def do_steps_based(base_args,
cores,
name,
steps,
runs,
options=None,
tasks={},
starting_task=''):
args = base_args.copy()
args['steps'] = steps
if options:
suffix = ''
if options['balancing_tf']:
suffix += '1_' + options['balancing_tf']
if options['balancing']:
suffix += '2_' + options['balancing']
if options['walking']:
suffix += '3_' + options['walking']
if suffix:
name += '-' + suffix
args['options'] = options
hp = Helper(args, 'cl', name, tasks, starting_task, cores, use_mp=True)
# generate configurations
solutions = [None] * len(runs)
begin = runs[0]
mp_cfgs = hp.gen_cfg(solutions, 1, begin=begin)
return mp_cfgs
def do_network_based_leo(base_args, cores, name, nn_params, runs, tasks,
starting_task):
args = base_args.copy()
args['rb_min_size'] = 1000
args['default_damage'] = 4035.00
args['perf_td_error'] = True
args['perf_l2_reg'] = True
args['steps'] = 300000
args["cl_batch_norm"] = False
args['cl_structure'] = 'rnnc:gru_tanh_6_dropout;fc_linear_3'
args['cl_stages'] = 'balancing_tf;balancing;walking:monotonic'
args['cl_depth'] = 2
args['cl_pt_shape'] = (2, 3)
args["cl_pt_load"] = nn_params[1]
cl_load = nn_params[0]
hp = Helper(args, 'cl', name, tasks, starting_task, cores, use_mp=True)
# Weights of the NN
solutions = [None] * len(runs)
begin = runs[0]
mp_cfgs = hp.gen_cfg(solutions, 1, begin=begin)
mp_cfgs_new = []
for cfg in mp_cfgs:
config, tasks, starting_task = cfg
copy_config = config.copy()
copy_config["cl_load"] = cl_load
mp_cfgs_new.append((copy_config, tasks, starting_task))
return mp_cfgs_new
def do_reach_timeout_based(base_args,
cores,
name,
reach_timeout,
runs,
options=None,
tasks={},
starting_task=''):
args = base_args.copy()
args['reach_timeout'] = reach_timeout
steps = 300000
args['steps'] = steps
args['rb_max_size'] = steps
if options:
suffix = ''
if options['balancing_tf']:
suffix += '1_' + options['balancing_tf'] + '_'
if options['balancing']:
suffix += '2_' + options['balancing'] + '_'
if options['walking']:
suffix += '3_' + options['walking']
if suffix:
name += '-' + suffix
args['options'] = options
hp = Helper(args, 'cl', name, tasks, starting_task, cores, use_mp=True)
# Weights of the NN
solutions = [None] * len(runs)
begin = runs[0]
mp_cfgs = hp.gen_cfg(solutions, 1, begin=begin)
return mp_cfgs
def do_steps_based(base_args, cores, name, steps, runs, tasks, starting_task):
args = base_args.copy()
args['steps'] = steps
hp = Helper(args, 'cl', name, tasks, starting_task, cores, use_mp=True)
# Weights of the NN
solutions = [None] * len(runs)
begin = runs[0]
mp_cfgs = hp.gen_cfg(solutions, 1, begin=begin)
return mp_cfgs
def do_network_based_mujoco(base_args, cores, name, nn_params, options, runs,
tasks, starting_task):
args = base_args.copy()
args['env_td_error_scale'] = 600.0
args['env_timeout'] = 16.5
args['steps'] = 700000
args["rb_max_size"] = args['steps']
args['rb_min_size'] = 1000
args['default_damage'] = 4035.00
args['perf_td_error'] = True
args['perf_l2_reg'] = True
args["cl_batch_norm"] = False
args['cl_structure'] = 'rnnc:gru_tanh_6_dropout;fc_linear_3'
args['cl_stages'] = 'balancing_tf;balancing;walking:monotonic'
args['cl_pt_shape'] = (args['cl_depth'], 3)
args["cl_pt_load"] = nn_params[1]
cl_load = nn_params[0]
if options:
suffix = ''
if options['balancing_tf']:
suffix += '1_' + options['balancing_tf'] + '_'
if options['balancing']:
suffix += '2_' + options['balancing'] + '_'
if options['walking']:
suffix += '3_' + options['walking']
if suffix:
name += '-' + suffix
args['options'] = options
hp = Helper(args, 'cl', name, tasks, starting_task, cores, use_mp=True)
# Weights of the NN
solutions = [None] * len(runs)
begin = runs[0]
mp_cfgs = hp.gen_cfg(solutions, 1, begin=begin)
mp_cfgs_new = []
for cfg in mp_cfgs:
config, tasks, starting_task = cfg
copy_config = config.copy()
copy_config["cl_load"] = cl_load
mp_cfgs_new.append((copy_config, tasks, starting_task))
return mp_cfgs_new
args['perf_l2_reg'] = True
args['steps'] = 300000
args["rb_max_size"] = args['steps']
#args["cl_batch_norm"] = True
#args['cl_structure'] = 'ffcritic:fc_relu_4;fc_relu_3;fc_relu_3'
args["cl_batch_norm"] = False
args['cl_structure'] = 'rnnc:gru_tanh_6_dropout;fc_linear_3'
args["cl_stages"] = "balancing_tf;balancing;walking:monotonic"
args['cl_depth'] = 2
args['cl_pt_shape'] = (args['cl_depth'], 3)
args['test_interval'] = 30
#args["cl_target"] = True
export_names = "eq_curriculum_network_depth_" + str(args['cl_depth'])
nn_params = (export_names, "{}_stat.pkl".format(export_names))
args["cl_pt_load"] = nn_params[1]
# Parameters
tasks = {
'balancing_tf': 'cfg/leo_balancing_tf.yaml',
'balancing': 'cfg/leo_balancing.yaml',
'walking': 'cfg/leo_walking.yaml'
}
starting_task = 'balancing_tf'
hp = Helper(args, 'cl', 'ddpg', tasks, starting_task, 1, use_mp=False)
# Run actual script.
config, tasks, starting_task = hp.gen_cfg([None], 1)[0]
config["cl_load"] = nn_params[0]
cl_run(tasks, starting_task, **config)