Feature : Add Perplexity metric to ignite.metrics.nlp

[steaphenai]

🚀 Feature

I’d like to add a new NLP metric: Perplexity under ignite.metrics.nlp, and expose it in top-level ignite.metrics.

Motivation

Perplexity is a standard metric for language modeling, and having it built into Ignite would make evaluation pipelines easier and more consistent with other built-in metrics.

Proposal

• Add Perplexity metric implementation in ignite.metrics.nlp.perplexity.
• Export it from:
  • ignite.metrics.nlp
  • ignite.metrics (top-level import path)
• Include tests for:
  • correctness vs manual computation
  • token-weighted accumulation across updates
  • reset behavior
  • input shape validation
  • edge cases (single token / empty compute handling)

Alternatives considered

Users can currently compute perplexity manually from cross-entropy, but that is repetitive and error-prone across projects.

Additional context

I have a working branch and tests for this implementation and can open a PR right after this issue is created.