args = parser.parse_args()
cfg = Config(args.cfg, create_dirs=False)
dtype = torch.float64
torch.set_default_dtype(dtype)
np.random.seed(cfg.seed)
torch.manual_seed(cfg.seed)
torch.set_grad_enabled(False)
logger = create_logger(os.path.join(cfg.log_dir, 'log_eval_wild.txt'))
"""test data"""
cnn_feat_file = '%s/features/cnn_feat_%s.p' % (cfg.data_dir, args.test_feat)
cnn_feat_dict, _ = pickle.load(open(cnn_feat_file, 'rb'))
takes = list(cnn_feat_dict.keys())
"""environment"""
env = HumanoidEnv(cfg)
env_vis = HumanoidVisEnv(cfg.mujoco_model_file, 10)
env.seed(cfg.seed)
cnn_feat_dim = cnn_feat_dict[takes[0]].shape[-1]
actuators = env.model.actuator_names
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
body_qposaddr = get_body_qposaddr(env.model)
"""load policy net"""
policy_vs_net = VideoStateNet(cnn_feat_dim, cfg.policy_v_hdim, cfg.fr_margin,
cfg.policy_v_net, cfg.policy_v_net_param,
cfg.causal)
value_vs_net = VideoStateNet(cnn_feat_dim, cfg.value_v_hdim, cfg.fr_margin,
cfg.value_v_net, cfg.value_v_net_param,
cfg.causal)
policy_net = PolicyGaussian(MLP(state_dim + cfg.policy_v_hdim,
cfg.policy_hsize, cfg.policy_htype),
env_vis.sim_forward()
def load_take():
global traj_pred, traj_orig
algo_res, algo = algos[algo_ind]
if algo_res is None:
return
take = takes[take_ind]
print('%s ---------- %s' % (algo, take))
traj_pred = algo_res['traj_pred'][take]
traj_orig = algo_res['traj_orig'][take]
traj_pred = None
traj_orig = None
vis_model_file = 'assets/mujoco_models/%s.xml' % (args.multi_vis_model if args.multi else args.vis_model)
env_vis = HumanoidVisEnv(vis_model_file, 1, focus=not args.multi)
env_vis.set_custom_key_callback(key_callback)
takes = meta[args.data]
algos = [(em_res, 'ego mimic'), (sr_res, 'state reg'), (pp_res, 'path pose'), (vg_res, 'vgail')]
algo_ind = args.algo_ind
take_ind = args.take_ind
ss_ind = 0
mfr_int = 14
show_gt = False
load_take()
"""render or select part of the clip"""
T = 10
fr = 0
paused = False
stop = False
parser.add_argument('--data', default='test')
parser.add_argument('--show-noise', action='store_true', default=False)
parser.add_argument('--fail-safe', default='valuefs')
args = parser.parse_args()
cfg = Config(args.cfg, create_dirs=False)
dtype = torch.float64
torch.set_default_dtype(dtype)
np.random.seed(cfg.seed)
torch.manual_seed(cfg.seed)
torch.set_grad_enabled(False)
logger = create_logger(os.path.join(cfg.log_dir, 'log_eval.txt'))
"""environment"""
env = HumanoidEnv(cfg)
env.load_experts(cfg.takes[args.data], cfg.expert_feat_file, cfg.cnn_feat_file)
env_vis = HumanoidVisEnv(cfg.vis_model_file, 10)
env.seed(cfg.seed)
epos = None
cnn_feat_dim = env.cnn_feat[0].shape[-1]
actuators = env.model.actuator_names
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
body_qposaddr = get_body_qposaddr(env.model)
if args.fail_safe == 'naivefs':
env.set_fix_head_lb(0.3)
"""load policy net"""
policy_vs_net = VideoStateNet(cnn_feat_dim, cfg.policy_v_hdim, cfg.fr_margin,
cfg.policy_v_net, cfg.policy_v_net_param,
cfg.causal)
value_vs_net = VideoStateNet(cnn_feat_dim, cfg.value_v_hdim, cfg.fr_margin,
cfg.value_v_net, cfg.value_v_net_param,
dtype = torch.float64
torch.set_default_dtype(dtype)
np.random.seed(cfg.seed)
torch.manual_seed(cfg.seed)
torch.set_grad_enabled(False)
logger = create_logger(os.path.join(cfg.log_dir, 'log_eval_wild.txt'))
"""test data"""
cnn_feat_file = '%s/features/cnn_feat_%s.p' % (cfg.data_dir, args.test_feat)
cnn_feat_dict, _ = pickle.load(open(cnn_feat_file, 'rb'))
takes = list(cnn_feat_dict.keys())
"""environment"""
env = HumanoidEnv(cfg)
env_vis = HumanoidVisEnv('assets/mujoco_models/%s.xml' % args.vis_model, 10)
env.seed(cfg.seed)
cnn_feat_dim = cnn_feat_dict[takes[0]].shape[-1]
actuators = env.model.actuator_names
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
body_qposaddr = get_body_qposaddr(env.model)
"""load policy net"""
policy_vs_net = VideoForecastNet(cnn_feat_dim, state_dim, cfg.policy_v_hdim, cfg.fr_margin, cfg.policy_v_net,
cfg.policy_v_net_param, cfg.policy_s_hdim, cfg.policy_s_net)
value_vs_net = VideoForecastNet(cnn_feat_dim, state_dim, cfg.value_v_hdim, cfg.fr_margin, cfg.value_v_net,
cfg.value_v_net_param, cfg.value_s_hdim, cfg.value_s_net)
policy_net = PolicyGaussian(MLP(policy_vs_net.out_dim, cfg.policy_hsize, cfg.policy_htype), action_dim,
log_std=cfg.log_std, fix_std=cfg.fix_std)
value_net = Value(MLP(value_vs_net.out_dim, cfg.value_hsize, cfg.value_htype))
args = parser.parse_args()
cfg = Config(args.cfg, create_dirs=False)
if hasattr(cfg, 'random_cur_t'):
cfg.random_cur_t = False
dtype = torch.float64
torch.set_default_dtype(dtype)
np.random.seed(cfg.seed)
torch.manual_seed(cfg.seed)
torch.set_grad_enabled(False)
logger = create_logger(os.path.join(cfg.log_dir, 'log_eval.txt'))
"""environment"""
env = HumanoidEnv(cfg)
env.load_experts(cfg.takes[args.data], cfg.expert_feat_file, cfg.cnn_feat_file)
env_vis = HumanoidVisEnv(cfg.vis_model_file, 10)
env.seed(cfg.seed)
cnn_feat_dim = env.cnn_feat[0].shape[-1]
actuators = env.model.actuator_names
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
body_qposaddr = get_body_qposaddr(env.model)
"""load policy net"""
policy_vs_net = VideoForecastNet(cnn_feat_dim, state_dim, cfg.policy_v_hdim,
cfg.fr_margin, cfg.policy_v_net,
cfg.policy_v_net_param, cfg.policy_s_hdim,
cfg.policy_s_net)
value_vs_net = VideoForecastNet(cnn_feat_dim, state_dim, cfg.value_v_hdim,
cfg.fr_margin, cfg.value_v_net,
cfg.value_v_net_param, cfg.value_s_hdim,
cfg.value_s_net)
if args.tpv:
frame_folder = 'datasets/tpv/s_frames/%s' % take
frame_files = glob.glob(os.path.join(frame_folder, '*.jpg'))
frame_files.sort()
traj_tpv = [cv2.imread(file) for file in frame_files]
tpv_offset = meta['tpv_offset'][take]
print(len(traj_fpv),
len(traj_tpv) if args.tpv else 0, traj_pred.shape[0])
print('%s ---------- %s' % (algo, take))
traj_pred = None
traj_fpv = None
traj_tpv = None
vis_model_file = 'assets/mujoco_models/%s.xml' % (
args.multi_vis_model if args.multi else args.vis_model)
env_vis = HumanoidVisEnv(vis_model_file, 1)
env_vis.set_custom_key_callback(key_callback)
takes = list(ef_res['traj_pred'].keys())
algos = [(ef_res, 'ego forecast')]
algo_ind = 0
take_ind = args.take_ind
traj_ind = 0
tpv_offset = 0
ss_ind = 0
mfr_int = 10
load_take()
"""render or select part of the clip"""
cv2.namedWindow('FPV')
cv2.moveWindow('FPV', 150, 400)
cv2.namedWindow('TPV')
glfw.set_window_size(env_vis.viewer.window, 1000, 960)
def __init__(self, cfg, gt_comparison=False):
if gt_comparison:
self.vis_model = 'humanoid_1205_vis_single_v1'
self.vis_model_file = 'assets/mujoco_models/%s.xml' % self.vis_model
self.multi = False
self.env = HumanoidVisEnv(self.vis_model_file,
1,
focus=not self.multi)
else:
self.env = HumanoidEnv(cfg)
self.body2id = self.env.model._body_name2id
self.body_names = self.env.model.body_names[1:]
self.body_set = {
'LeftForeArm', 'RightForeArm', 'LeftHand', 'RightHand', 'LeftArm',
'RightArm', 'LeftUpLeg', 'RightUpLeg', 'LeftLeg', 'RightLeg',
'LeftFoot', 'RightFoot'
}
self.nbody = len(self.body_set)
self.body_filter = np.zeros((len(self.body_names), ), dtype=bool)
for body in self.body2id.keys():
if body in self.body_set:
self.body_filter[self.body2id[body] - 1] = True
self.body_names = [
self.body_names[i] for i in range(len(self.body_filter))
if self.body_filter[i]
]
self.body2id = {body: i for i, body in enumerate(self.body_names)}
self.conn = [('RightUpLeg', 'RightArm', (255, 255, 0)),
('RightArm', 'RightForeArm', (255, 191, 0)),
('RightForeArm', 'RightHand', (255, 191, 0)),
('RightUpLeg', 'RightLeg', (255, 64, 0.0)),
('RightLeg', 'RightFoot', (255, 64, 0.0)),
('LeftUpLeg', 'LeftArm', (0, 255, 128)),
('LeftArm', 'LeftForeArm', (0, 255, 255)),
('LeftForeArm', 'LeftHand', (0, 255, 255)),
('LeftUpLeg', 'LeftLeg', (0, 64, 255)),
('LeftLeg', 'LeftFoot', (0, 64, 255))]
self.joints_map = [(2, self.body2id['RightArm']),
(3, self.body2id['RightForeArm']),
(4, self.body2id['RightHand']),
(5, self.body2id['LeftArm']),
(6, self.body2id['LeftForeArm']),
(7, self.body2id['LeftHand']),
(9, self.body2id['RightUpLeg']),
(10, self.body2id['RightLeg']),
(11, self.body2id['RightFoot']),
(12, self.body2id['LeftUpLeg']),
(13, self.body2id['LeftLeg']),
(14, self.body2id['LeftFoot'])]
"""
Since number of joints between
- human_dynamics (25 joints)
- ours (14 joints)
are different, the following mapping variable defines
which joint to be evaluated.
0: indices for human_dynamics
1: indices for ours
The name after # shows corresponded name of human_dynamics
"""
self.joints_map1 = [
(8, self.body2id['RightArm']), # Right shoulder
(7, self.body2id['RightForeArm']), # Right elbow
(6, self.body2id['RightHand']), # Right wrist
(9, self.body2id['LeftArm']), # Left shoulder
(10, self.body2id['LeftForeArm']), # Left elbow
(11, self.body2id['LeftHand']), # Left wrist
(2, self.body2id['RightUpLeg']), # Right hip
(1, self.body2id['RightLeg']), # Right knee
(0, self.body2id['RightFoot']), # Right heel
(3, self.body2id['LeftUpLeg']), # Left hip
(4, self.body2id['LeftLeg']), # Left knee
(5, self.body2id['LeftFoot'])
] # Left heel