def main(env, ctrl_type, ctrl_args, overrides, logdir, mode):
ctrl_args = DotMap(**{key: val for (key, val) in ctrl_args})
overrides.append(["exp_cfg.log_cfg.nrecord", "0"])
cfg = create_config(env, ctrl_type, ctrl_args, overrides, logdir, mode)
if cfg.exp_cfg.exp_cfg.use_value:
cfg.exp_cfg.exp_cfg.value = DeepValueFunction("value", cfg.val_cfg)
cfg.exp_cfg.exp_cfg.value_target = DeepValueFunction(
"target", cfg.val_cfg)
if not cfg.exp_cfg.exp_cfg.load_samples:
cfg.exp_cfg.exp_cfg.teacher = cfg.exp_cfg.sim_cfg.env.teacher(
cfg.exp_cfg.exp_cfg.value.model.sess)
cfg.exp_cfg.exp_cfg.teacher.env.set_mode(mode)
cfg.ctrl_cfg.value_func = cfg.exp_cfg.exp_cfg.value
cfg.ctrl_cfg.target_value_func = cfg.exp_cfg.exp_cfg.value_target
cfg.ctrl_cfg.use_value = cfg.exp_cfg.exp_cfg.use_value
if ctrl_type == "MPC":
cfg.exp_cfg.exp_cfg.policy = MPC(cfg.ctrl_cfg)
exp = MBExperiment(cfg.exp_cfg)
cfg.pprint()
os.makedirs(exp.logdir)
with open(os.path.join(exp.logdir, "config.txt"), "w") as f:
f.write(pprint.pformat(cfg.toDict()))
exp.run_experiment()
def main(env, ctrl_type, ctrl_args, overrides, logdir):
ctrl_args = DotMap(**{key: val for (key, val) in ctrl_args})
cfg = create_config(env, ctrl_type, ctrl_args, overrides, logdir)
cfg.pprint()
if ctrl_type == "MPC":
cfg.exp_cfg.exp_cfg.policy = MPC(cfg.ctrl_cfg)
exp = MBExperiment(cfg.exp_cfg)
os.makedirs(exp.logdir)
with open(os.path.join(exp.logdir, "config.txt"), "w") as f:
f.write(pprint.pformat(cfg.toDict()))
exp.run_experiment()
def exp(steps_needed_to_solve, planning_horizon, logdir):
tf.reset_default_graph()
env = "pointmass_u_wall"
ctrl_type = "MPC"
ctrl_args = []
overrides = [
# ["exp_cfg.log_cfg.nrecord", 1],
# ["exp_cfg.log_cfg.neval", 5],
# ["exp_cfg.log_cfg.nrecord_eval_mode", 1],
# ["exp_cfg.log_cfg.neval_eval_mode", 5],
# ["exp_cfg.log_cfg.nrecord", 2],
["exp_cfg.log_cfg.neval", 5],
# ["exp_cfg.log_cfg.nrecord_eval_mode", 1],
["exp_cfg.log_cfg.neval_eval_mode", 5],
["exp_cfg.exp_cfg.ntrain_iters", 3],
["ctrl_cfg.opt_cfg.plan_hor", 1],
["ctrl_cfg.opt_cfg.cfg.popsize", 5],
["ctrl_cfg.opt_cfg.cfg.num_elites", 2],
["ctrl_cfg.opt_cfg.cfg.max_iters", 1],
]
config_module_kwargs = {
'steps_needed_to_solve': steps_needed_to_solve,
'planning_horizon': planning_horizon,
'task_horizon_factor': 4,
}
ctrl_args = DotMap(**{key: val for (key, val) in ctrl_args})
cfg = create_config(env, ctrl_type, ctrl_args, overrides, logdir,
config_module_kwargs)
cfg.pprint()
if ctrl_type == "MPC":
cfg.exp_cfg.exp_cfg.policy = MPC(cfg.ctrl_cfg)
exp = MBExperiment(cfg.exp_cfg)
os.makedirs(exp.logdir)
config_dict = cfg.toDict()
config_dict['config_module_kwargs'] = config_module_kwargs
with open(os.path.join(exp.logdir, "config.txt"), "w") as f:
f.write(pprint.pformat(config_dict))
with open(os.path.join(exp.logdir, "variant.json"), "w") as f:
json.dump(config_dict, f, indent=2, sort_keys=True, cls=MyEncoder)
save_git_info(exp.logdir)
logger.set_snapshot_dir(exp.logdir)
logger.add_tabular_output(os.path.join(exp.logdir, 'progress.csv'))
logger.log_variant(os.path.join(exp.logdir, 'variant.json'), config_dict)
print("log dir:", exp.logdir)
exp.run_experiment()
def main(env, ctrl_type, ctrl_args, overrides, logdir, args):
ctrl_args = DotMap(**{key: val for (key, val) in ctrl_args})
cfg = create_config(env, ctrl_type, ctrl_args, overrides, logdir)
logger.info('\n' + pprint.pformat(cfg))
# add the part of popsize
if ctrl_type == "MPC":
cfg.exp_cfg.exp_cfg.policy = MPC(cfg.ctrl_cfg)
cfg.exp_cfg.misc = copy.copy(cfg)
exp = MBExperiment(cfg.exp_cfg, train_policy=bool(args.train_policy))
if not os.path.exists(exp.logdir):
os.makedirs(exp.logdir)
with open(os.path.join(exp.logdir, "config.txt"), "w") as f:
f.write(pprint.pformat(cfg.toDict()))
exp.run_experiment()
def main(env, ctrl_type, ctrl_args, overrides, logdir):
ctrl_args = DotMap(**{key: val for (key, val) in ctrl_args})
cfg = create_config(env, ctrl_type, ctrl_args, overrides, logdir)
cfg.pprint()
if ctrl_type == "MPC":
cfg.exp_cfg.exp_cfg.policy = MPC(cfg.ctrl_cfg)
exp = MBExperiment(cfg.exp_cfg)
os.makedirs(exp.logdir)
config_dict = cfg.toDict()
with open(os.path.join(exp.logdir, "config.txt"), "w") as f:
f.write(pprint.pformat(config_dict))
with open(os.path.join(exp.logdir, "variant.json"), "w") as f:
json.dump(config_dict, f, indent=2, sort_keys=True, cls=MyEncoder)
save_git_info(exp.logdir)
exp.run_experiment()
def main(
env,
ctrl_type,
ctrl_args,
overrides,
model_dir,
logdir,
init_iter,
last_iter,
nrecord,
rawdir,
):
ctrl_args = DotMap(**{key: val for (key, val) in ctrl_args})
overrides.append(["ctrl_cfg.prop_cfg.model_init_cfg.model_dir", model_dir])
overrides.append(["ctrl_cfg.prop_cfg.model_init_cfg.load_model", "True"])
overrides.append(["ctrl_cfg.prop_cfg.model_pretrained", "True"])
overrides.append(["exp_cfg.exp_cfg.ninit_rollouts", "0"])
overrides.append(["exp_cfg.exp_cfg.init_iter", str(init_iter)])
overrides.append(["exp_cfg.exp_cfg.ntrain_iters", str(last_iter)])
overrides.append(["exp_cfg.log_cfg.nrecord", str(nrecord)])
overrides.append(["exp_cfg.log_cfg.rawdir", str(rawdir)])
cfg = create_config(env, ctrl_type, ctrl_args, overrides, logdir)
cfg.pprint()
if ctrl_type == "MPC":
cfg.exp_cfg.exp_cfg.policy = MPC(cfg.ctrl_cfg)
exp = MBExperiment(cfg.exp_cfg)
if os.path.exists(exp.logdir):
overwrite = user_prompt("{} already exists. Overwrite?".format(
exp.logdir))
if not overwrite:
return
else:
os.makedirs(exp.logdir)
with open(os.path.join(exp.logdir, "config.txt"), "w") as f:
f.write(pprint.pformat(cfg.toDict()))
exp.run_experiment()
print("Saved to")
print(exp.logdir)
def main(env, ctrl_type, ctrl_args, overrides, model_dir, logdir):
ctrl_args = DotMap(**{key: val for (key, val) in ctrl_args})
overrides.append(["ctrl_cfg.prop_cfg.model_init_cfg.model_dir", model_dir])
overrides.append(["ctrl_cfg.prop_cfg.model_init_cfg.load_model", "True"])
overrides.append(["ctrl_cfg.prop_cfg.model_pretrained", "True"])
overrides.append(["exp_cfg.exp_cfg.ninit_rollouts", "0"])
overrides.append(["exp_cfg.exp_cfg.ntrain_iters", "1"])
overrides.append(["exp_cfg.log_cfg.nrecord", "1"])
cfg = create_config(env, ctrl_type, ctrl_args, overrides, logdir)
cfg.pprint()
if ctrl_type == "MPC":
cfg.exp_cfg.exp_cfg.policy = MPC(cfg.ctrl_cfg)
exp = MBExperiment(cfg.exp_cfg)
os.makedirs(exp.logdir)
with open(os.path.join(exp.logdir, "config.txt"), "w") as f:
f.write(pprint.pformat(cfg.toDict()))
exp.run_experiment()
def main(env, ctrl_type, ctrl_args, overrides, model_dir, logdir):
ctrl_args = DotMap(**{key: val for (key, val) in ctrl_args})
overrides.append(["ctrl_cfg.prop_cfg.model_init_cfg.model_dir", model_dir])
overrides.append(["ctrl_cfg.prop_cfg.model_init_cfg.load_model", "True"])
overrides.append(["ctrl_cfg.prop_cfg.model_pretrained", "True"])
overrides.append(["exp_cfg.exp_cfg.ninit_rollouts", "0"])
overrides.append(["exp_cfg.exp_cfg.ntrain_iters", "1"])
overrides.append(["exp_cfg.log_cfg.nrecord", "0"])
overrides.append(["exp_cfg.exp_cfg.nrollouts_per_iter", "200"])
overrides.append(["exp_cfg.log_cfg.neval", "200"])
cfg = create_config(env, ctrl_type, ctrl_args, overrides, logdir)
cfg.val_cfg.model_init_cfg.load_model = True
cfg.val_cfg.model_init_cfg.model_dir = model_dir
cfg.exp_cfg.exp_cfg.use_teacher = False
if cfg.exp_cfg.exp_cfg.use_teacher:
cfg.exp_cfg.exp_cfg.teacher = cfg.exp_cfg.sim_cfg.env.teacher()
cfg.exp_cfg.exp_cfg.use_value = False
if cfg.exp_cfg.exp_cfg.use_value:
cfg.exp_cfg.exp_cfg.value = DeepValueFunction(cfg.val_cfg)
cfg.exp_cfg.exp_cfg.ninit_rollouts = 0
cfg.ctrl_cfg.value_func = cfg.exp_cfg.exp_cfg.value
cfg.ctrl_cfg.use_value = cfg.exp_cfg.exp_cfg.use_value
cfg.pprint()
if ctrl_type == "MPC":
cfg.exp_cfg.exp_cfg.policy = MPC(cfg.ctrl_cfg)
exp = MBExperiment(cfg.exp_cfg)
os.makedirs(exp.logdir)
with open(os.path.join(exp.logdir, "config.txt"), "w") as f:
f.write(pprint.pformat(cfg.toDict()))
exp.run_experiment()