from typing import List, Dict, Any, Tuple, Union, Optional from collections import namedtuple import torch import copy from ding.torch_utils import Adam, to_device from ding.rl_utils import q_nstep_td_data, q_nstep_td_error, q_nstep_td_error_with_rescale, get_nstep_return_data, \ get_train_sample from ding.model import model_wrap from ding.utils import POLICY_REGISTRY from ding.utils.data import timestep_collate, default_collate, default_decollate from .base_policy import Policy @POLICY_REGISTRY.register('ngu') class NGUPolicy(Policy): r""" Overview: Policy class of NGU. The corresponding paper is `never give up: learning directed exploration strategies`. Config: == ==================== ======== ============== ======================================== ======================= ID Symbol Type Default Value Description Other(Shape) == ==================== ======== ============== ======================================== ======================= 1 ``type`` str dqn | RL policy register name, refer to | This arg is optional, | registry ``POLICY_REGISTRY`` | a placeholder 2 ``cuda`` bool False | Whether to use cuda for network | This arg can be diff- | erent from modes 3 ``on_policy`` bool False | Whether the RL algorithm is on-policy | or off-policy 4 ``priority`` bool False | Whether use priority(PER) | Priority sample, | update priority 5 | ``priority_IS`` bool False | Whether use Importance Sampling Weight | ``_weight`` | to correct biased update. If True, | priority must be True. 6 | ``discount_`` float 0.997, | Reward's future discount factor, aka. | May be 1 when sparse | ``factor`` [0.95, 0.999] | gamma | reward env 7 ``nstep`` int 3, | N-step reward discount sum for target [3, 5] | q_value estimation 8 ``burnin_step`` int 2 | The timestep of burnin operation, | which is designed to RNN hidden state | difference caused by off-policy 9 | ``learn.update`` int 1 | How many updates(iterations) to train | This args can be vary | ``per_collect`` | after collector's one collection. Only | from envs. Bigger val | valid in serial training | means more off-policy 10 | ``learn.batch_`` int 64 | The number of samples of an iteration | ``size`` 11 | ``learn.learning`` float 0.001 | Gradient step length of an iteration. | ``_rate`` 12 | ``learn.value_`` bool True | Whether use value_rescale function for | ``rescale`` | predicted value 13 | ``learn.target_`` int 100 | Frequence of target network update. | Hard(assign) update | ``update_freq`` 14 | ``learn.ignore_`` bool False | Whether ignore done for target value | Enable it for some | ``done`` | calculation. | fake termination env 15 ``collect.n_sample`` int [8, 128] | The number of training samples of a | It varies from | call of collector. | different envs 16 | ``collect.unroll`` int 1 | unroll length of an iteration | In RNN, unroll_len>1 | ``_len`` == ==================== ======== ============== ======================================== ======================= """ config = dict( # (str) RL policy register name (refer to function "POLICY_REGISTRY"). type='ngu', # (bool) Whether to use cuda for network. cuda=False, # (bool) Whether the RL algorithm is on-policy or off-policy. on_policy=False, # (bool) Whether use priority(priority sample, IS weight, update priority) priority=True, # (bool) Whether use Importance Sampling Weight to correct biased update. If True, priority must be True. priority_IS_weight=True, # ============================================================== # The following configs are algorithm-specific # ============================================================== # (float) Reward's future discount factor, aka. gamma. discount_factor=0.997, # (int) N-step reward for target q_value estimation nstep=5, # (int) the timestep of burnin operation, which is designed to RNN hidden state difference # caused by off-policy burnin_step=20, # (int) is the total length of [sequence sample] minus # the length of burnin part in [sequence sample], # i.e., = = + learn_unroll_len=80, # set this key according to the episode length learn=dict( update_per_collect=1, batch_size=64, learning_rate=0.0001, # ============================================================== # The following configs are algorithm-specific # ============================================================== # (float type) target_update_theta: Used for soft update of the target network, # aka. Interpolation factor in polyak averaging for target networks. target_update_theta=0.001, # (bool) whether use value_rescale function for predicted value value_rescale=True, ignore_done=False, ), collect=dict( # NOTE: It is important that set key traj_len_inf=True here, # to make sure self._traj_len=INF in serial_sample_collector.py. # In sequence-based policy, for each collect_env, # we want to collect data of length self._traj_len=INF # unless the episode enters the 'done' state. # In each collect phase, we collect a total of sequence samples. n_sample=32, traj_len_inf=True, # `env_num` is used in hidden state, should equal to that one in env config. # User should specify this value in user config. env_num=None, ), eval=dict( # `env_num` is used in hidden state, should equal to that one in env config. # User should specify this value in user config. env_num=None, ), other=dict( eps=dict( type='exp', start=0.95, end=0.05, decay=10000, ), replay_buffer=dict(replay_buffer_size=10000, ), ), ) def default_model(self) -> Tuple[str, List[str]]: return 'ngu', ['ding.model.template.ngu'] def _init_learn(self) -> None: r""" Overview: Init the learner model of R2D2Policy Arguments: .. note:: The _init_learn method takes the argument from the self._cfg.learn in the config file - learning_rate (:obj:`float`): The learning rate fo the optimizer - gamma (:obj:`float`): The discount factor - nstep (:obj:`int`): The num of n step return - value_rescale (:obj:`bool`): Whether to use value rescaled loss in algorithm - burnin_step (:obj:`int`): The num of step of burnin """ self._priority = self._cfg.priority self._priority_IS_weight = self._cfg.priority_IS_weight self._optimizer = Adam(self._model.parameters(), lr=self._cfg.learn.learning_rate) self._gamma = self._cfg.discount_factor self._nstep = self._cfg.nstep self._burnin_step = self._cfg.burnin_step self._value_rescale = self._cfg.learn.value_rescale self._target_model = copy.deepcopy(self._model) # here we should not adopt the 'assign' mode of target network here because the reset bug # self._target_model = model_wrap( # self._target_model, # wrapper_name='target', # update_type='assign', # update_kwargs={'freq': self._cfg.learn.target_update_freq} # ) self._target_model = model_wrap( self._target_model, wrapper_name='target', update_type='momentum', update_kwargs={'theta': self._cfg.learn.target_update_theta} ) self._target_model = model_wrap( self._target_model, wrapper_name='hidden_state', state_num=self._cfg.learn.batch_size, save_prev_state=True ) self._learn_model = model_wrap( self._model, wrapper_name='hidden_state', state_num=self._cfg.learn.batch_size, save_prev_state=True ) self._learn_model = model_wrap(self._learn_model, wrapper_name='argmax_sample') self._learn_model.reset() self._target_model.reset() def _data_preprocess_learn(self, data: List[Dict[str, Any]]) -> dict: r""" Overview: Preprocess the data to fit the required data format for learning Arguments: - data (:obj:`List[Dict[str, Any]]`): the data collected from collect function Returns: - data (:obj:`Dict[str, Any]`): the processed data, including at least \ ['main_obs', 'target_obs', 'burnin_obs', 'action', 'reward', 'done', 'weight'] - data_info (:obj:`dict`): the data info, such as replay_buffer_idx, replay_unique_id """ # data preprocess data = timestep_collate(data) if self._cuda: data = to_device(data, self._device) if self._priority_IS_weight: assert self._priority, "Use IS Weight correction, but Priority is not used." if self._priority and self._priority_IS_weight: data['weight'] = data['IS'] else: data['weight'] = data.get('weight', None) bs = self._burnin_step # data['done'], data['weight'], data['value_gamma'] is used in def _forward_learn() to calculate # the q_nstep_td_error, should be length of [self._sequence_len-self._burnin_step] ignore_done = self._cfg.learn.ignore_done if ignore_done: data['done'] = [None for _ in range(self._sequence_len - bs - self._nstep)] else: data['done'] = data['done'][bs:].float() # for computation of online model self._learn_model # NOTE that after the proprocessing of get_nstep_return_data() in _get_train_sample # the data['done'] [t] is already the n-step done # if the data don't include 'weight' or 'value_gamma' then fill in None in a list # with length of [self._sequence_len-self._burnin_step], # below is two different implementation ways if 'value_gamma' not in data: data['value_gamma'] = [None for _ in range(self._sequence_len - bs)] else: data['value_gamma'] = data['value_gamma'][bs:] if 'weight' not in data: data['weight'] = [None for _ in range(self._sequence_len - bs)] else: data['weight'] = data['weight'] * torch.ones_like(data['done']) # every timestep in sequence has same weight, which is the _priority_IS_weight in PER # the burnin_nstep_obs is used to calculate the init hidden state of rnn for the calculation of the q_value, # target_q_value, and target_q_action data['burnin_nstep_obs'] = data['obs'][:bs + self._nstep] data['burnin_nstep_action'] = data['action'][:bs + self._nstep] data['burnin_nstep_reward'] = data['reward'][:bs + self._nstep] data['burnin_nstep_beta'] = data['beta'][:bs + self._nstep] # split obs into three parts 'burnin_obs' [0:bs], 'main_obs' [bs:bs+nstep], 'target_obs' [bs+nstep:] # data['burnin_obs'] = data['obs'][:bs] data['main_obs'] = data['obs'][bs:-self._nstep] data['target_obs'] = data['obs'][bs + self._nstep:] # data['burnin_action'] = data['action'][:bs] data['main_action'] = data['action'][bs:-self._nstep] data['target_action'] = data['action'][bs + self._nstep:] # data['burnin_reward'] = data['reward'][:bs] data['main_reward'] = data['reward'][bs:-self._nstep] data['target_reward'] = data['reward'][bs + self._nstep:] # data['burnin_beta'] = data['beta'][:bs] data['main_beta'] = data['beta'][bs:-self._nstep] data['target_beta'] = data['beta'][bs + self._nstep:] # Note that Must be here after the previous slicing operation data['action'] = data['action'][bs:-self._nstep] data['reward'] = data['reward'][bs:-self._nstep] return data def _forward_learn(self, data: dict) -> Dict[str, Any]: r""" Overview: Forward and backward function of learn mode. Acquire the data, calculate the loss and optimize learner model. Arguments: - data (:obj:`dict`): Dict type data, including at least \ ['main_obs', 'target_obs', 'burnin_obs', 'action', 'reward', 'done', 'weight'] Returns: - info_dict (:obj:`Dict[str, Any]`): Including cur_lr and total_loss - cur_lr (:obj:`float`): Current learning rate - total_loss (:obj:`float`): The calculated loss """ # forward data = self._data_preprocess_learn(data) self._learn_model.train() self._target_model.train() # use the hidden state in timestep=0 self._learn_model.reset(data_id=None, state=data['prev_state'][0]) self._target_model.reset(data_id=None, state=data['prev_state'][0]) if len(data['burnin_nstep_obs']) != 0: with torch.no_grad(): inputs = { 'obs': data['burnin_nstep_obs'], 'action': data['burnin_nstep_action'], 'reward': data['burnin_nstep_reward'], 'beta': data['burnin_nstep_beta'], 'enable_fast_timestep': True } tmp = self._learn_model.forward( inputs, saved_state_timesteps=[self._burnin_step, self._burnin_step + self._nstep] ) tmp_target = self._target_model.forward( inputs, saved_state_timesteps=[self._burnin_step, self._burnin_step + self._nstep] ) inputs = { 'obs': data['main_obs'], 'action': data['main_action'], 'reward': data['main_reward'], 'beta': data['main_beta'], 'enable_fast_timestep': True } self._learn_model.reset(data_id=None, state=tmp['saved_state'][0]) q_value = self._learn_model.forward(inputs)['logit'] self._learn_model.reset(data_id=None, state=tmp['saved_state'][1]) self._target_model.reset(data_id=None, state=tmp_target['saved_state'][1]) next_inputs = { 'obs': data['target_obs'], 'action': data['target_action'], 'reward': data['target_reward'], 'beta': data['target_beta'], 'enable_fast_timestep': True } with torch.no_grad(): target_q_value = self._target_model.forward(next_inputs)['logit'] # argmax_action double_dqn target_q_action = self._learn_model.forward(next_inputs)['action'] action, reward, done, weight = data['action'], data['reward'], data['done'], data['weight'] value_gamma = [ None for _ in range(self._sequence_len - self._burnin_step) ] # NOTE this is important, because we use diffrent gamma according to their beta in NGU alg. # T, B, nstep -> T, nstep, B reward = reward.permute(0, 2, 1).contiguous() loss = [] td_error = [] self._gamma = [self.index_to_gamma[int(i)] for i in data['main_beta'][0]] # T, B -> B, e.g. 75,64 -> 64 # reward torch.Size([4, 5, 64]) for t in range(self._sequence_len - self._burnin_step - self._nstep): # here t=0 means timestep in the original sample sequence, we minus self._nstep # because for the last timestep in the sequence, we don't have their target obs td_data = q_nstep_td_data( q_value[t], target_q_value[t], action[t], target_q_action[t], reward[t], done[t], weight[t] ) if self._value_rescale: l, e = q_nstep_td_error_with_rescale(td_data, self._gamma, self._nstep, value_gamma=value_gamma[t]) loss.append(l) td_error.append(e.abs()) else: l, e = q_nstep_td_error(td_data, self._gamma, self._nstep, value_gamma=value_gamma[t]) loss.append(l) td_error.append(e.abs()) loss = sum(loss) / (len(loss) + 1e-8) # using the mixture of max and mean absolute n-step TD-errors as the priority of the sequence td_error_per_sample = 0.9 * torch.max( torch.stack(td_error), dim=0 )[0] + (1 - 0.9) * (torch.sum(torch.stack(td_error), dim=0) / (len(td_error) + 1e-8)) # td_error shape list(, B), # for example, (75,64) # torch.sum(torch.stack(td_error), dim=0) can also be replaced with sum(td_error) # update self._optimizer.zero_grad() loss.backward() self._optimizer.step() # after update self._target_model.update(self._learn_model.state_dict()) # the information for debug batch_range = torch.arange(action[0].shape[0]) q_s_a_t0 = q_value[0][batch_range, action[0]] target_q_s_a_t0 = target_q_value[0][batch_range, target_q_action[0]] return { 'cur_lr': self._optimizer.defaults['lr'], 'total_loss': loss.item(), 'priority': td_error_per_sample.abs().tolist(), # the first timestep in the sequence, may not be the start of episode 'q_s_taken-a_t0': q_s_a_t0.mean().item(), 'target_q_s_max-a_t0': target_q_s_a_t0.mean().item(), 'q_s_a-mean_t0': q_value[0].mean().item(), } def _reset_learn(self, data_id: Optional[List[int]] = None) -> None: self._learn_model.reset(data_id=data_id) def _state_dict_learn(self) -> Dict[str, Any]: return { 'model': self._learn_model.state_dict(), 'target_model': self._target_model.state_dict(), 'optimizer': self._optimizer.state_dict(), } def _load_state_dict_learn(self, state_dict: Dict[str, Any]) -> None: self._learn_model.load_state_dict(state_dict['model']) self._target_model.load_state_dict(state_dict['target_model']) self._optimizer.load_state_dict(state_dict['optimizer']) def _init_collect(self) -> None: r""" Overview: Collect mode init method. Called by ``self.__init__``. Init traj and unroll length, collect model. """ assert 'unroll_len' not in self._cfg.collect, "ngu use default " self._nstep = self._cfg.nstep self._burnin_step = self._cfg.burnin_step self._gamma = self._cfg.discount_factor self._sequence_len = self._cfg.learn_unroll_len + self._cfg.burnin_step self._unroll_len = self._sequence_len self._collect_model = model_wrap( self._model, wrapper_name='hidden_state', state_num=self._cfg.collect.env_num, save_prev_state=True ) self._collect_model = model_wrap(self._collect_model, wrapper_name='eps_greedy_sample') self._collect_model.reset() self.index_to_gamma = { # NOTE i: 1 - torch.exp( ( (self._cfg.collect.env_num - 1 - i) * torch.log(torch.tensor(1 - 0.997)) + i * torch.log(torch.tensor(1 - 0.99)) ) / (self._cfg.collect.env_num - 1) ) for i in range(self._cfg.collect.env_num) } # NOTE: for NGU policy collect phase self.beta_index = { i: torch.randint(0, self._cfg.collect.env_num, [1]) for i in range(self._cfg.collect.env_num) } # epsilon=0.4, alpha=9 self.eps = {i: 0.4 ** (1 + 8 * i / (self._cfg.collect.env_num - 1)) for i in range(self._cfg.collect.env_num)} def _forward_collect(self, data: dict) -> dict: r""" Overview: Collect output according to eps_greedy plugin Arguments: - data (:obj:`dict`): Dict type data, including at least ['obs']. Returns: - data (:obj:`dict`): The collected data """ data_id = list(data.keys()) data = default_collate(list(data.values())) obs = data['obs'] prev_action = data['prev_action'].long() prev_reward_extrinsic = data['prev_reward_extrinsic'] beta_index = default_collate(list(self.beta_index.values())) if len(data_id) != self._cfg.collect.env_num: # in case, some env is in reset state and only return part data beta_index = beta_index[data_id] if self._cuda: obs = to_device(obs, self._device) beta_index = to_device(beta_index, self._device) prev_action = to_device(prev_action, self._device) prev_reward_extrinsic = to_device(prev_reward_extrinsic, self._device) # TODO(pu): add prev_reward_intrinsic to network input, # reward uses some kind of embedding instead of 1D value data = { 'obs': obs, 'prev_action': prev_action, 'prev_reward_extrinsic': prev_reward_extrinsic, 'beta': beta_index } self._collect_model.eval() with torch.no_grad(): output = self._collect_model.forward(data, data_id=data_id, eps=self.eps, inference=True) if self._cuda: output = to_device(output, 'cpu') output = default_decollate(output) return {i: d for i, d in zip(data_id, output)} def _reset_collect(self, data_id: Optional[List[int]] = None) -> None: self._collect_model.reset(data_id=data_id) # NOTE: for NGU policy, in collect phase, each episode, we sample a new beta for each env if data_id is not None: self.beta_index[data_id[0]] = torch.randint(0, self._cfg.collect.env_num, [1]) def _process_transition(self, obs: Any, model_output: dict, timestep: namedtuple, env_id) -> dict: r""" Overview: Generate dict type transition data from inputs. Arguments: - obs (:obj:`Any`): Env observation - model_output (:obj:`dict`): Output of collect model, including at least ['action', 'prev_state'] - timestep (:obj:`namedtuple`): Output after env step, including at least ['reward', 'done'] \ (here 'obs' indicates obs after env step). Returns: - transition (:obj:`dict`): Dict type transition data. """ if hasattr(timestep, 'null'): transition = { 'beta': self.beta_index[env_id], 'obs': obs['obs'], # NOTE: input obs including obs, prev_action, prev_reward_extrinsic 'action': model_output['action'], 'prev_state': model_output['prev_state'], 'reward': timestep.reward, 'done': timestep.done, 'null': timestep.null, } else: transition = { 'beta': self.beta_index[env_id], 'obs': obs['obs'], # NOTE: input obs including obs, prev_action, prev_reward_extrinsic 'action': model_output['action'], 'prev_state': model_output['prev_state'], 'reward': timestep.reward, 'done': timestep.done, 'null': False, } return transition def _get_train_sample(self, data: list) -> Union[None, List[Any]]: r""" Overview: Get the trajectory and the n step return data, then sample from the n_step return data Arguments: - data (:obj:`list`): The trajectory's cache Returns: - samples (:obj:`dict`): The training samples generated """ data = get_nstep_return_data(data, self._nstep, gamma=self.index_to_gamma[int(data[0]['beta'])].item()) return get_train_sample(data, self._sequence_len) def _init_eval(self) -> None: r""" Overview: Evaluate mode init method. Called by ``self.__init__``. Init eval model with argmax strategy. """ self._eval_model = model_wrap(self._model, wrapper_name='hidden_state', state_num=self._cfg.eval.env_num) self._eval_model = model_wrap(self._eval_model, wrapper_name='argmax_sample') self._eval_model.reset() # NOTE: for NGU policy eval phase # beta_index = 0 -> beta is approximately 0 self.beta_index = {i: torch.tensor([0]) for i in range(self._cfg.eval.env_num)} def _forward_eval(self, data: dict) -> dict: r""" Overview: Forward function of collect mode, similar to ``self._forward_collect``. Arguments: - data (:obj:`dict`): Dict type data, including at least ['obs']. Returns: - output (:obj:`dict`): Dict type data, including at least inferred action according to input obs. """ data_id = list(data.keys()) data = default_collate(list(data.values())) obs = data['obs'] prev_action = data['prev_action'].long() prev_reward_extrinsic = data['prev_reward_extrinsic'] beta_index = default_collate(list(self.beta_index.values())) if len(data_id) != self._cfg.collect.env_num: # in case, some env is in reset state and only return part data beta_index = beta_index[data_id] if self._cuda: obs = to_device(obs, self._device) beta_index = to_device(beta_index, self._device) prev_action = to_device(prev_action, self._device) prev_reward_extrinsic = to_device(prev_reward_extrinsic, self._device) # TODO(pu): add prev_reward_intrinsic to network input, # reward uses some kind of embedding instead of 1D value data = { 'obs': obs, 'prev_action': prev_action, 'prev_reward_extrinsic': prev_reward_extrinsic, 'beta': beta_index } self._eval_model.eval() with torch.no_grad(): output = self._eval_model.forward(data, data_id=data_id, inference=True) if self._cuda: output = to_device(output, 'cpu') output = default_decollate(output) return {i: d for i, d in zip(data_id, output)} def _reset_eval(self, data_id: Optional[List[int]] = None) -> None: self._eval_model.reset(data_id=data_id) def _monitor_vars_learn(self) -> List[str]: return super()._monitor_vars_learn() + [ 'total_loss', 'priority', 'q_s_taken-a_t0', 'target_q_s_max-a_t0', 'q_s_a-mean_t0' ]