def predict(self, data_set: DataSet) -> Mapping[str, int]:
super().predict(data_set)
if self._scaler is None:
raise RuntimeError(
"no model has been built yet. Invoke fit before predict")
# no upsampling during prediction - we may not even have labels at this point
# standardize data using coefficients computed during training
data_set = data_set.scaled(self._scaler)
# get predictions
chunked_predictions = self._learner.predict(data_set)
if self._majority_vote:
return _majority_vote(data_set, chunked_predictions)
else:
return dict(list(zip(data_set.filenames, chunked_predictions)))
def fit(self, data_set: DataSet):
# generic parameter checks
super().fit(data_set)
if self._upsample:
data_set = upsample(data_set)
# shuffle data set after upsampling
data_set = data_set.shuffled()
# standardize features and remember coefficients for prediction
self._scaler = StandardScaler()
self._scaler.fit(data_set.features)
data_set = data_set.scaled(self._scaler)
# train model
self._learner.fit(data_set)