s = env.reset()
while True:
a = policy.get_best_action(s)
s, r, d, info = env.step(a)
rewards[-1] += r
if env.env.env.env.env.was_real_done:
rewards.append(0)
break
if d:
s = env.reset()
return rewards
if __name__ == '__main__':
logger.configure(f'{C.env_id}/logs_{time_stamp}')
for k, v in C._asdict().items():
logger.record_tabular(k, v)
logger.dump_tabular()
max_reward = tf.placeholder(tf.float32, name='max_reward')
mean_reward = tf.placeholder(tf.float32, name='mean_reward')
max_summary = tf.summary.scalar('max_rew', max_reward)
mean_summary = tf.summary.scalar('mean_rew', mean_reward)
with create_session(0) as sess:
eval_env = make_atari(C.env_id, 113, 'eval')()
envs = SubprocVecEnv(
[make_atari(C.env_id, r + 1, 'train') for r in range(4)])
model = Model(eval_env.observation_space.shape,
eval_env.action_space.n)
runner = Runner(envs, model.policy, nb_rollout=C.nb_rollout)
sess.run(tf.global_variables_initializer())
