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
on_traj = ef.add_next_obs(on_traj)
on_traj.register_epis()
off_traj.add_traj(on_traj)
total_epi += on_traj.num_epi
step = on_traj.num_step
total_step += step
result_dict = svg.train(
off_traj,
pol,
targ_pol,
qf,
targ_qf,
optim_pol,
optim_qf,
step,
args.batch_size,
args.tau,
args.gamma,
args.sampling,
)
rewards = [np.sum(epi['rews']) for epi in epis]
mean_rew = np.mean(rewards)
logger.record_results(args.log,
result_dict,
score_file,
total_epi,
step,
total_step,