def _save_test_set_predictions(self, agent: BanditAgent) -> None:
print("Saving test set predictions...")
obs: List[Dict[str, Any]] = self.test_data_frame.to_dict("records")
self.clean()
for ob in tqdm(obs, total=len(obs)):
if self.embeddings_for_metadata is not None:
ob[ITEM_METADATA_KEY] = self.embeddings_for_metadata
else:
ob[ITEM_METADATA_KEY] = None
if (self.project_config.available_arms_column_name
and len(ob[self.project_config.available_arms_column_name]) == 0):
ob[self.project_config.available_arms_column_name] = [
ob[self.project_config.item_column.name]
]
(
arm_contexts_list,
arms_list,
arm_indices_list,
arm_scores_list,
) = self._prepare_for_agent(agent, obs)
sorted_actions_list = []
proba_actions_list = []
for arm_contexts, arms, arm_indices, arm_scores in tqdm(
zip(arm_contexts_list, arms_list, arm_indices_list, arm_scores_list),
total=len(arm_contexts_list),
):
sorted_actions, proba_actions = agent.rank(
arms, arm_indices, arm_contexts, arm_scores
)
sorted_actions_list.append(sorted_actions)
proba_actions_list.append(proba_actions)
action_scores_list = [
list(reversed(sorted(action_scores))) for action_scores in arm_scores_list
]
del obs
# Create evaluation file
df = pd.read_csv(self.test_data_frame_path)
df["sorted_actions"] = sorted_actions_list
df["prob_actions"] = proba_actions_list
df["action_scores"] = action_scores_list
# join with train interaction information
df_train = self.get_data_frame_interactions()
df_train['trained'] = 1
df = df.merge(df_train, on = [self.project_config.user_column.name, self.project_config.item_column.name], how='left')
df['trained'] = df['trained'].fillna(0)
self.plot_scores([score for arm_scores in arm_scores_list for score in arm_scores])
self._to_csv_test_set_predictions(df)
def _act(self, agent: BanditAgent, ob: dict) -> int:
(
arm_contexts_list,
_,
arm_indices_list,
arm_scores_list,
) = self._prepare_for_agent(agent, [ob])
return agent.act(arm_indices_list[0], arm_contexts_list[0], arm_scores_list[0])
def create_agent(self) -> BanditAgent:
bandit_class = load_attr(self.bandit_policy_class, Type[BanditPolicy])
bandit = bandit_class(reward_model=self.get_trained_module(),
**self.bandit_policy_params)
return BanditAgent(bandit)
def _save_test_set_predictions(self, agent: BanditAgent) -> None:
print("Saving test set predictions...")
if self.sample_size_eval and len(
self.test_data_frame) > self.sample_size_eval:
obs: List[Dict[str, Any]] = self.test_data_frame.sample(
self.sample_size_eval,
random_state=self.seed).to_dict("records")
else:
obs: List[Dict[str, Any]] = self.test_data_frame.to_dict("records")
for ob in tqdm(obs, total=len(obs)):
if self.embeddings_for_metadata is not None:
ob[ITEM_METADATA_KEY] = self.embeddings_for_metadata
else:
ob[ITEM_METADATA_KEY] = None
print("...prepare_for_agent")
(
arm_contexts_list,
arms_list,
arm_indices_list,
arm_scores_list,
) = self._prepare_for_agent(agent, obs)
print("...")
sorted_actions_list = []
proba_actions_list = []
for arm_contexts, arms, arm_indices, arm_scores in tqdm(
zip(arm_contexts_list, arms_list, arm_indices_list,
arm_scores_list),
total=len(arm_contexts_list),
):
sorted_actions, proba_actions = agent.rank(arms, arm_indices,
arm_contexts,
arm_scores)
sorted_actions_list.append(sorted_actions)
proba_actions_list.append(proba_actions)
action_scores_list = [
list(reversed(sorted(action_scores)))
for action_scores in arm_scores_list
]
del obs
# Create evaluation file
df = pd.read_csv(self.test_data_frame_path)
if self.sample_size_eval and len(
self.test_data_frame) > self.sample_size_eval:
df = df.sample(self.sample_size_eval, random_state=self.seed)
df["sorted_actions"] = sorted_actions_list
df["prob_actions"] = proba_actions_list
df["action_scores"] = action_scores_list
# join with train interaction information
df_train = self.get_data_frame_interactions()[[
self.project_config.user_column.name,
self.project_config.item_column.name
]]
df_train['trained'] = 1
df = df.merge(df_train,
on=[
self.project_config.user_column.name,
self.project_config.item_column.name
],
how='left')
df['trained'] = df['trained'].fillna(0)
# Add indexed information
df['item_indexed'] = df[self.project_config.item_column.name].apply(
lambda i: self.index_mapping[self.project_config.item_column.name][
str(i)] > 0)
scores = [
score for arm_scores in arm_scores_list for score in arm_scores
]
self.plot_scores(scores)
self._to_csv_test_set_predictions(df)