o }oi @sHddlZddlmmZdejdejdedeejejffddZdS)Nlogitstargetsblank_idreturncCs|j}|jd}tj|dd}|d|f}tj|tj|g|j|ddfdd}tj |d|d |jdddd d}d |dddddf<||fS) a Given logits, calculate log probabilities for blank and target labels needed for transducer loss calculation. Naive implementation in PyTorch, for testing and prototyping purposes. Args: logits: Joint tensor of size [B, T, U+1, D] targets: Targets of size [B, U] blank_id: id of the blank output Returns: Tuple of tensors with log probabilities for targets and blank labels, both of size [B, T, U+1]. For the last non-existent target (U+1) output is zero. r)dim.)dtypedeviceN)rindexg) r shapeF log_softmaxtorchcatzerosr unsqueezegatherexpandsqueeze)rrrr batch_size log_probs blank_scores target_scoresr_/home/ubuntu/.local/lib/python3.10/site-packages/nemo/collections/asr/parts/k2/rnnt_logprobs.pyrnnt_logprobs_torchs  ("r) rtorch.nn.functionalnn functionalr Tensorinttuplerrrrrs