feature: faster calc_pscore

[usaito]

feature

• implemented a faster calc_pscore function _calc_pscore_given_policy_logit (instead of the previous _calc_pscore_given_action_list)
  • In the previous implementation, _calc_pscore_given_action_list( calculates the pscore of given policy_logit_i_ and action_list, meaning that we had to iterate all the rounds and combinatorial actions to calculate the pscores.
  • In fact, we can calculate the pscore of all possible slates at the same time, and this will make some functions, such as sample_action_and_obtain_pscore( and calc_ground_truth_policy_value(, much faster

before

zr-obp/obp/dataset/synthetic_slate.py

Lines 324 to 347 in 6dc904c

	def _calc_pscore_given_action_list(
	self, action_list: List[int], policy_logit_i_: np.ndarray
	) -> float:
	"""Calculate the propensity score given combinatorial set of actions.
	Parameters
	------------
	action_list: List[int], len=len_list
	List of combinatorial set of slate actions.
	policy_logit_i_: array-like, (n_unique_action, )
	Logit values given context (:math:`x`), i.e., :math:`\\f: \\mathcal{X} \\rightarrow \\mathbb{R}^{\\mathcal{A}}`.
	"""
	unique_action_set = np.arange(self.n_unique_action)
	pscore_ = 1.0
	for action in action_list:
	score_ = softmax(policy_logit_i_[:, unique_action_set])[0]
	action_index = np.where(unique_action_set == action)[0][0]
	pscore_ *= score_[action_index]
	unique_action_set = np.delete(
	unique_action_set, unique_action_set == action
	)
	return pscore_

after

zr-obp/obp/dataset/synthetic_slate.py

Lines 324 to 361 in 1187a89

	def _calc_pscore_given_policy_logit(
	self, all_slate_actions: np.ndarray, policy_logit_i_: np.ndarray
	) -> np.ndarray:
	"""Calculate the propensity score of each of the possible slate actions given policy_logit.
	Parameters
	------------
	all_slate_actions: array-like, (n_action, len_list)
	All possible slate actions.
	policy_logit_i_: array-like, (n_unique_action, )
	Logit values given context (:math:`x`), i.e., :math:`\\f: \\mathcal{X} \\rightarrow \\mathbb{R}^{\\mathcal{A}}`.
	Returns
	------------
	pscores: array-like, (n_action, )
	Propensity scores of all the possible slate actions given policy_logit.
	"""
	n_actions = len(all_slate_actions)
	unique_action_set_2d = np.tile(np.arange(self.n_unique_action), (n_actions, 1))
	pscores = np.ones(n_actions)
	for position_ in np.arange(self.len_list):
	action_index = np.where(
	unique_action_set_2d == all_slate_actions[:, position_][:, np.newaxis]
	)[1]
	pscores *= softmax(policy_logit_i_[unique_action_set_2d])[
	np.arange(n_actions), action_index
	]
	# delete actions
	if position_ + 1 != self.len_list:
	mask = np.ones((n_actions, self.n_unique_action - position_))
	mask[np.arange(n_actions), action_index] = 0
	unique_action_set_2d = unique_action_set_2d[mask.astype(bool)].reshape(
	(-1, self.n_unique_action - position_ - 1)
	)
	return pscores