def initializeGPs(self):
set_random_seeds(self.ini_seeds)
x_ninit = self.x_ninit
x = LatinDesign(self.xbounds, x_ninit)
self.logger.debug('init x shape ={}'.format(x.shape))
self.logger.debug('init x={}'.format(x))
z = np.tile(self.zbounds[1, :], (x.shape[0], 1))
y = self.evaluate(x,
z,
resources=self.mff.getCost(z, mode="c"),
mffname='mff')
self.batch_idx += 1
self.logger.debug('y_init:{}'.format(y))
self.logger.debug("budget used after initialization:{}".format(
self.budget_current))
self.logger.debug(
"budget used after initializaiton(from mff):{}".format(
self.mff.currentBudget))
self.logger.debug('y_init shape={}'.format(y.shape[0]))
# init a surrogate
self.gps = model_list[self.config['surrogate_name']](
X_init=x, Y_init=y, config=self.config['GPRS'])
self.logger.debug(
'********************check the init surrogate={}'.format(
self.gps.model))
self.gps.update()
self.verbose()
self.logger.debug("end model initialization")
self.iterations += 1
self.record(x, y, z)
def optimize(self):
self.logger.debug("start optimization")
if self.iterations == 0:
self.initializeGPs()
self._configure_acq()
set_random_seeds(self.config['seeds'])
while self.mff.currentBudget < self.budget:
if self.checkUpdateGPs():
self.gps.update()
self.verbose()
self.iterations += 1
if self.iterations % self.warping_update_interval == 0:
self.gps.unfix_warping_functions()
else:
self.gps.fix_warping_functions()
xz = self.optimize_acq_f(n_iter=self.config['optniter'],
method=self.method)
self.logger.debug("new next to evaluate x = {}, z={}".format(
xz[:, :-self.zdim], xz[:, -self.zdim:]))
y = self.evaluate(xz[:, :-self.zdim],
xz[:, -self.zdim:],
resources=self.mff.getCost(xz[:, -self.zdim:],
mode="c"),
mffname='mff')
self.logger.debug("new evaluate y = {}".format(y))
self.record(xz[:, :-self.zdim], xz[:, -self.zdim:], y)
self.gps.add_data(xz, y)
self.batch_idx += 1
def optimize(self):
self.logger.debug("start optimization")
if self.iterations == 0:
self.initializeGPs()
self._configure_acq()
set_random_seeds(self.config['seeds'])
while self.mff.currentBudget < self.budget:
if self.checkUpdateGps():
self.gps.update()
self.verbose()
# produce a batch of K designs
x = []
k = 0
gps_temp = deepcopy(self.gps)
while k < self.K:
x_k = self.optimize_acq_f(n_iter=self.config['optniter'],
method=self.method)
x.append(x_k)
fantasies, _ = self.gps.predict(x_k)
self.gps.add_data(x_k, fantasies)
self.gps.update()
k += 1
self.gps = deepcopy(gps_temp)
if self.acq_name == 'ES' or self.acq_name == 'MES':
self.acquisition.gps = self.gps
self.acquisition._config_proposal_function()
x = np.concatenate(x)
self.logger.debug('new next to evaluate batch x={}'.format(x))
z = np.tile(self.zbounds[1, :], (x.shape[0], 1))
y = self.evaluate(x,
z,
resources=self.mff.getCost(z, mode='c'),
mffname='mff')
self.logger.debug('new evaluate batch y={}'.format(y))
for idx in range(x.shape[0]):
self.iterations += 1
self.record(x[idx:idx + 1, :], y[idx:idx + 1, :],
z[idx:idx + 1, :])
self.gps.add_data(x, y)
self.batch_idx += 1
def optimize(self):
self.logger.debug('start optimization')
if self.iterations == 0:
self.initialize()
set_random_seeds(self.config['seeds'])
while self.mff.currentBudget < self.budget:
self.iterations += 1
np.random.seed(int(time.time() * 10**8 - int(time.time()) * 10**8))
x = np.random.uniform(low=self.xbounds[0, :],
high=self.xbounds[1, :],
size=(1, self.xdim))
np.random.seed(self.seeds)
self.logger.debug('new next to evaluate x={}'.format(x))
z = np.tile(self.zbounds[1, :], (x.shape[0], 1))
y = self.evaluate(x, z, resources=self.mff.getCost(z, mode='c'))
self.logger.debug("new evaluate y={}".format(y))
self.record(x, y, z)
self.batch_idx += 1
def initialize(self):
set_random_seeds(self.ini_seeds)
x_ninit = self.x_ninit
x = LatinDesign(self.xbounds, x_ninit)
self.logger.debug('init x shape ={}'.format(x.shape))
self.logger.debug('init x={}'.format(x))
z = np.tile(self.zbounds[1, :], (x.shape[0], 1))
y = self.mff.eval(x, z, mode="c").reshape((-1, 1))
self.batch_idx += 1
self.logger.debug('y_init:{}'.format(y))
self.logger.debug('budget used after initialization:{}'.format(
self.budget_current))
self.logger.debug(
'budget used after initialization(from mff):{}'.format(
self.mff.currentBudget))
self.logger.debug('y_init shape={}'.format(y.shape[0]))
self.logger.debug('end initialization')
self.iterations += 1
self.record(x, y, z)
def optimize(self):
self.logger.debug("start optimization")
if self.iterations == 0:
self.initializeGPs()
self._configure_acq()
set_random_seeds(self.config['seeds'])
while self.mff.currentBudget < self.budget:
if self.checkUpdateGPs():
self.gps.update()
self.verbose()
self.iterations += 1
x = self.optimize_acq_f(n_iter=self.config['optniter'],
method=self.method)
self.logger.debug('new next to evaluate x={}'.format(x))
z = np.tile(self.zbounds[1, :], (x.shape[0], 1))
y = self.evaluate(x,
z,
resources=self.mff.getCost(z, mode="c"),
mffname='mff')
self.logger.debug("new evaluate y= {}".format(y))
self.record(x, y, z)
self.gps.add_data(x, y)
self.batch_idx += 1
def main_eval(args):
eval_c_file = os.path.join(args.exp_dir, 'config.json')
eval_a_file = os.path.join(args.exp_dir, 'args.json')
eval_p_file = os.path.join(args.exp_dir, 'model', args.model_dir,
'checkpoint_' + args.model_name + '.tar')
# Read args from saved args file
with open(eval_a_file) as f:
eval_args = json.load(f)
# Read configs from saved config file
default_config = default_cfgs.config_dict[eval_args['m']]
with open(eval_c_file) as f:
eval_config = json.load(f)
config = deep_update_dict(eval_config, default_config)
set_random_seeds(eval_args['seeds'])
file_str = create_save_path(
eval_args, ['seeds'], 'eval',
'exp_dir=' + args.method + args.exp_dir.replace('/', '-'))
config['data_folder_experiment'] = file_str
# Create experiment folder
if not os.path.exists(file_str):
os.makedirs(file_str)
# configure and create logging
logger = create_logging('output.log', file_str, 'w', args.debug)
logger.info('Current config content is :{}'.format(config))
writer = SummaryWriter(log_dir=file_str)
env = CLEnv(select_environment(config['env']['env_name'])(config, 6))
env.set_new_design(args.design)
# Store args
save_args2json('args.json', file_str, args)
total_steps = config["op_algorithm_config"]["time_limits"] + 49
env.set_task_t(total_steps)
s_norm, actor, critic = load_model(eval_p_file)
acc_step = 0
acc_rwd = 0
for i in range(args.ei):
if args.save_video:
video_path = os.path.join(file_str, 'videos',
str(i) + '-th-iteration')
os.makedirs(video_path, exist_ok=True)
obs, info = env.reset(return_info=True)
obs = torch.Tensor(obs).float()
done = False
ite_step = 0
ite_rwd = 0
vis_info = {}
while True:
vis_info['design'] = env.get_current_design()
vis_info['steps'] = ite_step
for vis in vis_checklist:
if vis in info:
vis_info[vis] = info[vis]
if args.save_video:
rgb = env.render('rgb_array')
rgb = Image.fromarray(rgb)
traj = Image.new('RGB', (rgb.size[0] + 130, rgb.size[1]),
(255, 255, 255))
if 'pose(m)' in info:
x, y, z = info['pose(m)']
x = 8 * x
y = 8 * y
vis_info.setdefault("traj", []).append(
(x + 10, traj.size[1] - (y + traj.size[1] / 3)))
elif 'pos' in info:
x, y, z = info['pos']
x = 8 * x
y = 8 * y
vis_info.setdefault("traj", []).append(
(x + 10, traj.size[1] - (y + traj.size[1] / 3)))
all = Image.new('RGB',
(traj.size[0] + rgb.size[0], rgb.size[1]))
draw_rgb = ImageDraw.Draw(rgb)
draw_traj = ImageDraw.Draw(traj)
add_text(draw_traj, vis_info, ite_step, total_steps)
if 'traj' in vis_info:
draw_traj.point(vis_info['traj'], fill='red')
all.paste(rgb, (0, 0))
all.paste(traj, (rgb.size[0], 0))
all.save(
os.path.join(video_path, '{:05d}.png'.format(ite_step)))
obst = torch.FloatTensor(obs)
s_norm.record(obst)
obst_norm = s_norm(obst)
with torch.no_grad():
ac = actor.act_deterministic(obst_norm)
ac = ac.cpu().numpy()
if done:
break
obs, rwd, done, info = env.step(ac)
time.sleep(0.0165)
ite_step += 1
ite_rwd += rwd
info['rewards'] = ite_rwd
info['ave_rewards'] = ite_rwd / ite_step
acc_rwd += rwd
acc_step += 1
logger.info(
'return of {}-th iteration is {}, and num of steps is {}, return per step={}'
.format(i, ite_rwd, ite_step, ite_rwd / 250))
if args.save_video:
sp.call([
'ffmpeg', '-loglevel', 'panic', '-r', '60', '-f', 'image2',
'-i',
os.path.join(video_path, '%05d.png'), '-vcodec', 'libx264',
'-pix_fmt', 'yuv420p',
os.path.join(
video_path,
str(i) + '-th-ite_rwd=' + str(int(ite_rwd)) + '.mp4')
])
avg_step = acc_step / args.ei
avg_rwd = acc_rwd / args.ei
logger.info(
'average return of {} iterations is {}, and num of steps is {}, average return per step={}'
.format(args.ei, avg_rwd, avg_step, avg_rwd / avg_step))
def initializeGPs(self):
set_random_seeds(self.ini_seeds)
x_ninit = self.x_ninit
x = LatinDesign(self.xbounds, x_ninit)
self.logger.debug('init x shape={}'.format(x.shape))
self.logger.debug('init x={}'.format(x))
if self.config['ini_manner'] == 'latin_cross':
z_ninit = self.z_ninit
z = LatinDesign(self.zbounds, z_ninit)
z = self._roundfidelity(z)
z = z[np.where(z.sum(axis=1) != sum(self.zbounds[1, :]))[0], :]
self.logger.debug('init z shape={}'.format(z.shape))
self.logger.debug('init z={}'.format(z))
z_target = np.concatenate(
[self.zbounds[1:, :] for i in range(x.shape[0])])
x_ = np.random.permutation(
LatinDesign(self.xbounds, min(50,
self.z_ninit)))[:z.shape[0], :]
x = np.concatenate([x, x_])
z = np.concatenate([z_target, z])
self.logger.debug('final ini z={}'.format(z))
self.logger.debug('final ini x={}'.format(x))
else:
raise NotImplementedError
xz = np.concatenate((x, z), axis=1)
self.xz_ninit = xz.shape[0]
self.logger.debug("xz_ini cost={}".format(
sum(self.mff.getCost(xz[:, -self.zdim:]))))
self.logger.debug("xz_init shape={}".format(xz.shape))
self.logger.debug("xz_init={}".format(xz))
y = self.evaluate(xz[:, :-self.zdim],
xz[:, -self.zdim:],
resources=self.mff.getCost(xz[:, -self.zdim:],
mode="c"),
mffname='mff')
self.batch_idx += 1
self.logger.info("y_init:{}".format(y))
self.logger.info("budget used after initialization:{}".format(
self.budget_current))
self.logger.info(
"budget used after initialization(from mff):{}".format(
self.mff.currentBudget))
self.logger.debug("y_init shape={}".format(y.shape[0]))
#init a surrogate
self.gps = model_list[self.config['surrogate_name']](
X_init=xz,
Y_init=y,
z_min=self.zbounds[0, :],
z_max=self.zbounds[1, :],
config=self.config['GPRS'])
self.logger.debug('**************Check the init surrogate={}'.format(
self.gps.model))
self.gps.update()
self.gps.fix_warping_functions()
self.verbose()
self.logger.debug("end model initialization")
self.iterations += 1
self.record(xz[:, :-self.zdim], xz[:, -self.zdim:], y)
def main(args):
# Read config from config files
default_config = default_cfgs.config_dict[args.m]
if args.config:
new_config_file = os.path.join('./nfwbo/configs', args.robot,
args.config + '.json')
with open(new_config_file) as f:
new_config = json.load(f)
config = deep_update_dict(new_config, default_config)
else:
config = default_config
# check possible bugs in config settings
check_config(config, args)
set_random_seeds(args.seeds)
# change config file by arguments from args
config["op_algorithm_config"]["seeds"] = args.seeds
if args.dev:
config["op_algorithm_config"]["max_num_epochs_per_design"] = 1
config["op_algorithm_config"]["time_limits"] = 1
config["op_algorithm_config"]["x_ninit"] = 1
config["op_algorithm_config"]["z_ninit"] = 1
config["op_algorithm_config"]["optniter"] = 1
config["op_algorithm_config"]["optninit"] = 1
config["op_algorithm_config"]["batch_size"] = 2
config['rl_algorithm_config']["algo_params"]["checkpoint_batch"] = -1
config['force_manual_design'] = args.h
file_str = create_save_path(args, ['seeds']) + '_' + args.extra
if args.h:
file_str = file_str + '_human_design'
config['rl_algorithm_config']["algo_params"]["checkpoint_batch"] = 240
config['op_algorithm_config']["set_mff_inference"] = False
config['data_folder_experiment'] = file_str
plt_dir = os.path.join(file_str, 'plt')
config['op_algorithm_config']['plt_dir'] = plt_dir
model_path = file_str + "/model"
config["rl_algorithm_config"]["algo_params"]["save_dir"] = model_path
# Create experiment folder
if not os.path.exists(file_str):
os.makedirs(file_str)
#create model's folder
if not os.path.exists(model_path):
os.makedirs(model_path)
if not os.path.exists(plt_dir):
os.makedirs(plt_dir)
# Store config
with open(os.path.join(file_str, 'config.json'), 'w') as fd:
fd.write(json.dumps(config, indent=2))
# Store args
save_args2json('args.json', file_str, args)
# configure and create logging
logger = create_logging('output.log', file_str, 'w', args.debug)
logger.info('Current config content is :{}'.format(config))
writer = SummaryWriter(log_dir=file_str)
do = DesignOptimize(logger, writer, config)
do.optimize()
writer.close()