diff --git a/python/prophet/forecaster.py b/python/prophet/forecaster.py index 261776c85..d54e7f358 100644 --- a/python/prophet/forecaster.py +++ b/python/prophet/forecaster.py @@ -32,6 +32,19 @@ logger.setLevel(logging.INFO) NANOSECONDS_TO_SECONDS = 1000 * 1000 * 1000 + +def _to_datetime(values: pd.Series) -> pd.Series: + """pd.to_datetime that also converts pyarrow-backed timestamps to numpy datetime64. + + pd.to_datetime leaves some pyarrow timestamp columns (for example + timestamp[ns][pyarrow] from pd.read_parquet(..., dtype_backend="pyarrow")) + as pyarrow, and the numpy-based code downstream fails on them. + """ + ds = pd.to_datetime(values) + if isinstance(ds.dtype, getattr(pd, "ArrowDtype", ())): + ds = pd.Series(pd.DatetimeIndex(ds), index=ds.index, name=ds.name) + return ds + class Prophet: """Prophet forecaster. @@ -353,7 +366,7 @@ def setup_dataframe(self, df: pd.DataFrame, initialize_scales: bool = False) -> raise ValueError('Found infinity in column y.') if df['ds'].dtype == np.int64: df['ds'] = df['ds'].astype(str) - df['ds'] = pd.to_datetime(df['ds']) + df['ds'] = _to_datetime(df['ds']) if df['ds'].dt.tz is not None: raise ValueError( 'Column ds has timezone specified, which is not supported. ' @@ -1350,7 +1363,7 @@ def preprocess(self, df: pd.DataFrame, **kwargs: Any) -> ModelInputData: history = df[df['y'].notnull()].copy() if history.shape[0] < 2: raise ValueError('Dataframe has less than 2 non-NaN rows.') - self.history_dates = pd.to_datetime(pd.Series(df['ds'].unique(), name='ds')).sort_values() + self.history_dates = _to_datetime(pd.Series(df['ds'].unique(), name='ds')).sort_values() self.history = self.setup_dataframe(history, initialize_scales=True) self.set_auto_seasonalities() diff --git a/python/prophet/tests/test_prophet.py b/python/prophet/tests/test_prophet.py index 0c5e9d0e2..065bd5081 100644 --- a/python/prophet/tests/test_prophet.py +++ b/python/prophet/tests/test_prophet.py @@ -591,6 +591,29 @@ def test_auto_daily_seasonality(self, daily_univariate_ts, subdaily_univariate_t m.fit(daily_univariate_ts) assert "daily" not in m.seasonalities + @pytest.mark.parametrize("unit", ["ns", "us"]) + def test_fit_predict_pyarrow_ds(self, daily_univariate_ts, backend, unit): + # ds backed by pyarrow, e.g. from pd.read_parquet(..., dtype_backend="pyarrow") + pytest.importorskip("pyarrow") + arrow_dtype = f"timestamp[{unit}][pyarrow]" + train = daily_univariate_ts.head(100) + m_numpy = Prophet(stan_backend=backend).fit(train) + m_arrow = Prophet(stan_backend=backend).fit(train.astype({"ds": arrow_dtype})) + assert m_arrow.seasonalities == m_numpy.seasonalities + future = m_numpy.make_future_dataframe(10) + expected = m_numpy.predict(future)["yhat"] + np.testing.assert_allclose(m_arrow.predict(future)["yhat"], expected) + np.testing.assert_allclose( + m_arrow.predict(future.astype({"ds": arrow_dtype}))["yhat"], expected + ) + + def test_pyarrow_ds_with_timezone(self, daily_univariate_ts, backend): + pytest.importorskip("pyarrow") + train = daily_univariate_ts.head(100).copy() + train["ds"] = train["ds"].dt.tz_localize("UTC").astype("timestamp[ns, tz=UTC][pyarrow]") + with pytest.raises(ValueError, match="timezone"): + Prophet(stan_backend=backend).fit(train) + def test_set_seasonality_mode(self, backend): # Setting attribute m = Prophet(stan_backend=backend)