o pi@s4ddlZddlmZddlmZGdddeZdS)N)common_functions)ModuleWithRecordsAndDistancecs6eZdZfddZddZddZd dd ZZS) BaseMinerc s(tjdi||jgddddS)N) num_pos_pairs num_neg_pairs num_tripletsT) list_of_namesis_stat)super__init__add_to_recordable_attributes)selfkwargs __class__r ]/home/ubuntu/.local/lib/python3.10/site-packages/pytorch_metric_learning/miners/base_miner.pyr s  zBaseMiner.__init__cCst)N)NotImplementedError)r embeddingslabelsref_emb ref_labelsr r rmineszBaseMiner.minecCst|dkr&t|d|_|jt|dkr!t|dks$JJdSt|dkrRt|d|_t|d|_|jt|dksEJ|jt|dksPJdSt)a Args: output: the output of self.mine This asserts that the mining function is outputting properly formatted indices. The default is to require a tuple representing a,p,n indices or a1,p,a2,n indices within a batch of embeddings. For example, a tuple of (anchors, positives, negatives) will be (torch.tensor, torch.tensor, torch.tensor) rrN)lenrrr TypeError)routputr r routput_assertions 2 zBaseMiner.output_assertionNcCsz|t&t||t||}t||||\}}|||||}Wdn1s1wY|||S)a Args: embeddings: tensor of size (batch_size, embedding_size) labels: tensor of size (batch_size) Does any necessary preprocessing, then does mining, and then checks the shape of the mining output before returning it N) reset_statstorchno_gradc_f check_shapes to_device set_ref_embrr )rrrrr mining_outputr r rforward's    zBaseMiner.forward)NN)__name__ __module__ __qualname__r rr r) __classcell__r r rrrs  r)r"utilsrr$%utils.module_with_records_and_reducerrrr r r rs