def score_documents(user_obs,
doc_obs,
no_click_mass=1.0,
is_mnl=False,
min_normalizer=-1.0):
"""Computes unnormalized scores given both user and document observations.
Similar to score_documents_tf but works on NumPy objects.
Args:
user_obs: An instance of AbstractUserState.
doc_obs: A numpy array that represents the observation of all documents in
the candidate set.
no_click_mass: a float indicating the mass given to a no click option
is_mnl: whether to use a multinomial logit model instead of a multinomial
proportional model.
min_normalizer: A float (<= 0) used to offset the scores to be positive when
using multinomial proportional model.
Returns:
A float array that stores unnormalzied scores of documents and a float
number that represents the score for the action of picking no document.
"""
scores = np.array([])
for doc in doc_obs:
scores = np.append(scores, np.dot(user_obs, doc))
all_scores = np.append(scores, no_click_mass)
if is_mnl:
all_scores = choice_model.softmax(all_scores)
else:
all_scores = all_scores - min_normalizer
assert not all_scores[
all_scores < 0.0], 'Normalized scores have non-positive elements.'
return all_scores[:-1], all_scores[-1]
def test_softmax_negative_floats(self):
self.assertAllClose(
choice_model.softmax(-1.0 * np.log(np.arange(1, 5))),
np.array([0.48, 0.24, 0.16, 0.12]))
def test_softmax_equal_ints(self):
self.assertAllClose(choice_model.softmax(np.ones(4)),
np.array([0.25, 0.25, 0.25, 0.25]))
def test_softmax_positive_floats(self):
self.assertAllClose(choice_model.softmax(np.log(np.arange(1, 5))),
np.array([0.1, 0.2, 0.3, 0.4]))
def test_softmax_single_int(self):
self.assertAllClose(choice_model.softmax([0]), [1.0])
