训练 Hugging Face Transformer 时出现 KeyErrors

Question

我通常遵循本教程（ https://huggingface.co/docs/transformers/training#:~:text=%F0%9F%A4%97%20Transformers%20provides%20access%20to,an%20incredibly%20powerful%20training %20technique. ) 在预训练的变压器上实现微调。 主要区别在于我使用的是我自己的自定义数据集，该数据集来源于一个 JSON 文件，该文件具有文档的文本和它应该属于的标签。 为了能够做到这一点，我需要创建自己的类，该类基于 pytorch 的 Dataset 类。 这就是那个类的样子：

class PDFsDataset(torch.utils.data.Dataset):
    def __init__(self, encodings, labels):
        self.encodings = encodings
        self.labels = labels

    def __getitem__(self, idx):
        
        print("\n\n\n\nindex",idx)
        item = {key: torch.tensor(val[idx]) for key, val in self.encodings.items()}
        item['labels'] = torch.tensor(self.labels[idx])
        return item

    def __len__(self):
        return len(self.labels)

我在训练变压器时遇到错误，上面写着

Traceback (most recent call last):
  File "C:\Users\e417922\AppData\Roaming\Python\Python39\site-packages\pandas\core\indexes\base.py", line 3621, in get_loc
    return self._engine.get_loc(casted_key)
  File "pandas\_libs\index.pyx", line 136, in pandas._libs.index.IndexEngine.get_loc
  File "pandas\_libs\index.pyx", line 163, in pandas._libs.index.IndexEngine.get_loc
  File "pandas\_libs\hashtable_class_helper.pxi", line 2131, in pandas._libs.hashtable.Int64HashTable.get_item
  File "pandas\_libs\hashtable_class_helper.pxi", line 2140, in pandas._libs.hashtable.Int64HashTable.get_item
KeyError: 19

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "c:\Users\e417922\Downloads\enwiki-20220601-pages-meta-history1.xml-p1p857\HF_Transformer.py", line 147, in <module>
    transformer.train_transformer()
  File "c:\Users\e417922\Downloads\enwiki-20220601-pages-meta-history1.xml-p1p857\HF_Transformer.py", line 135, in train_transformer
    trainer.train()
  File "C:\Users\e417922\AppData\Roaming\Python\Python39\site-packages\transformers\trainer.py", line 1409, in train
    return inner_training_loop(
  File "C:\Users\e417922\AppData\Roaming\Python\Python39\site-packages\transformers\trainer.py", line 1625, in _inner_training_loop
    for step, inputs in enumerate(epoch_iterator):
  File "C:\Users\e417922\AppData\Roaming\Python\Python39\site-packages\torch\utils\data\dataloader.py", line 530, in __next__
    data = self._next_data()
  File "C:\Users\e417922\AppData\Roaming\Python\Python39\site-packages\torch\utils\data\dataloader.py", line 570, in _next_data
    data = self._dataset_fetcher.fetch(index)  # may raise StopIteration
  File "C:\Users\e417922\AppData\Roaming\Python\Python39\site-packages\torch\utils\data\_utils\fetch.py", line 49, in fetch
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "C:\Users\e417922\AppData\Roaming\Python\Python39\site-packages\torch\utils\data\_utils\fetch.py", line 49, in <listcomp>
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "c:\Users\e417922\Downloads\enwiki-20220601-pages-meta-history1.xml-p1p857\HF_Transformer.py", line 42, in __getitem__
    for key in self.encodings[idx]:
  File "C:\Users\e417922\AppData\Roaming\Python\Python39\site-packages\pandas\core\series.py", line 958, in __getitem__
    return self._get_value(key)
  File "C:\Users\e417922\AppData\Roaming\Python\Python39\site-packages\pandas\core\series.py", line 1069, in _get_value
    loc = self.index.get_loc(label)
  File "C:\Users\e417922\AppData\Roaming\Python\Python39\site-packages\pandas\core\indexes\base.py", line 3623, in get_loc
    raise KeyError(key) from err
KeyError: 19

每次我运行它时它失败的 KeyError 都会发生变化。 我是 Transformers 和 HuggingFace 的初学者，所以我不知道是什么导致了这个问题。

编辑：示例输入是一个 JSON 文件，其中元素看起来像这样：{“text_clean”：[“有几百个单词的文章”，另一个有很多单词的文章”，“又一篇文章”]，“most_similar_label”： [《量子》《人工智能》《材料》]}

完整代码：

import tkinter as tk
from tkinter import filedialog
import json
import pandas as pd
from transformers import AutoTokenizer
from transformers import AutoModelForSequenceClassification
from transformers import TrainingArguments
from transformers import TrainingArguments, Trainer
import numpy as np
from datasets import load_metric
from sklearn.model_selection import train_test_split
import torch

class PDFsDataset(torch.utils.data.Dataset):
    def __init__(self, encodings, labels):
        self.encodings = encodings
        self.labels = labels

    def __getitem__(self, idx):
        
        print("\n\n\n\nindex",idx)
        item = {key: torch.tensor(val[idx]) for key, val in self.encodings.items()}
        item['labels'] = torch.tensor(self.labels[idx])
        return item

    def __len__(self):
        return len(self.labels)

class HFTransformer:
    def __init__ (self):
        pass

    def import_from_json(self):
        #Prompts user to select json file
        root = tk.Tk()
        root.withdraw()
        self.json_file_path = filedialog.askopenfile().name
        #opens json file and loads data
        with open(self.json_file_path, "r") as json_file:
                try:
                    json_load = json.load(json_file)
                except:
                    raise ValueError("No PDFs to convert to JSON")
        self.pdfs = json_load
        #converts json file data to dataframe for easier manipulation
        self.pdfs = pd.DataFrame.from_dict(self.pdfs)

        for index in range(len(self.pdfs["new_tags"])):
            if self.pdfs["new_tags"][index] == "":
                self.pdfs["new_tags"][index] = self.pdfs["most_similar_label"][index]

        self.pdfs["labels"] = self.pdfs["new_tags"].apply(lambda val: self.change_tag_to_num(val))
        # for label in self.data["labels"]:
     
    def change_tag_to_num(self, value):
        if value == "Quantum":
            return 0
        elif value == "Artificial intelligence":
            return 1
        elif value == "Materials":
            return 2
        elif value == "Energy":
            return 3
        elif value == "Defense":
            return 4
        elif value == "Satellite":
            return 5
        elif value == "Other":
            return 6

    def tokenize_dataset(self):
        tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")

        X_train, X_test, y_train, y_test = train_test_split(self.pdfs["text_clean"], self.pdfs["labels"],test_size=0.2)

        train_encodings = X_train.apply(lambda string: tokenizer(string, truncation=True, padding=True,max_length=10))
        test_encodings = X_test.apply(lambda string: tokenizer(string, truncation=True, padding=True,max_length=10))
    
        
        self.train_dataset = PDFsDataset(train_encodings, y_train)
    
        data_to_add = {"input_ids": [], "token_type_ids": [], "attention_mask": []}

        for i in self.train_dataset.encodings:
            data_to_add["input_ids"].append(i["input_ids"])
            data_to_add["token_type_ids"].append(i["token_type_ids"])
            data_to_add["attention_mask"].append(i["attention_mask"])

        self.train_dataset.encodings = data_to_add

        self.eval_dataset = PDFsDataset(test_encodings,y_test)
        data_to_add = {"input_ids": [], "token_type_ids": [], "attention_mask": []}

        for i in self.eval_dataset.encodings:
            data_to_add["input_ids"].append(i["input_ids"])
            data_to_add["token_type_ids"].append(i["token_type_ids"])
            data_to_add["attention_mask"].append(i["attention_mask"])
        
        self.eval_dataset.encodings = data_to_add

    def train_transformer(self):
        model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=7)
        training_args = TrainingArguments(output_dir="test_trainer")
        self.metric = load_metric("accuracy")
        training_args = TrainingArguments(output_dir="test_trainer", evaluation_strategy="epoch")
    

        trainer = Trainer(
            model=model,
            args=training_args,
            train_dataset=self.train_dataset,
            eval_dataset=self.eval_dataset,
            compute_metrics=self.compute_metrics
        )
        trainer.train()
    def compute_metrics(self, eval_pred):
        logits, labels = eval_pred
        predictions = np.argmax(logits, axis=-1)
        return self.metric.compute(predictions=predictions, references=labels)


if __name__  == "__main__":
    transformer = HFTransformer()
    transformer.import_from_json()
    transformer.tokenize_dataset()
    transformer.train_transformer()

Answer 1

将pandas.Series转换为一个简单的 python 列表并去掉一些额外的材料可以解决这个问题

class PDFsDataset(torch.utils.data.Dataset):
    def __init__(self, encodings, labels):
        self.encodings = encodings
        self.labels = labels

    def __getitem__(self, idx):
        item = {key: torch.tensor(val[idx]) for key, val in self.encodings.items()}
        item['labels'] = torch.tensor(self.labels[idx])
        return item

    def __len__(self):
        return len(self.labels)

class HFTransformer:
    def __init__ (self):
        pass

    def import_from_json(self):
        #Prompts user to select json file
        self.json_file_path = '/content/truncated_data.json'
        #opens json file and loads data
        with open(self.json_file_path, "r") as json_file:
                try:
                    json_load = json.load(json_file)
                except:
                    raise ValueError("No PDFs to convert to JSON")
        self.pdfs = json_load
        #converts json file data to dataframe for easier manipulation
        self.pdfs = pd.DataFrame.from_dict(self.pdfs)

        for index in range(len(self.pdfs["new_tags"])):
            if self.pdfs["new_tags"][index] == "":
                self.pdfs["new_tags"][index] = self.pdfs["most_similar_label"][index]

        self.pdfs["labels"] = self.pdfs["new_tags"].apply(lambda val: self.change_tag_to_num(val))
        # for label in self.data["labels"]:
     
    def change_tag_to_num(self, value):
        if value == "Quantum":
            return 0
        elif value == "Artificial intelligence":
            return 1
        elif value == "Materials":
            return 2
        elif value == "Energy":
            return 3
        elif value == "Defense":
            return 4
        elif value == "Satellite":
            return 5
        elif value == "Other":
            return 6

    def tokenize_dataset(self):
        tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")

        X_train, X_test, y_train, y_test = train_test_split(self.pdfs["text_clean"].to_list(), self.pdfs["labels"].to_list(),test_size=0.2)
        train_encodings = tokenizer(X_train, truncation=True, padding=True,max_length=100)
        test_encodings = tokenizer(X_test, truncation=True, padding=True,max_length=100)
        self.train_dataset = PDFsDataset(train_encodings, y_train)
    
        self.eval_dataset = PDFsDataset(test_encodings,y_test)
    
    def train_transformer(self):
        model = AutoModelForSequenceClassification.from_pretrained("bert-base-cased", num_labels=7)
        training_args = TrainingArguments(output_dir="test_trainer")
        self.metric = load_metric("accuracy")
        training_args = TrainingArguments(output_dir="test_trainer", evaluation_strategy="epoch", )
    
        trainer = Trainer(
            model=model,
            args=training_args,
            train_dataset=self.train_dataset,
            eval_dataset=self.eval_dataset,
            compute_metrics=self.compute_metrics
        )
        trainer.train()
    def compute_metrics(self, eval_pred):
        logits, labels = eval_pred
        predictions = np.argmax(logits, axis=-1)
        return self.metric.compute(predictions=predictions, references=labels)


if __name__  == "__main__":
    tr = HFTransformer()
    tr.import_from_json()
    tr.tokenize_dataset()
    tr.train_transformer()