def plot_feature_importance(df: pd.DataFrame, title: str) -> AxPlotConfig:
if df.empty:
raise NoDataError("No Data on Feature Importances found.")
df.set_index(df.columns[0], inplace=True)
data = [
go.Bar(y=df.index,
x=df[column_name],
name=column_name,
orientation="h") for column_name in df.columns
]
fig = subplots.make_subplots(
rows=len(df.columns),
cols=1,
subplot_titles=df.columns,
print_grid=False,
shared_xaxes=True,
)
for idx, item in enumerate(data):
fig.append_trace(item, idx + 1, 1)
fig.layout.showlegend = False
fig.layout.margin = go.layout.Margin(
l=8 * min(max(len(idx) for idx in df.index), 75) # noqa E741
)
fig.layout.title = title
return AxPlotConfig(data=fig, plot_type=AxPlotTypes.GENERIC)
def _set_current_model(self, data: Optional[Data]) -> None:
"""Instantiate the current model with all available data.
"""
model_kwargs = self._curr.model_kwargs or {}
# If last generator run's index matches the current step, extract
# model state from last generator run and pass it to the model
# being instantiated in this function.
lgr = self.last_generator_run
if (
lgr is not None
and lgr._generation_step_index == self._curr.index
and lgr._model_state_after_gen
):
model_kwargs = _combine_model_kwargs_and_state(
model_kwargs=model_kwargs,
generator_run=lgr,
model_class=not_none(not_none(self.model).model.__class__),
)
if data is None:
if self._curr.use_update:
# If the new step is using `update`, it's important to instantiate
# the model with data for completed trials only, so later we can
# update it with data for new trials as they become completed.
# `experiment.fetch_data` can fetch all available data, including
# for non-completed trials (depending on how the experiment's metrics
# implement `fetch_experiment_data`). We avoid fetching data for
# trials with statuses other than `COMPLETED`, by fetching specifically
# for `COMPLETED` trials.
data = self.experiment.fetch_trials_data(
self.experiment.trial_indices_by_status[TrialStatus.COMPLETED]
)
else:
data = self.experiment.fetch_data()
# By the time we get here, we will have already transitioned
# to a new step, but if previou step required observed data,
# we should raise an error even if enough trials were completed.
# Such an empty data case does indicate an invalid state; this
# check is to improve the experience of detecting and debugging
# the invalid state that led to this.
previous_step_required_observations = (
self._curr.index > 0
and self._steps[self._curr.index - 1].min_trials_observed > 0
)
if data.df.empty and previous_step_required_observations:
raise NoDataError(
f"Observed data is required for generation step #{self._curr.index} "
f"(model {self._curr.model_name}), but fetched data was empty. "
"Something is wrong with experiment setup -- likely metrics do not "
"implement fetching logic (check your metrics) or no data was "
"attached to experiment for completed trials."
)
if isinstance(self._curr.model, Models):
self._set_current_model_from_models_enum(data=data, **model_kwargs)
else:
# If model was not specified as Models member, it was specified as a
# factory function.
self._set_current_model_from_factory_function(data=data, **model_kwargs)
def _get_data_for_fit(self, passed_in_data: Optional[Data]) -> Data:
if passed_in_data is None:
if self._curr.use_update:
# If the new step is using `update`, it's important to instantiate
# the model with data for completed trials only, so later we can
# update it with data for new trials as they become completed.
# `experiment.lookup_data` can lookup all available data, including
# for non-completed trials (depending on how the experiment's metrics
# implement `fetch_experiment_data`). We avoid fetching data for
# trials with statuses other than `COMPLETED`, by fetching specifically
# for `COMPLETED` trials.
avail_while_running_metrics = {
m.name
for m in self.experiment.metrics.values()
if m.is_available_while_running()
}
if avail_while_running_metrics:
raise NotImplementedError(
f"Metrics {avail_while_running_metrics} are available while "
"trial is running, but use of `update` functionality in "
"generation strategy relies on new data being available upon "
"trial completion.")
data = self.experiment.lookup_data(
trial_indices=self.experiment.trial_indices_by_status[
TrialStatus.COMPLETED])
else:
data = self.experiment.lookup_data()
else:
data = passed_in_data
# By the time we get here, we will have already transitioned
# to a new step, but if previous step required observed data,
# we should raise an error even if enough trials were completed.
# Such an empty data case does indicate an invalid state; this
# check is to improve the experience of detecting and debugging
# the invalid state that led to this.
previous_step_required_observations = (
self._curr.index > 0
and self._steps[self._curr.index - 1].min_trials_observed > 0)
if data.df.empty and previous_step_required_observations:
raise NoDataError(
f"Observed data is required for generation step #{self._curr.index} "
f"(model {self._curr.model_name}), but fetched data was empty. "
"Something is wrong with experiment setup -- likely metrics do not "
"implement fetching logic (check your metrics) or no data was "
"attached to experiment for completed trials.")
return data
def _set_current_model(self, data: Optional[Data]) -> None:
"""Instantiate the current model with all available data."""
model_kwargs = self._curr.model_kwargs or {}
# If last generator run's index matches the current step, extract
# model state from last generator run and pass it to the model
# being instantiated in this function.
lgr = self.last_generator_run
if (lgr is not None and lgr._generation_step_index == self._curr.index
and lgr._model_state_after_gen):
model_kwargs = _combine_model_kwargs_and_state(
model_kwargs=model_kwargs,
generator_run=lgr,
model_class=not_none(not_none(self.model).model.__class__),
)
if data is None:
if self._curr.use_update:
# If the new step is using `update`, it's important to instantiate
# the model with data for completed trials only, so later we can
# update it with data for new trials as they become completed.
# `experiment.lookup_data` can lookup all available data, including
# for non-completed trials (depending on how the experiment's metrics
# implement `fetch_experiment_data`). We avoid fetching data for
# trials with statuses other than `COMPLETED`, by fetching specifically
# for `COMPLETED` trials.
avail_while_running_metrics = {
m.name
for m in self.experiment.metrics.values()
if m.is_available_while_running()
}
if avail_while_running_metrics:
raise NotImplementedError(
f"Metrics {avail_while_running_metrics} are available while "
"trial is running, but use of `update` functionality in "
"generation strategy relies on new data being available upon "
"trial completion.")
data = self.experiment.lookup_data(
trial_indices=self.experiment.trial_indices_by_status[
TrialStatus.COMPLETED])
else:
data = self.experiment.lookup_data()
# By the time we get here, we will have already transitioned
# to a new step, but if previou step required observed data,
# we should raise an error even if enough trials were completed.
# Such an empty data case does indicate an invalid state; this
# check is to improve the experience of detecting and debugging
# the invalid state that led to this.
previous_step_required_observations = (
self._curr.index > 0
and self._steps[self._curr.index - 1].min_trials_observed > 0)
if data.df.empty and previous_step_required_observations:
raise NoDataError(
f"Observed data is required for generation step #{self._curr.index} "
f"(model {self._curr.model_name}), but fetched data was empty. "
"Something is wrong with experiment setup -- likely metrics do not "
"implement fetching logic (check your metrics) or no data was "
"attached to experiment for completed trials.")
if not data.df.empty:
trial_indices_in_data = sorted(data.df["trial_index"].unique())
logger.debug(
f"Setting model with data for trials: {trial_indices_in_data}")
# TODO(jej)[T87591836] Support non-`Data` data types.
if isinstance(self._curr.model, ModelRegistryBase):
# pyre-fixme [6]: Incompat param: Expect `Data` got `AbstractDataFrameData`
self._set_current_model_from_models_enum(data=data, **model_kwargs)
else:
# If model was not specified as Models member, it was specified as a
# factory function.
# pyre-fixme [6]: Incompat param: Expect `Data` got `AbstractDataFrameData`
self._set_current_model_from_factory_function(data=data,
**model_kwargs)