o }oim @s:ddlZddlmZddlmZdgZGdddeZdS)N) Categorical)Metric Perplexitycs8eZdZdZdZd fdd Zd ddZd d ZZS) ra This class computes mean perplexity of distributions in the last dimension of inputs. It is a wrapper around :doc:`torch.distributions.Categorical.perplexity` method. You have to provide either ``probs`` or ``logits`` to the :meth:`update` method. The class computes perplexities for distributions passed to :meth:`update` method in ``probs`` or ``logits`` arguments and averages the perplexities. Reducing results between all workers is done via SUM operations. See the `TorchMetrics in PyTorch Lightning guide `_ for the metric usage instructions. Args: dist_sync_on_step: Synchronize metric state across processes at each ``forward()`` before returning the value at the step. process_group: Specify the process group on which synchronization is called. default: ``None`` (which selects the entire world) validate_args: If ``True`` values of :meth:`update` method parameters are checked. ``logits`` has to not contain NaNs and ``probs`` last dim has to be valid probability distribution. TFNcsRtj||d||_|jdtjdtjddd|jdtjdtjddddS) N)dist_sync_on_step process_groupperplexities_sumg)dtypesum)dist_reduce_fxnum_distributionsr)super__init__ validate_args add_statetorchtensorfloat64int64)selfrrr __class__^/home/ubuntu/.local/lib/python3.10/site-packages/nemo/collections/common/metrics/perplexity.pyr /s zPerplexity.__init__cCs`t|durdn||durdn||jd}|}|j|7_|j|7_dS)z Updates :attr:`perplexities_sum` and :attr:`num_distributions`. Args: probs: A ``torch.Tensor`` which innermost dimension is valid probability distribution. logits: A ``torch.Tensor`` without NaNs. N)r)rdetachr perplexityr numelrr )rprobslogitsdpplrrrupdate6szPerplexity.updatecCs|jdrdS|j|jS)z Returns perplexity across all workers and resets to 0 :attr:`perplexities_sum` and :attr:`num_distributions`. rN)r eqr)rrrrcomputeFs  zPerplexity.compute)FNT)NN) __name__ __module__ __qualname____doc__full_state_updater r r" __classcell__rrrrrs  )rtorch.distributions.categoricalr torchmetricsr__all__rrrrrs