid_curves(client, namespaces,
variable_names, data)
ready_configs = get_ready_configs(hpos_ready, configs, to_replicate)
replicates = generate_replicates(ready_configs,
data,
variables,
objective,
hpo_budget,
num_replicates,
early_stopping=False)
register(client, foo, NAMESPACE, replicates)
worker = TrialWorker(URI, DATABASE, 0, None)
worker.max_retry = 0
worker.timeout = 0.02
worker.run()
print(fetch_vars_stats(client, NAMESPACE))
data = fetch_hpos_replicates(client, configs, replicates, variable_names,
space, data)
data = consolidate_results(data)
assert len(data) == 1
assert len(data['random_search']) == 4
hpo_reps = data['random_search']
assert hpo_reps['ideal'].obj.shape == (6, surrogate_budget,
num_experiments)
assert hpo_reps['biased'].obj.shape == (6, num_replicates, num_experiments)
assert hpo_reps['simul-fix'].obj.shape == (6, num_replicates,
num_experiments)
assert hpo_reps['simul-free'].obj.shape == (6, num_replicates,
num_experiments)
def count_unique(attr):
return len(set(attr.values.reshape(-1).tolist()))
# Test sources of variation
# NOTE: In ideal, source of variation will vary across ideal after consolidation
# but it stays fixed during the HPO itself
assert count_unique(hpo_reps['ideal']['d']) == num_experiments
assert count_unique(hpo_reps['biased']['d']) == (num_replicates *
num_experiments)
assert count_unique(hpo_reps['simul-free']['d']) == (num_replicates *
num_experiments)
assert count_unique(hpo_reps['simul-fix']['d']) == num_experiments
# Test HPs
assert count_unique(hpo_reps['ideal']['a']) == (num_experiments *
surrogate_budget)
assert count_unique(hpo_reps['biased']['a']) == num_experiments
assert count_unique(hpo_reps['simul-free']['a']) == (num_replicates *
num_experiments)
assert count_unique(hpo_reps['simul-fix']['a']) == (num_replicates *
num_experiments)
assert numpy.allclose(hpo_reps['simul-free']['a'].values,
hpo_reps['simul-fix']['a'].values)
