def from_logits(behaviour_policy_logits, target_policy_logits, actions, discounts, rewards, values, bootstrap_value, dist_class=TorchCategorical, model=None, clip_rho_threshold=1.0, clip_pg_rho_threshold=1.0): """multi_from_logits wrapper used only for tests""" res = multi_from_logits( [behaviour_policy_logits], [target_policy_logits], [actions], discounts, rewards, values, bootstrap_value, dist_class, model, clip_rho_threshold=clip_rho_threshold, clip_pg_rho_threshold=clip_pg_rho_threshold) assert len(res.behaviour_action_log_probs) == 1 assert len(res.target_action_log_probs) == 1 return VTraceFromLogitsReturns( vs=res.vs, pg_advantages=res.pg_advantages, log_rhos=res.log_rhos, behaviour_action_log_probs=res.behaviour_action_log_probs[0], target_action_log_probs=res.target_action_log_probs[0], )
def multi_from_logits(behaviour_policy_logits, target_policy_logits, actions, discounts, rewards, values, bootstrap_value, dist_class, model, behaviour_action_log_probs=None, clip_rho_threshold=1.0, clip_pg_rho_threshold=1.0): """V-trace for softmax policies. Calculates V-trace actor critic targets for softmax polices as described in "IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures" by Espeholt, Soyer, Munos et al. Target policy refers to the policy we are interested in improving and behaviour policy refers to the policy that generated the given rewards and actions. In the notation used throughout documentation and comments, T refers to the time dimension ranging from 0 to T-1. B refers to the batch size and ACTION_SPACE refers to the list of numbers each representing a number of actions. Args: behaviour_policy_logits: A list with length of ACTION_SPACE of float32 tensors of shapes [T, B, ACTION_SPACE[0]], ..., [T, B, ACTION_SPACE[-1]] with un-normalized log-probabilities parameterizing the softmax behavior policy. target_policy_logits: A list with length of ACTION_SPACE of float32 tensors of shapes [T, B, ACTION_SPACE[0]], ..., [T, B, ACTION_SPACE[-1]] with un-normalized log-probabilities parameterizing the softmax target policy. actions: A list with length of ACTION_SPACE of tensors of shapes [T, B, ...], ..., [T, B, ...] with actions sampled from the behavior policy. discounts: A float32 tensor of shape [T, B] with the discount encountered when following the behavior policy. rewards: A float32 tensor of shape [T, B] with the rewards generated by following the behavior policy. values: A float32 tensor of shape [T, B] with the value function estimates wrt. the target policy. bootstrap_value: A float32 of shape [B] with the value function estimate at time T. dist_class: action distribution class for the logits. model: backing ModelV2 instance behaviour_action_log_probs: Precalculated values of the behavior actions. clip_rho_threshold: A scalar float32 tensor with the clipping threshold for importance weights (rho) when calculating the baseline targets (vs). rho^bar in the paper. clip_pg_rho_threshold: A scalar float32 tensor with the clipping threshold on rho_s in: \rho_s \delta log \pi(a|x) (r + \gamma v_{s+1} - V(x_s)). Returns: A `VTraceFromLogitsReturns` namedtuple with the following fields: vs: A float32 tensor of shape [T, B]. Can be used as target to train a baseline (V(x_t) - vs_t)^2. pg_advantages: A float 32 tensor of shape [T, B]. Can be used as an estimate of the advantage in the calculation of policy gradients. log_rhos: A float32 tensor of shape [T, B] containing the log importance sampling weights (log rhos). behaviour_action_log_probs: A float32 tensor of shape [T, B] containing behaviour policy action log probabilities (log \mu(a_t)). target_action_log_probs: A float32 tensor of shape [T, B] containing target policy action probabilities (log \pi(a_t)). """ behaviour_policy_logits = convert_to_torch_tensor( behaviour_policy_logits, device="cpu") target_policy_logits = convert_to_torch_tensor( target_policy_logits, device="cpu") actions = convert_to_torch_tensor(actions, device="cpu") for i in range(len(behaviour_policy_logits)): # Make sure tensor ranks are as expected. # The rest will be checked by from_action_log_probs. assert len(behaviour_policy_logits[i].size()) == 3 assert len(target_policy_logits[i].size()) == 3 target_action_log_probs = multi_log_probs_from_logits_and_actions( target_policy_logits, actions, dist_class, model) if (len(behaviour_policy_logits) > 1 or behaviour_action_log_probs is None): # can't use precalculated values, recompute them. Note that # recomputing won't work well for autoregressive action dists # which may have variables not captured by 'logits' behaviour_action_log_probs = (multi_log_probs_from_logits_and_actions( behaviour_policy_logits, actions, dist_class, model)) behaviour_action_log_probs = force_list(behaviour_action_log_probs) log_rhos = get_log_rhos(target_action_log_probs, behaviour_action_log_probs) vtrace_returns = from_importance_weights( log_rhos=log_rhos, discounts=discounts, rewards=rewards, values=values, bootstrap_value=bootstrap_value, clip_rho_threshold=clip_rho_threshold, clip_pg_rho_threshold=clip_pg_rho_threshold) return VTraceFromLogitsReturns( log_rhos=log_rhos, behaviour_action_log_probs=behaviour_action_log_probs, target_action_log_probs=target_action_log_probs, **vtrace_returns._asdict())