run_pretrain.py

import argparse
import os
# os.environ['CUDA_VISIBLE_DEVICES'] = "2"
import torch
from transformers import BertTokenizer,  \
    get_linear_schedule_with_warmup, get_constant_schedule
import numpy as np
from data_pretrain import load_entities,get_mention_loader,PretrainDataset
from pretrain import *
from utils import *
from loss import *
from datetime import datetime
from torch.optim import AdamW
from tqdm import tqdm

def set_seeds(args):
    random.seed(args.seed)
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)


def count_parameters(model):
    return sum(p.numel() for p in model.parameters() if p.requires_grad)


def strtime(datetime_checkpoint):
    diff = datetime.now() - datetime_checkpoint
    return str(diff).rsplit('.')[0]  # Ignore below seconds


def load_model(is_init, device, tokenizer, args):
    model = MaskLMEncoder(args.pretrained_model, tokenizer, device)
    # model = CaEncoder(args.pretrained_model, device, type_loss)
    if not is_init:
        state_dict = torch.load(args.model) if device.type == 'cuda' else \
            torch.load(args.model, map_location=torch.device('cpu'))
        model.load_state_dict(state_dict['sd'])
    return model


def configure_optimizer(args, model, num_train_examples):
    # https://github.com/google-research/bert/blob/master/optimization.py#L25
    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}
    ]
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.lr,
                      eps=args.adam_epsilon)

    num_train_steps = int(num_train_examples / args.batch /
                          args.gradient_accumulation_steps * args.epochs)
    num_warmup_steps = int(num_train_steps * args.warmup_proportion)

    scheduler = get_linear_schedule_with_warmup(
        optimizer, num_warmup_steps=num_warmup_steps,
        num_training_steps=num_train_steps)

    return optimizer, scheduler, num_train_steps, num_warmup_steps


def configure_optimizer_simple(args, model, num_train_examples):
    optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
    num_train_steps = int(num_train_examples / args.B /
                          args.gradient_accumulation_steps * args.epochs)
    num_warmup_steps = 0

    scheduler = get_constant_schedule(optimizer)

    return optimizer, scheduler, num_train_steps, num_warmup_steps


def get_hit_scores(indices, labels):
    hit = 0
    nums = len(labels)
    for i in range(nums):
        indice = indices[i]
        label = labels[i]
        hit += any([label[index] for index in indice])
    return hit / nums


def evaluate(model,data_loader,device):
    data_loader = tqdm(data_loader)
    scores = []
    labels = []
    for step,batch in enumerate(data_loader):
        model.eval()
        batch = tuple(t.to(device) for t in batch)
        input_ids, attention_mask, positions, label = batch
        score = model(input_ids, attention_mask, positions, label,"val")
        scores += score.tolist()
        labels += label.tolist()

    scores = torch.tensor(scores)

    top1_indices = torch.topk(scores, k=1).indices.tolist()
    hit1 = get_hit_scores(top1_indices, labels)
    top5_indices = torch.topk(scores, k=5).indices.tolist()
    hit5 = get_hit_scores(top5_indices, labels)

    return hit1, hit5


def train(args):
    set_seeds(args)
    best_val_perf = float('-inf')
    logger = Logger(args.model + '.log', on=True)
    logger.log(str(args))
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    args.device = device
    logger.log(f'Using device: {str(device)}', force=True)
    entities = load_entities(args.dataset + args.kb_path)
    # logger.log('number of entities {:d}'.format(len(entities)))

    tokenizer = BertTokenizer.from_pretrained(args.pretrained_model)
    special_tokens = ["[E1]", "[/E1]", '[or]', "[NIL]"]
    sel_tokens = [f"[info{i}]" for i in range(args.cand_num)]
    special_tokens += sel_tokens
    tokenizer.add_special_tokens({'additional_special_tokens': special_tokens})

    model = load_model(True,device,tokenizer,args)
    num_train_samples = len(entities.keys())
    if args.simpleoptim:
        optimizer, scheduler, num_train_steps, num_warmup_steps \
            = configure_optimizer_simple(args, model, num_train_samples)
    else:
        optimizer, scheduler, num_train_steps, num_warmup_steps \
            = configure_optimizer(args, model, num_train_samples)
    args.n_gpu = torch.cuda.device_count()
    model.to(device)
    dp = args.n_gpu > 1
    if dp:
        logger.log('Data parallel across {:d} GPUs {:s}'
                   ''.format(len(args.gpus.split(',')), args.gpus))
        model = nn.DataParallel(model)

    train_loader = get_mention_loader(tokenizer,entities,args)

    model.train()
    effective_bsz = args.batch * args.gradient_accumulation_steps
    # train
    logger.log('***** train *****')
    logger.log('# train samples: {:d}'.format(num_train_samples))

    logger.log('# epochs: {:d}'.format(args.epochs))
    logger.log(' batch size : {:d}'.format(args.batch))
    logger.log(' gradient accumulation steps {:d}'
               ''.format(args.gradient_accumulation_steps))
    logger.log(
        ' effective training batch size with accumulation: {:d}'
        ''.format(effective_bsz))
    logger.log(' # training steps: {:d}'.format(num_train_steps))
    logger.log(' # warmup steps: {:d}'.format(num_warmup_steps))
    logger.log(' learning rate: {:g}'.format(args.lr))
    logger.log(' # parameters: {:d}'.format(count_parameters(model)))

    step_num = 0
    tr_loss, logging_loss = 0.0, 0.0
    start_epoch = 1

    model.zero_grad()

    for epoch in range(start_epoch, args.epochs + 1):
        logger.log('\nEpoch {:d}'.format(epoch))

        epoch_start_time = datetime.now()

        epoch_train_start_time = datetime.now()
        train_loader = tqdm(train_loader)
        for step, batch in enumerate(train_loader):

            model.train()
            bsz = batch[0].size(0)
            batch = tuple(t.to(device) for t in batch)
            input_ids,attention_mask,mask_index,label_ids,weights = batch
            # input_ids, attention_mask,labels = batch
            loss = model(input_ids,attention_mask,mask_index,label_ids,weights)
            if dp:
                loss = loss.sum() / bsz
            else:
                loss /= bsz

            loss_avg = loss / args.gradient_accumulation_steps

            loss_avg.backward()
            tr_loss += loss_avg.item()

            if (step + 1) % args.gradient_accumulation_steps == 0:

                torch.nn.utils.clip_grad_norm_(model.parameters(),
                                               args.clip)
                optimizer.step()
                scheduler.step()
                model.zero_grad()
                step_num += 1

                if step_num % args.logging_steps == 0:
                    avg_loss = (tr_loss - logging_loss) / args.logging_steps
                    logger.log('Step {:10d}/{:d} | Epoch {:3d} | '
                               'Batch {:5d}/{:5d} | '
                               'Average Loss {:8.4f}'
                               ''.format(step_num, num_train_steps,
                                         epoch, step + 1,
                                         len(train_loader), avg_loss))
                    logging_loss = tr_loss
        logger.log('training time for epoch {:3d} '
                   'is {:s}'.format(epoch, strtime(epoch_train_start_time)))
        logger.log("train loss {:8.4f}".format(tr_loss / step_num))

        save_model = True
        if save_model:
            torch.save({'opt': args,
                        'sd': model.module.state_dict() if dp else model.state_dict(),
                        'perf': best_val_perf, 'epoch': epoch,
                        'opt_sd': optimizer.state_dict(),
                        'scheduler_sd': scheduler.state_dict(),
                        'tr_loss': tr_loss, 'step_num': step_num,
                        'logging_loss': logging_loss},
                       args.model)
        else:
            logger.log('')




def main(args):


    train(args)


if __name__ == '__main__':
    parser = argparse.ArgumentParser()

    parser.add_argument("--dataset",
                        default="dataset/bc5cdr/")
    parser.add_argument("--model",
                        default="model_pretrain/bc5cdr_pretrain.pt")
    parser.add_argument("--pretrained_model",
                        default="cambridgeltl/SapBERT-from-PubMedBERT-fulltext")
    parser.add_argument("--type_loss", type=str,
                        default="sum_log_nce",
                        choices=["log_sum", "sum_log", "sum_log_nce",
                                 "max_min","bce_loss"])
    parser.add_argument("--max_length", default=256)


    parser.add_argument("--kb_path", default="entity_kb.json")

    parser.add_argument("--batch", default=64)
    parser.add_argument("--lr", default=5e-6,type=float)
    parser.add_argument("--epochs", default=10,type=int)
    parser.add_argument("--cand_num", default=6)
    parser.add_argument("--warmup_proportion", default=0.2)
    parser.add_argument("--weight_decay", default=0.01)
    parser.add_argument("--adam_epsilon", default=1e-6,type=float)
    parser.add_argument("--gradient_accumulation_steps", default=2)
    parser.add_argument("--seed", default=42)
    parser.add_argument("--num_workers", default=0)
    parser.add_argument("--simpleoptim", default=False)
    parser.add_argument("--clip", default=1)

    parser.add_argument("--gpus", default="0")
    parser.add_argument("--logging_steps", default=100)

    args = parser.parse_args()
    # Set environment variables before all else.
      # Sets torch.cuda behavior
    os.environ['CUDA_VISIBLE_DEVICES'] = args.gpus
    main(args)