How to calculate perplexity of a LM trained from scratch?

Question

I created a language model from scratch with BertForMaskedLM using my own domain dataset.

Now I want to assess whether the model is good so I would like to calculate