local_coh_ppl.py

"""
adapted to support pegasus-xsum / local files
"""

# Copyright 2022 The HuggingFace Datasets Authors and the current dataset script contributor.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Perplexity Metric."""

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

import evaluate
from evaluate import logging
import pdb

WINDOW_SIZE = 3


def prepare_coh_sents(predictions):
    blocks = []
    lens = []
    for pred in predictions:
        sents = pred.split("\n")
        if len(sents)<=WINDOW_SIZE:
            blocks.append(pred)
            lens.append(1)
        else:
            _block = []
            for i in range(0,len(sents)-WINDOW_SIZE+1):
                _block.append("\n".join(sents[i:i+WINDOW_SIZE]))
            lens.append(len(_block))
            blocks.extend(_block)
    #
    return blocks,lens




_CITATION = """\

"""

_DESCRIPTION = """
Perplexity (PPL) is one of the most common metrics for evaluating language models.
It is defined as the exponentiated average negative log-likelihood of a sequence, calculated with exponent base `e`.

For more information, see https://huggingface.co/docs/transformers/perplexity
"""

_KWARGS_DESCRIPTION = """
Args:
    model_id (str): model used for calculating Perplexity
            NOTE: Perplexity can only be calculated for causal language models.
                    This includes models such as gpt2, causal variations of bert,
                    causal versions of t5, and more (the full list can be found
                    in the AutoModelForCausalLM documentation here:
                    https://huggingface.co/docs/transformers/master/en/model_doc/auto#transformers.AutoModelForCausalLM )

    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.
Examples:
    Example 1:
        >>> perplexity = evaluate.load("perplexity", module_type="metric")
        >>> input_texts = ["lorem ipsum", "Happy Birthday!", "Bienvenue"]
        >>> results = perplexity.compute(model_id='gpt2',
        ...                              add_start_token=False,
        ...                              predictions=input_texts) # doctest:+ELLIPSIS
        >>> print(list(results.keys()))
        ['perplexities', 'mean_perplexity']
        >>> print(round(results["mean_perplexity"], 0))
        647.0
        >>> print(round(results["perplexities"][0], 0))
        32.0

    Example 2:
        >>> from datasets import load_dataset
        >>> perplexity = evaluate.load("perplexity", module_type="metric")
        >>> input_texts = load_dataset("wikitext", "wikitext-2-raw-v1", split="test")["text"][:10] # doctest: +SKIP
        >>> input_texts = [s for s in input_texts if s!='']
        >>> results = perplexity.compute(model_id='gpt2',
        ...                              predictions=input_texts)
        >>> print(list(results.keys()))
        ['perplexities', 'mean_perplexity']
        >>> print(round(results["mean_perplexity"], 2)) # doctest: +SKIP
        576.76
        >>> print(round(results["perplexities"][0], 2)) # doctest: +SKIP
        889.28
"""


@evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
class LocalCohPPL(evaluate.Measurement):
    def _info(self):
        return evaluate.MetricInfo(
            module_type="measurement",
            description=_DESCRIPTION,
            citation=_CITATION,
            inputs_description=_KWARGS_DESCRIPTION,
            features=datasets.Features(
                {
                    "predictions": datasets.Value("string"),
                }
            ),
            reference_urls=["https://huggingface.co/spaces/ronaldahmed/local_coh_ppl"],
        )

    ## PEDICTIONS: [str] sentences joined by "\n"
    def _compute(self, predictions, model_id, batch_size: int = 16, add_start_token: bool = True, device=None):
        MODEL_CACHE_DIR = "/home/rcardena/.cache/huggingface/"
        if getpass.getuser() == "s1987051":
            MODEL_CACHE_DIR="/disk/ocean/rcardenas/tools/huggingface/"
        elif getpass.getuser() == "rcardena":
            MODEL_CACHE_DIR="/gfs/team/nlp/users/rcardena/tools/huggingface/"
      
        if device is not None:
            assert device in ["gpu", "cpu", "cuda"], "device should be either gpu or cpu."
            if device == "gpu":
                device = "cuda"
        else:
            device = "cuda" if torch.cuda.is_available() else "cpu"

        model = AutoModelForCausalLM.from_pretrained(model_id,cache_dir=MODEL_CACHE_DIR)
        model = model.to(device)

        tokenizer = AutoTokenizer.from_pretrained(
                            model_id,
                            cache_dir=MODEL_CACHE_DIR,
                            use_fast="cnn_dailymail" not in model_id,
                            )

        # if batch_size > 1 (which generally leads to padding being required), and
        # if there is not an already assigned pad_token, assign an existing
        # special token to also be the padding token
        if tokenizer.pad_token is None and batch_size > 1:
            existing_special_tokens = list(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
            tokenizer.add_special_tokens({"pad_token": existing_special_tokens[0]})

        model.config.max_length = 512 if "scibert" in model_id else model.config.max_length
        if add_start_token:
            # leave room for <BOS> token to be added:
            assert (
                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 = model.config.max_length - 1
        else:
            max_tokenized_len = model.config.max_length
        
        loss_fct = CrossEntropyLoss(reduction="none")

        blocks,blens = prepare_coh_sents(predictions)
        all_norm_ppl = []
        for start_index in logging.tqdm(range(0, len(blocks), batch_size)):
            end_index = min(start_index + batch_size, len(blocks))
            batch_sents = blocks[start_index:end_index]
            
            encodings = tokenizer(
                batch_sents,
                add_special_tokens=False,
                padding=True,
                truncation=True,
                max_length=max_tokenized_len,
                return_tensors="pt",
                return_attention_mask=True,
            ).to(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."

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

            # tokenize by sentence
            sent_tok_lens = []
            for pred in batch_sents:
                ss = pred.split("\n")
                sslens = [len(tokenizer(y,add_special_tokens=False,padding=False).input_ids) for y in ss]
                offset = int(add_start_token)
                sspos = [offset]
                for sslen in sslens:
                    offset = min(offset + sslen,511 + int(add_start_token))
                    if offset == sspos[-1]: break # reached length limit
                    sspos.append(offset)
                sent_tok_lens.append(sspos)

            labels = encoded_texts

            with torch.no_grad():
                out_logits = model(encoded_texts, attention_mask=attn_masks).logits

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

            loss_out = loss_fct(shift_logits.transpose(1, 2), shift_labels) * shift_attention_mask_batch
            perplexity_all = torch.exp(
                loss_out.sum(1)
                / shift_attention_mask_batch.sum(1)
            ).detach().cpu().numpy().tolist()

            norm_ppl = []
            for b,stl in enumerate(sent_tok_lens):
                indv = []
                for i in range(1,len(stl)):
                    ppl = torch.exp( loss_out[b,stl[i-1]:stl[i]].sum() / shift_attention_mask_batch[b,stl[i-1]:stl[i]].sum() ).detach().cpu().item()
                    indv.append(ppl)
                nppl = perplexity_all[b] / sum(indv) if len(indv)>1 else 0.0
                norm_ppl.append(nppl)
            #
            all_norm_ppl.extend(norm_ppl)

            if any(np.isnan(norm_ppl)):
                print("[compute ppl] nan ...")
                pdb.set_trace()
                print(">>")

        #
        avg_ppl = []
        offset = 0
        for _len in blens:
            avg_ppl.append( float(np.mean(all_norm_ppl[offset:offset+_len])) )
            offset += _len
        if any(np.isnan(avg_ppl)):
            print("[compute ppl] nan ...")
            pdb.set_trace()
            print(">>")

        return {"local_coh_ppl": avg_ppl}