o $i@sddlZddlZddlZddlmZmZddlmZddlm Z ddl m Z m Z ddl mZeeZe GdddejZdS) N)AnyDict)Dataset)Policy) DeveloperAPIExperimentalAPI)SampleBatchTypec @seZdZdZedefddZejede de e e ffddZ ede de e e ffd d Zeed d ed ede e e ffddZdS)OfflineEvaluatorz.Interface for an offline evaluator of a policypolicycKs ||_dS)zInitializes an OffPolicyEstimator instance. Args: policy: Policy to evaluate. kwargs: forward compatibility placeholder. N)r )selfr kwargsr `/home/ubuntu/veenaModal/venv/lib/python3.10/site-packages/ray/rllib/offline/offline_evaluator.py__init__s zOfflineEvaluator.__init__batchreturncKst)aPReturns the evaluation results for the given batch of episodes. Args: batch: The batch to evaluate. kwargs: forward compatibility placeholder. Returns: The evaluation done on the given batch. The returned dict can be any arbitrary mapping of strings to metrics. )NotImplementedErrorr rr r r restimates zOfflineEvaluator.estimatecKsiS)a6Sometimes you need to train a model inside an evaluator. This method abstracts the training process. Args: batch: SampleBatch to train on kwargs: forward compatibility placeholder. Returns: Any optional metrics to return from the evaluator r rr r rtrain+s zOfflineEvaluator.train) n_parallelismdatasetrcCsdS)aCalculates the estimate of the metrics based on the given offline dataset. Typically, the dataset is passed through only once via n_parallel tasks in mini-batches to improve the run-time of metric estimation. Args: dataset: The ray dataset object to do offline evaluation on. n_parallelism: The number of parallelism to use for the computation. Returns: Dict[str, Any]: A dictionary of the estimated values. Nr )r rrr r restimate_on_dataset9sz$OfflineEvaluator.estimate_on_datasetN)__name__ __module__ __qualname____doc__rrrabcabstractmethodrrstrrrrros cpu_countrintrr r r rr s$    r )rloggingr typingrrray.datarray.rllib.policyrray.rllib.utils.annotationsrrray.rllib.utils.typingr getLoggerrloggerABCr r r r rs