o yiG @sddlmZmZddlZddlmmZddlmZej ej fZ dededdfddZ ddeded ee deeeffd d Zd ed edefddZddeded ee defddZdS))OptionalTupleN)TensorpredstargetreturncCst|jdkrtdt|jdt|jdkr$tdt|jd|jdd|jkr?td|jddd|jd|jtvrPtd td |jd|jtjkrctd tjd |jddS) aCheck shape and type consistency of input vectors. Args: preds: Probabilities assigned to each token in a sequence with shape [batch_size, seq_len, vocab_size]. target: Ground truth values with a shape [batch_size, seq_len]. Raises: ValueError: If ``preds`` tensor has no 3 dimensions. ValueError: If ``target`` tensor has no 2 dimensions. ValueError: If the first two dimensions of ``preds`` and ``target`` do not equal. TypeError: If ``preds`` dtype is not one of ``(torch.float16, torch.float32, torch.float64)`` TypeError: If ``target`` is not of a type LongTensor (torch.int64) zbInput tensor `preds` is expected to have 3 dimensions, [batch_size, seq_len, vocab_size], but got .zWInput tensor `target` is expected to have 2 dimensions, [batch_size, seq_len], but got NzvInput tensors `preds` and `target` are expected to have equaling first two dimensions, [batch_size, seq_len], but got z and z8Input tensor `preds` is expected to be of a type one of z but got z2Input tensor `target` is expected to be of a type )lenshape ValueErrordtype_TORCH_FLOAT_OR_DOUBLE TypeErrortorchint64)rrr[/home/ubuntu/.local/lib/python3.10/site-packages/torchmetrics/functional/text/perplexity.py!_check_shape_and_type_consistencys6   r ignore_indexcCst||tj|d|jddd}|d}|dur0||}|||ktjd|j d}ntj |tj d}|dd|f |}|  }|}||fS)a Compute intermediate statistics for Perplexity. Args: preds: Probabilities assigned to each token in a sequence with shape [batch_size, seq_len, vocab_size]. target: Ground truth values with a shape [batch_size, seq_len]. ignore_index: Integer specifying a target class to ignore. If given, this class index does not contribute to the returned score. Returns: Log probabilities, summed over all samples Number of samples )dimNr)device)r)rFsoftmaxreshaper newherertensorr ones_likebooldiagonallogsum)rrrprobsmasktotal_log_probscountrrr_perplexity_updateDs   r*totalr)cCst||S)zCompute the Perplexity. Args: total: Log probabilities, summed over all samples count: Number of samples Returns: Perplexity )rexp)r+r)rrr_perplexity_computefs r-cCst|||\}}t||S)aPerplexity measures how well a language model predicts a text sample. It's calculated as the average number of bits per word a model needs to represent the sample. Args: preds: Log probabilities assigned to each token in a sequence with shape [batch_size, seq_len, vocab_size]. target: Ground truth values with a shape [batch_size, seq_len]. ignore_index: Integer specifying a target class to ignore. If given, this class index does not contribute to the returned score. Returns: Perplexity value Examples: >>> import torch >>> preds = torch.rand(2, 8, 5, generator=torch.manual_seed(22)) >>> target = torch.randint(5, (2, 8), generator=torch.manual_seed(22)) >>> target[0, 6:] = -100 >>> perplexity(preds, target, ignore_index=-100) tensor(5.2545) )r*r-)rrrr+r)rrr perplexityrs r.)N)typingrrrtorch.nn.functionalnn functionalrrfloat32float64rrintr*r-r.rrrrs  (,"$