def probs_metric(inverse=False):
rand_p = Vec2(random()*table.width+table.min_point.x, random()*table.height+table.min_point.y)
try:
bestmeaning, bestsentence = generate_sentence(rand_p, False, scene, speaker, usebest=True, golden=inverse, printing=printing)
sampled_landmark, sampled_relation = bestmeaning.args[0], bestmeaning.args[3]
golden_posteriors = get_all_sentence_posteriors(bestsentence, meanings, golden=(not inverse), printing=printing)
# lmk_prior = speaker.get_landmark_probability(sampled_landmark, landmarks, PointRepresentation(rand_p))[0]
all_lmk_probs = speaker.all_landmark_probs(landmarks, Landmark(None, PointRepresentation(rand_p), None))
all_lmk_probs = dict(zip(landmarks, all_lmk_probs))
lmk_prior = all_lmk_probs[sampled_landmark]
head_on = speaker.get_head_on_viewpoint(sampled_landmark)
rel_prior = speaker.get_probabilities_points( np.array([rand_p]), sampled_relation, head_on, sampled_landmark)
lmk_post = golden_posteriors[sampled_landmark]
rel_post = golden_posteriors[sampled_relation]
ps = np.array([golden_posteriors[lmk]*golden_posteriors[rel] for lmk, rel in meanings])
rank = None
for i,p in enumerate(ps):
lmk,rel = meanings[i]
# logger( '%f, %s' % (p, m2s(lmk,rel)))
head_on = speaker.get_head_on_viewpoint(lmk)
# ps[i] *= speaker.get_landmark_probability(lmk, landmarks, PointRepresentation(rand_p))[0]
ps[i] *= all_lmk_probs[lmk]
ps[i] *= speaker.get_probabilities_points( np.array([rand_p]), rel, head_on, lmk)
if lmk == sampled_landmark and rel == sampled_relation:
idx = i
ps += epsilon
ps = ps/ps.sum()
prob = ps[idx]
rank = sorted(ps, reverse=True).index(prob)
entropy = entropy_of_probs(ps)
except (ParseError,RuntimeError) as e:
logger( e )
lmk_prior = 0
rel_prior = 0
lmk_post = 0
rel_post = 0
prob = 0
rank = len(meanings)-1
entropy = 0
distances = [[None]]
head_on = speaker.get_head_on_viewpoint(sampled_landmark)
all_descs = speaker.get_all_meaning_descriptions(trajector, scene, sampled_landmark, sampled_relation, head_on, 1)
distances = []
for desc in all_descs:
distances.append([edit_distance( bestsentence, desc ), desc])
distances.sort()
return lmk_prior,rel_prior,lmk_post,rel_post,\
prob,entropy,rank,distances[0][0],type(sampled_relation)
def probs_metric():
meaning, sentence = generate_sentence(rand_p, consistent, scene, speaker, usebest=True, printing=printing)
sampled_landmark,sampled_relation = meaning.args[0],meaning.args[3]
print meaning.args[0],meaning.args[3], len(sentence)
if sentence == "":
prob = 0
entropy = 0
else:
logger( 'Generated sentence: %s' % sentence)
try:
golden_posteriors = get_all_sentence_posteriors(sentence, meanings, golden=True, printing=printing)
epsilon = 1e-15
ps = np.array([golden_posteriors[lmk]*golden_posteriors[rel] for lmk, rel in meanings])
temp = None
for i,p in enumerate(ps):
lmk,rel = meanings[i]
# logger( '%f, %s' % (p, m2s(lmk,rel)))
head_on = speaker.get_head_on_viewpoint(lmk)
ps[i] *= speaker.get_landmark_probability(lmk, landmarks, PointRepresentation(rand_p))[0]
ps[i] *= speaker.get_probabilities_points( np.array([rand_p]), rel, head_on, lmk)
if lmk == meaning.args[0] and rel == meaning.args[3]:
temp = i
ps += epsilon
ps = ps/ps.sum()
prob = ps[temp]
rank = sorted(ps, reverse=True).index(prob)
entropy = entropy_of_probs(ps)
except ParseError as e:
logger( e )
prob = 0
rank = len(meanings)-1
entropy = 0
head_on = speaker.get_head_on_viewpoint(sampled_landmark)
all_descs = speaker.get_all_meaning_descriptions(trajector, scene, sampled_landmark, sampled_relation, head_on, 1)
distances = []
for desc in all_descs:
distances.append([edit_distance( sentence, desc ), desc])
distances.sort()
return prob,entropy,rank,distances[0][0]
