def get_reward(mean_sample_vote, sound_extended, i_target, speaker, loudness_factor, softmax=False):
if softmax:
reward = exp(mean_sample_vote[i_target])
other_exp = 0.0
for i in xrange(len(mean_sample_vote)):
if i!=i_target:
other_exp += exp(mean_sample_vote[i])
reward /= reward + other_exp
else:
reward = mean_sample_vote[i_target]
for i in xrange(len(mean_sample_vote)):
if i!=i_target:
reward -= mean_sample_vote[i]
if speaker == 'adult':
target_loudness = [72.77, 65.20, 66.04, 68.37, 68.47]
else:
target_loudness = [73.78, 68.68, 69.78]
level = float(sound_extended.level)
if isinf(level):
level = 0.0
loudness_reward = level - target_loudness[i_target]
if loudness_reward > 0.0:
loudness_reward = 0.0
reward += loudness_factor * loudness_reward
return reward
def get_confidences(self,mean_sample_vote): n_classes = len(mean_sample_vote) confidences = np.zeros(n_classes) norm_sum = 0.0 for i in xrange(n_classes): confidence_i = exp(mean_sample_vote[i]) confidences[i] = confidence_i norm_sum += confidence_i confidences /= norm_sum return confidences
def get_confidences(mean_sample_vote):
n_classes = len(mean_sample_vote)
confidences = np.zeros(n_classes)
norm_sum = 0.0
for i in xrange(n_classes):
confidence_i = exp(mean_sample_vote[i])
confidences[i] = confidence_i
norm_sum += confidence_i
confidences /= norm_sum
return confidences
def get_reward(mean_sample_vote,
sound_extended,
i_target,
speaker,
loudness_factor,
softmax=False):
if softmax:
reward = exp(mean_sample_vote[i_target])
other_exp = 0.0
for i in xrange(len(mean_sample_vote)):
if i != i_target:
other_exp += exp(mean_sample_vote[i])
reward /= reward + other_exp
else:
reward = mean_sample_vote[i_target]
for i in xrange(len(mean_sample_vote)):
if i != i_target:
reward -= mean_sample_vote[i]
if speaker == 'adult':
target_loudness = [72.77, 65.20, 66.04, 68.37, 68.47]
else:
target_loudness = [73.78, 68.68, 69.78]
level = float(sound_extended.level)
if isinf(level):
level = 0.0
loudness_reward = level - target_loudness[i_target]
if loudness_reward > 0.0:
loudness_reward = 0.0
reward += loudness_factor * loudness_reward
return reward
