Skip to content

⚡️ Allow FillGapsAR to perform prediction on data used for fitting. #17

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
BaptisteDE merged 1 commit into from
Feb 7, 2025
Merged

⚡️ Allow FillGapsAR to perform prediction on data used for fitting. #17

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
35 changes: 29 additions & 6 deletions tests/test_processing.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -669,12 +669,17 @@ def test_pd_fill_gap(self):
for gap in holes_pairs:
toy_df_gaps.loc[gap[0], gap[1]] = np.nan

filler = FillGapsAR()
filler = FillGapsAR(recursive_fill=False)
res = filler.fit_transform(toy_df_gaps.copy())

for gap in holes_pairs[1:]:
# Skip the first one. r2_score doesn't work for only value
assert r2_score(toy_df.loc[gap[0], gap[1]], res.loc[gap[0], gap[1]]) > 0.99
assert r2_score(toy_df.loc[gap[0], gap[1]], res.loc[gap[0], gap[1]]) > 0.80

filler = FillGapsAR(model_name="STL", recursive_fill=True)
res = filler.fit_transform(toy_df_gaps.copy())

for gap in holes_pairs[1:]:
assert r2_score(toy_df.loc[gap[0], gap[1]], res.loc[gap[0], gap[1]]) > 0.80

toy_df_15min = toy_df.resample("15min").mean().interpolate()
hole_backast = pd.date_range(
Expand All @@ -689,22 +694,40 @@ def test_pd_fill_gap(self):
toy_df_15min_hole.iloc[:12, 0] = np.nan
toy_df_15min_hole.iloc[-12:, 0] = np.nan

filler = FillGapsAR(resample_at_td="1h")
filler = FillGapsAR(resample_at_td="1h", recursive_fill=False)
res = filler.fit_transform(toy_df_15min_hole.copy())

assert (
r2_score(
res.loc[hole_backast, "Temp_1"],
toy_df_15min.loc[hole_backast, "Temp_1"],
)
> 0.80
)
assert (
r2_score(
res.loc[hole_forecast, "Temp_1"],
toy_df_15min.loc[hole_forecast, "Temp_1"],
)
> 0.80
)

filler = FillGapsAR(model_name="STL", resample_at_td="1h", recursive_fill=True)
res = filler.fit_transform(toy_df_15min_hole.copy())

assert (
r2_score(
res.loc[hole_backast, "Temp_1"],
toy_df_15min.loc[hole_backast, "Temp_1"],
)
> 0.95
> 0.80
)
assert (
r2_score(
res.loc[hole_forecast, "Temp_1"],
toy_df_15min.loc[hole_forecast, "Temp_1"],
)
> 0.95
> 0.80
)

def test_combiner(self):
Expand Down
133 changes: 73 additions & 60 deletions tide/processing.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1147,11 +1147,12 @@ class FillGapsAR(BaseFiller, BaseProcessing):

def __init__(
self,
model_name: str = "STL",
model_name: str = "Prophet",
model_kwargs: dict = {},
gaps_lte: str | dt.datetime | pd.Timestamp = None,
gaps_gte: str | dt.datetime | pd.Timestamp = None,
resample_at_td: str | dt.timedelta | pd.Timedelta = None,
recursive_fill: bool = False,
):
BaseFiller.__init__(self, gaps_lte, gaps_gte)
BaseProcessing.__init__(self)
Expand All @@ -1173,79 +1174,91 @@ def __init__(
f"Cannot predict data for gaps LTE to {gaps_lte} with data"
f"at a {resample_at_td} timestep"
)
self.recursive_fill = recursive_fill

def _fit_and_fill_x(self, X, biggest_group, col, idx, backcast):
check_is_fitted(self, attributes=["model_"])
bc_model = self.model_(backcast=backcast, **self.model_kwargs)
def _check_forecast_horizon(self, idx):
idx_dt = idx[-1] - idx[0]
if idx_dt == dt.timedelta(0):
idx_dt = idx.freq
if idx_dt < pd.to_timedelta(self.resample_at_td):
raise ValueError(
f"Forecaster is asked to predict at {idx_dt} in the future "
f"or in the past."
f" But data used for fitting have a {self.resample_at_td} frequency"
)

def _get_x_and_idx_at_freq(self, x, idx, backcast):
if self.resample_at_td is not None:
idx_dt = idx[-1] - idx[0]
if idx_dt == dt.timedelta(0):
idx_dt = idx.freq
if idx_dt < pd.to_timedelta(self.resample_at_td):
raise ValueError(
f"Forecaster is asked to predict at {idx_dt} in the future "
f"or in the past."
f" But data used for fitting have a {self.resample_at_td} frequency"
)
x_fit = X.loc[biggest_group, col].resample(self.resample_at_td).mean()
idx_origin = idx
if backcast:
idx = pd.date_range(
idx[0],
x_fit.index[0] - pd.Timedelta(self.resample_at_td),
freq=self.resample_at_td,
)
else:
idx = pd.date_range(
x_fit.index[-1] + pd.Timedelta(self.resample_at_td),
idx[-1],
freq=self.resample_at_td,
)
self._check_forecast_horizon(idx)
x_out = x.resample(self.resample_at_td).mean()
idx_out = pd.date_range(idx[0], idx[-1], freq=self.resample_at_td).floor(
self.resample_at_td
)
idx_out.freq = idx_out.inferred_freq
else:
x_fit = X.loc[biggest_group, col]
idx_origin = None

x_out = x
idx_out = idx

return x_out, idx_out

def _fill_up_sampling(self, X, idx, col):
beg = idx[0] - idx.freq
end = idx[-1] + idx.freq
# Interpolate linearly between inferred values and using neighbor data
X.loc[idx, col] = X.loc[beg:end, col].interpolate()
# If gap is at boundaries
if beg < X.index[0]:
X.loc[idx, col] = X.loc[idx, col].bfill()
if end > X.index[-1]:
X.loc[idx, col] = X.loc[idx, col].ffill()

def fill_x(self, X, group, col, idx, backcast):
check_is_fitted(self, attributes=["model_"])
bc_model = self.model_(backcast=backcast, **self.model_kwargs)
if self.resample_at_td:
self._check_forecast_horizon(idx)
x_fit, idx_pred = self._get_x_and_idx_at_freq(X.loc[group, col], idx, backcast)
bc_model.fit(x_fit)
to_predict = idx.to_series()
to_predict = idx_pred.to_series()
to_predict.name = col
X.loc[idx, col] = bc_model.predict(to_predict).to_numpy().flatten()
# Here a bit dirty. STL doesn't allow forecast on its fitting set
if self.model_name == "STL":
to_predict = to_predict[~to_predict.isin(x_fit.index)]

X.loc[to_predict, col] = bc_model.predict(to_predict).to_numpy().flatten()

if self.resample_at_td is not None:
beg = idx_origin[0] - idx_origin.freq
end = idx_origin[-1] + idx_origin.freq
# Interpolate linearly between inferred values and using neighbor data
X.loc[idx_origin, col] = X.loc[beg:end, col].interpolate()
# If gap is at boundaries
if beg < X.index[0]:
X.loc[idx_origin, col] = X.loc[idx_origin, col].bfill()
if end > X.index[-1]:
X.loc[idx_origin, col] = X.loc[idx_origin, col].ffill()
self._fill_up_sampling(X, idx, col)

def _fit_implementation(self, X: pd.Series | pd.DataFrame, y=None):
self.model_ = MODEL_MAP[self.model_name]
return self

def _transform_implementation(self, X: pd.Series | pd.DataFrame):
check_is_fitted(self, attributes=["model_"])
gaps = self.get_gaps_dict_to_fill(X)
for col in X:
while gaps[col]:
data_blocks = get_data_blocks(X[col], return_combination=False)[col]
data_timedelta = [block[-1] - block[0] for block in data_blocks]
biggest_group = data_blocks[data_timedelta.index(max(data_timedelta))]
start, end = get_outer_timestamps(biggest_group, X.index)

indices_to_delete = []
for i, idx in enumerate(gaps[col]):
if start in idx:
self._fit_and_fill_x(X, biggest_group, col, idx, backcast=True)
indices_to_delete.append(i)
elif end in idx:
self._fit_and_fill_x(X, biggest_group, col, idx, backcast=False)
indices_to_delete.append(i)

for i in sorted(indices_to_delete, reverse=True):
del gaps[col][i]

if not self.recursive_fill:
for idx in gaps[col]:
self.fill_x(X, X.index, col, idx, backcast=None)
else:
while gaps[col]:
data_blocks = get_data_blocks(X[col], return_combination=False)[col]
data_timedelta = [block[-1] - block[0] for block in data_blocks]
biggest_group = data_blocks[
data_timedelta.index(max(data_timedelta))
]
start, end = get_outer_timestamps(biggest_group, X.index)
indices_to_delete = []
for i, idx in enumerate(gaps[col]):
if start in idx:
self.fill_x(X, biggest_group, col, idx, backcast=True)
indices_to_delete.append(i)
elif end in idx:
self.fill_x(X, biggest_group, col, idx, backcast=False)
indices_to_delete.append(i)

for i in sorted(indices_to_delete, reverse=True):
del gaps[col][i]
return X


Expand Down