def draw_reject_ratio(settings, save_path):
project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), os.pardir))
log_name = "log"
parser_func = parse_reject_ratio
path_formatters = []
config_files, runs, num_datapoints, labels = [], [], [], []
for cfg, param_setting, num_runs, num_steps, label in settings:
config_files.append((cfg, param_setting))
runs.append(num_runs)
num_datapoints.append(num_steps)
labels.append(label)
for cf, best_setting in config_files:
swp = Sweeper(os.path.join(project_root, '/'.join(cf.split("/")[:-1])))
cfg = swp.parse(best_setting) # Creating a cfg with an arbitrary run id
cfg.data_root = os.path.join(project_root, 'data', 'output')
logdir_format = cfg.get_logdir_format()
path_format = os.path.join(logdir_format, log_name
)
path_formatters.append(path_format + '/' + cf.split("/")[-1])
v = RunLines(path_formatters, runs, num_datapoints,
labels, parser_func=parser_func,
save_path=save_path, xlabel="Number of steps in 10000s", ylabel="Reject Ratio",
interval=10000, ylim=(0, 1.2))
v.draw()
def draw_lunar_lander_dqn(settings, save_path, max_steps=2000000):
project_root = os.path.abspath(
os.path.join(os.path.dirname(__file__), os.pardir))
log_name = "log"
parser_func = parse_steps_log
path_formatters = []
config_files, runs, num_datapoints, labels = [], [], [], []
for cfg, param_setting, num_runs, num_steps, label in settings:
config_files.append((cfg, param_setting))
runs.append(num_runs)
num_datapoints.append(num_steps)
labels.append(label)
for cf, best_setting in config_files:
swp = Sweeper(os.path.join(project_root, cf))
cfg = swp.parse(
best_setting) # Creating a cfg with an arbitrary run id
cfg.data_root = os.path.join(project_root, 'data', 'output')
logdir_format = cfg.get_logdir_format()
path_format = os.path.join(logdir_format, log_name)
path_formatters.append(path_format)
v = RunLines(path_formatters,
runs,
num_datapoints,
labels,
parser_func=parser_func,
save_path=save_path,
xlabel="Number of steps",
ylabel="Reward",
interval=10000)
v.draw()
def _eval_lines(config_file, start_idx, end_idx, max_steps, interval=10000):
print('config_file: {}'.format(config_file))
project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), os.pardir))
sweeper = Sweeper(os.path.join(project_root, config_file))
eval = []
eval_lines = []
for k in range(sweeper.total_combinations):
eval.append([])
eval_lines.append([])
for idx in range(start_idx, end_idx):
cfg = sweeper.parse(idx)
cfg.data_root = os.path.join(project_root, 'data', 'output')
log_dir = cfg.get_logdir()
log_path = os.path.join(log_dir, 'log')
try:
with open(log_path, "r") as f:
lines = f.readlines()
except FileNotFoundError:
continue
if len(lines) == 0:
continue
# ugly parse based on the log_file format
try:
num_steps = get_max_steps(lines)
if num_steps >= max_steps:
assert idx % sweeper.total_combinations == cfg.param_setting
avg_eval_steps = extract_line(lines, max_steps, interval=interval)
eval[idx % sweeper.total_combinations].append(np.mean(avg_eval_steps[-int(len(avg_eval_steps)/2):]))
except IndexError:
print(idx)
raise
summary = list(map(lambda x: (x[0], np.mean(x[1]), np.std(x[1]), len(x[1])), enumerate(eval)))
summary = [x for x in summary if np.isnan(x[1]) == False]
# new_summary = []
# for s in summary:
# if np.isnan(s[1]) == False:
# new_summary.append(s)
# print(summary[0])
# print(new_summary[0])
# quit()
summary = sorted(summary, key=lambda s: s[1], reverse=True)
for idx, mean, std, num_runs in summary:
print("Param Setting # {:>3d} | Rewards: {:>10.10f} +/- {:>5.2f} ({:>2d} runs) {} | ".format(
idx, mean, std, num_runs, sweeper.param_setting_from_id(idx)))
def draw_lunar_lander_traces_h512(save_path, max_steps=2000000):
project_root = os.path.abspath(
os.path.join(os.path.dirname(__file__), os.pardir))
log_name = "log"
parser_func = parse_steps_log
path_formatters = []
settings = [
("experiment/config_files/lunar_lander/sarsa_lmbda/sweep_h512.json", 1,
5, max_steps, "Sarsa(0)"),
("experiment/config_files/lunar_lander/sarsa_lmbda/sweep_h512.json", 4,
5, max_steps, "Sarsa(0.5)"),
("experiment/config_files/lunar_lander/sarsa_lmbda/sweep_h512.json", 7,
5, max_steps, "Sarsa(0.7)"),
("experiment/config_files/lunar_lander/sarsa_lmbda/sweep_h512.json",
11, 5, max_steps, "Sarsa(0.85)"),
("experiment/config_files/lunar_lander/sarsa_lmbda/sweep_h512.json",
14, 5, max_steps, "Sarsa(0.9)")
]
config_files, runs, num_datapoints, labels = [], [], [], []
for cfg, param_setting, num_runs, num_steps, label in settings:
config_files.append((cfg, param_setting))
runs.append(num_runs)
num_datapoints.append(num_steps)
labels.append(label)
for cf, best_setting in config_files:
swp = Sweeper(os.path.join(project_root, cf))
cfg = swp.parse(
best_setting) # Creating a cfg with an arbitrary run id
cfg.data_root = os.path.join(project_root, 'data', 'output')
logdir_format = cfg.get_logdir_format()
path_format = os.path.join(logdir_format, log_name)
path_formatters.append(path_format)
v = RunLines(path_formatters,
runs,
num_datapoints,
labels,
parser_func=parser_func,
save_path=save_path,
xlabel="Number of steps",
ylabel="Episodes Completed",
interval=10000)
v.draw()
def run():
start = time.time()
parser = argparse.ArgumentParser(description="run_file")
parser.add_argument('--idx',
default=0,
type=int,
help='identifies run number and configuration')
parser.add_argument('--config-file',
default='config_files/actor_critic.json')
args = parser.parse_args()
project_root = os.path.abspath(os.path.join(os.path.dirname(__file__)))
sweeper = Sweeper(os.path.join(project_root, args.config_file))
cfg = sweeper.parse(args.idx)
cfg.env_instance = MountainCar(cfg)
agent_class = getattr(agents.agent, cfg.agent_class)
agent = agent_class(cfg)
log_dir = cfg.get_logdir()
ensure_dirs([log_dir])
steps_log = os.path.join(log_dir, 'steps_log')
steps_logger = setup_logger(steps_log, stdout=True)
cfg.log_config(steps_logger)
ep_log = os.path.join(log_dir, 'ep_log')
ep_logger = setup_logger(ep_log, stdout=False)
cfg.log_config(ep_logger)
exp = Experiment(agent,
cfg.env_instance,
max_steps=cfg.max_steps,
seed=args.idx,
steps_log=steps_log,
ep_log=ep_log)
exp.run_step_mode()
print("Memory used: {:5} MB".format(memory_usage_psutil()))
print("Time elapsed: {:5.2} minutes".format((time.time() - start) / 60))
def draw_mountain_car(save_path):
project_root = os.path.abspath(
os.path.join(os.path.dirname(__file__), os.pardir))
log_name = "log"
parser_func = parse_steps_log
path_formatters = []
settings = [
("experiment/config_files/mountain_car/sarsa/sweep.json", 12, 3,
300000, "Sarsa(0)"),
]
config_files, runs, num_datapoints, labels = [], [], [], []
for cfg, param_setting, num_runs, num_steps, label in settings:
config_files.append((cfg, param_setting))
runs.append(num_runs)
num_datapoints.append(num_steps)
labels.append(label)
for cf, best_setting in config_files:
swp = Sweeper(os.path.join(project_root, cf))
cfg = swp.parse(
best_setting) # Creating a cfg with an arbitrary run id
cfg.data_root = os.path.join(project_root, 'data', 'output')
logdir_format = cfg.get_logdir_format()
path_format = os.path.join(logdir_format, log_name)
path_formatters.append(path_format)
v = RunLines(path_formatters,
runs,
num_datapoints,
labels,
parser_func=parser_func,
save_path=save_path,
xlabel="Number of steps",
ylabel="Reward",
interval=10000)
v.draw()
if __name__ == '__main__':
start_idx = 0
end_idx = 640
config_file = 'config_files/sw_actor_critic.json'
max_steps = 40000
project_root = os.path.abspath(os.path.join(os.path.dirname(__file__)))
sweeper = Sweeper(os.path.join(project_root, config_file))
eval = []
for k in range(sweeper.total_combinations):
eval.append([])
for idx in range(start_idx, end_idx):
cfg = sweeper.parse(idx)
log_dir = cfg.get_logdir()
log_path = os.path.join(log_dir, 'steps_log')
with open(log_path, "r") as f:
lines = f.readlines()
if len(lines) == 0:
continue
# ugly parse based on the log_file format
num_steps = int(lines[-1].split("|")[1].split(":")[1])
episodes = int(lines[-1].split("|")[2].split(":")[1])
if cfg.max_steps == num_steps:
assert idx % sweeper.total_combinations == cfg.param_setting
eval[idx % sweeper.total_combinations].append(episodes)
summary = list(
return eps_over_time
if __name__ == '__main__':
project_root = os.path.abspath(
os.path.join(os.path.dirname(__file__), os.pardir))
log_name = "ep_log"
parser_func = parse_ep_log if log_name == "ep_log" else parse_steps_log
path_formatters = []
config_files = [("config_files/sw_actor_critic.json", 13)]
for cf, best_setting in config_files:
swp = Sweeper(os.path.join(project_root, cf))
cfg = swp.parse(
best_setting) # Creating a cfg with an arbitrary run id
logdir_format = cfg.get_logdir_format()
path_format = os.path.join(logdir_format, log_name)
path_formatters.append(path_format)
runs = [10]
num_datapoints = [40000]
labels = ['Actor-Critic']
v = RunLines(path_formatters,
runs,
num_datapoints,
labels,
parser_func=parser_func,
save_path="mountain_car.png",
xlabel="Number of steps",
ylabel="Episodes Completed")
