Add geometric mean of per-token-Perplexities

[LuggiStruggi]

Hi I'm new to this and might to be overlooking smth, but it seems to me that more often than the arithmetic mean over perplexity per batch, the geometric mean over all per-token-Perplexities is used. I.e here in huggingface's Transformers docs. So basically just what you calculate as a single perplexity per batch but for the whole dataset. Would be helpful to have that returned here additionally. So smth along the lines of (just to understand what i mean. Didn't test the code):

ppls = []
nll_sum = 0
total_tokens = 0

for start_index in logging.tqdm(range(0, len(encoded_texts), batch_size)):
          
          [...]
            
           negative_log_likelihood_batch = loss_fct(shift_logits.transpose(1, 2), shift_labels) * shift_attention_mask_batch).sum(1)
          
           perplexity_batch = torch.exp(negative_log_likelihood_batch / shift_attention_mask_batch.sum(1))

           ppls += perplexity_batch.tolist()
           nll_sum += negative_log_likelihood_batch.sum().item()
           total_tokens += shift_attention_mask_batch.sum().item()

        return {"perplexities": ppls,
                "mean_perplexity": np.mean(ppls),
                "geometric_mean_perplexity": torch.exp(nll_sum / total_tokens)}