Beispiel #1
0
class AWACTrainer(TorchTrainer):
    def __init__(
        self,
        env,
        policy,
        qf1,
        qf2,
        target_qf1,
        target_qf2,
        buffer_policy=None,
        discount=0.99,
        reward_scale=1.0,
        beta=1.0,
        beta_schedule_kwargs=None,
        policy_lr=1e-3,
        qf_lr=1e-3,
        policy_weight_decay=0,
        q_weight_decay=0,
        optimizer_class=optim.Adam,
        soft_target_tau=1e-2,
        target_update_period=1,
        plotter=None,
        render_eval_paths=False,
        use_automatic_entropy_tuning=True,
        target_entropy=None,
        bc_num_pretrain_steps=0,
        q_num_pretrain1_steps=0,
        q_num_pretrain2_steps=0,
        bc_batch_size=128,
        alpha=1.0,
        policy_update_period=1,
        q_update_period=1,
        weight_loss=True,
        compute_bc=True,
        use_awr_update=True,
        use_reparam_update=False,
        bc_weight=0.0,
        rl_weight=1.0,
        reparam_weight=1.0,
        awr_weight=1.0,
        post_pretrain_hyperparams=None,
        post_bc_pretrain_hyperparams=None,
        awr_use_mle_for_vf=False,
        vf_K=1,
        awr_sample_actions=False,
        buffer_policy_sample_actions=False,
        awr_min_q=False,
        brac=False,
        reward_transform_class=None,
        reward_transform_kwargs=None,
        terminal_transform_class=None,
        terminal_transform_kwargs=None,
        pretraining_logging_period=1000,
        train_bc_on_rl_buffer=False,
        use_automatic_beta_tuning=False,
        beta_epsilon=1e-10,
        normalize_over_batch=True,
        normalize_over_state="advantage",
        Z_K=10,
        clip_score=None,
        validation_qlearning=False,
        mask_positive_advantage=False,
        buffer_policy_reset_period=-1,
        num_buffer_policy_train_steps_on_reset=100,
        advantage_weighted_buffer_loss=True,
    ):
        super().__init__()
        self.env = env
        self.policy = policy
        self.qf1 = qf1
        self.qf2 = qf2
        self.target_qf1 = target_qf1
        self.target_qf2 = target_qf2
        self.buffer_policy = buffer_policy
        self.soft_target_tau = soft_target_tau
        self.target_update_period = target_update_period

        self.use_awr_update = use_awr_update
        self.use_automatic_entropy_tuning = use_automatic_entropy_tuning
        if self.use_automatic_entropy_tuning:
            if target_entropy:
                self.target_entropy = target_entropy
            else:
                self.target_entropy = -np.prod(
                    self.env.action_space.shape).item(
                    )  # heuristic value from Tuomas
            self.log_alpha = ptu.zeros(1, requires_grad=True)
            self.alpha_optimizer = optimizer_class(
                [self.log_alpha],
                lr=policy_lr,
            )

        self.awr_use_mle_for_vf = awr_use_mle_for_vf
        self.vf_K = vf_K
        self.awr_sample_actions = awr_sample_actions
        self.awr_min_q = awr_min_q

        self.plotter = plotter
        self.render_eval_paths = render_eval_paths

        self.qf_criterion = nn.MSELoss()
        self.vf_criterion = nn.MSELoss()

        self.optimizers = {}

        self.policy_optimizer = optimizer_class(
            self.policy.parameters(),
            weight_decay=policy_weight_decay,
            lr=policy_lr,
        )
        self.optimizers[self.policy] = self.policy_optimizer
        self.qf1_optimizer = optimizer_class(
            self.qf1.parameters(),
            weight_decay=q_weight_decay,
            lr=qf_lr,
        )
        self.qf2_optimizer = optimizer_class(
            self.qf2.parameters(),
            weight_decay=q_weight_decay,
            lr=qf_lr,
        )

        if buffer_policy and train_bc_on_rl_buffer:
            self.buffer_policy_optimizer = optimizer_class(
                self.buffer_policy.parameters(),
                weight_decay=policy_weight_decay,
                lr=policy_lr,
            )
            self.optimizers[self.buffer_policy] = self.buffer_policy_optimizer
            self.optimizer_class = optimizer_class
            self.policy_weight_decay = policy_weight_decay
            self.policy_lr = policy_lr

        self.use_automatic_beta_tuning = use_automatic_beta_tuning and buffer_policy and train_bc_on_rl_buffer
        self.beta_epsilon = beta_epsilon
        if self.use_automatic_beta_tuning:
            self.log_beta = ptu.zeros(1, requires_grad=True)
            self.beta_optimizer = optimizer_class(
                [self.log_beta],
                lr=policy_lr,
            )
        else:
            self.beta = beta
            self.beta_schedule_kwargs = beta_schedule_kwargs
            if beta_schedule_kwargs is None:
                self.beta_schedule = ConstantSchedule(beta)
            else:
                schedule_class = beta_schedule_kwargs.pop(
                    "schedule_class", PiecewiseLinearSchedule)
                self.beta_schedule = schedule_class(**beta_schedule_kwargs)

        self.discount = discount
        self.reward_scale = reward_scale
        self.eval_statistics = OrderedDict()
        self._n_train_steps_total = 0
        self._need_to_update_eval_statistics = True

        self.bc_num_pretrain_steps = bc_num_pretrain_steps
        self.q_num_pretrain1_steps = q_num_pretrain1_steps
        self.q_num_pretrain2_steps = q_num_pretrain2_steps
        self.bc_batch_size = bc_batch_size
        self.rl_weight = rl_weight
        self.bc_weight = bc_weight
        self.eval_policy = MakeDeterministic(self.policy)
        self.compute_bc = compute_bc
        self.alpha = alpha
        self.q_update_period = q_update_period
        self.policy_update_period = policy_update_period
        self.weight_loss = weight_loss

        self.reparam_weight = reparam_weight
        self.awr_weight = awr_weight
        self.post_pretrain_hyperparams = post_pretrain_hyperparams
        self.post_bc_pretrain_hyperparams = post_bc_pretrain_hyperparams
        self.update_policy = True
        self.pretraining_logging_period = pretraining_logging_period
        self.normalize_over_batch = normalize_over_batch
        self.normalize_over_state = normalize_over_state
        self.Z_K = Z_K

        self.reward_transform_class = reward_transform_class or LinearTransform
        self.reward_transform_kwargs = reward_transform_kwargs or dict(m=1,
                                                                       b=0)
        self.terminal_transform_class = terminal_transform_class or LinearTransform
        self.terminal_transform_kwargs = terminal_transform_kwargs or dict(m=1,
                                                                           b=0)
        self.reward_transform = self.reward_transform_class(
            **self.reward_transform_kwargs)
        self.terminal_transform = self.terminal_transform_class(
            **self.terminal_transform_kwargs)
        self.use_reparam_update = use_reparam_update
        self.clip_score = clip_score
        self.buffer_policy_sample_actions = buffer_policy_sample_actions

        self.train_bc_on_rl_buffer = train_bc_on_rl_buffer and buffer_policy
        self.validation_qlearning = validation_qlearning
        self.brac = brac
        self.mask_positive_advantage = mask_positive_advantage
        self.buffer_policy_reset_period = buffer_policy_reset_period
        self.num_buffer_policy_train_steps_on_reset = num_buffer_policy_train_steps_on_reset
        self.advantage_weighted_buffer_loss = advantage_weighted_buffer_loss

    def get_batch_from_buffer(self, replay_buffer, batch_size):
        batch = replay_buffer.random_batch(batch_size)
        batch = np_to_pytorch_batch(batch)
        return batch

    def run_bc_batch(self, replay_buffer, policy):
        batch = self.get_batch_from_buffer(replay_buffer, self.bc_batch_size)
        o = batch["observations"]
        u = batch["actions"]
        # g = batch["resampled_goals"]
        # og = torch.cat((o, g), dim=1)
        og = o
        # pred_u, *_ = self.policy(og)
        dist = policy(og)
        pred_u, log_pi = dist.rsample_and_logprob()
        stats = dist.get_diagnostics()

        mse = (pred_u - u)**2
        mse_loss = mse.mean()

        policy_logpp = dist.log_prob(u, )
        logp_loss = -policy_logpp.mean()
        policy_loss = logp_loss

        return policy_loss, logp_loss, mse_loss, stats

    def pretrain_policy_with_bc(
        self,
        policy,
        train_buffer,
        test_buffer,
        steps,
        label="policy",
    ):
        logger.remove_tabular_output(
            'progress.csv',
            relative_to_snapshot_dir=True,
        )
        logger.add_tabular_output(
            'pretrain_%s.csv' % label,
            relative_to_snapshot_dir=True,
        )

        optimizer = self.optimizers[policy]
        prev_time = time.time()
        for i in range(steps):
            train_policy_loss, train_logp_loss, train_mse_loss, train_stats = self.run_bc_batch(
                train_buffer, policy)
            train_policy_loss = train_policy_loss * self.bc_weight

            optimizer.zero_grad()
            train_policy_loss.backward()
            optimizer.step()

            test_policy_loss, test_logp_loss, test_mse_loss, test_stats = self.run_bc_batch(
                test_buffer, policy)
            test_policy_loss = test_policy_loss * self.bc_weight

            if i % self.pretraining_logging_period == 0:
                stats = {
                    "pretrain_bc/batch":
                    i,
                    "pretrain_bc/Train Logprob Loss":
                    ptu.get_numpy(train_logp_loss),
                    "pretrain_bc/Test Logprob Loss":
                    ptu.get_numpy(test_logp_loss),
                    "pretrain_bc/Train MSE":
                    ptu.get_numpy(train_mse_loss),
                    "pretrain_bc/Test MSE":
                    ptu.get_numpy(test_mse_loss),
                    "pretrain_bc/train_policy_loss":
                    ptu.get_numpy(train_policy_loss),
                    "pretrain_bc/test_policy_loss":
                    ptu.get_numpy(test_policy_loss),
                    "pretrain_bc/epoch_time":
                    time.time() - prev_time,
                }

                logger.record_dict(stats)
                logger.dump_tabular(with_prefix=True, with_timestamp=False)
                pickle.dump(
                    self.policy,
                    open(logger.get_snapshot_dir() + '/bc_%s.pkl' % label,
                         "wb"))
                prev_time = time.time()

        logger.remove_tabular_output(
            'pretrain_%s.csv' % label,
            relative_to_snapshot_dir=True,
        )
        logger.add_tabular_output(
            'progress.csv',
            relative_to_snapshot_dir=True,
        )

        if self.post_bc_pretrain_hyperparams:
            self.set_algorithm_weights(**self.post_bc_pretrain_hyperparams)

    def pretrain_q_with_bc_data(self):
        logger.remove_tabular_output('progress.csv',
                                     relative_to_snapshot_dir=True)
        logger.add_tabular_output('pretrain_q.csv',
                                  relative_to_snapshot_dir=True)

        self.update_policy = False
        # first train only the Q function
        for i in range(self.q_num_pretrain1_steps):
            self.eval_statistics = dict()

            train_data = self.replay_buffer.random_batch(self.bc_batch_size)
            train_data = np_to_pytorch_batch(train_data)
            obs = train_data['observations']
            next_obs = train_data['next_observations']
            # goals = train_data['resampled_goals']
            train_data['observations'] = obs  # torch.cat((obs, goals), dim=1)
            train_data[
                'next_observations'] = next_obs  # torch.cat((next_obs, goals), dim=1)
            self.train_from_torch(train_data, pretrain=True)
            if i % self.pretraining_logging_period == 0:
                stats_with_prefix = add_prefix(self.eval_statistics,
                                               prefix="trainer/")
                logger.record_dict(stats_with_prefix)
                logger.dump_tabular(with_prefix=True, with_timestamp=False)

        self.update_policy = True
        # then train policy and Q function together
        prev_time = time.time()
        for i in range(self.q_num_pretrain2_steps):
            self.eval_statistics = dict()
            if i % self.pretraining_logging_period == 0:
                self._need_to_update_eval_statistics = True
            train_data = self.replay_buffer.random_batch(self.bc_batch_size)
            train_data = np_to_pytorch_batch(train_data)
            obs = train_data['observations']
            next_obs = train_data['next_observations']
            # goals = train_data['resampled_goals']
            train_data['observations'] = obs  # torch.cat((obs, goals), dim=1)
            train_data[
                'next_observations'] = next_obs  # torch.cat((next_obs, goals), dim=1)
            self.train_from_torch(train_data, pretrain=True)

            if i % self.pretraining_logging_period == 0:
                self.eval_statistics["batch"] = i
                self.eval_statistics["epoch_time"] = time.time() - prev_time
                stats_with_prefix = add_prefix(self.eval_statistics,
                                               prefix="trainer/")
                logger.record_dict(stats_with_prefix)
                logger.dump_tabular(with_prefix=True, with_timestamp=False)
                prev_time = time.time()

        logger.remove_tabular_output(
            'pretrain_q.csv',
            relative_to_snapshot_dir=True,
        )
        logger.add_tabular_output(
            'progress.csv',
            relative_to_snapshot_dir=True,
        )

        self._need_to_update_eval_statistics = True
        self.eval_statistics = dict()

        if self.post_pretrain_hyperparams:
            self.set_algorithm_weights(**self.post_pretrain_hyperparams)

    def set_algorithm_weights(self, **kwargs):
        for key in kwargs:
            self.__dict__[key] = kwargs[key]

    def test_from_torch(self, batch):
        rewards = batch['rewards']
        terminals = batch['terminals']
        obs = batch['observations']
        actions = batch['actions']
        next_obs = batch['next_observations']
        weights = batch.get('weights', None)
        if self.reward_transform:
            rewards = self.reward_transform(rewards)

        if self.terminal_transform:
            terminals = self.terminal_transform(terminals)
        """
        Policy and Alpha Loss
        """
        dist = self.policy(obs)
        new_obs_actions, log_pi = dist.rsample_and_logprob()
        policy_mle = dist.mle_estimate()

        if self.use_automatic_entropy_tuning:
            alpha_loss = -(self.log_alpha *
                           (log_pi + self.target_entropy).detach()).mean()
            alpha = self.log_alpha.exp()
        else:
            alpha_loss = 0
            alpha = self.alpha

        q1_pred = self.qf1(obs, actions)
        q2_pred = self.qf2(obs, actions)
        # Make sure policy accounts for squashing functions like tanh correctly!
        next_dist = self.policy(next_obs)
        new_next_actions, new_log_pi = next_dist.rsample_and_logprob()
        target_q_values = torch.min(
            self.target_qf1(next_obs, new_next_actions),
            self.target_qf2(next_obs, new_next_actions),
        ) - alpha * new_log_pi

        q_target = self.reward_scale * rewards + (
            1. - terminals) * self.discount * target_q_values
        qf1_loss = self.qf_criterion(q1_pred, q_target.detach())
        qf2_loss = self.qf_criterion(q2_pred, q_target.detach())

        qf1_new_actions = self.qf1(obs, new_obs_actions)
        qf2_new_actions = self.qf2(obs, new_obs_actions)
        q_new_actions = torch.min(
            qf1_new_actions,
            qf2_new_actions,
        )

        policy_loss = (log_pi - q_new_actions).mean()

        self.eval_statistics['validation/QF1 Loss'] = np.mean(
            ptu.get_numpy(qf1_loss))
        self.eval_statistics['validation/QF2 Loss'] = np.mean(
            ptu.get_numpy(qf2_loss))
        self.eval_statistics['validation/Policy Loss'] = np.mean(
            ptu.get_numpy(policy_loss))
        self.eval_statistics.update(
            create_stats_ordered_dict(
                'validation/Q1 Predictions',
                ptu.get_numpy(q1_pred),
            ))
        self.eval_statistics.update(
            create_stats_ordered_dict(
                'validation/Q2 Predictions',
                ptu.get_numpy(q2_pred),
            ))
        self.eval_statistics.update(
            create_stats_ordered_dict(
                'validation/Q Targets',
                ptu.get_numpy(q_target),
            ))
        self.eval_statistics.update(
            create_stats_ordered_dict(
                'validation/Log Pis',
                ptu.get_numpy(log_pi),
            ))
        policy_statistics = add_prefix(dist.get_diagnostics(),
                                       "validation/policy/")
        self.eval_statistics.update(policy_statistics)

    def train_from_torch(
        self,
        batch,
        train=True,
        pretrain=False,
    ):
        rewards = batch['rewards']
        terminals = batch['terminals']
        obs = batch['observations']
        actions = batch['actions']
        next_obs = batch['next_observations']
        weights = batch.get('weights', None)
        if self.reward_transform:
            rewards = self.reward_transform(rewards)

        if self.terminal_transform:
            terminals = self.terminal_transform(terminals)
        """
        Policy and Alpha Loss
        """
        dist = self.policy(obs)
        new_obs_actions, log_pi = dist.rsample_and_logprob()
        policy_mle = dist.mle_estimate()

        if self.brac:
            buf_dist = self.buffer_policy(obs)
            buf_log_pi = buf_dist.log_prob(actions)
            rewards = rewards + buf_log_pi

        if self.use_automatic_entropy_tuning:
            alpha_loss = -(self.log_alpha *
                           (log_pi + self.target_entropy).detach()).mean()
            self.alpha_optimizer.zero_grad()
            alpha_loss.backward()
            self.alpha_optimizer.step()
            alpha = self.log_alpha.exp()
        else:
            alpha_loss = 0
            alpha = self.alpha
        """
        QF Loss
        """
        q1_pred = self.qf1(obs, actions)
        q2_pred = self.qf2(obs, actions)
        # Make sure policy accounts for squashing functions like tanh correctly!
        next_dist = self.policy(next_obs)
        new_next_actions, new_log_pi = next_dist.rsample_and_logprob()
        target_q_values = torch.min(
            self.target_qf1(next_obs, new_next_actions),
            self.target_qf2(next_obs, new_next_actions),
        ) - alpha * new_log_pi

        q_target = self.reward_scale * rewards + (
            1. - terminals) * self.discount * target_q_values
        qf1_loss = self.qf_criterion(q1_pred, q_target.detach())
        qf2_loss = self.qf_criterion(q2_pred, q_target.detach())
        """
        Policy Loss
        """
        qf1_new_actions = self.qf1(obs, new_obs_actions)
        qf2_new_actions = self.qf2(obs, new_obs_actions)
        q_new_actions = torch.min(
            qf1_new_actions,
            qf2_new_actions,
        )

        # Advantage-weighted regression
        if self.awr_use_mle_for_vf:
            v1_pi = self.qf1(obs, policy_mle)
            v2_pi = self.qf2(obs, policy_mle)
            v_pi = torch.min(v1_pi, v2_pi)
        else:
            if self.vf_K > 1:
                vs = []
                for i in range(self.vf_K):
                    u = dist.sample()
                    q1 = self.qf1(obs, u)
                    q2 = self.qf2(obs, u)
                    v = torch.min(q1, q2)
                    # v = q1
                    vs.append(v)
                v_pi = torch.cat(vs, 1).mean(dim=1)
            else:
                # v_pi = self.qf1(obs, new_obs_actions)
                v1_pi = self.qf1(obs, new_obs_actions)
                v2_pi = self.qf2(obs, new_obs_actions)
                v_pi = torch.min(v1_pi, v2_pi)

        if self.awr_sample_actions:
            u = new_obs_actions
            if self.awr_min_q:
                q_adv = q_new_actions
            else:
                q_adv = qf1_new_actions
        elif self.buffer_policy_sample_actions:
            buf_dist = self.buffer_policy(obs)
            u, _ = buf_dist.rsample_and_logprob()
            qf1_buffer_actions = self.qf1(obs, u)
            qf2_buffer_actions = self.qf2(obs, u)
            q_buffer_actions = torch.min(
                qf1_buffer_actions,
                qf2_buffer_actions,
            )
            if self.awr_min_q:
                q_adv = q_buffer_actions
            else:
                q_adv = qf1_buffer_actions
        else:
            u = actions
            if self.awr_min_q:
                q_adv = torch.min(q1_pred, q2_pred)
            else:
                q_adv = q1_pred

        policy_logpp = dist.log_prob(u)

        if self.use_automatic_beta_tuning:
            buffer_dist = self.buffer_policy(obs)
            beta = self.log_beta.exp()
            kldiv = torch.distributions.kl.kl_divergence(dist, buffer_dist)
            beta_loss = -1 * (beta *
                              (kldiv - self.beta_epsilon).detach()).mean()

            self.beta_optimizer.zero_grad()
            beta_loss.backward()
            self.beta_optimizer.step()
        else:
            beta = self.beta_schedule.get_value(self._n_train_steps_total)

        if self.normalize_over_state == "advantage":
            score = q_adv - v_pi
            if self.mask_positive_advantage:
                score = torch.sign(score)
        elif self.normalize_over_state == "Z":
            buffer_dist = self.buffer_policy(obs)
            K = self.Z_K
            buffer_obs = []
            buffer_actions = []
            log_bs = []
            log_pis = []
            for i in range(K):
                u = buffer_dist.sample()
                log_b = buffer_dist.log_prob(u)
                log_pi = dist.log_prob(u)
                buffer_obs.append(obs)
                buffer_actions.append(u)
                log_bs.append(log_b)
                log_pis.append(log_pi)
            buffer_obs = torch.cat(buffer_obs, 0)
            buffer_actions = torch.cat(buffer_actions, 0)
            p_buffer = torch.exp(torch.cat(log_bs, 0).sum(dim=1, ))
            log_pi = torch.cat(log_pis, 0)
            log_pi = log_pi.sum(dim=1, )
            q1_b = self.qf1(buffer_obs, buffer_actions)
            q2_b = self.qf2(buffer_obs, buffer_actions)
            q_b = torch.min(q1_b, q2_b)
            q_b = torch.reshape(q_b, (-1, K))
            adv_b = q_b - v_pi
            # if self._n_train_steps_total % 100 == 0:
            #     import ipdb; ipdb.set_trace()
            # Z = torch.exp(adv_b / beta).mean(dim=1, keepdim=True)
            # score = torch.exp((q_adv - v_pi) / beta) / Z
            # score = score / sum(score)
            logK = torch.log(ptu.tensor(float(K)))
            logZ = torch.logsumexp(adv_b / beta - logK, dim=1, keepdim=True)
            logS = (q_adv - v_pi) / beta - logZ
            # logZ = torch.logsumexp(q_b/beta - logK, dim=1, keepdim=True)
            # logS = q_adv/beta - logZ
            score = F.softmax(logS, dim=0)  # score / sum(score)
        else:
            error

        if self.clip_score is not None:
            score = torch.clamp(score, max=self.clip_score)

        if self.weight_loss and weights is None:
            if self.normalize_over_batch == True:
                weights = F.softmax(score / beta, dim=0)
            elif self.normalize_over_batch == "whiten":
                adv_mean = torch.mean(score)
                adv_std = torch.std(score) + 1e-5
                normalized_score = (score - adv_mean) / adv_std
                weights = torch.exp(normalized_score / beta)
            elif self.normalize_over_batch == "exp":
                weights = torch.exp(score / beta)
            elif self.normalize_over_batch == "step_fn":
                weights = (score > 0).float()
            elif self.normalize_over_batch == False:
                weights = score
            else:
                error
        weights = weights[:, 0]

        policy_loss = alpha * log_pi.mean()

        if self.use_awr_update and self.weight_loss:
            policy_loss = policy_loss + self.awr_weight * (
                -policy_logpp * len(weights) * weights.detach()).mean()
        elif self.use_awr_update:
            policy_loss = policy_loss + self.awr_weight * (
                -policy_logpp).mean()

        if self.use_reparam_update:
            policy_loss = policy_loss + self.reparam_weight * (
                -q_new_actions).mean()

        policy_loss = self.rl_weight * policy_loss
        if self.compute_bc:
            train_policy_loss, train_logp_loss, train_mse_loss, _ = self.run_bc_batch(
                self.demo_train_buffer, self.policy)
            policy_loss = policy_loss + self.bc_weight * train_policy_loss

        if not pretrain and self.buffer_policy_reset_period > 0 and self._n_train_steps_total % self.buffer_policy_reset_period == 0:
            del self.buffer_policy_optimizer
            self.buffer_policy_optimizer = self.optimizer_class(
                self.buffer_policy.parameters(),
                weight_decay=self.policy_weight_decay,
                lr=self.policy_lr,
            )
            self.optimizers[self.buffer_policy] = self.buffer_policy_optimizer
            for i in range(self.num_buffer_policy_train_steps_on_reset):
                if self.train_bc_on_rl_buffer:
                    if self.advantage_weighted_buffer_loss:
                        buffer_dist = self.buffer_policy(obs)
                        buffer_u = actions
                        buffer_new_obs_actions, _ = buffer_dist.rsample_and_logprob(
                        )
                        buffer_policy_logpp = buffer_dist.log_prob(buffer_u)
                        buffer_policy_logpp = buffer_policy_logpp[:, None]

                        buffer_q1_pred = self.qf1(obs, buffer_u)
                        buffer_q2_pred = self.qf2(obs, buffer_u)
                        buffer_q_adv = torch.min(buffer_q1_pred,
                                                 buffer_q2_pred)

                        buffer_v1_pi = self.qf1(obs, buffer_new_obs_actions)
                        buffer_v2_pi = self.qf2(obs, buffer_new_obs_actions)
                        buffer_v_pi = torch.min(buffer_v1_pi, buffer_v2_pi)

                        buffer_score = buffer_q_adv - buffer_v_pi
                        buffer_weights = F.softmax(buffer_score / beta, dim=0)
                        buffer_policy_loss = self.awr_weight * (
                            -buffer_policy_logpp * len(buffer_weights) *
                            buffer_weights.detach()).mean()
                    else:
                        buffer_policy_loss, buffer_train_logp_loss, buffer_train_mse_loss, _ = self.run_bc_batch(
                            self.replay_buffer.train_replay_buffer,
                            self.buffer_policy)

                    self.buffer_policy_optimizer.zero_grad()
                    buffer_policy_loss.backward(retain_graph=True)
                    self.buffer_policy_optimizer.step()

        if self.train_bc_on_rl_buffer:
            if self.advantage_weighted_buffer_loss:
                buffer_dist = self.buffer_policy(obs)
                buffer_u = actions
                buffer_new_obs_actions, _ = buffer_dist.rsample_and_logprob()
                buffer_policy_logpp = buffer_dist.log_prob(buffer_u)
                buffer_policy_logpp = buffer_policy_logpp[:, None]

                buffer_q1_pred = self.qf1(obs, buffer_u)
                buffer_q2_pred = self.qf2(obs, buffer_u)
                buffer_q_adv = torch.min(buffer_q1_pred, buffer_q2_pred)

                buffer_v1_pi = self.qf1(obs, buffer_new_obs_actions)
                buffer_v2_pi = self.qf2(obs, buffer_new_obs_actions)
                buffer_v_pi = torch.min(buffer_v1_pi, buffer_v2_pi)

                buffer_score = buffer_q_adv - buffer_v_pi
                buffer_weights = F.softmax(buffer_score / beta, dim=0)
                buffer_policy_loss = self.awr_weight * (
                    -buffer_policy_logpp * len(buffer_weights) *
                    buffer_weights.detach()).mean()
            else:
                buffer_policy_loss, buffer_train_logp_loss, buffer_train_mse_loss, _ = self.run_bc_batch(
                    self.replay_buffer.train_replay_buffer, self.buffer_policy)
        """
        Update networks
        """
        if self._n_train_steps_total % self.q_update_period == 0:
            self.qf1_optimizer.zero_grad()
            qf1_loss.backward()
            self.qf1_optimizer.step()

            self.qf2_optimizer.zero_grad()
            qf2_loss.backward()
            self.qf2_optimizer.step()

        if self._n_train_steps_total % self.policy_update_period == 0 and self.update_policy:
            self.policy_optimizer.zero_grad()
            policy_loss.backward()
            self.policy_optimizer.step()

        if self.train_bc_on_rl_buffer and self._n_train_steps_total % self.policy_update_period == 0:
            self.buffer_policy_optimizer.zero_grad()
            buffer_policy_loss.backward()
            self.buffer_policy_optimizer.step()
        """
        Soft Updates
        """
        if self._n_train_steps_total % self.target_update_period == 0:
            ptu.soft_update_from_to(self.qf1, self.target_qf1,
                                    self.soft_target_tau)
            ptu.soft_update_from_to(self.qf2, self.target_qf2,
                                    self.soft_target_tau)
        """
        Save some statistics for eval
        """
        if self._need_to_update_eval_statistics:
            self._need_to_update_eval_statistics = False
            """
            Eval should set this to None.
            This way, these statistics are only computed for one batch.
            """
            policy_loss = (log_pi - q_new_actions).mean()

            self.eval_statistics['QF1 Loss'] = np.mean(ptu.get_numpy(qf1_loss))
            self.eval_statistics['QF2 Loss'] = np.mean(ptu.get_numpy(qf2_loss))
            self.eval_statistics['Policy Loss'] = np.mean(
                ptu.get_numpy(policy_loss))
            self.eval_statistics.update(
                create_stats_ordered_dict(
                    'Q1 Predictions',
                    ptu.get_numpy(q1_pred),
                ))
            self.eval_statistics.update(
                create_stats_ordered_dict(
                    'Q2 Predictions',
                    ptu.get_numpy(q2_pred),
                ))
            self.eval_statistics.update(
                create_stats_ordered_dict(
                    'Q Targets',
                    ptu.get_numpy(q_target),
                ))
            self.eval_statistics.update(
                create_stats_ordered_dict(
                    'Log Pis',
                    ptu.get_numpy(log_pi),
                ))
            self.eval_statistics.update(
                create_stats_ordered_dict(
                    'rewards',
                    ptu.get_numpy(rewards),
                ))
            self.eval_statistics.update(
                create_stats_ordered_dict(
                    'terminals',
                    ptu.get_numpy(terminals),
                ))
            policy_statistics = add_prefix(dist.get_diagnostics(), "policy/")
            self.eval_statistics.update(policy_statistics)
            self.eval_statistics.update(
                create_stats_ordered_dict(
                    'Advantage Weights',
                    ptu.get_numpy(weights),
                ))
            self.eval_statistics.update(
                create_stats_ordered_dict(
                    'Advantage Score',
                    ptu.get_numpy(score),
                ))

            if self.normalize_over_state == "Z":
                self.eval_statistics.update(
                    create_stats_ordered_dict(
                        'logZ',
                        ptu.get_numpy(logZ),
                    ))

            if self.use_automatic_entropy_tuning:
                self.eval_statistics['Alpha'] = alpha.item()
                self.eval_statistics['Alpha Loss'] = alpha_loss.item()

            if self.compute_bc:
                test_policy_loss, test_logp_loss, test_mse_loss, _ = self.run_bc_batch(
                    self.demo_test_buffer, self.policy)
                self.eval_statistics.update({
                    "bc/Train Logprob Loss":
                    ptu.get_numpy(train_logp_loss),
                    "bc/Test Logprob Loss":
                    ptu.get_numpy(test_logp_loss),
                    "bc/Train MSE":
                    ptu.get_numpy(train_mse_loss),
                    "bc/Test MSE":
                    ptu.get_numpy(test_mse_loss),
                    "bc/train_policy_loss":
                    ptu.get_numpy(train_policy_loss),
                    "bc/test_policy_loss":
                    ptu.get_numpy(test_policy_loss),
                })
            if self.train_bc_on_rl_buffer:
                _, buffer_train_logp_loss, _, _ = self.run_bc_batch(
                    self.replay_buffer.train_replay_buffer, self.buffer_policy)

                _, buffer_test_logp_loss, _, _ = self.run_bc_batch(
                    self.replay_buffer.validation_replay_buffer,
                    self.buffer_policy)
                buffer_dist = self.buffer_policy(obs)
                kldiv = torch.distributions.kl.kl_divergence(dist, buffer_dist)

                _, train_offline_logp_loss, _, _ = self.run_bc_batch(
                    self.demo_train_buffer, self.buffer_policy)

                _, test_offline_logp_loss, _, _ = self.run_bc_batch(
                    self.demo_test_buffer, self.buffer_policy)

                self.eval_statistics.update({
                    "buffer_policy/Train Online Logprob":
                    -1 * ptu.get_numpy(buffer_train_logp_loss),
                    "buffer_policy/Test Online Logprob":
                    -1 * ptu.get_numpy(buffer_test_logp_loss),
                    "buffer_policy/Train Offline Logprob":
                    -1 * ptu.get_numpy(train_offline_logp_loss),
                    "buffer_policy/Test Offline Logprob":
                    -1 * ptu.get_numpy(test_offline_logp_loss),
                    "buffer_policy/train_policy_loss":
                    ptu.get_numpy(buffer_policy_loss),
                    # "buffer_policy/test_policy_loss": ptu.get_numpy(buffer_test_policy_loss),
                    "buffer_policy/kl_div":
                    ptu.get_numpy(kldiv.mean()),
                })
            if self.use_automatic_beta_tuning:
                self.eval_statistics.update({
                    "adaptive_beta/beta":
                    ptu.get_numpy(beta.mean()),
                    "adaptive_beta/beta loss":
                    ptu.get_numpy(beta_loss.mean()),
                })

            if self.validation_qlearning:
                train_data = self.replay_buffer.validation_replay_buffer.random_batch(
                    self.bc_batch_size)
                train_data = np_to_pytorch_batch(train_data)
                obs = train_data['observations']
                next_obs = train_data['next_observations']
                # goals = train_data['resampled_goals']
                train_data[
                    'observations'] = obs  # torch.cat((obs, goals), dim=1)
                train_data[
                    'next_observations'] = next_obs  # torch.cat((next_obs, goals), dim=1)
                self.test_from_torch(train_data)

        self._n_train_steps_total += 1

    def get_diagnostics(self):
        stats = super().get_diagnostics()
        stats.update(self.eval_statistics)
        return stats

    def end_epoch(self, epoch):
        self._need_to_update_eval_statistics = True

    @property
    def networks(self):
        nets = [
            self.policy,
            self.qf1,
            self.qf2,
            self.target_qf1,
            self.target_qf2,
        ]
        if self.buffer_policy:
            nets.append(self.buffer_policy)
        return nets

    def get_snapshot(self):
        return dict(
            policy=self.policy,
            qf1=self.qf1,
            qf2=self.qf2,
            target_qf1=self.qf1,
            target_qf2=self.qf2,
            buffer_policy=self.buffer_policy,
        )
Beispiel #2
0
class ACAITrainer():
    def __init__(
        self,
        train_dataset,
        test_dataset,
        model,
        batch_size=128,
        log_interval=0,
        beta=0.5,
        beta_schedule=None,
        imsize=84,
        lr=1e-3,
        do_scatterplot=False,
        normalize=False,
        state_sim_debug=False,
        mse_weight=0.1,
        is_auto_encoder=False,
        lmbda=0.5,
        mu=1,
        gamma=0.2,
    ):
        self.log_interval = log_interval
        self.batch_size = batch_size
        self.beta = beta
        if is_auto_encoder:
            self.beta = 0
        self.beta_schedule = beta_schedule
        if self.beta_schedule is None:
            self.beta_schedule = ConstantSchedule(self.beta)
        self.imsize = imsize
        self.do_scatterplot = do_scatterplot
        self.lmbda = lmbda
        self.mu = mu
        self.gamma = gamma
        """
        I think it's a bit nicer if the caller makes this call, i.e.
        ```
        m = ConvVAE(representation_size)
        if ptu.gpu_enabled():
            m.cuda()
        t = ConvVAETrainer(train_data, test_data, m)
        ```
        However, I'll leave this here for backwards-compatibility.
        """
        if ptu.gpu_enabled():
            model.cuda()

        self.model = model
        self.representation_size = model.representation_size
        self.input_channels = model.input_channels
        self.imlength = model.imlength

        self.optimizer = optim.Adam(self.model.parameters(), lr=lr)
        self.train_dataset, self.test_dataset = train_dataset, test_dataset
        self.normalize = normalize
        self.state_sim_debug = state_sim_debug
        self.mse_weight = mse_weight
        self.x_next_index = self.input_channels * self.imsize**2

        if self.normalize:
            self.train_data_mean = np.mean(self.train_dataset, axis=0)
        # self.train_dataset = ((self.train_dataset - self.train_data_mean)) + 1 / 2
        # self.test_dataset = ((self.test_dataset - self.train_data_mean)) + 1 / 2

    def get_batch(self, train=True):
        dataset = self.train_dataset if train else self.test_dataset
        ind = np.random.randint(0, len(dataset), self.batch_size)
        samples = dataset[ind, :]
        samples = normalize_image(samples)
        if self.normalize:
            samples = ((samples - self.train_data_mean) + 1) / 2
        return ptu.np_to_var(samples)

    def get_debug_batch(self, train=True):
        dataset = self.train_dataset if train else self.test_dataset
        X, Y = dataset
        ind = np.random.randint(0, Y.shape[0], self.batch_size)
        X = X[ind, :]
        Y = Y[ind, :]
        return ptu.np_to_var(X), ptu.np_to_var(Y)

    def get_batch_smooth(self, train=True):
        dataset = self.train_dataset if train else self.test_dataset
        ind = np.random.randint(0, len(dataset), self.batch_size)
        samples = dataset[ind, :]
        samples = normalize_image(samples)
        if self.normalize:
            samples = ((samples - self.train_data_mean) + 1) / 2
        x_next, x = samples[:, :self.x_next_index], samples[:,
                                                            self.x_next_index:]
        return ptu.np_to_var(x_next), ptu.np_to_var(x)

    def logprob(self, recon_x, x, mu, logvar):

        # Divide by batch_size rather than setting size_average=True because
        # otherwise the averaging will also happen across dimension 1 (the
        # pixels)
        return F.binary_cross_entropy(
            recon_x,
            x.narrow(start=0, length=self.imlength,
                     dimension=1).contiguous().view(-1, self.imlength),
            size_average=False,
        ) / self.batch_size

    def kl_divergence(self, recon_x, x, mu, logvar):
        return -torch.sum(1 + logvar - mu.pow(2) - logvar.exp(), dim=1).mean()

    def state_similarity_loss(self, model, encoded_x, states):
        output = self.model.fc6(F.relu(self.model.fc5(encoded_x)))
        return torch.norm(output - states)**2 / self.batch_size

    def train_epoch(self,
                    epoch,
                    sample_batch=None,
                    batches=100,
                    from_rl=False):
        self.model.train()
        losses = []
        bces = []
        kles = []
        mses = []
        losses_c = []
        beta = self.beta_schedule.get_value(epoch)
        for batch_idx in range(batches):
            data = self.get_batch()
            if sample_batch is not None:
                data = sample_batch(self.batch_size)

            self.optimizer.zero_grad()
            recon_batch, mu, logvar, predicted_alpha, alpha = self.model(data)
            bce = self.logprob(recon_batch, data, mu, logvar)
            kle = self.kl_divergence(recon_batch, data, mu, logvar)
            loss = bce + beta * kle + self.lmbda * torch.norm(
                predicted_alpha, 2)
            regularizer_a = self.model.critic(self.gamma * data +
                                              (1 - self.gamma) * recon_batch)
            loss_c = torch.norm(predicted_alpha - alpha,
                                2) + self.mu * torch.norm(regularizer_a, 2)

            loss.backward(retain_graph=True)
            loss_c.backward()
            losses.append(loss.data[0])
            losses_c.append(loss_c.data[0])
            bces.append(bce.data[0])
            kles.append(kle.data[0])
            if self.state_sim_debug:
                mses.append(sim_loss.data[0])

            self.optimizer.step()
            if self.log_interval and batch_idx % self.log_interval == 0:
                print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
                    epoch, batch_idx * len(data),
                    len(self.train_loader.dataset),
                    100. * batch_idx / len(self.train_loader),
                    loss.data[0] / len(data)))

        if not from_rl:
            logger.record_tabular("train/epoch", epoch)
            logger.record_tabular("train/critic_loss", np.mean(losses_c))
            logger.record_tabular("train/BCE", np.mean(bces))
            logger.record_tabular("train/KL", np.mean(kles))
            if self.state_sim_debug:
                logger.record_tabular("train/mse", np.mean(mses))
            logger.record_tabular("train/loss", np.mean(losses))

    def test_epoch(
        self,
        epoch,
        save_reconstruction=True,
        save_scatterplot=True,
        save_vae=True,
        from_rl=False,
    ):
        self.model.eval()
        losses = []
        losses_c = []
        bces = []
        kles = []
        zs = []
        mses = []
        beta = self.beta_schedule.get_value(epoch)
        for batch_idx in range(10):
            data = self.get_batch(train=False)
            recon_batch, mu, logvar, predicted_alpha, alpha = self.model(data)
            bce = self.logprob(recon_batch, data, mu, logvar)
            kle = self.kl_divergence(recon_batch, data, mu, logvar)
            loss = bce + beta * kle + self.lmbda * torch.norm(
                predicted_alpha, 2)
            regularizer_a = self.model.critic(self.gamma * data +
                                              (1 - self.gamma) * recon_batch)
            loss_c = torch.norm(predicted_alpha - alpha,
                                2) + self.mu * torch.norm(regularizer_a, 2)

            z_data = ptu.get_numpy(mu.cpu())
            for i in range(len(z_data)):
                zs.append(z_data[i, :])
            losses.append(loss.data[0])
            losses_c.append(loss_c.data[0])
            bces.append(bce.data[0])
            kles.append(kle.data[0])
            if self.state_sim_debug:
                mses.append(sim_loss.data[0])

            if batch_idx == 0 and save_reconstruction:
                n = min(data.size(0), 8)
                comparison = torch.cat([
                    data[:n].narrow(start=0, length=self.imlength,
                                    dimension=1).contiguous().view(
                                        -1, self.input_channels, self.imsize,
                                        self.imsize),
                    recon_batch.view(
                        self.batch_size,
                        self.input_channels,
                        self.imsize,
                        self.imsize,
                    )[:n]
                ])
                save_dir = osp.join(logger.get_snapshot_dir(),
                                    'r%d.png' % epoch)
                save_image(comparison.data.cpu(), save_dir, nrow=n)

        zs = np.array(zs)
        self.model.dist_mu = zs.mean(axis=0)
        self.model.dist_std = zs.std(axis=0)
        if self.do_scatterplot and save_scatterplot:
            self.plot_scattered(np.array(zs), epoch)

        if not from_rl:
            for key, value in self.debug_statistics().items():
                logger.record_tabular(key, value)

            logger.record_tabular("test/BCE", np.mean(bces))
            logger.record_tabular("test/KL", np.mean(kles))
            logger.record_tabular("test/loss", np.mean(losses))
            logger.record_tabular("test/critic_loss", np.mean(losses_c))
            logger.record_tabular("beta", beta)
            if self.state_sim_debug:
                logger.record_tabular("test/MSE", np.mean(mses))

            logger.dump_tabular()
            if save_vae:
                logger.save_itr_params(epoch, self.model)  # slow...

    # logdir = logger.get_snapshot_dir()
    # filename = osp.join(logdir, 'params.pkl')
    # torch.save(self.model, filename)

    def debug_statistics(self):
        """
        Given an image $$x$$, samples a bunch of latents from the prior
        $$z_i$$ and decode them $$\hat x_i$$.
        Compare this to $$\hat x$$, the reconstruction of $$x$$.
        Ideally
         - All the $$\hat x_i$$s do worse than $$\hat x$$ (makes sure VAE
           isn’t ignoring the latent)
         - Some $$\hat x_i$$ do better than other $$\hat x_i$$ (tests for
           coverage)
        """
        debug_batch_size = 64

        data = self.get_batch(train=False)
        recon_batch, mu, logvar, predicted_alpha, alpha = self.model(data)
        img = data[0]
        recon_mse = ((recon_batch[0] - img)**2).mean()

        img_repeated = img.expand((debug_batch_size, img.shape[0]))

        samples = ptu.Variable(
            torch.randn(debug_batch_size, self.representation_size))
        random_imgs = self.model.decode(samples)
        random_mse = ((random_imgs - img_repeated)**2).mean(dim=1)

        mse_improvement = ptu.get_numpy(random_mse - recon_mse)
        stats = create_stats_ordered_dict(
            'debug/MSE improvement over random',
            mse_improvement,
        )
        stats['debug/MSE of random reconstruction'] = ptu.get_numpy(
            recon_mse)[0]
        return stats

    def dump_samples(self, epoch):
        self.model.eval()
        sample = ptu.Variable(torch.randn(64, self.representation_size))
        sample = self.model.decode(sample).cpu()
        save_dir = osp.join(logger.get_snapshot_dir(), 's%d.png' % epoch)
        save_image(
            sample.data.view(64, self.input_channels, self.imsize,
                             self.imsize), save_dir)

    def plot_scattered(self, z, epoch):
        try:
            import matplotlib.pyplot as plt
        except ImportError:
            logger.log(__file__ + ": Unable to load matplotlib. Consider "
                       "setting do_scatterplot to False")
            return
        dim_and_stds = [(i, np.std(z[:, i])) for i in range(z.shape[1])]
        dim_and_stds = sorted(dim_and_stds, key=lambda x: x[1])
        dim1 = dim_and_stds[-1][0]
        dim2 = dim_and_stds[-2][0]
        plt.figure(figsize=(8, 8))
        plt.scatter(z[:, dim1], z[:, dim2], marker='o', edgecolor='none')
        if self.model.dist_mu is not None:
            x1 = self.model.dist_mu[dim1:dim1 + 1]
            y1 = self.model.dist_mu[dim2:dim2 + 1]
            x2 = self.model.dist_mu[dim1:dim1 +
                                    1] + self.model.dist_std[dim1:dim1 + 1]
            y2 = self.model.dist_mu[dim2:dim2 +
                                    1] + self.model.dist_std[dim2:dim2 + 1]
        plt.plot([x1, x2], [y1, y2], color='k', linestyle='-', linewidth=2)
        axes = plt.gca()
        axes.set_xlim([-6, 6])
        axes.set_ylim([-6, 6])
        axes.set_title('dim {} vs dim {}'.format(dim1, dim2))
        plt.grid(True)
        save_file = osp.join(logger.get_snapshot_dir(),
                             'scatter%d.png' % epoch)
        plt.savefig(save_file)
Beispiel #3
0
class VAETrainer(LossFunction):
    def __init__(
        self,
        model,
        batch_size=128,
        log_interval=0,
        beta=0.5,
        beta_schedule=None,
        lr=None,
        do_scatterplot=False,
        normalize=False,
        mse_weight=0.1,
        is_auto_encoder=False,
        background_subtract=False,
        linearity_weight=0.0,
        distance_weight=0.0,
        loss_weights=None,
        use_linear_dynamics=False,
        use_parallel_dataloading=False,
        train_data_workers=2,
        skew_dataset=False,
        skew_config=None,
        priority_function_kwargs=None,
        start_skew_epoch=0,
        weight_decay=0,
        key_to_reconstruct='observations',
        num_epochs=None,
    ):
        #TODO:steven fix pickling
        assert not use_parallel_dataloading, "Have to fix pickling the dataloaders first"

        if skew_config is None:
            skew_config = {}
        self.log_interval = log_interval
        self.batch_size = batch_size
        self.beta = beta
        if is_auto_encoder:
            self.beta = 0
        if lr is None:
            if is_auto_encoder:
                lr = 1e-2
            else:
                lr = 1e-3
        self.beta_schedule = beta_schedule
        self.num_epochs = num_epochs
        if self.beta_schedule is None or is_auto_encoder:
            self.beta_schedule = ConstantSchedule(self.beta)
        self.imsize = model.imsize
        self.do_scatterplot = do_scatterplot
        model.to(ptu.device)

        self.model = model
        self.representation_size = model.representation_size
        self.input_channels = model.input_channels
        self.imlength = model.imlength

        self.lr = lr
        params = list(self.model.parameters())
        self.optimizer = optim.Adam(
            params,
            lr=self.lr,
            weight_decay=weight_decay,
        )

        self.key_to_reconstruct = key_to_reconstruct
        self.use_parallel_dataloading = use_parallel_dataloading
        self.train_data_workers = train_data_workers
        self.skew_dataset = skew_dataset
        self.skew_config = skew_config
        self.start_skew_epoch = start_skew_epoch
        if priority_function_kwargs is None:
            self.priority_function_kwargs = dict()
        else:
            self.priority_function_kwargs = priority_function_kwargs

        if use_parallel_dataloading:
            self.train_dataset_pt = ImageDataset(train_dataset,
                                                 should_normalize=True)
            self.test_dataset_pt = ImageDataset(test_dataset,
                                                should_normalize=True)

            if self.skew_dataset:
                base_sampler = InfiniteWeightedRandomSampler(
                    self.train_dataset, self._train_weights)
            else:
                base_sampler = InfiniteRandomSampler(self.train_dataset)
            self.train_dataloader = DataLoader(
                self.train_dataset_pt,
                sampler=InfiniteRandomSampler(self.train_dataset),
                batch_size=batch_size,
                drop_last=False,
                num_workers=train_data_workers,
                pin_memory=True,
            )
            self.test_dataloader = DataLoader(
                self.test_dataset_pt,
                sampler=InfiniteRandomSampler(self.test_dataset),
                batch_size=batch_size,
                drop_last=False,
                num_workers=0,
                pin_memory=True,
            )
            self.train_dataloader = iter(self.train_dataloader)
            self.test_dataloader = iter(self.test_dataloader)

        self.normalize = normalize
        self.mse_weight = mse_weight
        self.background_subtract = background_subtract

        if self.normalize or self.background_subtract:
            self.train_data_mean = np.mean(self.train_dataset, axis=0)
            self.train_data_mean = normalize_image(
                np.uint8(self.train_data_mean))
        self.linearity_weight = linearity_weight
        self.distance_weight = distance_weight
        self.loss_weights = loss_weights

        self.use_linear_dynamics = use_linear_dynamics
        self._extra_stats_to_log = None

        # stateful tracking variables, reset every epoch
        self.eval_statistics = collections.defaultdict(list)
        self.eval_data = collections.defaultdict(list)
        self.num_batches = 0

    @property
    def log_dir(self):
        return logger.get_snapshot_dir()

    def get_dataset_stats(self, data):
        torch_input = ptu.from_numpy(normalize_image(data))
        mus, log_vars = self.model.encode(torch_input)
        mus = ptu.get_numpy(mus)
        mean = np.mean(mus, axis=0)
        std = np.std(mus, axis=0)
        return mus, mean, std

    def _kl_np_to_np(self, np_imgs):
        torch_input = ptu.from_numpy(normalize_image(np_imgs))
        mu, log_var = self.model.encode(torch_input)
        return ptu.get_numpy(
            -torch.sum(1 + log_var - mu.pow(2) - log_var.exp(), dim=1))

    def _reconstruction_squared_error_np_to_np(self, np_imgs):
        torch_input = ptu.from_numpy(normalize_image(np_imgs))
        recons, *_ = self.model(torch_input)
        error = torch_input - recons
        return ptu.get_numpy((error**2).sum(dim=1))

    def set_vae(self, vae):
        self.model = vae
        self.optimizer = optim.Adam(self.model.parameters(), lr=self.lr)

    def get_batch(self, test_data=False, epoch=None):
        if self.use_parallel_dataloading:
            if test_data:
                dataloader = self.test_dataloader
            else:
                dataloader = self.train_dataloader
            samples = next(dataloader).to(ptu.device)
            return samples

        dataset = self.test_dataset if test_data else self.train_dataset
        skew = False
        if epoch is not None:
            skew = (self.start_skew_epoch < epoch)
        if not test_data and self.skew_dataset and skew:
            probs = self._train_weights / np.sum(self._train_weights)
            ind = np.random.choice(
                len(probs),
                self.batch_size,
                p=probs,
            )
        else:
            ind = np.random.randint(0, len(dataset), self.batch_size)
        samples = normalize_image(dataset[ind, :])
        if self.normalize:
            samples = ((samples - self.train_data_mean) + 1) / 2
        if self.background_subtract:
            samples = samples - self.train_data_mean
        return ptu.from_numpy(samples)

    def get_debug_batch(self, train=True):
        dataset = self.train_dataset if train else self.test_dataset
        X, Y = dataset
        ind = np.random.randint(0, Y.shape[0], self.batch_size)
        X = X[ind, :]
        Y = Y[ind, :]
        return ptu.from_numpy(X), ptu.from_numpy(Y)

    def train_epoch(self, epoch, dataset, batches=100):
        start_time = time.time()
        for b in range(batches):
            self.train_batch(epoch, dataset.random_batch(self.batch_size))
        self.eval_statistics["train/epoch_duration"].append(time.time() -
                                                            start_time)

    def test_epoch(self, epoch, dataset, batches=10):
        start_time = time.time()
        for b in range(batches):
            self.test_batch(epoch, dataset.random_batch(self.batch_size))
        self.eval_statistics["test/epoch_duration"].append(time.time() -
                                                           start_time)

    def compute_loss(self, batch, epoch=-1, test=False):
        prefix = "test/" if test else "train/"

        beta = float(self.beta_schedule.get_value(epoch))
        obs = batch[self.key_to_reconstruct]
        reconstructions, obs_distribution_params, latent_distribution_params = self.model(
            obs)
        log_prob = self.model.logprob(obs, obs_distribution_params)
        kle = self.model.kl_divergence(latent_distribution_params)
        loss = -1 * log_prob + beta * kle

        self.eval_statistics['epoch'] = epoch
        self.eval_statistics['beta'] = beta
        self.eval_statistics[prefix + "losses"].append(loss.item())
        self.eval_statistics[prefix + "log_probs"].append(log_prob.item())
        self.eval_statistics[prefix + "kles"].append(kle.item())
        self.eval_statistics["num_train_batches"].append(self.num_batches)

        encoder_mean = self.model.get_encoding_from_latent_distribution_params(
            latent_distribution_params)
        z_data = ptu.get_numpy(encoder_mean.cpu())
        for i in range(len(z_data)):
            self.eval_data[prefix + "zs"].append(z_data[i, :])
        self.eval_data[prefix + "last_batch"] = (obs, reconstructions)

        return loss

    def train_batch(self, epoch, batch):
        self.num_batches += 1
        self.model.train()
        self.optimizer.zero_grad()

        loss = self.compute_loss(batch, epoch, False)
        loss.backward()

        self.optimizer.step()
        #self.scheduler.step()

    def test_batch(
        self,
        epoch,
        batch,
    ):
        self.model.eval()
        loss = self.compute_loss(batch, epoch, True)

    def end_epoch(self, epoch):
        self.eval_statistics = collections.defaultdict(list)
        self.test_last_batch = None

    def get_diagnostics(self):
        stats = OrderedDict()
        for k in sorted(self.eval_statistics.keys()):
            stats[k] = np.mean(self.eval_statistics[k])
        return stats

    def dump_scatterplot(self, z, epoch):
        try:
            import matplotlib.pyplot as plt
        except ImportError:
            logger.log(__file__ + ": Unable to load matplotlib. Consider "
                       "setting do_scatterplot to False")
            return
        dim_and_stds = [(i, np.std(z[:, i])) for i in range(z.shape[1])]
        dim_and_stds = sorted(dim_and_stds, key=lambda x: x[1])
        dim1 = dim_and_stds[-1][0]
        dim2 = dim_and_stds[-2][0]
        plt.figure(figsize=(8, 8))
        plt.scatter(z[:, dim1], z[:, dim2], marker='o', edgecolor='none')
        if self.model.dist_mu is not None:
            x1 = self.model.dist_mu[dim1:dim1 + 1]
            y1 = self.model.dist_mu[dim2:dim2 + 1]
            x2 = (self.model.dist_mu[dim1:dim1 + 1] +
                  self.model.dist_std[dim1:dim1 + 1])
            y2 = (self.model.dist_mu[dim2:dim2 + 1] +
                  self.model.dist_std[dim2:dim2 + 1])
        plt.plot([x1, x2], [y1, y2], color='k', linestyle='-', linewidth=2)
        axes = plt.gca()
        axes.set_xlim([-6, 6])
        axes.set_ylim([-6, 6])
        axes.set_title('dim {} vs dim {}'.format(dim1, dim2))
        plt.grid(True)
        save_file = osp.join(self.log_dir, 'scatter%d.png' % epoch)
        plt.savefig(save_file)