def _evaluate(params):
e, base, heuristic, strategy = params
session = Session(get_engine())
model = session.query(Model).filter_by(file_name=e.model_file.name).first()
session.close()
# Generate evaluation frames
y_pred = e.binarizer.binarize(base=base, heuristic=heuristic, strategy=strategy, go_backwards=False)
y_true = e.binarizer.get_targets()
evaluations = []
for axis in [0, 1, 2]:
for i, attribute_name in enumerate(e.dataset.attribute_keys):
def get_evaluation(label, precision, recall, f1):
return Evaluation(model_id=model.id, file_name=model.file_name,
label=label, perspective=perspective, attribute_name=attribute_name,
axis=axis, base=base, heuristic=heuristic, strategy=strategy,
precision=precision, recall=recall, f1=f1)
perspective = 'Control Flow' if i == 0 else 'Data'
if i > 0 and not e.ad_.supports_attributes:
evaluations.append(get_evaluation('Normal', 0.0, 0.0, 0.0))
evaluations.append(get_evaluation('Anomaly', 0.0, 0.0, 0.0))
else:
yp = label_collapse(y_pred[:, :, i:i + 1], axis=axis).compressed()
yt = label_collapse(y_true[:, :, i:i + 1], axis=axis).compressed()
p, r, f, _ = metrics.precision_recall_fscore_support(yt, yp, labels=[0, 1])
evaluations.append(get_evaluation('Normal', p[0], r[0], f[0]))
evaluations.append(get_evaluation('Anomaly', p[1], r[1], f[1]))
return evaluations
def fit_and_save(dataset_name, ad, ad_kwargs=None, fit_kwargs=None):
if ad_kwargs is None:
ad_kwargs = {}
if fit_kwargs is None:
fit_kwargs = {}
# Save start time
start_time = arrow.now()
# Dataset
dataset = Dataset(dataset_name)
# AD
ad = ad(**ad_kwargs)
# Train and save
ad.fit(dataset, **fit_kwargs)
file_name = f'{dataset_name}_{ad.abbreviation}_{start_time.format(DATE_FORMAT)}'
model_file = ad.save(file_name)
# Save end time
end_time = arrow.now()
# Cache result
#print(model_file.str_path)
Evaluator(model_file.str_path).cache_result()
# Calculate training time in seconds
training_time = (end_time - start_time).total_seconds()
# Write to database
engine = get_engine()
session = Session(engine)
session.add(
Model(creation_date=end_time.datetime,
algorithm=ad.name,
training_duration=training_time,
file_name=model_file.file,
training_event_log_id=EventLog.get_id_by_name(dataset_name),
training_host=socket.gethostname(),
hyperparameters=str(dict(**ad_kwargs, **fit_kwargs))))
session.commit()
session.close()
if isinstance(ad, NNAnomalyDetector):
from keras.backend import clear_session
clear_session()
if base is not None and base not in e.ad_.supported_bases:
continue
_params.append([e, base, heuristic, strategy])
return [_e for p in _params for _e in _evaluate(p)]
models = sorted([m.name for m in get_model_files()])
evaluations = []
for i in range(len(models)):
e = evaluate(models[i])
evaluations.append(e)
# Write to database
session = Session(get_engine())
for e in evaluations:
session.bulk_save_objects(e)
session.commit()
session.close()
out_dir = PLOT_DIR / 'isj-2019'
eval_file = out_dir / 'eval.pkl'
session = Session(get_engine())
evaluations = session.query(Evaluation).all()
rows = []
for ev in tqdm(evaluations):
m = ev.model
el = ev.model.training_event_log
rows.append([
