def test_model_raises_if_not_fit():
model = GroupedProphet()
with pytest.raises(
DivinerException,
match="The model has not been fit. Please fit the model first.",
):
model.forecast(30, "days")
def test_model_raises_if_already_fit():
train = data_generator.generate_test_data(2, 1, 1000, "2020-01-01", 1)
model = GroupedProphet().fit(train.df, train.key_columns)
with pytest.raises(
DivinerException,
match="The model has already been fit. Create a new instance to fit the model again.",
):
model.fit(train.df, train.key_columns)
def test_prophet_extract_params():
train = data_generator.generate_test_data(4, 6, 1000, "2020-01-01", 1)
model = GroupedProphet(uncertainty_samples=0).fit(train.df, train.key_columns)
params = model.extract_model_params()
assert len(params) == 6
def grouped_prophet_example(locations, start_dt, artifact_path):
print("Generating data...\n")
data = generate_data(location_data=locations, start_dt=start_dt)
grouping_keys = ["country", "city"]
print("Data Generated.\nBuilding GroupedProphet Model...")
model = GroupedProphet(n_changepoints=96, uncertainty_samples=0).fit(
df=data, group_key_columns=grouping_keys, y_col="watts", datetime_col="datetime"
)
print("GroupedProphet model built.\n")
params = model.extract_model_params()
print(f"Params: \n{params.to_string()}")
print("Running Cross Validation on all groups...\n")
metrics = model.cross_validate_and_score(
horizon="120 hours",
period="480 hours",
initial="960 hours",
parallel="threads",
rolling_window=0.05,
monthly=False,
)
print(f"Cross Validation Metrics: \n{metrics.to_string()}")
mlflow.diviner.log_model(diviner_model=model, artifact_path=artifact_path)
# As an Alternative to saving metrics and params directly with a `log_dict()` function call,
# Serializing the DataFrames to local as a .csv can be done as well, without requiring
# column or object manipulation as shown below this block, utilizing a temporary directory
# with a context wrapper to clean up the files from the local OS after the artifact logging
# is complete:
with tempfile.TemporaryDirectory() as tmpdir:
params.to_csv(f"{tmpdir}/params.csv", index=False, header=True)
metrics.to_csv(f"{tmpdir}/metrics.csv", index=False, header=True)
mlflow.log_artifacts(tmpdir, artifact_path="run_data")
# Saving the parameters and metrics as json without having to serialize to local
# NOTE: this requires casting of fields that cannot be serialized to JSON
# NOTE: Do not use both of these methods. These are shown as an either/or alternative based
# on how you would choose to consume, view, or analyze the per-group metrics and parameters.
# NB: There are object references present in the Prophet model parameters. Coerce to string if
# using a JSON serialization approach with ``mlflow.log_dict()``.
params = params.astype(dtype=str, errors="ignore")
mlflow.log_dict(params.to_dict(), "params.json")
mlflow.log_dict(metrics.to_dict(), "metrics.json")
return mlflow.get_artifact_uri(artifact_path=artifact_path)
def test_prophet_forecast_correct_start():
train = data_generator.generate_test_data(2, 5, 1000, "2020-01-01", 1)
expected_start_of_forecast = max(train.df["ds"]) + timedelta(days=1)
model = GroupedProphet().fit(train.df, train.key_columns)
forecasted_data = model.forecast(10, "D")
# check that the first date in the forecasted df for the first model is 1 day after last date.
min_forecast = min(forecasted_data["ds"])
assert expected_start_of_forecast == min_forecast
assert len(forecasted_data) == 50
def test_prophet_save_and_load():
# Tests serialization, deserialization, and utilization of forecasting API from loaded model
save_path = os.path.join("/tmp/grouped_prophet_test", "model")
train = data_generator.generate_test_data(2, 2, 1000, "2020-01-01", 1)
grouped_model = GroupedProphet().fit(train.df, train.key_columns)
grouped_model.save(save_path)
loaded_model = GroupedProphet.load(save_path)
forecasts = loaded_model.forecast(25, "D")
shutil.rmtree(os.path.dirname(save_path))
assert len(forecasts) == 50
def test_prophet_df_naming_overrides():
train = data_generator.generate_test_data(2, 1, 1000, "2020-01-01", 1)
train_df = train.df
train_df.rename(columns={"ds": "datetime", "y": "sales"}, inplace=True)
assert {"datetime", "sales"}.issubset(set(train_df.columns))
model = GroupedProphet().fit(train_df, train.key_columns, "sales", "datetime")
params = model.extract_model_params()
assert len(params) == 1
def test_prophet_cross_validation_extract():
train = data_generator.generate_test_data(4, 6, 1000, "2020-01-01", 1)
model = GroupedProphet(uncertainty_samples=0).fit(train.df, train.key_columns)
scores = model.cross_validate_and_score(
initial="100 days", period="90 days", horizon="15 days", parallel=None
)
assert all(scores["rmse"] > 0)
assert len(scores) == 6
assert "coverage" not in scores
def test_prophet_manual_predict():
train = data_generator.generate_test_data(2, 1, 1000, "2020-01-01", 1)
train_df = train.df
predict_df = train_df[["key1", "key0", "ds"]][-10:]
model = GroupedProphet().fit(train_df, train.key_columns)
prediction = model.predict(predict_df)
assert len(prediction) == 10
for _, row in prediction.iterrows():
assert row["yhat"] > 0
def test_prophet_group_subset_predict_raises_and_warns():
_rows_to_generate = 30
train = data_generator.generate_test_data(2, 1, 1000, "2020-01-01", 1)
train_df = train.df
model = GroupedProphet().fit(train_df, train.key_columns)
key_entries = []
for v in train_df[["key1", "key0"]].iloc[[0]].to_dict().values():
key_entries.append(list(v.values())[0])
groups = [(key_entries[0], key_entries[1]), ("missing", "key")]
with pytest.raises(
DivinerException, match="Cannot perform predictions due to submitted"
):
model.predict_groups(groups, _rows_to_generate, "D")
with pytest.warns(
UserWarning, match="Specified groups are unable to be predicted due to "
):
model.predict_groups(groups, _rows_to_generate, "D", on_error="warn")
with pytest.raises(
DivinerException, match="Groups specified for subset forecasting are not"
):
model.predict_groups(
("invalid", "invalid"), _rows_to_generate, "D", on_error="ignore"
)
def test_prophet_cross_validation_extract_custom_scores():
train = data_generator.generate_test_data(4, 2, 1000, "2020-01-01", 1)
model = GroupedProphet(uncertainty_samples=0).fit(train.df, train.key_columns)
scores = model.cross_validate_and_score(
initial="100 days",
period="90 days",
horizon="15 days",
parallel=None,
metrics=["rmse", "mape"],
disable_tqdm=False,
monthly=True,
)
assert all(scores["rmse"] > 0)
assert len(scores) == 2
assert "coverage" not in scores
def test_prophet_default_fit():
train = data_generator.generate_test_data(4, 2, 1000, "2020-01-01", 1)
model = GroupedProphet().fit(train.df, train.key_columns)
first_model = _get_individual_model(model, 0)
assert len(first_model.history) > 0
assert (
len(first_model.params["trend"][0]) == 1000
) # fit value for each value in series
assert len(list(model.model.keys())) == 2
def test_prophet_group_subset_predict():
_rows_to_generate = 30
train = data_generator.generate_test_data(2, 1, 1000, "2020-01-01", 1)
train_df = train.df
model = GroupedProphet().fit(train_df, train.key_columns)
key_entries = []
for v in train_df[["key1", "key0"]].iloc[[0]].to_dict().values():
key_entries.append(list(v.values())[0])
groups = [tuple(key_entries)]
group_prediction = model.predict_groups(groups, _rows_to_generate, "D")
assert len(group_prediction) == _rows_to_generate
_key1 = group_prediction["key1"].unique()
assert len(_key1) == 1
assert _key1[0] == groups[0][0]
_key0 = group_prediction["key0"].unique()
assert len(_key0) == 1
assert _key0[0] == groups[0][1]
def test_prophet_execution_with_kwargs_override_for_pystan():
train = data_generator.generate_test_data(4, 6, 1000, "2020-01-01", 1)
default_prophet_uncertainty_samples = Prophet().uncertainty_samples
model = GroupedProphet(uncertainty_samples=0).fit(
train.df, train.key_columns, algorithm="LBFGS"
)
last_model = _get_individual_model(model, 5)
assert last_model.uncertainty_samples == 0
assert default_prophet_uncertainty_samples != last_model.uncertainty_samples
def test_prophet_with_bad_group_data():
train = data_generator.generate_test_data(2, 1, 1000, "2020-01-01", 1)
train_df = train.df
bad_data = pd.DataFrame(
{
"ds": datetime.strptime("2021-01-01", "%Y-%M-%d"),
"y": -500.3,
"key1": "bad",
"key0": "data",
},
index=[1000],
)
train_df_add = pd.concat([train_df, bad_data])
with pytest.warns(RuntimeWarning, match="An error occurred while fitting group"):
model = GroupedProphet().fit(train_df_add, train.key_columns)
assert ("bad", "data") not in model.model.keys()
def test_prophet_save_load_override_object():
"""Test to ensure that deserialization updates object properly for all attributes"""
train1 = data_generator.generate_test_data(3, 2, 1000, "2020-01-01", 1)
train2 = data_generator.generate_test_data(2, 2, 500, "2021-01-01", 1)
model1 = GroupedProphet().fit(train1.df, train1.key_columns)
model2 = GroupedProphet().fit(train2.df, train2.key_columns)
model1_group_keys = deepcopy(model1._group_key_columns)
model1_model = deepcopy(model1.model)
# save model 2
save_path = os.path.join("/tmp/group_prophet_test", "model2serdetest.gpm")
model2.save(save_path)
# use model1 object to load model2
reloaded = model1.load(save_path)
assert set(reloaded._group_key_columns) != set(model1_group_keys)
assert reloaded.model.keys() == model2.model.keys()
assert reloaded.model.keys() != model1_model.keys()
def grouped_prophet(diviner_data):
return GroupedProphet(n_changepoints=20, uncertainty_samples=0).fit(
df=diviner_data.df, group_key_columns=diviner_data.key_columns, y_col="y", datetime_col="ds"
)
def execute_grouped_prophet():
"""
This function call will generate synthetic group time series data in a normalized format.
The structure will be of:
============ ====== =========== =========== ===========
ds y group_key_1 group_key_2 group_key_3
============ ====== =========== =========== ===========
"2016-02-01" 1234.5 A B C
============ ====== =========== =========== ===========
With the grouping key values that are generated per ``ds`` and ``y`` values assigned in a
non-deterministic fashion.
For utililzation of this API, the normalized representation of the data is required, such that
a particular target variables' data 'y' and the associated indexed datetime values in ``'ds'``
are 'stacked' (unioned) from a more traditional denormalized data storage paradigm.
For guidance on this data transposition from denormalized representations, see:
https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.melt.html
"""
generated_data = generate_example_data(
column_count=3,
series_count=10,
series_size=365 * 5,
start_dt="2016-02-01",
days_period=1,
)
# Extract the normalized grouped datetime series data
training_data = generated_data.df
# Extract the names of the grouping columns that define the unique series data
grouping_key_columns = generated_data.key_columns
# Create a GroupedProphet model instance
grouped_model = GroupedProphet(n_changepoints=20,
uncertainty_samples=0).fit(
training_data, grouping_key_columns)
# Save the model to the local file system
save_path = "/tmp/grouped_prophet.gpm"
grouped_model.save(path="/tmp/grouped_prophet.gpm")
# Load the model from the local storage location
retrieved_model = GroupedProphet.load(save_path)
# Score the model and print the results
model_scores = retrieved_model.cross_validate_and_score(
horizon="30 days",
period="180 days",
initial="365 days",
parallel="threads",
rolling_window=0.05,
monthly=False,
)
print(f"Model scores:\n{model_scores.to_string()}")
# Run a forecast for each group
forecasts = retrieved_model.forecast(horizon=20, frequency="D")
print(f"Forecasted data:\n{forecasts[:50].to_string()}")
# Extract the parameters from each model for logging
params = retrieved_model.extract_model_params()
print(f"Model parameters:\n{params.to_string()}")
column_count=3,
series_count=10,
series_size=365 * 5,
start_dt="2016-02-01",
days_period=1,
)
# Extract the normalized grouped datetime series data
training_data = generated_data.df
# Extract the names of the grouping columns that define the unique series data
group_key_columns = generated_data.key_columns
# Create a GroupedProphet model instance
grouped_model = GroupedProphet(n_changepoints=20, uncertainty_samples=0).fit(
training_data, group_key_columns
)
# Get a subset of group keys to generate forecasts for
group_df = training_data.copy()
group_df["groups"] = list(zip(*[group_df[c] for c in group_key_columns]))
distinct_groups = group_df["groups"].unique()
groups_to_predict = list(distinct_groups[:3])
print("-" * 65)
print(f"\nUnique groups that have been modeled: \n{distinct_groups}\n")
print(f"Subset of groups to generate predictions for: \n{groups_to_predict}\n")
print("-" * 65)
forecasts = grouped_model.predict_groups(
groups=groups_to_predict,