def play(env, agent, config, evaluation=False):
if evaluation:
print('\nEvaluation:')
mean_rewards = 0
max_reward = 0
max_mean_rewards = 0
num_episodes = config['eval_episodes'] if evaluation else config['train_episodes']
for episode in tqdm.tqdm(range(num_episodes)):
observations, infos = env.reset()
infos_array = dict_to_array(infos, config['environment_batch_size'])
rewards = [0] * config['environment_batch_size']
dones = [False] * config['environment_batch_size']
max_score = max([info['max_score'] for info in infos_array])
steps = 0
# TODO: maybe condition on max_steps as well.
while not all(dones):
win_factor = max_mean_rewards / float(max_score) if config['use_adaptive_epsilon'] else None
actions = agent.choose_actions(observations, infos_array, dones, evaluation,
win_factor=win_factor)
new_observations, new_rewards, new_dones, new_infos = env.step(actions)
new_infos_array = dict_to_array(new_infos, config['environment_batch_size'])
for idx, done in enumerate(dones):
if not done and not evaluation:
agent.add_state(observations[idx],
infos_array[idx],
actions[idx],
new_observations[idx],
new_infos_array[idx],
new_rewards[idx] - rewards[idx],
dones[idx])
observations = new_observations
infos_array = new_infos_array
rewards = new_rewards
dones = new_dones
if not evaluation and steps % config['update_frequency'] == 0:
agent.train()
steps += 1
mean_rewards = np.mean(rewards)
max_reward = max(max_reward, max(rewards))
max_mean_rewards = max(max_mean_rewards, np.mean(rewards))
wins_percentage = sum([info['has_won'] for info in infos_array]) * 100. / len(infos_array)
print('Mean rewards: {}({}), steps: {}, max reward: {}({}), wins percentage - {}'.format(
mean_rewards, max_mean_rewards, steps, max_reward, max_score, wins_percentage))
agent.end_episode()
if not evaluation and mean_rewards > 0.99 * max_score:
break
return
def __group_by_field(self, i, field):
'''Group a test index subset i by field (SNP=0, sample=1).'''
size = self.problem.genotype.data.shape[field]
group_count = util.dict_to_array(statutil.group_by_value(self.test_index[field][i]))
result = np.zeros((size,), dtype=int)
result[group_count['k']] = group_count['v']
return result
def __group_by_field(self, i, field):
'''Group a test index subset i by field (SNP=0, sample=1).'''
size = self.problem.genotype.data.shape[field]
group_count = util.dict_to_array(
statutil.group_by_value(self.test_index[field][i]))
result = np.zeros((size, ), dtype=int)
result[group_count['k']] = group_count['v']
return result
def __group_by_field(g, i, field):
'''Group a test index subset i by field (SNP=0, sample=1).'''
group_count = util.dict_to_array(statutil.group_by_value(i[field]))
result = np.zeros((g.shape[field],), dtype=int)
result[group_count['k']] = group_count['v']
return result
def __group_by_field(g, i, field):
'''Group a test index subset i by field (SNP=0, sample=1).'''
group_count = util.dict_to_array(statutil.group_by_value(i[field]))
result = np.zeros((g.shape[field], ), dtype=int)
result[group_count['k']] = group_count['v']
return result
