import torch from torch.utils import data from torch.utils.data import Dataset from datasets.arrow_dataset import Dataset as HFDataset from datasets.load import load_dataset, load_metric from transformers import ( AutoTokenizer, DataCollatorWithPadding, EvalPrediction, default_data_collator, ) import copy, math import os import numpy as np import logging, re from datasets.formatting.formatting import LazyRow, LazyBatch from tqdm import tqdm from tasks import utils task_to_keys = { "cola": ("sentence", None), "mnli": ("premise", "hypothesis"), "mrpc": ("sentence1", "sentence2"), "qnli": ("question", "sentence"), "qqp": ("question1", "question2"), "rte": ("sentence1", "sentence2"), "sst2": ("sentence", None), "stsb": ("sentence1", "sentence2"), "wnli": ("sentence1", "sentence2"), } logger = logging.getLogger(__name__) idx = 0 class GlueDataset(): def __init__(self, tokenizer: AutoTokenizer, data_args, training_args) -> None: super().__init__() self.tokenizer = tokenizer self.data_args = data_args #labels raw_datasets = load_dataset("glue", data_args.dataset_name) self.is_regression = data_args.dataset_name == "stsb" if not self.is_regression: self.label_list = raw_datasets["train"].features["label"].names self.num_labels = len(self.label_list) else: self.num_labels = 1 # Preprocessing the raw_datasets self.sentence1_key, self.sentence2_key = task_to_keys[data_args.dataset_name] sc_template = f'''{'{' + self.sentence1_key + '}'}''' \ if self.sentence2_key is None else f'''{'{' + self.sentence1_key + '}'}{'{' + self.sentence2_key + '}'}''' self.tokenizer.template = self.template = [sc_template] print(f"-> using template:{self.template}") # Padding strategy if data_args.pad_to_max_length: self.padding = "max_length" else: # We will pad later, dynamically at batch creation, to the max sequence length in each batch self.padding = False # Some models have set the order of the labels to use, so let's make sure we do use it. if not self.is_regression: self.label2id = {l: i for i, l in enumerate(self.label_list)} self.id2label = {id: label for label, id in self.label2id.items()} if data_args.max_seq_length > tokenizer.model_max_length: logger.warning( f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the" f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}." ) self.max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length) new_datasets = raw_datasets.map( self.preprocess_function, batched=True, load_from_cache_file=not data_args.overwrite_cache, desc="Running tokenizer on clean dataset", ) for key in new_datasets.keys(): if "idx" not in raw_datasets[key].column_names: idx = np.arange(len(raw_datasets[key])).tolist() raw_datasets[key] = raw_datasets[key].add_column("idx", idx) if training_args.do_train: self.train_dataset = new_datasets["train"] if data_args.max_train_samples is not None: data_args.max_train_samples = min(data_args.max_train_samples, len(self.train_dataset)) self.train_dataset = self.train_dataset.select(range(data_args.max_train_samples)) size = len(self.train_dataset) select = np.random.choice(size, math.ceil(size * training_args.poison_rate), replace=False) idx = torch.zeros([size]) idx[select] = 1 self.train_dataset.poison_idx = idx if training_args.do_eval: self.eval_dataset = new_datasets["validation_matched" if data_args.dataset_name == "mnli" else "validation"] if data_args.max_eval_samples is not None: data_args.max_eval_samples = min(data_args.max_eval_samples, len(self.eval_dataset)) self.eval_dataset = self.eval_dataset.select(range(data_args.max_eval_samples)) if training_args.do_predict or data_args.dataset_name is not None or data_args.test_file is not None: self.predict_dataset = new_datasets["test_matched" if data_args.dataset_name == "mnli" else "test"] if data_args.max_predict_samples is not None: data_args.max_predict_samples = min(data_args.max_predict_samples, len(self.predict_dataset)) self.predict_dataset = self.predict_dataset.select(range(data_args.max_predict_samples)) self.metric = load_metric("glue", data_args.dataset_name) if data_args.pad_to_max_length: self.data_collator = default_data_collator elif training_args.fp16: self.data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=8) def filter(self, examples, length=None): if type(examples) == list: return [self.filter(x, length) for x in examples] elif type(examples) == dict or type(examples) == LazyRow or type(examples) == LazyBatch: return {k: self.filter(v, length) for k, v in examples.items()} elif type(examples) == str: #txt = re.sub(r"[^a-zA-Z0-9\ \%#!.,]+", '', examples) txt = examples.replace(self.tokenizer.prompt_token, "T").replace(self.tokenizer.skey_token, "K").replace( self.tokenizer.predict_token, "P").replace("[X]", "Y").replace("[Y]", "Y") if length is not None: return txt[:length] return txt return examples def preprocess_function(self, examples, **kwargs): examples = self.filter(examples, length=200) # Tokenize the texts, args = [text1, text2, ...] _examples = copy.deepcopy(examples) args = ( (_examples[self.sentence1_key],) if self.sentence2_key is None else (_examples[self.sentence1_key], _examples[self.sentence2_key]) ) result = self.tokenizer(*args, padding=self.padding, max_length=self.max_seq_length, truncation=True) result["idx"] = examples["idx"] return result def compute_metrics(self, p: EvalPrediction): preds = p.predictions[0] if isinstance(p.predictions, tuple) else p.predictions preds = np.squeeze(preds) if self.is_regression else np.argmax(preds, axis=1) if self.data_args.dataset_name is not None: result = self.metric.compute(predictions=preds, references=p.label_ids) if len(result) > 1: result["combined_score"] = np.mean(list(result.values())).item() return result elif self.is_regression: return {"mse": ((preds - p.label_ids) ** 2).mean().item()} else: return {"accuracy": (preds == p.label_ids).astype(np.float32).mean().item()}