def train(self, core_id, config_data, start_date):
start = time.time()
cfg = Helpers.prepare_config(config_data)
walls, seeker, hiding, width, height = Helpers.prepare_map(cfg)
env, step_img_path, fps_batch, render_mode, wins_l = Helpers.create_env(
config=cfg,
width=width,
height=height,
walls=walls,
seeker=seeker,
hiding=hiding,
start_date=start_date,
core_id=core_id,
)
AGENTS = 2
algorithm = Helpers.pick_algorithm(cfg, env=env, agents=AGENTS)
algorithm.prepare_model()
for i in range(cfg['game']['episodes']):
algorithm.before_episode()
metadata = Helpers.update_celery_metadata(
core_id=core_id,
curr=i + 1,
total=cfg['game']['episodes'],
ep_iter=cfg['game']['duration'],
fps=None,
itera=0,
iter_perc=0,
time_elap=round(time.time() - start),
img_path=step_img_path,
eta=None,
rewards=[0, 0],
wins=[0, 0],
wins_moving=[0, 0],
)
self.update_state(state='PROGRESS', meta=metadata)
obs_n, reward_n, rewards_ep, done, fps_episode = Helpers.new_ep(env)
while True:
rewards_ep = [
rewards_ep[0] + reward_n[0], rewards_ep[1] + reward_n[1]
]
metadata['status'] = Helpers.update_metadata_status(
fps=env.clock.get_fps(),
itera=int(cfg['game']['duration']) - env.duration,
iter_perc=round(
((int(cfg['game']['duration']) - env.duration) /
int(cfg['game']['duration'])) * 100, 2),
time_elap=round(time.time() - start),
eta=round((env.duration / env.clock.get_fps()) +
int(cfg['game']['duration']) / env.clock.get_fps() *
(int(cfg['game']['episodes']) - i))
if env.clock.get_fps() else None,
img_path=step_img_path[8:],
rewards=rewards_ep,
wins=[sum(w) for w in wins_l],
wins_moving=[sum(w[-10:]) for w in wins_l],
)
fps_episode.append(env.clock.get_fps())
algorithm.before_action(obs_n=obs_n)
action_n = algorithm.take_action(obs_n=obs_n)
obs_old_n = copy.deepcopy(obs_n)
algorithm.before_step(action_n=action_n)
obs_n, reward_n, done, _ = env.step(action_n)
algorithm.after_step(reward_n=reward_n,
obs_old_n=obs_old_n,
obs_n=obs_n,
done=done)
Helpers.update_img_status(env, cfg['video']['monitoring'],
step_img_path, render_mode)
self.update_state(state='PROGRESS', meta=metadata)
if done[0]:
algorithm.handle_gameover(
obs_n=obs_n,
reward_n=reward_n,
ep_length=int(cfg['game']['duration']) - env.duration,
)
Helpers.handle_gameover(done[1], wins_l)
break
algorithm.after_episode()
fps_batch.append(statistics.fmean(fps_episode))
algorithm.before_cleanup()
Helpers.cleanup(env, core_id)
metadata = {'step': 'read meta file'}
self.update_state(state='PLOTTING', meta=metadata)
path_to_file = '/opt/app/monitor/' + \
start_date + "/core-" + str(core_id) + '/'
raw_stats = read_json(path_to_file)
metadata = {'step': 'parse meta file'}
self.update_state(state='PLOTTING', meta=metadata)
parsed_stats, _ = parse_json_2_dataframe(raw_stats)
metadata = {
'step': 'plot seeker',
}
self.update_state(state='PLOTTING', meta=metadata)
seeker_plot = plot_rewards(parsed_stats, 'SEEKER', path_to_file)
metadata = {
'step': 'plot hiding',
}
self.update_state(state='PLOTTING', meta=metadata)
hiding_plot = plot_rewards(parsed_stats, 'HIDING', path_to_file)
return Helpers.get_celery_success(
core_id=core_id,
time_elap=round(time.time() - start, 4),
fps_batch=fps_batch,
wins=[sum(w) for w in wins_l],
seeker_plot=seeker_plot,
hiding_plot=hiding_plot,
)
plugin.notify(_(30100).encode('utf-8'))
plugin.open_settings()
except exceptions.RequestError:
plugin.notify(_(30103).encode('utf-8'))
plugin.log.error(sys.stdout.getvalue())
except exceptions.ResourceNotFoundError:
plugin.notify(_(30104).encode('utf-8'))
plugin.log.error(sys.stdout.getvalue())
except exceptions.ResponseError:
plugin.notify(_(30105).encode('utf-8'))
plugin.log.error(sys.stdout.getvalue())
except exceptions.StreamingRightsError:
plugin.notify(_(30106).encode('utf-8'))
plugin.log.error(sys.stdout.getvalue())
except IOError as e:
plugin.log.error(sys.stdout.getvalue())
traceback.print_exc()
if e.errno == 2 and e.message.endswith('.tmp'):
pass
except Exception as e:
plugin.log.error(sys.stdout.getvalue())
traceback.print_exc()
plugin.notify(e)
finally:
helpers.cleanup()
stdout = sys.stdout.getvalue()
if len(stdout):
for line in stdout.split('\n'):
plugin.log.info(line.replace('\r', ''))