def preprocess_interactions_data_frame(
data_frame: pd.DataFrame, project_config: ProjectConfig
):
if len(data_frame) == 0:
return data_frame
data_frame[project_config.user_column.name] = data_frame[
project_config.user_column.name
].astype(str)
data_frame[project_config.item_column.name] = data_frame[
project_config.item_column.name
].astype(str)
literal_eval_array_columns(
data_frame,
[
project_config.user_column,
project_config.item_column,
project_config.output_column,
]
+ [input_column for input_column in project_config.other_input_columns]\
+ [input_column for input_column in project_config.auxiliar_output_columns],
)
#from IPython import embed; embed()
if project_config.available_arms_column_name and isinstance(
data_frame.iloc[0][project_config.available_arms_column_name], str
):
data_frame[project_config.available_arms_column_name] = parallel_literal_eval(
data_frame[project_config.available_arms_column_name]
)
return data_frame
def literal_eval_array_columns(data_frame: pd.DataFrame, columns: List[Column]):
for column in columns:
if (
column.type
in (IOType.FLOAT_ARRAY, IOType.INT_ARRAY, IOType.INDEXABLE_ARRAY)
and column.name in data_frame
):
data_frame[column.name] = parallel_literal_eval(data_frame[column.name])
def run(self):
os.makedirs(self.output().path)
# df: pd.DataFrame = preprocess_interactions_data_frame(
# pd.read_csv(
# get_test_set_predictions_path(self.model_training.output().path)
# ),
# self.model_training.project_config,
# ) # .sample(10000)
df: pd.DataFrame = pd.read_csv(
get_test_set_predictions_path(self.model_training.output().path),
dtype={self.model_training.project_config.item_column.name:
"str"}) # .sample(10000)
df["sorted_actions"] = parallel_literal_eval(df["sorted_actions"])
df["prob_actions"] = parallel_literal_eval(df["prob_actions"])
df["action_scores"] = parallel_literal_eval(df["action_scores"])
df["action"] = df["sorted_actions"].apply(
lambda sorted_actions: str(sorted_actions[0]))
with Pool(self.num_processes) as p:
print("Creating the relevance lists...")
# from IPython import embed; embed()
df["relevance_list"] = list(
tqdm(
p.starmap(
_create_relevance_list,
zip(
df["sorted_actions"],
df[self.model_training.project_config.item_column.
name],
df[self.model_training.project_config.
output_column.name],
),
),
total=len(df),
))
if self.model_training.metadata_data_frame is not None:
df = pd.merge(
df,
pd.read_csv(
self.model_training.metadata_data_frame_path,
dtype={
self.model_training.project_config.item_column.name:
"str"
}),
left_on="action",
right_on=self.model_training.project_config.item_column.name,
suffixes=("", "_action"),
)
ground_truth_df = df[~(
df[self.model_training.project_config.output_column.name] == 0)]
print("Rank Metrics...")
df_rank, dict_rank = self.rank_metrics(ground_truth_df)
gc.collect()
print("Fairness Metrics")
df_fairness, df_fairness_metrics = self.fairness_metrics(
ground_truth_df)
gc.collect()
print("Offpolice Metrics")
df_offpolicy, dict_offpolice = self.offpolice_metrics(df)
gc.collect()
# dict_offpolice = {}
# Save Logs
metrics = {**dict_rank, **dict_offpolice}
pprint.pprint(metrics)
with open(os.path.join(self.output().path, "metrics.json"),
"w") as metrics_file:
json.dump(metrics, metrics_file, cls=JsonEncoder, indent=4)
df_offpolicy.to_csv(os.path.join(self.output().path,
"df_offpolicy.csv"),
index=False)
df_rank.to_csv(os.path.join(self.output().path, "rank_metrics.csv"),
index=False)
df_fairness_metrics.to_csv(os.path.join(self.output().path,
"fairness_metrics.csv"),
index=False)
df_fairness.to_csv(os.path.join(self.output().path, "fairness_df.csv"),
index=False)