from collections import namedtuple from typing import List, Dict, Any, Tuple import torch import treetensor.torch as ttorch from ding.model import model_wrap from ding.rl_utils import vtrace_data, vtrace_error_discrete_action, vtrace_error_continuous_action, get_train_sample from ding.torch_utils import Adam, RMSprop, to_device from ding.utils import POLICY_REGISTRY from ding.utils.data import default_collate, default_decollate, ttorch_collate from ding.policy.base_policy import Policy @POLICY_REGISTRY.register('impala') class IMPALAPolicy(Policy): """ Overview: Policy class of IMPALA algorithm. Paper link: https://arxiv.org/abs/1802.01561. Config: == ==================== ======== ============== ======================================== ======================= ID Symbol Type Default Value Description Other(Shape) == ==================== ======== ============== ======================================== ======================= 1 ``type`` str impala | 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_`` bool False | Whether use Importance Sampling Weight | If True, priority | ``IS_weight`` | | must be True 6 ``unroll_len`` int 32 | trajectory length to calculate v-trace | target 7 | ``learn.update`` int 4 | 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 == ==================== ======== ============== ======================================== ======================= """ config = dict( # (str) RL policy register name (refer to function "POLICY_REGISTRY"). type='impala', # (bool) Whether to use cuda in policy. cuda=False, # (bool) Whether learning policy is the same as collecting data policy(on-policy). on_policy=False, # (bool) Whether to enable priority experience sample. priority=False, # (bool) Whether use Importance Sampling Weight to correct biased update. If True, priority must be True. priority_IS_weight=False, # (str) Which kind of action space used in IMPALAPolicy, ['discrete', 'continuous']. action_space='discrete', # (int) the trajectory length to calculate v-trace target. unroll_len=32, # (bool) Whether to need policy data in process transition. transition_with_policy_data=True, # learn_mode config learn=dict( # (int) collect n_sample data, train model update_per_collect times. update_per_collect=4, # (int) the number of data for a train iteration. batch_size=16, # (float) The step size of gradient descent. learning_rate=0.0005, # (float) loss weight of the value network, the weight of policy network is set to 1. value_weight=0.5, # (float) loss weight of the entropy regularization, the weight of policy network is set to 1. entropy_weight=0.0001, # (float) discount factor for future reward, defaults int [0, 1]. discount_factor=0.99, # (float) additional discounting parameter. lambda_=0.95, # (float) clip ratio of importance weights. rho_clip_ratio=1.0, # (float) clip ratio of importance weights. c_clip_ratio=1.0, # (float) clip ratio of importance sampling. rho_pg_clip_ratio=1.0, # (str) The gradient clip operation type used in IMPALA, ['clip_norm', clip_value', 'clip_momentum_norm']. grad_clip_type=None, # (float) The gradient clip target value used in IMPALA. # If ``grad_clip_type`` is 'clip_norm', then the maximum of gradient will be normalized to this value. clip_value=0.5, # (str) Optimizer used to train the network, ['adam', 'rmsprop']. optim='adam', ), # collect_mode config collect=dict( # (int) How many training samples collected in one collection procedure. # Only one of [n_sample, n_episode] shoule be set. # n_sample=16, ), eval=dict(), # for compatibility other=dict( replay_buffer=dict( # (int) Maximum size of replay buffer. Usually, larger buffer size is better. replay_buffer_size=1000, # (int) Maximum use times for a sample in buffer. If reaches this value, the sample will be removed. max_use=16, ), ), ) def default_model(self) -> Tuple[str, List[str]]: """ Overview: Return this algorithm default neural network model setting for demonstration. ``__init__`` method will \ automatically call this method to get the default model setting and create model. Returns: - model_info (:obj:`Tuple[str, List[str]]`): The registered model name and model's import_names. .. note:: The user can define and use customized network model but must obey the same inferface definition indicated \ by import_names path. For example about IMPALA , its registered name is ``vac`` and the import_names is \ ``ding.model.template.vac``. """ return 'vac', ['ding.model.template.vac'] def _init_learn(self) -> None: """ Overview: Initialize the learn mode of policy, including related attributes and modules. For IMPALA, it mainly \ contains optimizer, algorithm-specific arguments such as loss weight and gamma, main (learn) model. This method will be called in ``__init__`` method if ``learn`` field is in ``enable_field``. .. note:: For the member variables that need to be saved and loaded, please refer to the ``_state_dict_learn`` \ and ``_load_state_dict_learn`` methods. .. note:: For the member variables that need to be monitored, please refer to the ``_monitor_vars_learn`` method. .. note:: If you want to set some spacial member variables in ``_init_learn`` method, you'd better name them \ with prefix ``_learn_`` to avoid conflict with other modes, such as ``self._learn_attr1``. """ assert self._cfg.action_space in ["continuous", "discrete"], self._cfg.action_space self._action_space = self._cfg.action_space # Optimizer optim_type = self._cfg.learn.optim if optim_type == 'rmsprop': self._optimizer = RMSprop(self._model.parameters(), lr=self._cfg.learn.learning_rate) elif optim_type == 'adam': self._optimizer = Adam( self._model.parameters(), grad_clip_type=self._cfg.learn.grad_clip_type, clip_value=self._cfg.learn.clip_value, lr=self._cfg.learn.learning_rate ) else: raise NotImplementedError("Now only support rmsprop and adam, but input is {}".format(optim_type)) self._learn_model = model_wrap(self._model, wrapper_name='base') self._action_shape = self._cfg.model.action_shape self._unroll_len = self._cfg.unroll_len # Algorithm config self._priority = self._cfg.priority self._priority_IS_weight = self._cfg.priority_IS_weight self._value_weight = self._cfg.learn.value_weight self._entropy_weight = self._cfg.learn.entropy_weight self._gamma = self._cfg.learn.discount_factor self._lambda = self._cfg.learn.lambda_ self._rho_clip_ratio = self._cfg.learn.rho_clip_ratio self._c_clip_ratio = self._cfg.learn.c_clip_ratio self._rho_pg_clip_ratio = self._cfg.learn.rho_pg_clip_ratio # Main model self._learn_model.reset() def _data_preprocess_learn(self, data: List[Dict[str, Any]]): """ Overview: Data preprocess function of learn mode. Convert list trajectory data to to trajectory data, which is a dict of tensors. Arguments: - data (:obj:`List[Dict[str, Any]]`): List type data, a list of data for training. Each list element is a \ dict, whose values are torch.Tensor or np.ndarray or dict/list combinations, keys include at least \ 'obs', 'next_obs', 'logit', 'action', 'reward', 'done' Returns: - data (:obj:`dict`): Dict type data. Values are torch.Tensor or np.ndarray or dict/list combinations. \ ReturnsKeys: - necessary: 'logit', 'action', 'reward', 'done', 'weight', 'obs_plus_1'. - optional and not used in later computation: 'obs', 'next_obs'.'IS', 'collect_iter', 'replay_unique_id', \ 'replay_buffer_idx', 'priority', 'staleness', 'use'. ReturnsShapes: - obs_plus_1 (:obj:`torch.FloatTensor`): :math:`(T * B, obs_shape)`, where T is timestep, B is batch size \ and obs_shape is the shape of single env observation - logit (:obj:`torch.FloatTensor`): :math:`(T, B, N)`, where N is action dim - action (:obj:`torch.LongTensor`): :math:`(T, B)` - reward (:obj:`torch.FloatTensor`): :math:`(T+1, B)` - done (:obj:`torch.FloatTensor`): :math:`(T, B)` - weight (:obj:`torch.FloatTensor`): :math:`(T, B)` """ elem = data[0] if isinstance(elem, dict): # old pipeline data = default_collate(data) elif isinstance(elem, list): # new task pipeline data = default_collate(default_collate(data)) else: raise TypeError("not support element type ({}) in IMPALA".format(type(elem))) 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) if isinstance(elem, dict): # old pipeline for k in data: if isinstance(data[k], list): data[k] = default_collate(data[k]) data['obs_plus_1'] = torch.cat([data['obs'], data['next_obs'][-1:]], dim=0) # shape (T+1)*B,env_obs_shape return data def _forward_learn(self, data: List[Dict[str, Any]]) -> Dict[str, Any]: """ Overview: Policy forward function of learn mode (training policy and updating parameters). Forward means \ that the policy inputs some training batch data from the replay buffer and then returns the output \ result, including various training information such as loss and current learning rate. Arguments: - data (:obj:`List[Dict[int, Any]]`): The input data used for policy forward, including a batch of \ training samples. For each element in list, the key of the dict is the name of data items and the \ value is the corresponding data. Usually, the value is torch.Tensor or np.ndarray or there dict/list \ combinations. In the ``_forward_learn`` method, data often need to first be stacked in the batch \ dimension by some utility functions such as ``default_preprocess_learn``. \ For IMPALA, each element in list is a dict containing at least the following keys: ``obs``, \ ``action``, ``logit``, ``reward``, ``next_obs``, ``done``. Sometimes, it also contains other keys such \ as ``weight``. Returns: - info_dict (:obj:`Dict[str, Any]`): The information dict that indicated training result, which will be \ recorded in text log and tensorboard, values must be python scalar or a list of scalars. For the \ detailed definition of the dict, refer to the code of ``_monitor_vars_learn`` method. .. note:: The input value can be torch.Tensor or dict/list combinations and current policy supports all of them. \ For the data type that not supported, the main reason is that the corresponding model does not support it. \ You can implement you own model rather than use the default model. For more information, please raise an \ issue in GitHub repo and we will continue to follow up. .. note:: For more detailed examples, please refer to unittest for IMPALAPolicy: ``ding.policy.tests.test_impala``. """ data = self._data_preprocess_learn(data) # ==================== # IMPALA forward # ==================== self._learn_model.train() output = self._learn_model.forward( data['obs_plus_1'].view((-1, ) + data['obs_plus_1'].shape[2:]), mode='compute_actor_critic' ) target_logit, behaviour_logit, actions, values, rewards, weights = self._reshape_data(output, data) # Calculate vtrace error data = vtrace_data(target_logit, behaviour_logit, actions, values, rewards, weights) g, l, r, c, rg = self._gamma, self._lambda, self._rho_clip_ratio, self._c_clip_ratio, self._rho_pg_clip_ratio if self._action_space == 'continuous': vtrace_loss = vtrace_error_continuous_action(data, g, l, r, c, rg) elif self._action_space == 'discrete': vtrace_loss = vtrace_error_discrete_action(data, g, l, r, c, rg) wv, we = self._value_weight, self._entropy_weight total_loss = vtrace_loss.policy_loss + wv * vtrace_loss.value_loss - we * vtrace_loss.entropy_loss # ==================== # IMPALA update # ==================== self._optimizer.zero_grad() total_loss.backward() self._optimizer.step() return { 'cur_lr': self._optimizer.defaults['lr'], 'total_loss': total_loss.item(), 'policy_loss': vtrace_loss.policy_loss.item(), 'value_loss': vtrace_loss.value_loss.item(), 'entropy_loss': vtrace_loss.entropy_loss.item(), } def _reshape_data(self, output: Dict[str, Any], data: Dict[str, Any]) -> Tuple: """ Overview: Obtain weights for loss calculating, where should be 0 for done positions. Update values and rewards with \ the weight. Arguments: - output (:obj:`Dict[int, Any]`): Dict type data, output of learn_model forward. \ Values are torch.Tensor or np.ndarray or dict/list combinations,keys are value, logit. - data (:obj:`Dict[int, Any]`): Dict type data, input of policy._forward_learn Values are torch.Tensor or \ np.ndarray or dict/list combinations. Keys includes at least ['logit', 'action', 'reward', 'done']. Returns: - data (:obj:`Tuple[Any]`): Tuple of target_logit, behaviour_logit, actions, values, rewards, weights. ReturnsShapes: - target_logit (:obj:`torch.FloatTensor`): :math:`((T+1), B, Obs_Shape)`, where T is timestep,\ B is batch size and Obs_Shape is the shape of single env observation. - behaviour_logit (:obj:`torch.FloatTensor`): :math:`(T, B, N)`, where N is action dim. - actions (:obj:`torch.LongTensor`): :math:`(T, B)` - values (:obj:`torch.FloatTensor`): :math:`(T+1, B)` - rewards (:obj:`torch.FloatTensor`): :math:`(T, B)` - weights (:obj:`torch.FloatTensor`): :math:`(T, B)` """ if self._action_space == 'continuous': target_logit = {} target_logit['mu'] = output['logit']['mu'].reshape(self._unroll_len + 1, -1, self._action_shape)[:-1 ] # shape (T+1),B,env_action_shape target_logit['sigma'] = output['logit']['sigma'].reshape(self._unroll_len + 1, -1, self._action_shape )[:-1] # shape (T+1),B,env_action_shape elif self._action_space == 'discrete': target_logit = output['logit'].reshape(self._unroll_len + 1, -1, self._action_shape)[:-1] # shape (T+1),B,env_action_shape behaviour_logit = data['logit'] # shape T,B actions = data['action'] # shape T,B for discrete # shape T,B,env_action_shape for continuous values = output['value'].reshape(self._unroll_len + 1, -1) # shape T+1,B,env_action_shape rewards = data['reward'] # shape T,B weights_ = 1 - data['done'].float() # shape T,B weights = torch.ones_like(rewards) # shape T,B values[1:] = values[1:] * weights_ weights[1:] = weights_[:-1] rewards = rewards * weights # shape T,B return target_logit, behaviour_logit, actions, values, rewards, weights def _init_collect(self) -> None: """ Overview: Initialize the collect mode of policy, including related attributes and modules. For IMPALA, it contains \ the collect_model to balance the exploration and exploitation (e.g. the multinomial sample mechanism in \ discrete action space), and other algorithm-specific arguments such as unroll_len. This method will be called in ``__init__`` method if ``collect`` field is in ``enable_field``. .. note:: If you want to set some spacial member variables in ``_init_collect`` method, you'd better name them \ with prefix ``_collect_`` to avoid conflict with other modes, such as ``self._collect_attr1``. """ assert self._cfg.action_space in ["continuous", "discrete"] self._action_space = self._cfg.action_space if self._action_space == 'continuous': self._collect_model = model_wrap(self._model, wrapper_name='reparam_sample') elif self._action_space == 'discrete': self._collect_model = model_wrap(self._model, wrapper_name='multinomial_sample') self._collect_model.reset() def _forward_collect(self, data: Dict[int, Any]) -> Dict[int, Any]: """ Overview: Policy forward function of collect mode (collecting training data by interacting with envs). Forward means \ that the policy gets some necessary data (mainly observation) from the envs and then returns the output \ data, such as the action to interact with the envs. Arguments: - data (:obj:`Dict[int, Any]`): The input data used for policy forward, including at least the obs. The \ key of the dict is environment id and the value is the corresponding data of the env. Returns: - output (:obj:`Dict[int, Any]`): The output data of policy forward, including at least the action and \ other necessary data (action logit and value) for learn mode defined in ``self._process_transition`` \ method. The key of the dict is the same as the input data, i.e. environment id. .. tip:: If you want to add more tricks on this policy, like temperature factor in multinomial sample, you can pass \ related data as extra keyword arguments of this method. .. note:: The input value can be torch.Tensor or dict/list combinations and current policy supports all of them. \ For the data type that not supported, the main reason is that the corresponding model does not support it. \ You can implement you own model rather than use the default model. For more information, please raise an \ issue in GitHub repo and we will continue to follow up. .. note:: For more detailed examples, please refer to unittest for IMPALAPolicy: ``ding.policy.tests.test_impala``. """ data_id = list(data.keys()) data = default_collate(list(data.values())) if self._cuda: data = to_device(data, self._device) self._collect_model.eval() with torch.no_grad(): output = self._collect_model.forward(data, mode='compute_actor') if self._cuda: output = to_device(output, 'cpu') output = default_decollate(output) output = {i: d for i, d in zip(data_id, output)} return output def _get_train_sample(self, data: List[Dict[str, Any]]) -> List[Dict[str, Any]]: """ Overview: For a given trajectory (transitions, a list of transition) data, process it into a list of sample that \ can be used for training. In IMPALA, a train sample is processed transitions with unroll_len length. Arguments: - transitions (:obj:`List[Dict[str, Any]`): The trajectory data (a list of transition), each element is \ the same format as the return value of ``self._process_transition`` method. Returns: - samples (:obj:`List[Dict[str, Any]]`): The processed train samples, each element is the similar format \ as input transitions, but may contain more data for training. """ return get_train_sample(data, self._unroll_len) def _process_transition(self, obs: torch.Tensor, policy_output: Dict[str, torch.Tensor], timestep: namedtuple) -> Dict[str, torch.Tensor]: """ Overview: Process and pack one timestep transition data into a dict, which can be directly used for training and \ saved in replay buffer. For IMPALA, it contains obs, next_obs, action, reward, done, logit. Arguments: - obs (:obj:`torch.Tensor`): The env observation of current timestep, such as stacked 2D image in Atari. - policy_output (:obj:`Dict[str, torch.Tensor]`): The output of the policy network with the observation \ as input. For IMPALA, it contains the action and the logit of the action. - timestep (:obj:`namedtuple`): The execution result namedtuple returned by the environment step method, \ except all the elements have been transformed into tensor data. Usually, it contains the next obs, \ reward, done, info, etc. Returns: - transition (:obj:`Dict[str, torch.Tensor]`): The processed transition data of the current timestep. """ transition = { 'obs': obs, 'next_obs': timestep.obs, 'logit': policy_output['logit'], 'action': policy_output['action'], 'reward': timestep.reward, 'done': timestep.done, } return transition def _init_eval(self) -> None: """ Overview: Initialize the eval mode of policy, including related attributes and modules. For IMPALA, it contains the \ eval model to select optimial action (e.g. greedily select action with argmax mechanism in discrete action). This method will be called in ``__init__`` method if ``eval`` field is in ``enable_field``. .. note:: If you want to set some spacial member variables in ``_init_eval`` method, you'd better name them \ with prefix ``_eval_`` to avoid conflict with other modes, such as ``self._eval_attr1``. """ assert self._cfg.action_space in ["continuous", "discrete"], self._cfg.action_space self._action_space = self._cfg.action_space if self._action_space == 'continuous': self._eval_model = model_wrap(self._model, wrapper_name='deterministic_sample') elif self._action_space == 'discrete': self._eval_model = model_wrap(self._model, wrapper_name='argmax_sample') self._eval_model.reset() def _forward_eval(self, data: Dict[int, Any]) -> Dict[int, Any]: """ Overview: Policy forward function of eval mode (evaluation policy performance by interacting with envs). Forward \ means that the policy gets some necessary data (mainly observation) from the envs and then returns the \ action to interact with the envs. ``_forward_eval`` in IMPALA often uses deterministic sample to get \ actions while ``_forward_collect`` usually uses stochastic sample method for balance exploration and \ exploitation. Arguments: - data (:obj:`Dict[int, Any]`): The input data used for policy forward, including at least the obs. The \ key of the dict is environment id and the value is the corresponding data of the env. Returns: - output (:obj:`Dict[int, Any]`): The output data of policy forward, including at least the action. The \ key of the dict is the same as the input data, i.e. environment id. .. note:: The input value can be torch.Tensor or dict/list combinations and current policy supports all of them. \ For the data type that not supported, the main reason is that the corresponding model does not support it. \ You can implement you own model rather than use the default model. For more information, please raise an \ issue in GitHub repo and we will continue to follow up. .. note:: For more detailed examples, please refer to unittest for IMPALAPolicy: ``ding.policy.tests.test_impala``. """ data_id = list(data.keys()) data = default_collate(list(data.values())) if self._cuda: data = to_device(data, self._device) self._eval_model.eval() with torch.no_grad(): output = self._eval_model.forward(data, mode='compute_actor') if self._cuda: output = to_device(output, 'cpu') output = default_decollate(output) output = {i: d for i, d in zip(data_id, output)} return output def _monitor_vars_learn(self) -> List[str]: """ Overview: Return the necessary keys for logging the return dict of ``self._forward_learn``. The logger module, such \ as text logger, tensorboard logger, will use these keys to save the corresponding data. Returns: - necessary_keys (:obj:`List[str]`): The list of the necessary keys to be logged. """ return super()._monitor_vars_learn() + ['policy_loss', 'value_loss', 'entropy_loss']