def get_model_train_test_dataset_splits(self, dataset_df: pd.DataFrame) -> DatasetSplits:
return DatasetSplits.from_proportions(
df=dataset_df,
proportion_train=0.7,
proportion_test=0.2,
proportion_val=0.1,
)
def get_model_train_test_dataset_splits(
self, dataset_df: pd.DataFrame) -> DatasetSplits:
return DatasetSplits.from_proportions(dataset_df,
proportion_train=0.8,
proportion_val=0.05,
proportion_test=0.15,
random_seed=0)
def get_model_train_test_dataset_splits(self, dataset_df: pd.DataFrame) -> DatasetSplits:
if self.test_set_ids_csv:
test_set_ids_csv = self.local_dataset / self.test_set_ids_csv
test_series = pd.read_csv(test_set_ids_csv).series
all_series = dataset_df.series.values
check_all_test_series = all(test_series.isin(all_series))
if not check_all_test_series:
raise ValueError(f"Not all test series from {test_set_ids_csv} were found in the dataset.")
test_set_subjects = dataset_df[dataset_df.series.isin(test_series)].subject.values
train_and_val_series = dataset_df[~dataset_df.subject.isin(test_set_subjects)].series.values
random.seed(42)
random.shuffle(train_and_val_series)
num_val_samples = math.floor(len(train_and_val_series) / 9)
val_series = train_and_val_series[:num_val_samples]
train_series = train_and_val_series[num_val_samples:]
logging.info(f"Dropped {len(all_series) - (len(test_series) + len(train_and_val_series))} series "
f"due to subject overlap with test set.")
return DatasetSplits.from_subject_ids(dataset_df,
train_ids=train_series,
val_ids=val_series,
test_ids=test_series,
subject_column="series",
group_column="subject")
else:
return DatasetSplits.from_proportions(dataset_df,
proportion_train=0.8,
proportion_val=0.1,
proportion_test=0.1,
subject_column="series",
group_column="subject",
shuffle=True)
def get_model_train_test_dataset_splits(
self, dataset_df: pd.DataFrame) -> DatasetSplits:
return DatasetSplits.from_proportions(
df=dataset_df,
proportion_train=0.7,
proportion_test=0.2,
proportion_val=0.1,
random_seed=1,
subject_column=self.subject_column)
def test_grouped_splits(group_column: str) -> None:
test_df = _get_test_df()[0]
proportions = [0.5, 0.4, 0.1]
splits = DatasetSplits.from_proportions(test_df,
proportions[0],
proportions[1],
proportions[2],
group_column=group_column)
_check_is_partition(test_df, [splits.train, splits.test, splits.val],
CSV_SUBJECT_HEADER)
_check_is_partition(test_df, [splits.train, splits.test, splits.val],
group_column)
def test_grouped_k_fold_cross_validation_splits(group_column: str) -> None:
test_df = _get_test_df()[0]
proportions = [0.5, 0.4, 0.1]
splits = DatasetSplits.from_proportions(test_df,
proportions[0],
proportions[1],
proportions[2],
group_column=group_column)
n_splits = 7 # mutually prime with numbers of subjects and groups
val_folds = []
for fold in splits.get_k_fold_cross_validation_splits(n_splits):
_check_is_partition(test_df, [fold.train, fold.test, fold.val],
CSV_SUBJECT_HEADER)
_check_is_partition(test_df, [fold.train, fold.test, fold.val],
group_column)
assert fold.test.equals(splits.test)
val_folds.append(fold.val)
# ensure validation folds partition the original train+val set
train_val = pd.concat([splits.train, splits.val])
_check_is_partition(train_val, val_folds, CSV_SUBJECT_HEADER)
_check_is_partition(train_val, val_folds, group_column)
