def test_learning(self): ob_space = self.env.real_observation_space skill_space = self.env.skill_space ob_skill_space = self.env.observation_space ac_space = self.env.action_space ob_dim = ob_skill_space.shape[0] - 4 f_dim = ob_dim def discrim_f(x): return x pol_net = PolNet(ob_skill_space, ac_space) pol = GaussianPol(ob_skill_space, ac_space, pol_net) qf_net1 = QNet(ob_skill_space, ac_space) qf1 = DeterministicSAVfunc(ob_skill_space, ac_space, qf_net1) targ_qf_net1 = QNet(ob_skill_space, ac_space) targ_qf_net1.load_state_dict(qf_net1.state_dict()) targ_qf1 = DeterministicSAVfunc(ob_skill_space, ac_space, targ_qf_net1) qf_net2 = QNet(ob_skill_space, ac_space) qf2 = DeterministicSAVfunc(ob_skill_space, ac_space, qf_net2) targ_qf_net2 = QNet(ob_skill_space, ac_space) targ_qf_net2.load_state_dict(qf_net2.state_dict()) targ_qf2 = DeterministicSAVfunc(ob_skill_space, ac_space, targ_qf_net2) qfs = [qf1, qf2] targ_qfs = [targ_qf1, targ_qf2] log_alpha = nn.Parameter(torch.ones(())) high = np.array([np.finfo(np.float32).max]*f_dim) f_space = gym.spaces.Box(-high, high, dtype=np.float32) discrim_net = DiaynDiscrimNet( f_space, skill_space, h_size=100, discrim_f=discrim_f) discrim = DeterministicSVfunc(f_space, discrim_net) optim_pol = torch.optim.Adam(pol_net.parameters(), 1e-4) optim_qf1 = torch.optim.Adam(qf_net1.parameters(), 3e-4) optim_qf2 = torch.optim.Adam(qf_net2.parameters(), 3e-4) optim_qfs = [optim_qf1, optim_qf2] optim_alpha = torch.optim.Adam([log_alpha], 1e-4) optim_discrim = torch.optim.SGD(discrim.parameters(), lr=0.001, momentum=0.9) off_traj = Traj() sampler = EpiSampler(self.env, pol, num_parallel=1) epis = sampler.sample(pol, max_steps=200) on_traj = Traj() on_traj.add_epis(epis) on_traj = ef.add_next_obs(on_traj) on_traj = ef.compute_diayn_rews( on_traj, lambda x: diayn_sac.calc_rewards(x, 4, discrim)) on_traj.register_epis() off_traj.add_traj(on_traj) step = on_traj.num_step log_alpha = nn.Parameter(np.log(0.1)*torch.ones(())) # fix alpha result_dict = diayn_sac.train( off_traj, pol, qfs, targ_qfs, log_alpha, optim_pol, optim_qfs, optim_alpha, step, 128, 5e-3, 0.99, 1, discrim, 4, True) discrim_losses = diayn.train( discrim, optim_discrim, on_traj, 32, 100, 4) del sampler
def test_learning_rnn(self): def rew_func(next_obs, acs, mean_obs=0., std_obs=1., mean_acs=0., std_acs=1.): next_obs = next_obs * std_obs + mean_obs acs = acs * std_acs + mean_acs # Pendulum rews = -(torch.acos(next_obs[:, 0].clamp(min=-1, max=1))**2 + 0.1 * (next_obs[:, 2].clamp(min=-8, max=8)**2) + 0.001 * acs.squeeze(-1)**2) rews = rews.squeeze(0) return rews # init models dm_net = ModelNetLSTM(self.env.observation_space, self.env.action_space) dm = DeterministicSModel(self.env.observation_space, self.env.action_space, dm_net, rnn=True, data_parallel=False, parallel_dim=0) mpc_pol = MPCPol(self.env.observation_space, self.env.action_space, dm_net, rew_func, 1, 1, mean_obs=0., std_obs=1., mean_acs=0., std_acs=1., rnn=True) optim_dm = torch.optim.Adam(dm_net.parameters(), 1e-3) # sample with mpc policy sampler = EpiSampler(self.env, mpc_pol, num_parallel=1) epis = sampler.sample(mpc_pol, max_epis=1) traj = Traj() traj.add_epis(epis) traj = ef.add_next_obs(traj) traj = ef.compute_h_masks(traj) traj.register_epis() traj.add_traj(traj) # train result_dict = mpc.train_dm(traj, dm, optim_dm, epoch=1, batch_size=1) del sampler
def setUpClass(cls): cls.env = GymEnv('Pendulum-v0') pol = RandomPol(cls.env.observation_space, cls.env.action_space) sampler = EpiSampler(cls.env, pol, num_parallel=1) epis = sampler.sample(pol, max_steps=32) cls.traj = Traj() cls.traj.add_epis(epis) cls.traj.register_epis()
def test_learning(self): pol_net = PolNet(self.env.observation_space, self.env.action_space, h1=32, h2=32) pol = GaussianPol(self.env.observation_space, self.env.action_space, pol_net) vf_net = VNet(self.env.observation_space, h1=32, h2=32) vf = DeterministicSVfunc(self.env.observation_space, vf_net) sampler = EpiSampler(self.env, pol, num_parallel=1) optim_vf = torch.optim.Adam(vf_net.parameters(), 3e-4) epis = sampler.sample(pol, max_steps=32) traj = Traj() traj.add_epis(epis) traj = ef.compute_vs(traj, vf) traj = ef.compute_rets(traj, 0.99) traj = ef.compute_advs(traj, 0.99, 0.95) traj = ef.centerize_advs(traj) traj = ef.compute_h_masks(traj) traj.register_epis() result_dict = trpo.train(traj, pol, vf, optim_vf, 1, 24) del sampler
def test_learning(self): pol_net = PolNet(self.env.ob_space, self.env.ac_space, h1=32, h2=32) pol = GaussianPol(self.env.ob_space, self.env.ac_space, pol_net) targ_pol_net = PolNet(self.env.ob_space, self.env.ac_space, 32, 32) targ_pol_net.load_state_dict(pol_net.state_dict()) targ_pol = GaussianPol( self.env.ob_space, self.env.ac_space, targ_pol_net) qf_net = QNet(self.env.ob_space, self.env.ac_space, h1=32, h2=32) qf = DeterministicSAVfunc(self.env.ob_space, self.env.ac_space, qf_net) targ_qf_net = QNet(self.env.ob_space, self.env.ac_space, 32, 32) targ_qf_net.load_state_dict(targ_qf_net.state_dict()) targ_qf = DeterministicSAVfunc( self.env.ob_space, self.env.ac_space, targ_qf_net) sampler = EpiSampler(self.env, pol, num_parallel=1) optim_pol = torch.optim.Adam(pol_net.parameters(), 3e-4) optim_qf = torch.optim.Adam(qf_net.parameters(), 3e-4) epis = sampler.sample(pol, max_steps=32) traj = Traj() traj.add_epis(epis) traj = ef.add_next_obs(traj) traj.register_epis() result_dict = svg.train( traj, pol, targ_pol, qf, targ_qf, optim_pol, optim_qf, 1, 32, 0.01, 0.9, 1) del sampler
def test_learning(self): pol_net = PolNet(self.env.ob_space, self.env.ac_space, h1=32, h2=32) pol = CategoricalPol(self.env.ob_space, self.env.ac_space, pol_net) vf_net = VNet(self.env.ob_space, h1=32, h2=32) vf = DeterministicSVfunc(self.env.ob_space, vf_net) sampler = EpiSampler(self.env, pol, num_parallel=1) optim_pol = torch.optim.Adam(pol_net.parameters(), 3e-4) optim_vf = torch.optim.Adam(vf_net.parameters(), 3e-4) epis = sampler.sample(pol, max_steps=32) traj = Traj() traj.add_epis(epis) traj = ef.compute_vs(traj, vf) traj = ef.compute_rets(traj, 0.99) traj = ef.compute_advs(traj, 0.99, 0.95) traj = ef.centerize_advs(traj) traj = ef.compute_h_masks(traj) traj.register_epis() result_dict = ppo_clip.train(traj=traj, pol=pol, vf=vf, clip_param=0.2, optim_pol=optim_pol, optim_vf=optim_vf, epoch=1, batch_size=32) result_dict = ppo_kl.train(traj=traj, pol=pol, vf=vf, kl_beta=0.1, kl_targ=0.2, optim_pol=optim_pol, optim_vf=optim_vf, epoch=1, batch_size=32, max_grad_norm=10) del sampler
def test_learning(self): t_pol_net = PolNet(self.env.observation_space, self.env.action_space, h1=200, h2=100) s_pol_net = PolNet(self.env.observation_space, self.env.action_space, h1=190, h2=90) t_pol = GaussianPol( self.env.observation_space, self.env.action_space, t_pol_net) s_pol = GaussianPol( self.env.observation_space, self.env.action_space, s_pol_net) student_sampler = EpiSampler(self.env, s_pol, num_parallel=1) optim_pol = torch.optim.Adam(s_pol.parameters(), 3e-4) epis = student_sampler.sample(s_pol, max_steps=32) traj = Traj() traj.add_epis(epis) traj = ef.compute_h_masks(traj) traj.register_epis() result_dict = on_pol_teacher_distill.train( traj=traj, student_pol=s_pol, teacher_pol=t_pol, student_optim=optim_pol, epoch=1, batchsize=32) del student_sampler
def test_learning_rnn(self): pol_net = PolNetLSTM( self.env.observation_space, self.env.action_space, h_size=32, cell_size=32) pol = GaussianPol(self.env.observation_space, self.env.action_space, pol_net, rnn=True) vf_net = VNetLSTM(self.env.observation_space, h_size=32, cell_size=32) vf = DeterministicSVfunc(self.env.observation_space, vf_net, rnn=True) sampler = EpiSampler(self.env, pol, num_parallel=1) optim_pol = torch.optim.Adam(pol_net.parameters(), 3e-4) optim_vf = torch.optim.Adam(vf_net.parameters(), 3e-4) epis = sampler.sample(pol, max_steps=400) traj = Traj() traj.add_epis(epis) traj = ef.compute_vs(traj, vf) traj = ef.compute_rets(traj, 0.99) traj = ef.compute_advs(traj, 0.99, 0.95) traj = ef.centerize_advs(traj) traj = ef.compute_h_masks(traj) traj.register_epis() result_dict = ppo_clip.train(traj=traj, pol=pol, vf=vf, clip_param=0.2, optim_pol=optim_pol, optim_vf=optim_vf, epoch=1, batch_size=2) result_dict = ppo_kl.train(traj=traj, pol=pol, vf=vf, kl_beta=0.1, kl_targ=0.2, optim_pol=optim_pol, optim_vf=optim_vf, epoch=1, batch_size=2, max_grad_norm=20) del sampler
def train(self, epis): traj = Traj(ddp=True, traj_device=self.device) traj.add_epis(epis) traj = ef.compute_vs(traj, self.vf) traj = ef.compute_rets(traj, args.gamma) traj = ef.compute_advs(traj, args.gamma, args.lam) traj = ef.centerize_advs(traj) traj = ef.compute_h_masks(traj) traj.register_epis() result_dict = ppo_clip.train(traj=traj, pol=self.ddp_pol, vf=self.ddp_vf, clip_param=self.args.clip_param, optim_pol=self.optim_pol, optim_vf=self.optim_vf, epoch=self.args.epoch_per_iter, batch_size=self.args.batch_size, max_grad_norm=self.args.max_grad_norm, log_enable=self.rank == 0) result_dict["traj_num_step"] = traj.num_step result_dict["traj_num_epi"] = traj.num_epi return result_dict
def test_learning_rnn(self): pol_net = PolNetLSTM( self.env.observation_space, self.env.action_space, h_size=32, cell_size=32) pol = CategoricalPol( self.env.observation_space, self.env.action_space, pol_net, rnn=True) vf_net = VNetLSTM(self.env.observation_space, h_size=32, cell_size=32) vf = DeterministicSVfunc(self.env.observation_space, vf_net, rnn=True) sampler = EpiSampler(self.env, pol, num_parallel=1) optim_vf = torch.optim.Adam(vf_net.parameters(), 3e-4) epis = sampler.sample(pol, max_steps=400) traj = Traj() traj.add_epis(epis) traj = ef.compute_vs(traj, vf) traj = ef.compute_rets(traj, 0.99) traj = ef.compute_advs(traj, 0.99, 0.95) traj = ef.centerize_advs(traj) traj = ef.compute_h_masks(traj) traj.register_epis() result_dict = trpo.train(traj, pol, vf, optim_vf, 1, 2) del sampler
def test_learning(self): pol_net = PolNet(self.env.ob_space, self.env.ac_space, h1=32, h2=32) pol = GaussianPol(self.env.ob_space, self.env.ac_space, pol_net) qf_net1 = QNet(self.env.ob_space, self.env.ac_space) qf1 = DeterministicSAVfunc(self.env.ob_space, self.env.ac_space, qf_net1) targ_qf_net1 = QNet(self.env.ob_space, self.env.ac_space) targ_qf_net1.load_state_dict(qf_net1.state_dict()) targ_qf1 = DeterministicSAVfunc(self.env.ob_space, self.env.ac_space, targ_qf_net1) qf_net2 = QNet(self.env.ob_space, self.env.ac_space) qf2 = DeterministicSAVfunc(self.env.ob_space, self.env.ac_space, qf_net2) targ_qf_net2 = QNet(self.env.ob_space, self.env.ac_space) targ_qf_net2.load_state_dict(qf_net2.state_dict()) targ_qf2 = DeterministicSAVfunc(self.env.ob_space, self.env.ac_space, targ_qf_net2) qfs = [qf1, qf2] targ_qfs = [targ_qf1, targ_qf2] log_alpha = nn.Parameter(torch.zeros(())) sampler = EpiSampler(self.env, pol, num_parallel=1) optim_pol = torch.optim.Adam(pol_net.parameters(), 3e-4) optim_qf1 = torch.optim.Adam(qf_net1.parameters(), 3e-4) optim_qf2 = torch.optim.Adam(qf_net2.parameters(), 3e-4) optim_qfs = [optim_qf1, optim_qf2] optim_alpha = torch.optim.Adam([log_alpha], 3e-4) epis = sampler.sample(pol, max_steps=32) traj = Traj() traj.add_epis(epis) traj = ef.add_next_obs(traj) traj.register_epis() result_dict = sac.train( traj, pol, qfs, targ_qfs, log_alpha, optim_pol, optim_qfs, optim_alpha, 2, 32, 0.01, 0.99, 2, ) del sampler
def test_learning(self): pol_net = PolNet(self.env.observation_space, self.env.action_space, h1=32, h2=32, deterministic=True) noise = OUActionNoise(self.env.action_space) pol = DeterministicActionNoisePol(self.env.observation_space, self.env.action_space, pol_net, noise) targ_pol_net = PolNet(self.env.observation_space, self.env.action_space, 32, 32, deterministic=True) targ_pol_net.load_state_dict(pol_net.state_dict()) targ_noise = OUActionNoise(self.env.action_space) targ_pol = DeterministicActionNoisePol(self.env.observation_space, self.env.action_space, targ_pol_net, targ_noise) qf_net = QNet(self.env.observation_space, self.env.action_space, h1=32, h2=32) qf = DeterministicSAVfunc(self.env.observation_space, self.env.action_space, qf_net) targ_qf_net = QNet(self.env.observation_space, self.env.action_space, 32, 32) targ_qf_net.load_state_dict(targ_qf_net.state_dict()) targ_qf = DeterministicSAVfunc(self.env.observation_space, self.env.action_space, targ_qf_net) sampler = EpiSampler(self.env, pol, num_parallel=1) optim_pol = torch.optim.Adam(pol_net.parameters(), 3e-4) optim_qf = torch.optim.Adam(qf_net.parameters(), 3e-4) epis = sampler.sample(pol, max_steps=32) traj = Traj() traj.add_epis(epis) traj = ef.add_next_obs(traj) traj.register_epis() result_dict = ddpg.train(traj, pol, targ_pol, qf, targ_qf, optim_pol, optim_qf, 1, 32, 0.01, 0.9) del sampler
def test_learning(self): qf_net = QNet(self.env.observation_space, self.env.action_space, 32, 32) lagged_qf_net = QNet(self.env.observation_space, self.env.action_space, 32, 32) lagged_qf_net.load_state_dict(qf_net.state_dict()) targ_qf1_net = QNet(self.env.observation_space, self.env.action_space, 32, 32) targ_qf1_net.load_state_dict(qf_net.state_dict()) targ_qf2_net = QNet(self.env.observation_space, self.env.action_space, 32, 32) targ_qf2_net.load_state_dict(lagged_qf_net.state_dict()) qf = DeterministicSAVfunc(self.env.observation_space, self.env.action_space, qf_net) lagged_qf = DeterministicSAVfunc(self.env.observation_space, self.env.action_space, lagged_qf_net) targ_qf1 = CEMDeterministicSAVfunc(self.env.observation_space, self.env.action_space, targ_qf1_net, num_sampling=60, num_best_sampling=6, num_iter=2, multivari=False) targ_qf2 = DeterministicSAVfunc(self.env.observation_space, self.env.action_space, targ_qf2_net) pol = ArgmaxQfPol(self.env.observation_space, self.env.action_space, targ_qf1, eps=0.2) sampler = EpiSampler(self.env, pol, num_parallel=1) optim_qf = torch.optim.Adam(qf_net.parameters(), 3e-4) epis = sampler.sample(pol, max_steps=32) traj = Traj() traj.add_epis(epis) traj = ef.add_next_obs(traj) traj.register_epis() result_dict = qtopt.train(traj, qf, lagged_qf, targ_qf1, targ_qf2, optim_qf, 1000, 32, 0.9999, 0.995, 'mse') del sampler
def setUpClass(cls): env = GymEnv('Pendulum-v0') random_pol = RandomPol(cls.env.observation_space, cls.env.action_space) sampler = EpiSampler(cls.env, pol, num_parallel=1) epis = sampler.sample(pol, max_steps=32) traj = Traj() traj.add_epis(epis) traj.register_epis() cls.num_step = traj.num_step make_redis('localhost', '6379') cls.r = get_redis() cls.r.set('env', env) cls.r.set('traj', traj) pol_net = PolNet(env.observation_space, env.action_space) gpol = GaussianPol(env.observation_space, env.action_space, pol_net) pol_net = PolNet(env.observation_space, env.action_space, deterministic=True) dpol = DeterministicActionNoisePol( env.observation_space, env.action_space, pol_net) model_net = ModelNet(env.observation_space, env.action_space) mpcpol = MPCPol(env.observation_space, env.action_space, model_net, rew_func) q_net = QNet(env.observation_space, env.action_space) qfunc = DeterministicSAVfunc( env.observation_space, env.action_space, q_net) aqpol = ArgmaxQfPol(env.observation_space, env.action_space, qfunc) v_net = VNet(env.observation_space) vfunc = DeterministicSVfunc(env.observation_space, v_net) cls.r.set('gpol', cloudpickle.dumps(gpol)) cls.r.set('dpol', cloudpickle.dumps(dpol)) cls.r.set('mpcpol', cloudpickle.dumps(mpcpol)) cls.r.set('qfunc', cloudpickle.dumps(qfunc)) cls.r.set('aqpol', cloudpickle.dumps(aqpol)) cls.r.set('vfunc', cloudpickle.dumps(vfunc)) c2d = C2DEnv(env) pol_net = PolNet(c2d.observation_space, c2d.action_space) mcpol = MultiCategoricalPol( env.observation_space, env.action_space, pol_net) cls.r.set('mcpol', cloudpickle.dumps(mcpol))
print('optimizer') optim_qf = torch.optim.Adam(qf_net.parameters(), args.qf_lr) # epsilon-greedy policy print('Policy') pol = ArgmaxQfPol(flattend_observation_space, action_space, targ_qf1, eps=args.eps) # replay bufferからサンプリング? print('sampler') sampler = EpiSampler(env, pol, num_parallel=args.num_parallel, seed=args.seed) # off-policy experience. Traj=(s,a,r,s') off_traj = Traj(args.max_steps_off, traj_device='cpu') total_epi = 0 total_step = 0 total_grad_step = 0 # パラメータ更新回数 num_update_lagged = 0 # lagged netの更新回数 max_rew = -1000 print('start') while args.max_epis > total_epi: with measure('sample'): print('sampling') # policyにしたがって行動し、経験を貯める(env.stepをone_epiの__init__内で行っている) # off-policy epis = sampler.sample(pol, max_steps=args.max_steps_per_iter) with measure('train'):
optim_pol = torch.optim.Adam(s_pol_net.parameters(), args.pol_lr) total_epi = 0 total_step = 0 max_rew = -1e6 while args.max_epis > total_epi: with measure('sample'): if args.sampling_policy == 'teacher': epis = teacher_sampler.sample( t_pol, max_epis=args.max_epis_per_iter) else: epis = student_sampler.sample( s_pol, max_epis=args.max_epis_per_iter) with measure('train'): traj = Traj() traj.add_epis(epis) traj = ef.compute_h_masks(traj) traj.register_epis() result_dict = on_pol_teacher_distill.train( traj=traj, student_pol=s_pol, teacher_pol=t_pol, student_optim=optim_pol, epoch=args.epoch_per_iter, batchsize=args.batch_size) logger.log('Testing Student-policy') with measure('sample'): epis_measure = student_sampler.sample( s_pol, max_epis=args.max_epis_per_iter)
else: raise ValueError('Only Box, Discrete, and MultiDiscrete are supported') vf_net = VNet(observation_space) vf = DeterministicSVfunc(observation_space, vf_net, data_parallel=args.data_parallel) sampler = EpiSampler(env, pol, num_parallel=args.num_parallel, seed=args.seed) optim_pol = torch.optim.Adam(pol_net.parameters(), args.pol_lr) optim_vf = torch.optim.Adam(vf_net.parameters(), args.vf_lr) with open(os.path.join(args.expert_dir, args.expert_fname), 'rb') as f: expert_epis = pickle.load(f) expert_traj = Traj() expert_traj.add_epis(expert_epis) expert_traj = ef.add_next_obs(expert_traj) expert_traj.register_epis() expert_rewards = [np.sum(epi['rews']) for epi in expert_epis] expert_mean_rew = np.mean(expert_rewards) logger.log('expert_score={}'.format(expert_mean_rew)) logger.log('expert_num_epi={}'.format(expert_traj.num_epi)) total_epi = 0 total_step = 0 max_rew = -1e6 kl_beta = args.init_kl_beta if args.pretrain: with measure('bc pretrain'):
targ_pol = DeterministicActionNoisePol( ob_space, ac_space, targ_pol_net, targ_noise) qf_net = QNet(ob_space, ac_space, args.h1, args.h2) qf = DeterministicSAVfunc(ob_space, ac_space, qf_net) targ_qf_net = QNet(ob_space, ac_space, args.h1, args.h2) targ_qf_net.load_state_dict(qf_net.state_dict()) targ_qf = DeterministicSAVfunc(ob_space, ac_space, targ_qf_net) sampler = EpiSampler(env, pol, num_parallel=args.num_parallel, seed=args.seed) optim_pol = torch.optim.Adam(pol_net.parameters(), args.pol_lr) optim_qf = torch.optim.Adam(qf_net.parameters(), args.qf_lr) off_traj = Traj(args.max_steps_off) total_epi = 0 total_step = 0 max_rew = -1e6 while args.max_epis > total_epi: with measure('sample'): epis = sampler.sample(pol, max_steps=args.max_steps_per_iter) with measure('train'): on_traj = Traj() on_traj.add_epis(epis) on_traj = ef.add_next_obs(on_traj) on_traj.register_epis()
observation_space, action_space, targ_qf_net2, rnn=True) qfs = [qf1, qf2] targ_qfs = [targ_qf1, targ_qf2] log_alpha = nn.Parameter(torch.zeros((), device=device)) sampler = EpiSampler(env, pol, args.num_parallel, seed=args.seed) optim_pol = torch.optim.Adam(pol_net.parameters(), args.pol_lr) optim_qf1 = torch.optim.Adam(qf_net1.parameters(), args.qf_lr) optim_qf2 = torch.optim.Adam(qf_net2.parameters(), args.qf_lr) optim_qfs = [optim_qf1, optim_qf2] optim_alpha = torch.optim.Adam([log_alpha], args.pol_lr) off_traj = Traj(args.max_steps_off, traj_device='cpu') total_epi = 0 total_step = 0 max_rew = -1e6 while args.max_epis > total_epi: with measure('sample'): epis = sampler.sample(pol, max_steps=args.max_steps_per_iter) with measure('train'): on_traj = Traj(traj_device='cpu') on_traj.add_epis(epis) on_traj = ef.add_next_obs(on_traj) max_pri = on_traj.get_max_pri()
args.apex_keep_batchnorm_fp32, args.apex_sync_bn, args.apex_loss_scale, device_ids=[args.local_rank], output_device=args.local_rank) total_epi = 0 total_step = 0 max_rew = -1e6 kl_beta = args.init_kl_beta while args.max_epis > total_epi: with measure('sample', log_enable=rank == 0): if rank == 0: epis = sampler.sample(pol, max_steps=args.max_steps_per_iter) with measure('train', log_enable=rank == 0): traj = Traj(ddp=True, traj_device="cpu") if rank == 0: traj.add_epis(epis) traj = ef.compute_vs(traj, vf) traj = ef.compute_rets(traj, args.gamma) traj = ef.compute_advs(traj, args.gamma, args.lam) traj = ef.centerize_advs(traj) traj = ef.compute_h_masks(traj) traj.register_epis() traj = tf.sync(traj) if args.ppo_type == 'clip': result_dict = ppo_clip.train(traj=traj, pol=ddp_pol, vf=ddp_vf,
parallel_dim=1 if args.rnn else 0) sampler = EpiSampler(env, pol, num_parallel=args.num_parallel, seed=args.seed) optim_pol = torch.optim.Adam(pol_net.parameters(), args.pol_lr) optim_vf = torch.optim.Adam(vf_net.parameters(), args.vf_lr) total_epi = 0 total_step = 0 max_rew = -1e6 kl_beta = args.init_kl_beta while args.max_epis > total_epi: with measure('sample'): epis = sampler.sample(pol, max_steps=args.max_steps_per_iter) with measure('train'): traj = Traj() traj.add_epis(epis) traj = ef.compute_vs(traj, vf) traj = ef.compute_rets(traj, args.gamma) traj = ef.compute_advs(traj, args.gamma, args.lam) traj = ef.centerize_advs(traj) traj = ef.compute_h_masks(traj) traj.register_epis() if args.data_parallel: pol.dp_run = True vf.dp_run = True if args.ppo_type == 'clip': result_dict = ppo_clip.train(traj=traj,
###################### ### Model-Based RL ### ###################### ### Prepare the dataset D_RAND ### # Performing rollouts to collect training data rand_sampler = EpiSampler(env, random_pol, num_parallel=args.num_parallel, seed=args.seed) epis = rand_sampler.sample(random_pol, max_epis=args.num_random_rollouts) epis = add_noise_to_init_obs(epis, args.noise_to_init_obs) traj = Traj(traj_device='cpu') traj.add_epis(epis) traj = ef.add_next_obs(traj) traj = ef.compute_h_masks(traj) # obs, next_obs, and acs should become mean 0, std 1 traj, mean_obs, std_obs, mean_acs, std_acs = ef.normalize_obs_and_acs(traj) traj.register_epis() del rand_sampler ### Train Dynamics Model ### # initialize dynamics model and mpc policy if args.rnn: dm_net = ModelNetLSTM(ob_space, ac_space) else:
elif isinstance(action_space, gym.spaces.MultiDiscrete): pol = MultiCategoricalPol(observation_space, action_space, pol_net, data_parallel=args.data_parallel) else: raise ValueError('Only Box, Discrete, and MultiDiscrete are supported') sampler = EpiSampler(env, pol, num_parallel=args.num_parallel, seed=args.seed) optim_pol = torch.optim.Adam(pol_net.parameters(), args.pol_lr) with open(os.path.join(args.expert_dir, args.expert_fname), 'rb') as f: expert_epis = pickle.load(f) train_epis, test_epis = ef.train_test_split(expert_epis, train_size=args.train_size) train_traj = Traj() train_traj.add_epis(train_epis) train_traj.register_epis() test_traj = Traj() test_traj.add_epis(test_epis) test_traj.register_epis() expert_rewards = [np.sum(epi['rews']) for epi in expert_epis] expert_mean_rew = np.mean(expert_rewards) logger.log('expert_score={}'.format(expert_mean_rew)) logger.log('num_train_epi={}'.format(train_traj.num_epi)) max_rew = -1e6 for curr_epoch in range(args.epoch): if args.data_parallel: pol.dp_run = True
f_space, skill_space, h_size=args.discrim_h_size, discrim_f=discrim_f).to(device) discrim = DeterministicSVfunc( f_space, discrim_net, rnn=False, data_parallel=False, parallel_dim=0) # set optimizer to both models optim_pol = torch.optim.Adam(pol_net.parameters(), args.pol_lr) optim_qf1 = torch.optim.Adam(qf_net1.parameters(), args.qf_lr) optim_qf2 = torch.optim.Adam(qf_net2.parameters(), args.qf_lr) optim_qfs = [optim_qf1, optim_qf2] optim_alpha = torch.optim.Adam([log_alpha], args.pol_lr) optim_discrim = torch.optim.SGD(discrim.parameters( ), lr=args.discrim_lr, momentum=args.discrim_momentum) off_traj = Traj() sampler = EpiSampler( env, pol, num_parallel=args.num_parallel, seed=args.seed) if not os.path.exists(args.log): os.mkdir(args.log) os.mkdir(args.log+'/models') score_file = os.path.join(args.log, 'progress.csv') logger.add_tabular_output(score_file) np.random.seed(args.seed) torch.manual_seed(args.seed) # counter and record for loop total_epi = 0 total_step = 0 mean_rew = 0
def test_learning(self): pol_net = PolNetLSTM( self.env.observation_space, self.env.action_space, h_size=32, cell_size=32) pol = GaussianPol(self.env.observation_space, self.env.action_space, pol_net, rnn=True) qf_net1 = QNetLSTM(self.env.observation_space, self.env.action_space, h_size=32, cell_size=32) qf1 = DeterministicSAVfunc( self.env.observation_space, self.env.action_space, qf_net1, rnn=True) targ_qf_net1 = QNetLSTM( self.env.observation_space, self.env.action_space, h_size=32, cell_size=32) targ_qf_net1.load_state_dict(qf_net1.state_dict()) targ_qf1 = DeterministicSAVfunc( self.env.observation_space, self.env.action_space, targ_qf_net1, rnn=True) qf_net2 = QNetLSTM(self.env.observation_space, self.env.action_space, h_size=32, cell_size=32) qf2 = DeterministicSAVfunc( self.env.observation_space, self.env.action_space, qf_net2, rnn=True) targ_qf_net2 = QNetLSTM( self.env.observation_space, self.env.action_space, h_size=32, cell_size=32) targ_qf_net2.load_state_dict(qf_net2.state_dict()) targ_qf2 = DeterministicSAVfunc( self.env.observation_space, self.env.action_space, targ_qf_net2, rnn=True) qfs = [qf1, qf2] targ_qfs = [targ_qf1, targ_qf2] log_alpha = nn.Parameter(torch.zeros(())) sampler = EpiSampler(self.env, pol, num_parallel=1) optim_pol = torch.optim.Adam(pol_net.parameters(), 3e-4) optim_qf1 = torch.optim.Adam(qf_net1.parameters(), 3e-4) optim_qf2 = torch.optim.Adam(qf_net2.parameters(), 3e-4) optim_qfs = [optim_qf1, optim_qf2] optim_alpha = torch.optim.Adam([log_alpha], 3e-4) epis = sampler.sample(pol, max_steps=32) traj = Traj() traj.add_epis(epis) traj = ef.add_next_obs(traj) max_pri = traj.get_max_pri() traj = ef.set_all_pris(traj, max_pri) traj = ef.compute_seq_pris(traj, 4) traj = ef.compute_h_masks(traj) for i in range(len(qfs)): traj = ef.compute_hs( traj, qfs[i], hs_name='q_hs'+str(i), input_acs=True) traj = ef.compute_hs( traj, targ_qfs[i], hs_name='targ_q_hs'+str(i), input_acs=True) traj.register_epis() result_dict = r2d2_sac.train( traj, pol, qfs, targ_qfs, log_alpha, optim_pol, optim_qfs, optim_alpha, 2, 32, 4, 2, 0.01, 0.99, 2, ) del sampler
def test_learning(self): pol_net = PolNet(self.env.observation_space, self.env.action_space, h1=32, h2=32) pol = GaussianPol(self.env.observation_space, self.env.action_space, pol_net) vf_net = VNet(self.env.observation_space) vf = DeterministicSVfunc(self.env.observation_space, vf_net) rewf_net = VNet(self.env.observation_space, h1=32, h2=32) rewf = DeterministicSVfunc(self.env.observation_space, rewf_net) shaping_vf_net = VNet(self.env.observation_space, h1=32, h2=32) shaping_vf = DeterministicSVfunc( self.env.observation_space, shaping_vf_net) sampler = EpiSampler(self.env, pol, num_parallel=1) optim_vf = torch.optim.Adam(vf_net.parameters(), 3e-4) optim_discrim = torch.optim.Adam( list(rewf_net.parameters()) + list(shaping_vf_net.parameters()), 3e-4) with open(os.path.join('data/expert_epis', 'Pendulum-v0_2epis.pkl'), 'rb') as f: expert_epis = pickle.load(f) expert_traj = Traj() expert_traj.add_epis(expert_epis) expert_traj = ef.add_next_obs(expert_traj) expert_traj.register_epis() epis = sampler.sample(pol, max_steps=32) agent_traj = Traj() agent_traj.add_epis(epis) agent_traj = ef.add_next_obs(agent_traj) agent_traj = ef.compute_pseudo_rews( agent_traj, rew_giver=rewf, state_only=True) agent_traj = ef.compute_vs(agent_traj, vf) agent_traj = ef.compute_rets(agent_traj, 0.99) agent_traj = ef.compute_advs(agent_traj, 0.99, 0.95) agent_traj = ef.centerize_advs(agent_traj) agent_traj = ef.compute_h_masks(agent_traj) agent_traj.register_epis() result_dict = airl.train(agent_traj, expert_traj, pol, vf, optim_vf, optim_discrim, rewf=rewf, shaping_vf=shaping_vf, rl_type='trpo', epoch=1, batch_size=32, discrim_batch_size=32, discrim_step=1, pol_ent_beta=1e-3, gamma=0.99) del sampler
def test_learning(self): pol_net = PolNet(self.env.observation_space, self.env.action_space, h1=32, h2=32) pol = GaussianPol(self.env.observation_space, self.env.action_space, pol_net) sampler = EpiSampler(self.env, pol, num_parallel=1) optim_pol = torch.optim.Adam(pol_net.parameters(), 3e-4) with open(os.path.join('data/expert_epis', 'Pendulum-v0_2epis.pkl'), 'rb') as f: expert_epis = pickle.load(f) train_epis, test_epis = ef.train_test_split( expert_epis, train_size=0.7) train_traj = Traj() train_traj.add_epis(train_epis) train_traj.register_epis() test_traj = Traj() test_traj.add_epis(test_epis) test_traj.register_epis() result_dict = behavior_clone.train( train_traj, pol, optim_pol, 256 ) del sampler
pol, num_parallel=args.num_parallel, seed=args.seed) optim_pol = torch.optim.Adam(pol_net.parameters(), args.pol_lr) optim_vf = torch.optim.Adam(vf_net.parameters(), args.vf_lr) total_epi = 0 total_step = 0 max_rew = -1e6 while args.max_epis > total_epi: with measure('sample'): epis1 = sampler1.sample(pol, max_epis=args.max_epis_per_iter) epis2 = sampler2.sample(pol, max_epis=args.max_epis_per_iter) with measure('train'): traj1 = Traj() traj2 = Traj() traj1.add_epis(epis1) traj1 = ef.compute_vs(traj1, vf) traj1 = ef.compute_rets(traj1, args.gamma) traj1 = ef.compute_advs(traj1, args.gamma, args.lam) traj1 = ef.centerize_advs(traj1) traj1 = ef.compute_h_masks(traj1) traj1.register_epis() traj2.add_epis(epis2) traj2 = ef.compute_vs(traj2, vf) traj2 = ef.compute_rets(traj2, args.gamma) traj2 = ef.compute_advs(traj2, args.gamma, args.lam) traj2 = ef.centerize_advs(traj2)
def test_add_traj(self): new_traj = Traj() new_traj.add_traj(self.traj) assert new_traj.num_epi == self.traj.num_epi assert new_traj.num_step == self.traj.num_step
def test_learning(self): pol_net = PolNet(self.env.ob_space, self.env.ac_space, h1=32, h2=32) pol = GaussianPol(self.env.ob_space, self.env.ac_space, pol_net) vf_net = VNet(self.env.ob_space) vf = DeterministicSVfunc(self.env.ob_space, vf_net) discrim_net = DiscrimNet( self.env.ob_space, self.env.ac_space, h1=32, h2=32) discrim = DeterministicSAVfunc( self.env.ob_space, self.env.ac_space, discrim_net) sampler = EpiSampler(self.env, pol, num_parallel=1) optim_vf = torch.optim.Adam(vf_net.parameters(), 3e-4) optim_discrim = torch.optim.Adam(discrim_net.parameters(), 3e-4) with open(os.path.join('data/expert_epis', 'Pendulum-v0_2epis.pkl'), 'rb') as f: expert_epis = pickle.load(f) expert_traj = Traj() expert_traj.add_epis(expert_epis) expert_traj.register_epis() epis = sampler.sample(pol, max_steps=32) agent_traj = Traj() agent_traj.add_epis(epis) agent_traj = ef.compute_pseudo_rews(agent_traj, discrim) agent_traj = ef.compute_vs(agent_traj, vf) agent_traj = ef.compute_rets(agent_traj, 0.99) agent_traj = ef.compute_advs(agent_traj, 0.99, 0.95) agent_traj = ef.centerize_advs(agent_traj) agent_traj = ef.compute_h_masks(agent_traj) agent_traj.register_epis() result_dict = gail.train(agent_traj, expert_traj, pol, vf, discrim, optim_vf, optim_discrim, rl_type='trpo', epoch=1, batch_size=32, discrim_batch_size=32, discrim_step=1, pol_ent_beta=1e-3, discrim_ent_beta=1e-5) del sampler