def main(args, seed, episodes):
# Set seed for all randomness sources
utils.seed(seed)
# Define agent
env = gym.make(args.env)
env.seed(seed)
agent = utils.load_agent(env, args.model, None, None, args.argmax,
args.env)
if args.model is None and args.episodes > len(agent.demos):
# Set the number of episodes to be the number of demos
episodes = len(agent.demos)
# Evaluate
if isinstance(agent, utils.DemoAgent):
logs = evaluate_demo_agent(agent, episodes)
elif isinstance(agent, utils.BotAgent) or args.contiguous_episodes:
logs = evaluate(agent, env, episodes, False)
else:
logs = batch_evaluate(agent,
args.env,
seed,
episodes,
return_obss_actions=True)
return logs
def main_test(args, seed, episodes):
# Set seed for all randomness sources
utils.seed(seed)
# Define agent
# do test environment
env_name = args.env + "_Test-v0"
env = gym.make(env_name)
env.seed(seed)
agent = utils.load_agent(env,
args.model,
argmax=args.argmax,
env_name=env_name)
if args.model is None and args.episodes > len(agent.demos):
# Set the number of episodes to be the number of demos
episodes = len(agent.demos)
# Evaluate
if isinstance(agent, utils.DemoAgent):
logs = evaluate_demo_agent(agent, episodes)
elif isinstance(agent, utils.BotAgent):
logs = evaluate(agent, env, episodes, False)
else:
logs = batch_evaluate(agent, env_name, seed, episodes)
return logs
def main(args, seed, episodes):
# Set seed for all randomness sources
utils.seed(seed)
# Keep track of results per task.
results = {}
for env_name in args.env:
start_time = time.time()
env = gym.make(env_name)
env.seed(seed)
if args.model is None and args.episodes > len(agent.demos):
# Set the number of episodes to be the number of demos
episodes = len(agent.demos)
# Define agent
agent = utils.load_agent(env,
args.model,
args.demos,
args.demos_origin,
args.argmax,
env_name,
model_path=args.model_path)
# Evaluate
if isinstance(agent, utils.DemoAgent):
logs = evaluate_demo_agent(agent, episodes)
elif isinstance(agent, utils.BotAgent) or args.contiguous_episodes:
logs = evaluate(agent, env, episodes, False)
else:
logs = batch_evaluate(agent, env_name, seed, episodes)
end_time = time.time()
# Print logs
num_frames = sum(logs["num_frames_per_episode"])
fps = num_frames / (end_time - start_time)
ellapsed_time = int(end_time - start_time)
duration = datetime.timedelta(seconds=ellapsed_time)
if args.model is not None:
return_per_episode = utils.synthesize(logs["return_per_episode"])
success_per_episode = utils.synthesize(
[1 if r > 0 else 0 for r in logs["return_per_episode"]])
num_frames_per_episode = utils.synthesize(
logs["num_frames_per_episode"])
if args.model is not None:
print(
"F {} | FPS {:.0f} | D {} | R:xsmM {:.3f} {:.3f} {:.3f} {:.3f} | S {:.3f} | F:xsmM {:.1f} {:.1f} {} {}"
.format(num_frames, fps, duration,
*return_per_episode.values(),
success_per_episode['mean'],
*num_frames_per_episode.values()))
else:
print(
"F {} | FPS {:.0f} | D {} | F:xsmM {:.1f} {:.1f} {} {}".format(
num_frames, fps, duration,
*num_frames_per_episode.values()))
indexes = sorted(range(len(logs["num_frames_per_episode"])),
key=lambda k: -logs["num_frames_per_episode"][k])
n = args.worst_episodes_to_show
if n > 0:
print("{} worst episodes:".format(n))
for i in indexes[:n]:
if 'seed_per_episode' in logs:
print(logs['seed_per_episode'][i])
if args.model is not None:
print("- episode {}: R={}, F={}".format(
i, logs["return_per_episode"][i],
logs["num_frames_per_episode"][i]))
else:
print("- episode {}: F={}".format(
i, logs["num_frames_per_episode"][i]))
# Store results for this env.
logs['return_per_episode'] = return_per_episode
logs['success_per_episode'] = success_per_episode
logs['num_frames_per_episode'] = num_frames_per_episode
results[env_name] = logs
return results