about evaluate/perplexity

[KangkangStu]

I've found that the perplexity example in the evaluation function on huggingface is always in the Runtime error statehttps://huggingface.co/spaces/evaluate-metric/perplexity. When will it be fixed?

BTW,After referring to some tutorials, I successfully evaluated the model perplexity of GPT2, but when testing the local Llama-7B model, I encountered a lot of errors. Do you have any kind friends who can help me explain how to import models and datasets from local for testing, thank you very much!

[beyondguo]

For ppl, I strongly suggest your to rewrite their perplexity.py.

• for Llama, you should use LlamaTokenizer
• rewriting the init function of the Perplexity class. The current version will reload the model every time you call the compute function! which is ridiculous...

[beyondguo]

Usage：

from ppl import Perplexity
ppl = Perplexity(model_id='your_model_path')
texts = ['asdfasdf','apxl wndo aslewr sdf', 'hello world']
ppl._compute(texts)

ppl.py as follows:

import datasets
import numpy as np
import torch
from torch.nn import CrossEntropyLoss
from transformers import AutoModelForCausalLM, AutoTokenizer, LlamaTokenizer

import evaluate
from evaluate import logging


_CITATION = """"""

_DESCRIPTION = """"""

_KWARGS_DESCRIPTION = """
Args:
    model_id (str): model used for calculating Perplexity
    predictions (list of str): input text, each separate text snippet
        is one list entry.
    batch_size (int): the batch size to run texts through the model. Defaults to 16.
    add_start_token (bool): whether to add the start token to the texts,
        so the perplexity can include the probability of the first word. Defaults to True.
    device (str): device to run on, defaults to 'cuda' when available
Returns:
    perplexity: dictionary containing the perplexity scores for the texts
        in the input list, as well as the mean perplexity. If one of the input texts is
        longer than the max input length of the model, then it is truncated to the
        max length for the perplexity computation.

"""


# @evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
class Perplexity(evaluate.Metric):
    def __init__(self, model_id, device=None):
        super(evaluate.Metric, self).__init__()
        if device is not None:
            assert device in ["gpu", "cpu", "cuda"], "device should be either gpu or cpu."
            if device == "gpu":
                self.device = "cuda"
        else:
            self.device = "cuda" if torch.cuda.is_available() else "cpu"

        self.model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True).to(self.device)
        if 'llama' in model_id or 'alpaca' in model_id:
            self.tokenizer = LlamaTokenizer.from_pretrained(model_id)
        else:
            self.tokenizer = AutoTokenizer.from_pretrained(model_id)

    
    def _info(self):
        return evaluate.MetricInfo(
            module_type="metric",
            description=_DESCRIPTION,
            citation=_CITATION,
            inputs_description=_KWARGS_DESCRIPTION,
            features=datasets.Features(
                {
                    "predictions": datasets.Value("string"),
                }
            ),
            reference_urls=["https://huggingface.co/docs/transformers/perplexity"],
        )

    def _compute(
        self, predictions, batch_size: int = 16, add_start_token: bool = True,  max_length=None
    ):
        # special token to also be the padding token
        if self.tokenizer.pad_token is None and batch_size > 1:
            existing_special_tokens = list(self.tokenizer.special_tokens_map_extended.values())
            # check that the model already has at least one special token defined
            assert (
                len(existing_special_tokens) > 0
            ), "If batch_size > 1, model must have at least one special token to use for padding. Please use a different model or set batch_size=1."
            # assign one of the special tokens to also be the pad token
            self.tokenizer.add_special_tokens({"pad_token": existing_special_tokens[0]})

        if add_start_token and max_length:
            # leave room for <BOS> token to be added:
            assert (
                self.tokenizer.bos_token is not None
            ), "Input model must already have a BOS token if using add_start_token=True. Please use a different model, or set add_start_token=False"
            max_tokenized_len = max_length - 1
        else:
            max_tokenized_len = max_length

        encodings = self.tokenizer(
            predictions,
            add_special_tokens=False,
            padding=True,
            truncation=True if max_tokenized_len else False,
            max_length=max_tokenized_len,
            return_tensors="pt",
            return_attention_mask=True,
        ).to(self.device)

        encoded_texts = encodings["input_ids"]
        attn_masks = encodings["attention_mask"]

        # check that each input is long enough:
        if add_start_token:
            assert torch.all(torch.ge(attn_masks.sum(1), 1)), "Each input text must be at least one token long."
        else:
            assert torch.all(
                torch.ge(attn_masks.sum(1), 2)
            ), "When add_start_token=False, each input text must be at least two tokens long. Run with add_start_token=True if inputting strings of only one token, and remove all empty input strings."

        ppls = []
        loss_fct = CrossEntropyLoss(reduction="none")

        for start_index in logging.tqdm(range(0, len(encoded_texts), batch_size)):
            end_index = min(start_index + batch_size, len(encoded_texts))
            encoded_batch = encoded_texts[start_index:end_index]
            attn_mask = attn_masks[start_index:end_index]

            if add_start_token:
                bos_tokens_tensor = torch.tensor([[self.tokenizer.bos_token_id]] * encoded_batch.size(dim=0)).to(self.device)
                encoded_batch = torch.cat([bos_tokens_tensor, encoded_batch], dim=1)
                attn_mask = torch.cat(
                    [torch.ones(bos_tokens_tensor.size(), dtype=torch.int64).to(self.device), attn_mask], dim=1
                )

            labels = encoded_batch

            with torch.no_grad():
                out_logits = self.model(encoded_batch, attention_mask=attn_mask).logits

            shift_logits = out_logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            shift_attention_mask_batch = attn_mask[..., 1:].contiguous()

            perplexity_batch = torch.exp(
                (loss_fct(shift_logits.transpose(1, 2), shift_labels) * shift_attention_mask_batch).sum(1)
                / shift_attention_mask_batch.sum(1)
            )

            ppls += perplexity_batch.tolist()

        return {"perplexities": ppls, "mean_perplexity": np.mean(ppls)}