def compute_test_accuracy(model):
test_words, test_tags = zip(*[zip(*sentence) for sentence in test_data])
test_words, test_tags = to_matrix(test_words, word_to_id), to_matrix(
test_tags, tag_to_id)
# predict tag probabilities of shape [batch,time,n_tags]
predicted_tag_probabilities = model.predict(test_words, verbose=1)
predicted_tags = predicted_tag_probabilities.argmax(axis=-1)
# compute accurary excluding padding
numerator = np.sum(
np.logical_and((predicted_tags == test_tags), (test_words != 0)))
denominator = np.sum(test_words != 0)
return float(numerator) / denominator
def update(self, dt):
"""Update the building template.
This method just applies the current position of the entity to the
renderable transform.
:param dt: Time delta from last update.
:type dt: float
"""
x, y = self.pos
m_x, m_y = to_matrix(x, y, self.scale_factor)
if not in_matrix(self.matrix, m_x, m_y) or not self.matrix[m_y][m_x]:
self.props.material.color = self.NON_BUILDABLE_COLOR
else:
self.props.material.color = self.BUILDABLE_COLOR
self.obj.position = to_scene(x, y)
self.obj.scale = Vec(1.05, 1.05, 1.05)
def update(self, dt):
"""Update the building template.
This method just applies the current position of the entity to the
renderable transform.
:param dt: Time delta from last update.
:type dt: float
"""
x, y = self.pos
m_x, m_y = to_matrix(x, y, self.scale_factor)
if not in_matrix(self.matrix, m_x, m_y) or not self.matrix[m_y][m_x]:
self[Renderable].node.params.update(self.NON_BUILDABLE_COLOR)
else:
self[Renderable].node.params.update(self.BUILDABLE_COLOR)
t = self[Renderable].transform
t.identity()
t.translate(to_scene(x, y))
t.scale(Vec(1.05, 1.05, 1.05))
def set_cell_walkable(evt):
"""Set a cell as walkable in the debug terrain."""
context = evt.context
x, y = to_matrix(evt.pos[0], evt.pos[1], context.scale_factor)
if in_matrix(context.matrix, x, y):
context.matrix[y][x] = True
def set_cell_walkable(evt):
"""Set a cell as walkable in the debug terrain."""
context = evt.context
x, y = to_matrix(evt.pos[0], evt.pos[1], context.scale_factor)
if in_matrix(context.matrix, x, y):
context.matrix[y][x] = True
#q_values[(state, action)] += stats.alpha(state, action) * td_error
for i in trajectory:
st, ac = i
q_values[(st, ac)] += alpha * td_error * stats.trace[(st, ac)]
stats.decay_trace(st, ac, GAMMA, LAMBDA)
state = new_state
action = new_action
print('win: %d, lose: %d, draw: %d' % (sum(win), sum(lose), sum(draw)))
print('win: %0.1f%%, lose: %0.1f%%, draw: %0.1f%%' \
%(sum(win)/EPISODES * 100, sum(lose)/EPISODES * 100, sum(draw)/EPISODES * 100))
print('usable ace: %0.2f%%' % (has_ace / EPISODES * 100))
q_value_matrix = to_matrix(q_values.dump_values())
fn = save_data(data=q_value_matrix \
, prefix='sarsa-lambda' + ('-biased' if use_bias_env else ''), iterations=EPISODES)
print('Saved %s' % fn)
x = np.arange(len(win))
plt.scatter(x, win, label='WIN', color='green', s=1)
plt.scatter(x, lose, label='LOSE', color='red', s=1)
plt.scatter(x, draw, label='DRAW', color='blue', s=1)
l = u'λ'
plt.title('SARSA(%s=%0.1f)' % (u'λ', LAMBDA))
plt.legend()
