def main(args):
U.reset()
with open(os.path.join(args.logdir, 'hyps.json'), 'r') as f:
hyps = json.load(f)
train_args = namedtuple('Args', hyps.keys())(**hyps)
env_fn = make_env_fn(train_args)
model_fn = make_model_fn(train_args)
env = env_fn(0)
model = model_fn(env)
model.build('model', 1, 1)
model.sampler.build('model', 1, 1)
sess = U.make_session()
sess.__enter__()
U.initialize()
t = U.Experiment(args.logdir).load(args.ckpt)
ls = []
rs = []
for i in range(args.samples):
env.update_robot(model.sampler.sample(args.stochastic)[0])
l, r = eval_robot(args, env, model)
ls.append(l)
rs.append(r)
if not args.stochastic:
break
os.makedirs(os.path.join(args.logdir, 'eval'), exist_ok=True)
with open(os.path.join(args.logdir, 'eval', '{}.json'.format(t)),
'w') as f:
json.dump({'l': ls, 'r': rs}, f)
sess.__exit__(None, None, None)
def main(run_args):
with open(os.path.join(run_args.logdir, 'hyps.json'), 'r') as f:
hyps = json.load(f)
args = namedtuple('Args', hyps.keys())(**hyps)
env_fn = make_env_fn(args)
model_fn = make_model_fn(args)
time.sleep(1)
alg = Algorithm(run_args.logdir,
env_fn,
model_fn,
args.nenv,
args.rollout_length,
args.batchsize,
epochs_per_iter=args.epochs,
lr=args.lr,
momentum=args.momentum,
ent_coef=args.entcoeff,
gamma=args.gamma,
lambda_=args.lmbda,
clip_norm=args.grad_clip_norm,
clip_param=args.ppo_clip_param,
robot_lr=args.robot_lr,
robot_momentum=args.robot_momentum,
fixed_robot=args.fixed_robot,
steps_before_robot_update=args.steps_before_robot_update,
steps_after_robot_update=args.steps_after_robot_update,
chop_freq=args.chop_freq,
tmax=args.maxtimesteps)
alg.train(args.maxtimesteps, run_args.maxseconds, run_args.save_freq)
alg.close()
def main(args):
U.reset()
with open(os.path.join(args.logdir, 'hyps.json'), 'r') as f:
hyps = json.load(f)
train_args = namedtuple('Args', hyps.keys())(**hyps)
env_fn = make_env_fn(train_args)
model_fn = make_model_fn(train_args)
env = env_fn(0)
env.unwrapped.set_render_ground(True)
model = model_fn(env)
model.build('model', 1, 1)
model.sampler.build('model', 1, 1)
sess = U.make_session()
sess.__enter__()
U.initialize()
t = U.Experiment(args.logdir).load(args.ckpt)
# load mode of design distribution
env.update_robot(model.sampler.sample(stochastic=False)[0])
i = 0
if not args.save:
env.reset()
env.render()
else:
outdir = './video_tmp'
os.makedirs(outdir, exist_ok=True)
os.makedirs(os.path.join(args.logdir, 'video'), exist_ok=True)
for j in range(args.nepisodes):
done = False
ob = env.reset()
while not done:
if args.save:
rgb = env.render('rgb_array')
ac = model.actor.mode(ob[None])[0]
ob, rew, done, _ = env.step(ac)
if args.save:
imwrite(os.path.join(outdir, '{:05d}.png'.format(i)), rgb)
i += 1
if args.save:
outfile = str(t) + '.mp4'
sp.call([
'ffmpeg', '-r', '60', '-f', 'image2', '-i',
os.path.join(outdir, '%05d.png'), '-vcodec', 'libx264', '-pix_fmt',
'yuv420p',
os.path.join(outdir, 'out.mp4')
])
sp.call([
'mv',
os.path.join(outdir, 'out.mp4'),
os.path.join(args.logdir, 'video', outfile)
])
sp.call(['rm', '-rf', outdir])