main_bert.py

# coding=utf-8
"""Finetuning the library models for question-answering on Dureader2020 (DistilBERT, Bert, XLM, XLNet)."""

import argparse
import glob
import json
import logging
import os
import pickle
import random
import timeit
import numpy as np
import torch
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
from torch.utils.data.distributed import DistributedSampler
from tqdm import tqdm, trange

from transformers import (
    MODEL_FOR_QUESTION_ANSWERING_MAPPING,
    WEIGHTS_NAME,
    AdamW,
    BertConfig,
    BertTokenizer,
    BertForQuestionAnswering,
    AutoConfig,
    AutoModelForQuestionAnswering,
    AutoTokenizer,
    get_linear_schedule_with_warmup
)
from metrics import (
    compute_predictions_log_probs,
    compute_predictions_logits,
    squad_evaluate,
    baidu_evaluate
)
# from model.modeling_bert import BertForQuestionAnswering, BertConfig
from helper import SquadResult, MyProcessor, squad_convert_examples_to_features_orig
from optimizer import Adam_GC, AdamW_GC


try:
    from torch.utils.tensorboard import SummaryWriter
except ImportError:
    from tensorboardX import SummaryWriter


logger = logging.getLogger(__name__)

MODEL_CONFIG_CLASSES = list(MODEL_FOR_QUESTION_ANSWERING_MAPPING.keys())
MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in MODEL_CONFIG_CLASSES), (),)


def set_seed(args):
    random.seed(args.seed)
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    if args.n_gpu > 0:
        torch.cuda.manual_seed_all(args.seed)


def to_list(tensor):
    return tensor.detach().cpu().tolist()


class FGM():
    def __init__(self, model):
        self.model = model
        self.backup = {}

    def attack(self, epsilon=1., emb_name='word_embedding.'):
        # emb_name 这个参数要换成你模型中 embedding 的参数名
        for name, param in self.model.named_parameters():
            if param.requires_grad and emb_name in name:
                self.backup[name] = param.data.clone()
                norm = torch.norm(param.grad)
                if norm != 0 and not torch.isnan(norm):
                    r_at = epsilon * param.grad / norm
                    param.data.add_(r_at)

    def restore(self, emb_name='word_embedding.'):
        # emb_name 这个参数要换成你模型中 embedding 的参数名
        for name, param in self.model.named_parameters():
            if param.requires_grad and emb_name in name:
                assert name in self.backup
                param.data = self.backup[name]
        self.backup = {}


class PGD():
    def __init__(self, model, k=3):
        self.model = model
        self.emb_backup = {}
        self.grad_backup = {}
        self.k = k

    def attack(self, epsilon=1., alpha=0.33, emb_name='word_embedding.', is_first_attack=False):
        # emb_name 这个参数要换成你模型中 embedding 的参数名
        for name, param in self.model.named_parameters():
            if param.requires_grad and emb_name in name:
                if is_first_attack:
                    self.emb_backup[name] = param.data.clone()
                norm = torch.norm(param.grad)
                if norm != 0 and not torch.isnan(norm):
                    r_at = alpha * param.grad / norm
                    param.data.add_(r_at)
                    param.data = self.project(name, param.data, epsilon)

    def restore(self, emb_name='word_embedding.'):
        # emb_name 这个参数要换成你模型中 embedding 的参数名
        for name, param in self.model.named_parameters():
            if param.requires_grad and emb_name in name: 
                assert name in self.emb_backup
                param.data = self.emb_backup[name]
        self.emb_backup = {}

    def project(self, param_name, param_data, epsilon):
        r = param_data - self.emb_backup[param_name]
        if torch.norm(r) > epsilon:
            r = epsilon * r / torch.norm(r)
        return param_data + r

    def backup_grad(self):
        for name, param in self.model.named_parameters():
            if param.requires_grad:
                self.grad_backup[name] = param.grad

    def restore_grad(self):
        for name, param in self.model.named_parameters():
            if param.requires_grad:
                param.grad = self.grad_backup[name]


def train(args, train_dataset, model, tokenizer):
    """ Train the model """
    if args.do_fgm:
        fgm = FGM(model)
    if args.do_pgd:
        pgd = PGD(model, k=3)

    if args.local_rank in [-1, 0]:
        tb_writer = SummaryWriter(args.summary)

    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)

    if args.max_steps > 0:
        t_total = args.max_steps
        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
    else:
        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs

    warmup_steps = int(t_total * args.warmup_ratio)
    logging_steps = int(t_total * args.logging_ratio)
    save_steps = int(t_total * args.save_ratio)

    # Prepare optimizer and schedule (linear warmup and decay)
    no_decay = ["bias", "LayerNorm.weight"]
    optimizer_grouped_parameters = [
        {
            "params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
            "weight_decay": args.weight_decay,
        },
        {"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0},
    ]
    if args.gc:
        optimizer = AdamW_GC(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
    else:
        optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
    # optimizer = Lookahead(optimizer=optimizer, k=5, alpha=0.5)
    scheduler = get_linear_schedule_with_warmup(
        optimizer, num_warmup_steps=warmup_steps, num_training_steps=t_total
    )

    # Check if saved optimizer or scheduler states exist
    if os.path.isfile(os.path.join(args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
        os.path.join(args.model_name_or_path, "scheduler.pt")
    ):
        # Load in optimizer and scheduler states
        optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
        scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))

    if args.fp16:
        try:
            from apex import amp
        except ImportError:
            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")

        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)

    # multi-gpu training (should be after apex fp16 initialization)
    if args.n_gpu > 1:
        model = torch.nn.DataParallel(model)

    # Distributed training (should be after apex fp16 initialization)
    if args.local_rank != -1:
        model = torch.nn.parallel.DistributedDataParallel(
            model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True
        )

    # Train!
    logger.info("***** Running training *****")
    logger.info("  Num examples = %d", len(train_dataset))
    logger.info("  Num Epochs = %d", args.num_train_epochs)
    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
    logger.info(
        "  Total train batch size (w. parallel, distributed & accumulation) = %d",
        args.train_batch_size
        * args.gradient_accumulation_steps
        * (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
    )
    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
    logger.info("  Total optimization steps = %d", t_total)

    global_step = 1
    epochs_trained = 0
    steps_trained_in_current_epoch = 0
    # Check if continuing training from a checkpoint
    if os.path.exists(args.model_name_or_path):
        try:
            # set global_step to gobal_step of last saved checkpoint from model path
            checkpoint_suffix = args.model_name_or_path.split("-")[-1].split("/")[0]
            global_step = int(checkpoint_suffix)
            epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
            steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)

            logger.info("  Continuing training from checkpoint, will skip to saved global_step")
            logger.info("  Continuing training from epoch %d", epochs_trained)
            logger.info("  Continuing training from global step %d", global_step)
            logger.info("  Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
            steps_trained_in_current_epoch *= args.gradient_accumulation_steps
        except ValueError:
            logger.info("  Starting fine-tuning.")

    tr_loss, logging_loss = 0.0, 0.0
    model.zero_grad()
    train_iterator = trange(
        epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]
    )
    # Added here for reproductibility
    set_seed(args)

    for _ in train_iterator:
        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
        for step, batch in enumerate(epoch_iterator):

            # Skip past any already trained steps if resuming training
            if steps_trained_in_current_epoch > 0:
                steps_trained_in_current_epoch -= 1
                continue

            model.train()
            batch = tuple(t.to(args.device) for t in batch)

            inputs = {
                "input_ids": batch[0],
                "attention_mask": batch[1],
                "token_type_ids": batch[2],
                "start_positions": batch[3],
                "end_positions": batch[4],
            }

            if args.model_type in ["xlm", "roberta", "distilbert", "camembert"]:
                del inputs["token_type_ids"]

            if args.model_type in ["xlnet", "xlm"]:
                inputs.update({"cls_index": batch[5], "p_mask": batch[6]})
                if args.version_2_with_negative:
                    inputs.update({"is_impossible": batch[7]})
                if hasattr(model, "config") and hasattr(model.config, "lang2id"):
                    inputs.update(
                        {"langs": (torch.ones(batch[0].shape, dtype=torch.int64) * args.lang_id).to(args.device)}
                    )

            outputs = model(**inputs)
            # model outputs are always tuple in transformers (see doc)
            loss = outputs[0]

            if args.n_gpu > 1:
                loss = loss.mean()  # mean() to average on multi-gpu parallel (not distributed) training
            if args.gradient_accumulation_steps > 1:
                loss = loss / args.gradient_accumulation_steps

            if args.fp16:
                with amp.scale_loss(loss, optimizer) as scaled_loss:
                    scaled_loss.backward()
            else:
                loss.backward()

            tr_loss += loss.item()

            if args.do_fgm:
                # 对抗训练
                fgm.attack()
                outputs_adv = model(**inputs)
                loss_adv = outputs_adv[0]
                loss_adv = loss_adv.mean() / args.gradient_accumulation_steps
                loss_adv.backward()
                fgm.restore()

            if args.do_pgd:
                # 对抗训练-PGD
                pgd.backup_grad()
                for t in range(pgd.k):
                    pgd.attack(is_first_attack=(t==0)) # 在 embedding 上添加对抗扰动, first attack 时备份 param.data
                    if t != pgd.k-1:
                        model.zero_grad()
                    else:
                        pgd.restore_grad()
                    outputs_adv = model(**inputs)
                    loss_adv = outputs_adv[0]
                    loss_adv = loss_adv.mean() / args.gradient_accumulation_steps
                    loss_adv.backward() # 反向传播，并在正常的 grad 基础上，累加对抗训练的梯度
                pgd.restore() # 恢复 embedding 参数

            if (step + 1) % args.gradient_accumulation_steps == 0:
                # # 当使用 GC 或 FGM 训练时去掉梯度裁剪
                if not args.do_fgm and not args.gc:
                    if args.fp16:
                        torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
                    else:
                        torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)

                optimizer.step()
                scheduler.step()  # Update learning rate schedule
                model.zero_grad()
                global_step += 1

                # Log metrics
                if args.local_rank in [-1, 0] and logging_steps > 0 and global_step % logging_steps == 0:
                    # Only evaluate when single GPU otherwise metrics may not average well
                    if args.local_rank == -1 and args.evaluate_during_training:
                        results = evaluate(args, model, tokenizer, prefix='dev', step=global_step)
                        res_baidu = baidu_evaluate(f'{args.data_dir}/{args.predict_file}', f'{args.output_dir}/predictions_dev_{global_step}.json')
                        print(res_baidu)
                        if res_baidu['F1'] >= args.best_val_f1:
                            args.best_val_f1 = res_baidu['F1']
                            args.best_val_step = global_step
                        for key, value in results.items():
                            tb_writer.add_scalar("squad_eval_{}".format(key), value, global_step)
                        for key, value in res_baidu.items():
                            tb_writer.add_scalar("baidu_eval_{}".format(key), value, global_step)
                    tb_writer.add_scalar("lr", scheduler.get_lr()[0], global_step)
                    tb_writer.add_scalar("loss", (tr_loss - logging_loss) / logging_steps, global_step)
                    logging_loss = tr_loss

                # Save model checkpoint
                if args.local_rank in [-1, 0] and save_steps > 0 and global_step % save_steps == 0:
                    output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
                    if not os.path.exists(output_dir):
                        os.makedirs(output_dir)
                    # Take care of distributed/parallel training
                    model_to_save = model.module if hasattr(model, "module") else model
                    model_to_save.save_pretrained(output_dir)
                    tokenizer.save_pretrained(output_dir)

                    torch.save(args, os.path.join(output_dir, "training_args.bin"))
                    logger.info("Saving model checkpoint to %s", output_dir)

                    torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
                    torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
                    logger.info("Saving optimizer and scheduler states to %s", output_dir)

            if args.max_steps > 0 and global_step > args.max_steps:
                epoch_iterator.close()
                break
        if args.max_steps > 0 and global_step > args.max_steps:
            train_iterator.close()
            break

    if args.local_rank in [-1, 0]:
        tb_writer.close()

    return global_step, tr_loss / global_step


def evaluate(args, model, tokenizer, prefix="dev", step=0):
    dataset, examples, features = load_and_cache_examples(args, tokenizer, set_type=prefix, output_examples=True)

    if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
        os.makedirs(args.output_dir)

    args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)

    # Note that DistributedSampler samples randomly
    eval_sampler = SequentialSampler(dataset)
    eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)

    # multi-gpu evaluate
    if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel):
        model = torch.nn.DataParallel(model)

    # Eval!
    logger.info("***** Running evaluation {} *****".format(prefix))
    logger.info("  Num examples = %d", len(dataset))
    logger.info("  Batch size = %d", args.eval_batch_size)

    all_results = []
    start_time = timeit.default_timer()

    for batch in tqdm(eval_dataloader, desc="Evaluating"):
        model.eval()
        batch = tuple(t.to(args.device) for t in batch)

        with torch.no_grad():
            inputs = {
                "input_ids": batch[0],
                "attention_mask": batch[1],
                "token_type_ids": batch[2],
            }

            if args.model_type in ["xlm", "roberta", "distilbert", "camembert"]:
                del inputs["token_type_ids"]

            example_indices = batch[3]

            # XLNet and XLM use more arguments for their predictions
            if args.model_type in ["xlnet", "xlm"]:
                inputs.update({"cls_index": batch[4], "p_mask": batch[5]})
                # for lang_id-sensitive xlm models
                if hasattr(model, "config") and hasattr(model.config, "lang2id"):
                    inputs.update(
                        {"langs": (torch.ones(batch[0].shape, dtype=torch.int64) * args.lang_id).to(args.device)}
                    )

            outputs = model(**inputs)

        for i, example_index in enumerate(example_indices):
            eval_feature = features[example_index.item()]
            unique_id = int(eval_feature.unique_id)

            output = [to_list(output[i]) for output in outputs]

            # Some models (XLNet, XLM) use 5 arguments for their predictions, while the other "simpler"
            # models only use two.
            if len(output) >= 5:
                start_logits = output[0]
                start_top_index = output[1]
                end_logits = output[2]
                end_top_index = output[3]
                cls_logits = output[4]

                result = SquadResult(
                    unique_id,
                    start_logits,
                    end_logits,
                    start_top_index=start_top_index,
                    end_top_index=end_top_index,
                    cls_logits=cls_logits,
                )

            else:
                start_logits, end_logits = output
                result = SquadResult(unique_id, start_logits, end_logits)

            all_results.append(result)

    if prefix == 'test':
        with open(os.path.join(args.output_dir, args.test_prob_file), 'wb') as f:
            pickle.dump(all_results, f)

    evalTime = timeit.default_timer() - start_time
    logger.info("  Evaluation done in total %f secs (%f sec per example)", evalTime, evalTime / len(dataset))

    # Compute predictions
    output_prediction_file = os.path.join(args.output_dir, "predictions_{}_{}.json".format(prefix, step))
    output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}_{}.json".format(prefix, step))

    if args.version_2_with_negative:
        output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}_{}.json".format(prefix, step))
    else:
        output_null_log_odds_file = None

    # XLNet and XLM use a more complex post-processing procedure
    if args.model_type in ["xlnet", "xlm"]:
        start_n_top = model.config.start_n_top if hasattr(model, "config") else model.module.config.start_n_top
        end_n_top = model.config.end_n_top if hasattr(model, "config") else model.module.config.end_n_top

        predictions = compute_predictions_log_probs(
            examples,
            features,
            all_results,
            args.n_best_size,
            args.max_answer_length,
            output_prediction_file,
            output_nbest_file,
            output_null_log_odds_file,
            start_n_top,
            end_n_top,
            args.version_2_with_negative,
            tokenizer,
            args.verbose_logging
        )
    else:
        predictions = compute_predictions_logits(
            examples,
            features,
            all_results,
            args.n_best_size,
            args.max_answer_length,
            args.do_lower_case,
            output_prediction_file,
            output_nbest_file,
            output_null_log_odds_file,
            args.verbose_logging,
            args.version_2_with_negative,
            args.null_score_diff_threshold,
            tokenizer
        )

    if prefix == 'dev':
        # Compute the F1 and exact scores.
        results = squad_evaluate(examples, predictions)
        return results
    else:
        return None


def merge(args, tokenizer, prefix="test"):
    dataset, examples, features = load_and_cache_examples(args, tokenizer, set_type=prefix, output_examples=True)

    with open(f'{args.output_dir}/test_prob_k0.pkl', 'rb') as f:
        all_results_kf0 = pickle.load(f)
    with open(f'{args.output_dir}/test_prob_k1.pkl', 'rb') as f:
        all_results_kf1 = pickle.load(f)
    with open(f'{args.output_dir}/test_prob_k2.pkl', 'rb') as f:
        all_results_kf2 = pickle.load(f)
    with open(f'{args.output_dir}/test_prob_k3.pkl', 'rb') as f:
        all_results_kf3 = pickle.load(f)
    with open(f'{args.output_dir}/test_prob_k4.pkl', 'rb') as f:
        all_results_kf4 = pickle.load(f)

    all_results = []
    for r0,r1,r2,r3,r4 in zip(all_results_kf0, all_results_kf1, all_results_kf2, all_results_kf3, all_results_kf4):
        # print(r0.start_logits)
        unique_id = r0.unique_id
        start_logits = np.mean([r0.start_logits, r1.start_logits, r2.start_logits, r3.start_logits, r4.start_logits], axis=0)
        end_logits = np.mean([r0.end_logits, r1.end_logits, r2.end_logits, r3.end_logits, r4.end_logits], axis=0)
        result = SquadResult(unique_id, start_logits, end_logits)
        all_results.append(result)

    # Compute predictions
    output_prediction_file = f"{args.output_dir}/merge_predictions_{prefix}.json"
    output_nbest_file = f"{args.output_dir}/merge_nbest_predictions_{prefix}.json"

    if args.version_2_with_negative:
        output_null_log_odds_file = f"{args.output_dir}/merge_null_odds_{prefix}.json"
    else:
        output_null_log_odds_file = None

    predictions = compute_predictions_logits(
        examples,
        features,
        all_results,
        args.n_best_size,
        args.max_answer_length,
        args.do_lower_case,
        output_prediction_file,
        output_nbest_file,
        output_null_log_odds_file,
        args.verbose_logging,
        args.version_2_with_negative,
        args.null_score_diff_threshold,
        tokenizer
    )

    with open(f'{args.output_dir}/merge_test_prob.pkl', 'wb') as f:
        pickle.dump(all_results, f)


def load_and_cache_examples(args, tokenizer, set_type='train', output_examples=False):
    if args.local_rank not in [-1, 0] and set_type == 'train':
        # Make sure only the first process in distributed training process the dataset, and the others will use the cache
        torch.distributed.barrier()

    # Load data features from cache or dataset file
    input_dir = args.feature_dir if args.feature_dir else "."
    cached_features_file = os.path.join(
        input_dir,
        "cached_{}_{}".format(
            set_type,
            str(args.max_seq_length),
        ),
    )

    # Init features and dataset from cache if it exists
    if os.path.exists(cached_features_file) and not args.overwrite_cache:
        logger.info("Loading features from cached file %s", cached_features_file)
        features_and_dataset = torch.load(cached_features_file)
        features, dataset, examples = (
            features_and_dataset["features"],
            features_and_dataset["dataset"],
            features_and_dataset["examples"],
        )
    else:
        logger.info("Creating features from dataset file at %s", input_dir)

        if not args.data_dir and ((is_evaluate and not args.predict_file) or (not is_evaluate and not args.train_file)):
            try:
                import tensorflow_datasets as tfds
            except ImportError:
                raise ImportError("If not data_dir is specified, tensorflow_datasets needs to be installed.")

            if args.version_2_with_negative:
                logger.warn("tensorflow_datasets does not handle version 2 of SQuAD.")

            tfds_examples = tfds.load("squad")
            examples = SquadV1Processor().get_examples_from_dataset(tfds_examples, evaluate=set_type == 'train')
        else:
            processor = MyProcessor()
            if set_type == 'dev':
                examples = processor.get_dev_examples(args.data_dir, filename=args.predict_file)
            elif set_type == 'train':
                examples = processor.get_train_examples(args.data_dir, filename=args.train_file)
            elif set_type == 'test':
                examples = processor.get_test_examples(args.data_dir, filename=args.test_file)

        features, dataset = squad_convert_examples_to_features_orig(
            examples=examples,
            tokenizer=tokenizer,
            max_seq_length=args.max_seq_length,
            doc_stride=args.doc_stride,
            max_query_length=args.max_query_length,
            is_training=set_type == 'train',
            return_dataset="pt",
            threads=args.threads,
        )

        if args.local_rank in [-1, 0]:
            logger.info("Saving features into cached file %s", cached_features_file)
            torch.save({"features": features, "dataset": dataset, "examples": examples}, cached_features_file)

    if args.local_rank == 0 and not is_evaluate:
        # Make sure only the first process in distributed training process the dataset, and the others will use the cache
        torch.distributed.barrier()

    if output_examples:
        return dataset, examples, features
    return dataset


def main():
    parser = argparse.ArgumentParser()

    # Required parameters
    parser.add_argument(
        "--model_type",
        default=None,
        type=str,
        required=True,
        help="Model type selected in the list: " + ", ".join(MODEL_TYPES),
    )
    parser.add_argument(
        "--model_name_or_path",
        default=None,
        type=str,
        required=True,
        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS),
    )
    parser.add_argument(
        "--summary",
        default=None,
        type=str,
        help="Model summary",
    )
    parser.add_argument(
        "--output_dir",
        default=None,
        type=str,
        required=True,
        help="The output directory where the model checkpoints and predictions will be written.",
    )

    # Other parameters
    parser.add_argument(
        "--data_dir",
        default=None,
        type=str,
        help="The input data dir. Should contain the .json files for the task."
        + "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
    )
    parser.add_argument(
        "--feature_dir",
        default=None,
        type=str,
        help="The input feature dir. Should contain the cached_features_file for the task."
    )
    parser.add_argument(
        "--train_file",
        default=None,
        type=str,
        help="The input training file. If a data dir is specified, will look for the file there"
        + "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
    )
    parser.add_argument(
        "--predict_file",
        default=None,
        type=str,
        help="The input evaluation file. If a data dir is specified, will look for the file there"
        + "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
    )
    parser.add_argument(
        "--test_file",
        default=None,
        type=str,
        help="The input test file.",
    )
    parser.add_argument(
        "--test_prob_file",
        default=None,
        type=str,
        help="The output test_prob file.",
    )
    parser.add_argument(
        "--config_name", default="", type=str, help="Pretrained config name or path if not the same as model_name"
    )
    parser.add_argument(
        "--tokenizer_name",
        default="",
        type=str,
        help="Pretrained tokenizer name or path if not the same as model_name",
    )
    parser.add_argument(
        "--cache_dir",
        default="",
        type=str,
        help="Where do you want to store the pre-trained models downloaded from s3",
    )

    parser.add_argument(
        "--version_2_with_negative",
        action="store_true",
        help="If true, the SQuAD examples contain some that do not have an answer.",
    )
    parser.add_argument(
        "--null_score_diff_threshold",
        type=float,
        default=0.0,
        help="If null_score - best_non_null is greater than the threshold predict null.",
    )

    parser.add_argument(
        "--max_seq_length",
        default=512,
        type=int,
        help="The maximum total input sequence length after WordPiece tokenization. Sequences "
        "longer than this will be truncated, and sequences shorter than this will be padded.",
    )
    parser.add_argument(
        "--doc_stride",
        default=128,
        type=int,
        help="When splitting up a long document into chunks, how much stride to take between chunks.",
    )
    parser.add_argument(
        "--max_query_length",
        default=32,
        type=int,
        help="The maximum number of tokens for the question. Questions longer than this will "
        "be truncated to this length.",
    )
    parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
    parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.")
    parser.add_argument("--do_test", action="store_true", help="Whether to run test on the test set.")
    parser.add_argument("--do_merge", action="store_true", help="Whether to merge test prob.")
    parser.add_argument(
        "--evaluate_during_training", action="store_true", help="Run evaluation during training at each logging step."
    )
    parser.add_argument(
        "--do_lower_case", action="store_true", help="Set this flag if you are using an uncased model."
    )
    parser.add_argument("--do_fgm", action="store_true", help="Whether to run Adv-FGM training.")
    parser.add_argument("--do_pgd", action="store_true", help="Whether to run Adv-PGD training.")
    parser.add_argument("--gc", action="store_true", help="Whether to run optimizer-gc training.")

    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int, help="Batch size per GPU/CPU for training.")
    parser.add_argument(
        "--per_gpu_eval_batch_size", default=32, type=int, help="Batch size per GPU/CPU for evaluation."
    )
    parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
    parser.add_argument(
        "--gradient_accumulation_steps",
        type=int,
        default=1,
        help="Number of updates steps to accumulate before performing a backward/update pass.",
    )
    parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
    parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
    parser.add_argument(
        "--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform."
    )
    parser.add_argument(
        "--max_steps",
        default=-1,
        type=int,
        help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
    )
    parser.add_argument("--warmup_ratio", default=0.1, type=float, help="Linear warmup over warmup_ratio.")
    parser.add_argument(
        "--n_best_size",
        default=10,
        type=int,
        help="The total number of n-best predictions to generate in the nbest_predictions.json output file.",
    )
    parser.add_argument(
        "--max_answer_length",
        default=32,
        type=int,
        help="The maximum length of an answer that can be generated. This is needed because the start "
        "and end predictions are not conditioned on one another.",
    )
    parser.add_argument(
        "--verbose_logging",
        action="store_true",
        help="If true, all of the warnings related to data processing will be printed. "
        "A number of warnings are expected for a normal SQuAD evaluation.",
    )
    parser.add_argument(
        "--lang_id",
        default=0,
        type=int,
        help="language id of input for language-specific xlm models (see tokenization_xlm.PRETRAINED_INIT_CONFIGURATION)",
    )
    parser.add_argument("--best_val_f1", type=float, default=0., help="best_val_f1")
    parser.add_argument("--best_val_step", type=int, default=0, help="best_val_step")
    parser.add_argument("--logging_ratio", type=float, default=0.1, help="Log every X updates ratio.")
    parser.add_argument("--save_ratio", type=float, default=0.1, help="Save checkpoint every X updates ratio.")
    parser.add_argument(
        "--eval_all_checkpoints",
        action="store_true",
        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number",
    )
    parser.add_argument("--no_cuda", action="store_true", help="Whether not to use CUDA when available")
    parser.add_argument(
        "--overwrite_output_dir", action="store_true", help="Overwrite the content of the output directory"
    )
    parser.add_argument(
        "--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets"
    )
    parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")

    parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus")
    parser.add_argument(
        "--fp16",
        action="store_true",
        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
    )
    parser.add_argument(
        "--fp16_opt_level",
        type=str,
        default="O1",
        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
        "See details at https://nvidia.github.io/apex/amp.html",
    )
    parser.add_argument("--server_ip", type=str, default="", help="Can be used for distant debugging.")
    parser.add_argument("--server_port", type=str, default="", help="Can be used for distant debugging.")

    parser.add_argument("--threads", type=int, default=1, help="multiple threads for converting example to features")
    args = parser.parse_args()

    if args.doc_stride >= args.max_seq_length - args.max_query_length:
        logger.warning(
            "WARNING - You've set a doc stride which may be superior to the document length in some "
            "examples. This could result in errors when building features from the examples. Please reduce the doc "
            "stride or increase the maximum length to ensure the features are correctly built."
        )

    if (
        os.path.exists(args.output_dir)
        and os.listdir(args.output_dir)
        and args.do_train
        and not args.overwrite_output_dir
    ):
        raise ValueError(
            "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
                args.output_dir
            )
        )

    # Setup distant debugging if needed
    if args.server_ip and args.server_port:
        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
        import ptvsd

        print("Waiting for debugger attach")
        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
        ptvsd.wait_for_attach()

    # Setup CUDA, GPU & distributed training
    if args.local_rank == -1 or args.no_cuda:
        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
        torch.cuda.set_device(args.local_rank)
        device = torch.device("cuda", args.local_rank)
        torch.distributed.init_process_group(backend="nccl")
        args.n_gpu = 1
    args.device = device

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
    )
    logger.warning(
        "Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
        args.local_rank,
        device,
        args.n_gpu,
        bool(args.local_rank != -1),
        args.fp16,
    )

    # Set seed
    set_seed(args)

    # Load pretrained model and tokenizer
    if args.local_rank not in [-1, 0]:
        # Make sure only the first process in distributed training will download model & vocab
        torch.distributed.barrier()

    args.model_type = args.model_type.lower()
    config = BertConfig.from_pretrained(
        args.config_name if args.config_name else args.model_name_or_path,
        cache_dir=args.cache_dir if args.cache_dir else None,
    )
    tokenizer = BertTokenizer.from_pretrained(
        args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
        do_lower_case=args.do_lower_case,
        cache_dir=args.cache_dir if args.cache_dir else None,
    )
    model = AutoModelForQuestionAnswering.from_pretrained(
        args.model_name_or_path,
        from_tf=bool(".ckpt" in args.model_name_or_path),
        config=config,
        cache_dir=args.cache_dir if args.cache_dir else None,
    )

    if args.local_rank == 0:
        # Make sure only the first process in distributed training will download model & vocab
        torch.distributed.barrier()

    model.to(args.device)

    logger.info("Training/evaluation parameters %s", args)

    # Before we do anything with models, we want to ensure that we get fp16 execution of torch.einsum if args.fp16 is set.
    # Otherwise it'll default to "promote" mode, and we'll get fp32 operations. Note that running `--fp16_opt_level="O2"` will
    # remove the need for this code, but it is still valid.
    if args.fp16:
        try:
            import apex
            apex.amp.register_half_function(torch, "einsum")
        except ImportError:
            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")

    # Training
    if args.do_train:
        train_dataset = load_and_cache_examples(args, tokenizer, set_type='train', output_examples=False)
        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)

    # Save the trained model and the tokenizer
    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
        # Create output directory if needed
        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
            os.makedirs(args.output_dir)

        logger.info("Saving model checkpoint to %s", args.output_dir)
        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
        # They can then be reloaded using `from_pretrained()`
        # Take care of distributed/parallel training
        model_to_save = model.module if hasattr(model, "module") else model
        model_to_save.save_pretrained(args.output_dir)
        tokenizer.save_pretrained(args.output_dir)

        # Good practice: save your training arguments together with the trained model
        torch.save(args, os.path.join(args.output_dir, "training_args.bin"))

        # Load a trained model and vocabulary that you have fine-tuned
        model = AutoModelForQuestionAnswering.from_pretrained(args.output_dir)
        tokenizer = BertTokenizer.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
        model.to(args.device)

    # Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
    results = {}
    if args.do_eval and args.local_rank in [-1, 0]:
        if args.do_train:
            logger.info("Loading checkpoints saved during training for evaluation")
            checkpoints = [args.output_dir]
            if args.eval_all_checkpoints:
                checkpoints = list(
                    os.path.dirname(c)
                    for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
                )
                logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce model loading logs
        else:
            logger.info("Loading checkpoint %s for evaluation", args.model_name_or_path)
            checkpoints = [args.model_name_or_path]

        logger.info("Evaluate the following checkpoints: %s", checkpoints)

        for checkpoint in checkpoints:
            # Reload the model
            global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
            model = AutoModelForQuestionAnswering.from_pretrained(checkpoint)
            model.to(args.device)

            # Evaluate
            result = evaluate(args, model, tokenizer, prefix='dev', step=global_step)

            result = dict((k + ("_{}".format(global_step) if global_step else ""), v) for k, v in result.items())
            results.update(result)

    logger.info("Results: {}".format(results))

    if args.do_test and args.local_rank in [-1, 0]:
        if args.do_train:
            checkpoint = f'{args.output_dir}/checkpoint-{args.best_val_step}'
            model = AutoModelForQuestionAnswering.from_pretrained(checkpoint)
            model.to(args.device)
            evaluate(args, model, tokenizer, prefix='test', step=args.best_val_step)
        else:
            global_step = args.model_name_or_path.split("-")[-1]
            model = AutoModelForQuestionAnswering.from_pretrained(args.model_name_or_path)
            model.to(args.device)
            evaluate(args, model, tokenizer, prefix='test', step=global_step)
    if args.do_merge:
        merge(args, tokenizer, prefix="test")


if __name__ == "__main__":
    main()