def test_reward_matrix():
matrix = reward_matrix('abc$', 'abc$', 'abc$', eos_label=3)
should_be = numpy.array([[0, -1, -1, -3], [-1, 0, -1, -2], [-1, -1, 0, -1],
[-1, -1, -1, 0], [-1, -1, -1, -1]])
assert_equal(matrix, should_be)
matrix = reward_matrix('abc$', 'acb$', 'abc$', eos_label=3)
should_be = numpy.array([[0, -1, -1, -3], [-1, 0, -1, -2],
[-2, -1, -1, -1], [-2, -2, -1, -2],
[-3, -3, -2, -2]])
assert_equal(matrix, should_be)
def test_reward_matrix():
matrix = reward_matrix('abc$', 'abc$', 'abc$', eos_label=3)
should_be = numpy.array([[ 0, -1, -1, -3],
[-1, 0, -1, -2],
[-1, -1, 0, -1],
[-1, -1, -1, 0],
[-1, -1, -1, -1]])
assert_equal(matrix, should_be)
matrix = reward_matrix('abc$', 'acb$', 'abc$', eos_label=3)
should_be = numpy.array([[ 0, -1, -1, -3],
[-1, 0, -1, -2],
[-2, -1, -1, -1],
[-2, -2, -1, -2],
[-3, -3, -2, -2]])
assert_equal(matrix, should_be)
def perform(self, node, inputs, output_storage):
groundtruth, recognized = inputs
if (groundtruth.ndim != 2 or recognized.ndim != 2
or groundtruth.shape[1] != recognized.shape[1]):
raise ValueError
batch_size = groundtruth.shape[1]
all_rewards = numpy.zeros(recognized.shape + (self.alphabet_size, ),
dtype='int64')
all_gains = numpy.zeros(recognized.shape + (self.alphabet_size, ),
dtype='int64')
alphabet = list(range(self.alphabet_size))
for index in range(batch_size):
y = list(groundtruth[:, index])
y_hat = list(recognized[:, index])
try:
eos_pos = y.index(self.eos_label)
y = y[:eos_pos + 1]
except:
# Sometimes groundtruth is in fact also a prediction
# and in this case it might not have EOS label
pass
if self.eos_label in y_hat:
y_hat_eos_pos = y_hat.index(self.eos_label)
y_hat_trunc = y_hat[:y_hat_eos_pos + 1]
else:
y_hat_trunc = y_hat
rewards_trunc = reward_matrix(y, y_hat_trunc, alphabet,
self.eos_label)
# pass freshly computed rewards to gain_matrix to speed things up
# a bit
gains_trunc = gain_matrix(y,
y_hat_trunc,
alphabet,
given_reward_matrix=rewards_trunc)
gains = numpy.ones((len(y_hat), len(alphabet))) * -1000
gains[:(gains_trunc.shape[0] - 1), :] = gains_trunc[:-1, :]
rewards = numpy.ones((len(y_hat), len(alphabet))) * -1
rewards[:(rewards_trunc.shape[0] - 1), :] = rewards_trunc[:-1, :]
all_rewards[:, index, :] = rewards
all_gains[:, index, :] = gains
output_storage[0][0] = all_rewards
output_storage[1][0] = all_gains
def perform(self, node, inputs, output_storage):
groundtruth, recognized = inputs
if (groundtruth.ndim != 2 or recognized.ndim != 2
or groundtruth.shape[1] != recognized.shape[1]):
raise ValueError
batch_size = groundtruth.shape[1]
all_rewards = numpy.zeros(
recognized.shape + (self.alphabet_size,), dtype='int64')
all_gains = numpy.zeros(
recognized.shape + (self.alphabet_size,), dtype='int64')
alphabet = list(range(self.alphabet_size))
for index in range(batch_size):
y = list(groundtruth[:, index])
y_hat = list(recognized[:, index])
try:
eos_pos = y.index(self.eos_label)
y = y[:eos_pos + 1]
except:
# Sometimes groundtruth is in fact also a prediction
# and in this case it might not have EOS label
pass
if self.eos_label in y_hat:
y_hat_eos_pos = y_hat.index(self.eos_label)
y_hat_trunc = y_hat[:y_hat_eos_pos + 1]
else:
y_hat_trunc = y_hat
rewards_trunc = reward_matrix(
y, y_hat_trunc, alphabet, self.eos_label)
# pass freshly computed rewards to gain_matrix to speed things up
# a bit
gains_trunc = gain_matrix(y, y_hat_trunc, alphabet,
given_reward_matrix=rewards_trunc)
gains = numpy.ones((len(y_hat), len(alphabet))) * -1000
gains[:(gains_trunc.shape[0] - 1), :] = gains_trunc[:-1, :]
rewards = numpy.ones((len(y_hat), len(alphabet))) * -1
rewards[:(rewards_trunc.shape[0] - 1), :] = rewards_trunc[:-1, :]
all_rewards[:, index, :] = rewards
all_gains[:, index, :] = gains
output_storage[0][0] = all_rewards
output_storage[1][0] = all_gains
