def run_agent(player: MarioPlayer, env: Wrapper, record: bool, vids_path: str,
index):
if record:
rec_output_path = os.path.join(vids_path,
"{name}.mp4".format(name=index))
rec = monitor.video_recorder.VideoRecorder(env, path=rec_output_path)
state = env.reset()
done = False
for step in range(steps_limit):
if done:
break
action = player.act(state)
state, reward, done, info = env.step(action)
env.render()
if record:
rec.capture_frame()
player.update_info(info)
player.update_reward(reward)
if info['flag_get']: # if got to the flag - run is ended.
done = True
if record:
rec.close()
player.calculate_fitness()
outcome = player.get_run_info()
outcome['index'] = index
return outcome
def _test(id: int, env: gym.Wrapper, model: TD3Network, render: bool = False, recording_path=None,
save_video=False):
episode_rewards = []
action_repeats = []
state = env.reset()
done = False
episode_images = []
while not done:
# get action
state = torch.FloatTensor(state).unsqueeze(0)
action = model.actor(state)
repeat_q = model.critic_1(state, action)
repeat_idx = repeat_q.argmax(1).item()
action = action.data.cpu().numpy()[0]
repeat = model.action_repeats[repeat_idx]
action_repeats.append(repeat)
for _ in range(repeat):
if render:
if save_video:
img = env.render(mode='rgb_array')
episode_images.append(img)
else:
env.render(mode='human')
# step
state, reward, done, info = env.step(action)
episode_rewards.append(reward)
if done:
break
if render and save_video:
write_gif(episode_images, action_repeats, episode_rewards,
os.path.join(recording_path, 'ep_{}.gif'.format(id)))
return sum(episode_rewards), action_repeats