def test_read_pickle_read_preds(ensemble_backend):
"""
This procedure test that we save the read predictions before
destroying the ensemble builder and that we are able to read
them safely after
"""
ensbuilder = EnsembleBuilder(
backend=ensemble_backend,
dataset_name="TEST",
task_type=MULTILABEL_CLASSIFICATION, # Multilabel Classification
metric=roc_auc,
seed=0, # important to find the test files
ensemble_nbest=2,
max_models_on_disc=None,
)
ensbuilder.SAVE2DISC = False
ensbuilder.main(time_left=np.inf, iteration=1, return_predictions=False)
# Check that the memory was created
ensemble_memory_file = os.path.join(ensemble_backend.internals_directory,
'ensemble_read_preds.pkl')
assert os.path.exists(ensemble_memory_file)
# Make sure we pickle the correct read preads and hash
with (open(ensemble_memory_file, "rb")) as memory:
read_preds, last_hash = pickle.load(memory)
compare_read_preds(read_preds, ensbuilder.read_preds)
assert last_hash == ensbuilder.last_hash
ensemble_memory_file = os.path.join(ensemble_backend.internals_directory,
'ensemble_read_scores.pkl')
assert os.path.exists(ensemble_memory_file)
# Make sure we pickle the correct read scores
with (open(ensemble_memory_file, "rb")) as memory:
read_scores = pickle.load(memory)
compare_read_preds(read_scores, ensbuilder.read_scores)
# Then create a new instance, which should automatically read this file
ensbuilder2 = EnsembleBuilder(
backend=ensemble_backend,
dataset_name="TEST",
task_type=MULTILABEL_CLASSIFICATION, # Multilabel Classification
metric=roc_auc,
seed=0, # important to find the test files
ensemble_nbest=2,
max_models_on_disc=None,
)
compare_read_preds(ensbuilder2.read_preds, ensbuilder.read_preds)
compare_read_preds(ensbuilder2.read_scores, ensbuilder.read_scores)
assert ensbuilder2.last_hash == ensbuilder.last_hash
def testMain(self):
ensbuilder = EnsembleBuilder(backend=self.backend,
dataset_name="TEST",
task_type=1, #Binary Classification
metric=roc_auc,
limit=-1, # not used,
seed=0, # important to find the test files
ensemble_nbest=2,
max_iterations=1 # prevents infinite loop
)
ensbuilder.SAVE2DISC = False
ensbuilder.main()
self.assertEqual(len(ensbuilder.read_preds), 2)
self.assertIsNotNone(ensbuilder.last_hash)
self.assertIsNotNone(ensbuilder.y_true_ensemble)
def testMain(self):
ensbuilder = EnsembleBuilder(backend=self.backend,
dataset_name="TEST",
task_type=1, #Binary Classification
metric=roc_auc,
limit=-1, # not used,
seed=0, # important to find the test files
ensemble_nbest=2,
max_iterations=1 # prevents infinite loop
)
ensbuilder.SAVE2DISC = False
ensbuilder.main()
self.assertEqual(len(ensbuilder.read_preds), 2)
self.assertIsNotNone(ensbuilder.last_hash)
self.assertIsNotNone(ensbuilder.y_true_ensemble)
def test_main(ensemble_backend):
ensbuilder = EnsembleBuilder(
backend=ensemble_backend,
dataset_name="TEST",
task_type=MULTILABEL_CLASSIFICATION, # Multilabel Classification
metric=roc_auc,
seed=0, # important to find the test files
ensemble_nbest=2,
max_models_on_disc=None,
)
ensbuilder.SAVE2DISC = False
run_history, ensemble_nbest, _, _, _ = ensbuilder.main(
time_left=np.inf,
iteration=1,
return_predictions=False,
)
assert len(ensbuilder.read_preds) == 3
assert ensbuilder.last_hash is not None
assert ensbuilder.y_true_ensemble is not None
# Make sure the run history is ok
# We expect at least 1 element to be in the ensemble
assert len(run_history) > 0
# As the data loader loads the same val/train/test
# we expect 1.0 as score and all keys available
expected_performance = {
'ensemble_val_score': 1.0,
'ensemble_test_score': 1.0,
'ensemble_optimization_score': 1.0,
}
# Make sure that expected performance is a subset of the run history
assert all(item in run_history[0].items()
for item in expected_performance.items())
assert 'Timestamp' in run_history[0]
assert isinstance(run_history[0]['Timestamp'], pd.Timestamp)
assert os.path.exists(
os.path.join(ensemble_backend.internals_directory,
'ensemble_read_preds.pkl')), os.listdir(
ensemble_backend.internals_directory)
assert os.path.exists(
os.path.join(ensemble_backend.internals_directory,
'ensemble_read_scores.pkl')), os.listdir(
ensemble_backend.internals_directory)
def testMain(self):
ensbuilder = EnsembleBuilder(
backend=self.backend,
dataset_name="TEST",
task_type=3, # Multilabel Classification
metric=roc_auc,
limit=-1, # not used,
seed=0, # important to find the test files
ensemble_nbest=2,
max_iterations=1, # prevents infinite loop
max_models_on_disc=None,
)
ensbuilder.SAVE2DISC = False
ensbuilder.main()
self.assertEqual(len(ensbuilder.read_preds), 3)
self.assertIsNotNone(ensbuilder.last_hash)
self.assertIsNotNone(ensbuilder.y_true_ensemble)
# Make sure the run history is ok
run_history = ensbuilder.get_ensemble_history()
# We expect 1 element to be the ensemble
self.assertEqual(len(run_history), 1)
# As the data loader loads the same val/train/test
# we expect 1.0 as score and all keys available
expected_performance = {
'ensemble_val_score': 1.0,
'ensemble_test_score': 1.0,
'ensemble_optimization_score': 1.0,
}
self.assertDictContainsSubset(expected_performance, run_history[0])
self.assertIn('Timestamp', run_history[0])
self.assertIsInstance(run_history[0]['Timestamp'], pd.Timestamp)
def main(task_id, ensemble_dir, performance_range_threshold, ensemble_size,
max_keep_best, seed, only_portfolio_runs, call_from_cmd):
if max_keep_best > 1:
assert max_keep_best == int(max_keep_best)
max_keep_best = int(max_keep_best)
memory_limit = 4000
precision = 32
metric = make_scorer('balanced_accuracy_fast', BalancedAccuracy())
if not os.path.exists(ensemble_dir):
raise NotADirectoryError("%s does not exist")
if call_from_cmd:
assert str(task_id) in ensemble_dir
fl_name = "ensemble_results_%fthresh_%dsize_%fbest" % \
(performance_range_threshold, ensemble_size, max_keep_best)
if only_portfolio_runs:
fl_name += "_only_portfolio"
fl_name = os.path.join(ensemble_dir, fl_name)
if os.path.isfile(fl_name):
raise ValueError("Nothing left to do, %s already exists" % fl_name)
# figure out how many prediction files are in dir
if call_from_cmd:
pred_dir = os.path.join(ensemble_dir, "auto-sklearn-output",
".auto-sklearn", "predictions_ensemble")
n_models = glob.glob(pred_dir +
"/predictions_ensemble_%d_*.npy.gz" % seed)
else:
pred_dir = os.path.join(ensemble_dir, ".auto-sklearn",
"predictions_ensemble")
n_models = glob.glob(pred_dir +
"/predictions_ensemble_%d_*.npy" % seed)
n_models.sort(key=lambda x: int(float(x.split("_")[-2])))
print("\n".join(n_models))
print("Found %d ensemble predictions" % len(n_models))
if len(n_models) == 0:
raise ValueError("%s has no ensemble predictions" % pred_dir)
# Get start time of ensemble building: 1) load json 2) find key 3) get creation times
if call_from_cmd:
timestamps_fl = os.path.join(ensemble_dir, "auto-sklearn-output",
"timestamps.json")
else:
timestamps_fl = os.path.join(ensemble_dir, "timestamps.json")
with open(timestamps_fl, "r") as fh:
timestamps = json.load(fh)
model_timestamps = None
overall_start_time = None
for k in timestamps:
if "predictions_ensemble" in k:
model_timestamps = timestamps[k]
if "start_time_%d" % seed in timestamps[k]:
overall_start_time = timestamps[k]["start_time_%d" % seed]
timestamp_keys = list(model_timestamps.keys())
for timestamp_key in timestamp_keys:
if timestamp_key.endswith(
'lock') or 'predictions_ensemble' not in timestamp_key:
del model_timestamps[timestamp_key]
assert model_timestamps is not None and overall_start_time is not None
assert len(model_timestamps) == len(n_models), (len(model_timestamps),
len(n_models))
# Get overall timelimit
vanilla_results_fl = os.path.join(ensemble_dir, "result.json")
with open(vanilla_results_fl, "r") as fh:
vanilla_results = json.load(fh)
# If only portfolio configurations, read runhistory
if only_portfolio_runs:
if call_from_cmd:
runhistory_fl = os.path.join(ensemble_dir, "auto-sklearn-output",
"smac3-output", "run*",
"runhistory.json")
else:
runhistory_fl = os.path.join(ensemble_dir, "smac3-output", "run*",
"runhistory.json")
runhistory_fl = glob.glob(runhistory_fl)
assert len(runhistory_fl) == 1
with open(runhistory_fl[0], "r") as fh:
runhistory = json.load(fh)
init_design_num_runs = []
for i in runhistory["data"]:
if i[1][3]["configuration_origin"] == "Initial design":
if "error" in i[1][3]:
continue
init_design_num_runs.append(i[1][3]["num_run"])
print("Portfolio stopped after %s runs" % str(init_design_num_runs))
last_run = max(init_design_num_runs)
print("Cut down to only portfolio runs fom %d" % len(n_models))
for i, n in enumerate(n_models):
if int(float(n.split("_")[-2])) > last_run:
n_models = n_models[:i]
break
print("... to %d" % len(n_models))
# load data
X_train, y_train, X_test, y_test, cat = load_task(task_id)
if len(np.unique(y_test)) == 2:
task_type = BINARY_CLASSIFICATION
elif len(np.unique(y_test)) > 2:
task_type = MULTICLASS_CLASSIFICATION
else:
raise ValueError("Unknown task type for task %d" % task_id)
tmp_dir = tempfile.TemporaryDirectory()
loss_trajectory = []
# Construct ensemble builder
context = BackendContextMock(
temporary_directory=(ensemble_dir + "/auto-sklearn-output/"
if call_from_cmd else ensemble_dir),
output_directory=tmp_dir.name,
delete_tmp_folder_after_terminate=False,
delete_output_folder_after_terminate=False,
shared_mode=False)
backend = Backend(context)
ens_builder = EnsembleBuilder(
backend=backend,
dataset_name=str(task_id),
task_type=task_type,
metric=metric,
limit=np.inf,
ensemble_size=ensemble_size,
ensemble_nbest=max_keep_best,
performance_range_threshold=performance_range_threshold,
max_models_on_disc=None,
seed=seed,
shared_mode=False,
precision=precision,
max_iterations=1,
read_at_most=1,
memory_limit=memory_limit,
random_state=1,
sleep_duration=0)
try:
# iterate over all models, take construction time into account when creating new trajectory
current_ensemble_timestamp = 0
skipped = 1
for midx, model_path in enumerate(n_models):
tstamp = model_timestamps[model_path.split("/")[-1].replace(
'.gz', '')] - overall_start_time
if current_ensemble_timestamp > tstamp:
# while this model was built, the ensemble script was not yet done
skipped += 1
continue
# Do one ensemble building step
start = time.time()
ens_builder.random_state = check_random_state(1)
print("############## %d: Working on %s (skipped %d)" %
(midx + 1, model_path, skipped - 1))
logging.basicConfig(level=logging.DEBUG)
ens_builder.read_at_most = skipped
valid_pred, test_pred = ens_builder.main(return_pred=True)
last_dur = time.time() - start
current_ensemble_timestamp = tstamp + last_dur
if current_ensemble_timestamp >= vanilla_results["0"]["time_limit"]:
print("############## Went over time %f > %f; Stop here" %
(current_ensemble_timestamp,
vanilla_results["0"]["time_limit"]))
break
# Reset, since we have just read model files
skipped = 1
if test_pred is None:
# Adding this model did not change the ensemble, no new prediction
continue
if task_type == BINARY_CLASSIFICATION:
# Recreate nx2 array
test_pred = np.concatenate([
1 - test_pred.reshape([-1, 1]),
test_pred.reshape([-1, 1])
],
axis=1)
# Build trajectory entry
score = 1 - balanced_accuracy(y_true=y_test, y_pred=test_pred)
loss_trajectory.append((current_ensemble_timestamp, score))
print("############## Round %d took %g sec" %
(midx, time.time() - start))
except:
raise
finally:
tmp_dir.cleanup()
# Store results
result = dict()
result[ensemble_size] = {
'task_id': task_id,
'time_limit': vanilla_results["0"]["time_limit"],
'loss': loss_trajectory[-1][1],
'configuration': {
"n_models": n_models,
"performance_range_threshold": performance_range_threshold,
"ensemble_size": ensemble_size,
"max_keep_best": max_keep_best,
"seed": seed,
"memory_limit": memory_limit,
"precision": precision,
},
'n_models': len(n_models),
'trajectory': loss_trajectory,
}
with open(fl_name, 'wt') as fh:
json.dump(result, fh, indent=4)
print("Dumped to %s" % fl_name)
