d5b0cd8e7960c247bb7c5b7c832358f8831780fb,ch15/03_train_trpo.py,,,#,90

Before Change

    env = gym.make(args.env)
    test_env = gym.make(args.env)

    net_act = model.ModelActor(env.observation_space.shape[0], env.action_space.shape[0])
    net_crt = model.ModelCritic(env.observation_space.shape[0])
    if args.cuda:
        net_act.cuda()
        net_crt.cuda()
    print(net_act)
    print(net_crt)

    writer = SummaryWriter(comment="-trpo_" + args.name)
    agent = model.AgentA2C(net_act, cuda=args.cuda)
    exp_source = ptan.experience.ExperienceSource(env, agent, steps_count=1)

    opt_crt = optim.Adam(net_crt.parameters(), lr=LEARNING_RATE_CRITIC)

    trajectory = []
    best_reward = None
    with ptan.common.utils.RewardTracker(writer) as tracker:
        for step_idx, exp in enumerate(exp_source):
            rewards_steps = exp_source.pop_rewards_steps()
            if rewards_steps:
                rewards, steps = zip(*rewards_steps)
                writer.add_scalar("episode_steps", np.mean(steps), step_idx)
                tracker.reward(np.mean(rewards), step_idx)

            if step_idx % TEST_ITERS == 0:
                ts = time.time()
                rewards, steps = test_net(net_act, test_env, cuda=args.cuda)
                print("Test done in %.2f sec, reward %.3f, steps %d" % (
                    time.time() - ts, rewards, steps))
                writer.add_scalar("test_reward", rewards, step_idx)
                writer.add_scalar("test_steps", steps, step_idx)
                if best_reward is None or best_reward < rewards:
                    if best_reward is not None:
                        print("Best reward updated: %.3f -> %.3f" % (best_reward, rewards))
                        name = "best_%+.3f_%d.dat" % (rewards, step_idx)
                        fname = os.path.join(save_path, name)
                        torch.save(net_act.state_dict(), fname)
                    best_reward = rewards

            trajectory.append(exp)
            if len(trajectory) < TRAJECTORY_SIZE:
                continue

            traj_states = [t[0].state for t in trajectory]
            traj_actions = [t[0].action for t in trajectory]
            traj_states_v = Variable(torch.from_numpy(np.array(traj_states, dtype=np.float32)))
            traj_actions_v = Variable(torch.from_numpy(np.array(traj_actions, dtype=np.float32)))
            if args.cuda:
                traj_states_v = traj_states_v.cuda()
                traj_actions_v = traj_actions_v.cuda()

            traj_adv_v, traj_ref_v = calc_adv_ref(trajectory, net_crt, traj_states_v, cuda=args.cuda)
            mu_v = net_act(traj_states_v)
            old_logprob_v = calc_logprob(mu_v, net_act.logstd, traj_actions_v)

            // normalize advantages
            traj_adv_v = (traj_adv_v - torch.mean(traj_adv_v)) / torch.std(traj_adv_v)

            // drop last entry from the trajectory, an our adv and ref value calculated without it
            trajectory = trajectory[:-1]
            old_logprob_v = old_logprob_v[:-1].detach()
            traj_states_v = traj_states_v[:-1]
            traj_actions_v = traj_actions_v[:-1]
            sum_loss_value = 0.0
            sum_loss_policy = 0.0
            count_steps = 0

            // critic step
            opt_crt.zero_grad()
            value_v = net_crt(traj_states_v)
            loss_value_v = F.mse_loss(value_v, traj_ref_v)
            loss_value_v.backward()
            opt_crt.step()

After Change

    env = gym.make(args.env)
    test_env = gym.make(args.env)

    net_act = model.ModelActor(env.observation_space.shape[0], env.action_space.shape[0]).to(device)
    net_crt = model.ModelCritic(env.observation_space.shape[0]).to(device)
    print(net_act)
    print(net_crt)

    writer = SummaryWriter(comment="-trpo_" + args.name)
    agent = model.AgentA2C(net_act, device=device)
    exp_source = ptan.experience.ExperienceSource(env, agent, steps_count=1)

    opt_crt = optim.Adam(net_crt.parameters(), lr=LEARNING_RATE_CRITIC)

    trajectory = []
    best_reward = None
    with ptan.common.utils.RewardTracker(writer) as tracker:
        for step_idx, exp in enumerate(exp_source):
            rewards_steps = exp_source.pop_rewards_steps()
            if rewards_steps:
                rewards, steps = zip(*rewards_steps)
                writer.add_scalar("episode_steps", np.mean(steps), step_idx)
                tracker.reward(np.mean(rewards), step_idx)

            if step_idx % TEST_ITERS == 0:
                ts = time.time()
                rewards, steps = test_net(net_act, test_env, device=device)
                print("Test done in %.2f sec, reward %.3f, steps %d" % (
                    time.time() - ts, rewards, steps))
                writer.add_scalar("test_reward", rewards, step_idx)
                writer.add_scalar("test_steps", steps, step_idx)
                if best_reward is None or best_reward < rewards:
                    if best_reward is not None:
                        print("Best reward updated: %.3f -> %.3f" % (best_reward, rewards))
                        name = "best_%+.3f_%d.dat" % (rewards, step_idx)
                        fname = os.path.join(save_path, name)
                        torch.save(net_act.state_dict(), fname)
                    best_reward = rewards

            trajectory.append(exp)
            if len(trajectory) < TRAJECTORY_SIZE:
                continue

            traj_states = [t[0].state for t in trajectory]
            traj_actions = [t[0].action for t in trajectory]
            traj_states_v = torch.FloatTensor(traj_states).to(device)
            traj_actions_v = torch.FloatTensor(traj_actions).to(device)
            traj_adv_v, traj_ref_v = calc_adv_ref(trajectory, net_crt, traj_states_v, device=device)
            mu_v = net_act(traj_states_v)
            old_logprob_v = calc_logprob(mu_v, net_act.logstd, traj_actions_v)

            // normalize advantages
            traj_adv_v = (traj_adv_v - torch.mean(traj_adv_v)) / torch.std(traj_adv_v)

            // drop last entry from the trajectory, an our adv and ref value calculated without it
            trajectory = trajectory[:-1]
            old_logprob_v = old_logprob_v[:-1].detach()
            traj_states_v = traj_states_v[:-1]
            traj_actions_v = traj_actions_v[:-1]
            sum_loss_value = 0.0
            sum_loss_policy = 0.0
            count_steps = 0

            // critic step
            opt_crt.zero_grad()
            value_v = net_crt(traj_states_v)
            loss_value_v = F.mse_loss(value_v.squeeze(-1), traj_ref_v)
            loss_value_v.backward()
            opt_crt.step()

Instances