class PolicyEvaluator(object):
def __init__(self, architecture_file, model_file):
# Initialize objects
self.net = BaristaNet(architecture_file, model_file, None)
self.batch_size = self.net.batch_size
game = SnakeGame()
preprocessor = generate_preprocessor(self.net.state.shape[2:],
gray_scale)
self.engines = [ExpGain(self.net, ['w', 'a', 's', 'd'],
preprocessor, game.cpu_play,
None, game.encode_state())
for _ in range(self.batch_size)]
def evaluate(self, model, num_trials):
""" Runs |num_trials| games and returns average score. """
for eg in self.engines:
set_net_params(eg.net.net, model)
eg.reset_game()
total_score = 0
trials_completed = 0
scores = [0] * self.batch_size
while trials_completed < num_trials:
states = [eg.get_preprocessed_state() for eg in self.engines]
actions = self.net.select_action(states,
batch_size=self.batch_size)
for i, (action, eg) in enumerate(zip(actions, self.engines)):
scores[i] += eg.play_action(eg.actions[action])
if eg.game_over:
total_score += scores[i]
trials_completed += 1
if trials_completed == num_trials:
break
eg.reset_game()
scores[i] = 0
return float(total_score)/num_trials
