def run_ddpg()

in salina_examples/rl/ddpg/ddpg.py [0:0]


def run_ddpg(q_agent, action_agent, logger, cfg):
    q_agent.set_name("q_agent")
    action_agent.set_name("action_agent")
    modulo = 1

    env_agent = AutoResetGymAgent(
        get_class(cfg.algorithm.env),
        get_arguments(cfg.algorithm.env),
        n_envs=int(cfg.algorithm.n_envs / cfg.algorithm.n_processes),
    )

    q_target_agent = copy.deepcopy(q_agent)  # Create target agent
    action_target_agent = copy.deepcopy(action_agent)

    acq_agent = TemporalAgent(Agents(env_agent, copy.deepcopy(action_agent)))
    acq_remote_agent, acq_workspace = NRemoteAgent.create(
        acq_agent,
        num_processes=cfg.algorithm.n_processes,
        t=0,
        n_steps=cfg.algorithm.n_timesteps,
        epsilon=1.0,
    )
    acq_remote_agent.seed(cfg.algorithm.env_seed)

    # == Setting up the training agents
    train_temporal_q_agent = TemporalAgent(q_agent)
    train_temporal_action_agent = TemporalAgent(action_agent)
    train_temporal_q_target_agent = TemporalAgent(q_target_agent)
    train_temporal_action_target_agent = TemporalAgent(action_target_agent)

    train_temporal_q_agent.to(cfg.algorithm.loss_device)
    train_temporal_action_agent.to(cfg.algorithm.loss_device)
    train_temporal_q_target_agent.to(cfg.algorithm.loss_device)
    train_temporal_action_target_agent.to(cfg.algorithm.loss_device)

    replay_buffer = ReplayBuffer(cfg.algorithm.buffer_size)

    acq_remote_agent(
        acq_workspace,
        t=0,
        n_steps=cfg.algorithm.n_timesteps,
        epsilon=cfg.algorithm.action_noise,
    )
    replay_buffer.put(acq_workspace, time_size=cfg.algorithm.buffer_time_size)
    logger.message("[DDQN] Initializing replay buffer")
    while replay_buffer.size() < cfg.algorithm.initial_buffer_size:
        acq_workspace.copy_n_last_steps(cfg.algorithm.overlapping_timesteps)
        acq_remote_agent(
            acq_workspace,
            t=cfg.algorithm.overlapping_timesteps,
            n_steps=cfg.algorithm.n_timesteps - cfg.algorithm.overlapping_timesteps,
            epsilon=cfg.algorithm.action_noise,
        )
        replay_buffer.put(acq_workspace, time_size=cfg.algorithm.buffer_time_size)

    print("[DDPG] Learning")
    _epoch_start_time = time.time()
    optimizer_args = get_arguments(cfg.algorithm.optimizer)
    optimizer_q = get_class(cfg.algorithm.optimizer)(
        q_agent.parameters(), **optimizer_args
    )

    optimizer_action = get_class(cfg.algorithm.optimizer)(
        action_agent.parameters(), **optimizer_args
    )

    for epoch in range(cfg.algorithm.max_epoch):
        for a in acq_remote_agent.get_by_name("action_agent"):
            a.load_state_dict(_state_dict(action_agent, "cpu"))

        acq_workspace.copy_n_last_steps(cfg.algorithm.overlapping_timesteps)
        acq_remote_agent(
            acq_workspace,
            t=cfg.algorithm.overlapping_timesteps,
            n_steps=cfg.algorithm.n_timesteps - cfg.algorithm.overlapping_timesteps,
            epsilon=cfg.algorithm.action_noise,
        )
        replay_buffer.put(acq_workspace, time_size=cfg.algorithm.buffer_time_size)

        # Get cumulated reward over terminated episodes
        done, creward = acq_workspace["env/done", "env/cumulated_reward"]
        creward = creward[done]
        if creward.size()[0] > 0:
            logger.add_scalar("monitor/reward", creward.mean().item(), epoch)
        logger.add_scalar("monitor/replay_buffer_size", replay_buffer.size(), epoch)

        # Inner loop to minimize the TD
        for inner_epoch in range(cfg.algorithm.inner_epochs):
            batch_size = cfg.algorithm.batch_size
            replay_workspace = replay_buffer.get(batch_size).to(
                cfg.algorithm.loss_device
            )
            done, reward = replay_workspace["env/done", "env/reward"]

            train_temporal_q_agent(
                replay_workspace,
                t=0,
                n_steps=cfg.algorithm.buffer_time_size,
                detach_action=True,
            )
            q = replay_workspace["q"].squeeze(-1)

            with torch.no_grad():
                train_temporal_action_target_agent(
                    replay_workspace,
                    t=0,
                    n_steps=cfg.algorithm.buffer_time_size,
                    epsilon=0.0,
                )  # epsilon = cfg.algorithm.target_noise
                train_temporal_q_target_agent(
                    replay_workspace, t=0, n_steps=cfg.algorithm.buffer_time_size
                )
                q_target = replay_workspace["q"]

            q_target = q_target.squeeze(-1)
            target = (
                reward[1:]
                + cfg.algorithm.discount_factor
                * (1.0 - done[1:].float())
                * q_target[1:]
            )

            td = q[:-1] - target
            error = td ** 2

            # Add burning for the first timesteps in the trajectories => no gradient (commonly done in R2D2)
            burning = torch.zeros_like(error)
            burning[cfg.algorithm.burning_timesteps :] = 1.0
            error = error * burning
            loss = error.mean()
            logger.add_scalar("loss/td_loss", error.mean().item(), epoch)
            optimizer_q.zero_grad()
            loss.backward()

            if cfg.algorithm.clip_grad > 0:
                n = torch.nn.utils.clip_grad_norm_(
                    q_agent.parameters(), cfg.algorithm.clip_grad
                )
                logger.add_scalar("monitor/grad_norm_q", n.item(), epoch)

            optimizer_q.step()

            # Update policy
            train_temporal_action_agent(
                replay_workspace,
                epsilon=0.0,
                t=0,
                n_steps=cfg.algorithm.buffer_time_size,
            )
            train_temporal_q_agent(
                replay_workspace, t=0, n_steps=cfg.algorithm.buffer_time_size
            )
            q = replay_workspace["q"].squeeze(-1)
            burning = torch.zeros_like(q)
            burning[cfg.algorithm.burning_timesteps :] = 1.0
            q = q * burning
            q = q[:-1]
            optimizer_action.zero_grad()
            loss = -q.mean()
            loss.backward()

            if cfg.algorithm.clip_grad > 0:
                n = torch.nn.utils.clip_grad_norm_(
                    action_agent.parameters(), cfg.algorithm.clip_grad
                )
                logger.add_scalar("monitor/grad_norm_action", n.item(), epoch)

            logger.add_scalar("loss/q_loss", loss.item(), epoch)
            optimizer_action.step()

            # Soft update of target policy and q function
            tau = cfg.algorithm.update_target_tau
            soft_update_params(q_agent, q_target_agent, tau)
            soft_update_params(action_agent, action_target_agent, tau)