train.py

import argparse
import os
import math
import numpy as np
import torch
from apex import amp
import ujson as json
from torch.utils.data import DataLoader
from transformers import AutoConfig, AutoModel, AutoTokenizer
from transformers.optimization import AdamW, get_polynomial_decay_schedule_with_warmup, get_linear_schedule_with_warmup, get_cosine_with_hard_restarts_schedule_with_warmup
from torch.optim.lr_scheduler import CyclicLR, CosineAnnealingWarmRestarts,MultiStepLR
from model import DocREModel_KD
from utils import set_seed, collate_fn, collate_fn_kd, label_collate_fn, get_label_input_ids
from prepro import read_docred
from copy import deepcopy
from evaluation import to_official, official_evaluate
import wandb

def get_lr(optimizer):
    lm_lr = optimizer.param_groups[0]['lr']
    classifier_lr = optimizer.param_groups[1]['lr']
    return lm_lr, classifier_lr

def train(args, model, train_features, dev_features, test_features, label_loader):
    def finetune(features, optimizer, num_epoch, num_steps, args):
        n_e = 42
        train_scores = []
        dev_scores = []
        best_score = -1
        tmp_features = get_random_mask(features, args.drop_prob)
        train_dataloader = DataLoader(tmp_features, batch_size=args.train_batch_size, shuffle=True, collate_fn=collate_fn_kd, drop_last=False)
        train_iterator = range(int(num_epoch))
        total_steps = int(len(train_dataloader) * num_epoch // args.gradient_accumulation_steps)
        warmup_steps = int(total_steps * args.warmup_ratio)
        scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=warmup_steps, num_training_steps=total_steps)
        #scheduler = get_polynomial_decay_schedule_with_warmup(optimizer, num_warmup_steps=warmup_steps, power = 3.0, lr_end=1e-9, num_training_steps=total_steps)
        print("Total steps: {}".format(total_steps))
        print("Warmup steps: {}".format(warmup_steps))
        grad_norms = []
        for epoch in train_iterator:
            model.zero_grad()               
            for step, batch in enumerate(train_dataloader):
                model.train()
                inputs = {'input_ids': batch[0].to(args.device),
                          'attention_mask': batch[1].to(args.device),
                          'labels': batch[2],
                          'entity_pos': batch[3],
                          'hts': batch[4],
                          'mention_pos': batch[5],
                          'mention_hts': batch[6],
                          'padded_mention':batch[7],
                          'padded_mention_mask':batch[8],
                          'sentid_mask':batch[9],
                          'teacher_logits': batch[10],
                          'entity_types': batch[11],
                          'segment_spans': batch[12],
                          'negative_mask': batch[13].to(args.device),
                          'label_loader' : label_loader,
                          }
                #with torch.autograd.set_detect_anomaly(True):
                loss = model(**inputs)
                loss = loss / args.gradient_accumulation_steps
                #grad_norms.append(mean_grad_norm.detach().cpu().numpy())
                #with torch.autograd.set_detect_anomaly(True):
                with amp.scale_loss(loss, optimizer) as scaled_loss:
                    scaled_loss.backward()
                    #loss.backward()
                if step % args.gradient_accumulation_steps == 0:
                    if args.max_grad_norm > 0:
                        torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
                        #torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
                    optimizer.step()
                    lm_lr, classifier_lr = get_lr(optimizer)
                    #if epoch < 50:
                    scheduler.step()
                    model.zero_grad()
                    num_steps += 1
                wandb.log({"loss": loss.item()}, step=num_steps)
                if (step + 1) == len(train_dataloader) - 1 or ((args.evaluation_steps > 0 and num_steps % args.evaluation_steps == 0 and step % args.gradient_accumulation_steps == 0) and num_steps > args.start_steps ):
                    dev_score, dev_output = evaluate(args, model, dev_features, label_loader, tag="dev")
                    #train_score, train_output = evaluate(args, model, train_features, tag="train")
                    wandb.log(dev_output, step=num_steps)
                    print(dev_output)
                    train_score=0
                    train_scores.append(train_score)
                    dev_scores.append(dev_score)
                    lm_lr, classifier_lr = get_lr(optimizer)
                    print('Current Step: {:d},  Current LM lr: {:.5e}, Current Classifier lr: {:.5e}'.format(num_steps, lm_lr, classifier_lr))

                    if dev_score > best_score:
                        best_score = dev_score
                        pred, logits= report(args, model, test_features,label_loader)
                        with open("result.json", "w") as fh:
                            json.dump(pred, fh)
                        if args.save_path != "":
                            torch.save(model.state_dict(), args.save_path)
                    if args.save_last != "":
                        torch.save(model.state_dict(), args.save_last)
            
        return num_steps, train_scores, dev_scores

    new_layer = ["extractor", "bilinear", "classifier",  "projection"]
    optimizer_grouped_parameters = [
        {"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in new_layer)], },
        {"params": [p for n, p in model.named_parameters() if any(nd in n for nd in new_layer)], "lr": args.classifier_lr},
    ]

    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
    model, optimizer = amp.initialize(model, optimizer, opt_level="O1", verbosity=0)
    num_steps = 0
    set_seed(args)
    model.zero_grad()
    num_steps, train_scores, dev_scores = finetune(train_features, optimizer, args.num_train_epochs, num_steps, args)



def evaluate(args, model, features, label_loader, tag="dev"):

    dataloader = DataLoader(features, batch_size=args.test_batch_size, shuffle=False, collate_fn=collate_fn_kd, drop_last=False)
    preds = []
   
    for batch in dataloader:
        model.eval()

        inputs = {'input_ids': batch[0].to(args.device),
                  'attention_mask': batch[1].to(args.device),
                  'entity_pos': batch[3],
                  'hts': batch[4],
                  'mention_pos': batch[5],
                  'mention_hts': batch[6],
                  'padded_mention':batch[7],
                  'padded_mention_mask':batch[8],
                  'sentid_mask':batch[9],
                  'entity_types': batch[11],
                  'segment_spans': batch[12],
                  'label_loader': label_loader,
                  }
        with torch.no_grad():
            pred, *_ = model(**inputs)
            pred = pred.cpu().numpy()
            pred[np.isnan(pred)] = 0
            preds.append(pred)

    preds = np.concatenate(preds, axis=0).astype(np.float32)
    ans = to_official(preds, features)
    #if len(ans) > 0:
    if tag=='dev':
        best_f1, _, best_f1_ign, _, best_p, best_r = official_evaluate(ans, args.data_dir, args.train_file, args.dev_file)
    elif tag=='train':
        best_f1, _, best_f1_ign, _, best_p, best_r = official_evaluate(ans, args.data_dir, args.train_file, args.train_file)
    output = {
        tag + "_F1": best_f1 * 100,
        tag + "_F1_ign": best_f1_ign * 100,
        tag + "_P": best_p * 100,
        tag + "_R": best_r * 100,
    }
    return best_f1, output

def get_loss(args, model, features, tag="dev"):

    dataloader = DataLoader(features, batch_size=args.test_batch_size, shuffle=False, collate_fn=collate_fn_kd, drop_last=False)
    preds = []
    pos_loss = []
    neg_loss = []

    for batch in dataloader:
        model.eval()

        inputs = {'input_ids': batch[0].to(args.device),
                  'attention_mask': batch[1].to(args.device),
                  'labels': batch[2],
                  'entity_pos': batch[3],
                  'hts': batch[4],
                  'mention_pos': batch[5],
                  'mention_hts': batch[6],
                  'padded_mention':batch[7],
                  'padded_mention_mask':batch[8],
                  'sentid_mask':batch[9],
                  'entity_types': batch[11],
                  'segment_spans': batch[12],
                  'label_loader': label_loader,
                  }

        with torch.no_grad():
            loss, loss1, loss2 = model(**inputs)
            pos_loss.append(loss1)
            neg_loss.append(loss2)

    print('Postive Loss: {}'.format(torch.mean(torch.stack(pos_loss))))
    print('Negative Loss: {}'.format(torch.mean(torch.stack(neg_loss))))
    return pos_loss, neg_loss

def report(args, model, features, label_loader):

    dataloader = DataLoader(features, batch_size=args.test_batch_size, shuffle=False, collate_fn=collate_fn_kd, drop_last=False)
    preds = []
    logits = []
    for batch in dataloader:
        model.eval()

        inputs = {'input_ids': batch[0].to(args.device),
                  'attention_mask': batch[1].to(args.device),
                  'entity_pos': batch[3],
                  'hts': batch[4],
                  'mention_pos': batch[5],
                  'mention_hts': batch[6],
                  'padded_mention':batch[7],
                  'padded_mention_mask':batch[8],
                  'sentid_mask':batch[9],
                  'entity_types': batch[11],
                  'segment_spans': batch[12],
                  'label_loader': label_loader,
                  }

        with torch.no_grad():
            pred, logit = model(**inputs)
            pred = pred.cpu().numpy()
            pred[np.isnan(pred)] = 0
            preds.append(pred)
            logits.append(logit.detach().cpu())

    preds = np.concatenate(preds, axis=0).astype(np.float32)
    preds = to_official(preds, features)
    return preds, logits

def get_random_mask(train_features, drop_prob):
    new_features = []  
    n_e = 42
    for old_feature in train_features:
        feature = deepcopy(old_feature)
        neg_labels = torch.tensor(feature['labels'])[:, 0]
        neg_index = torch.where(neg_labels==1)[0]
        pos_index = torch.where(neg_labels==0)[0]
        perm = torch.randperm(neg_index.size(0))
        sampled_negative_index = neg_index[perm[:int(drop_prob * len(neg_index))]]
        neg_mask = torch.ones(len(feature['labels']))
        neg_mask[sampled_negative_index] = 0
        #feature['negative_mask'] = neg_mask        
        pad_neg = torch.zeros((n_e, n_e))
        num_e = int(math.sqrt(len(neg_mask)))
        pad_neg[:num_e,:num_e] = neg_mask.view(num_e,num_e)
        feature['negative_mask'] = pad_neg
        new_features.append(feature)
    return new_features


def add_logits_to_features(features, logits):
    new_features = []
    for i, old_feature in enumerate(features):
        new_feature = deepcopy(old_feature)
        new_feature['teacher_logits'] = logits[i]
        assert logits[i].shape[0] == len(new_feature['hts'])
        new_features.append(new_feature)

    return new_features

def create_negative_mask(train_features, drop_prob):
    n_e = 42
    new_features = []
    for old_feature in train_features:
        feature = deepcopy(old_feature)
        neg_labels = np.array(feature['labels'])[:, 0]
        neg_index = np.squeeze(np.argwhere(neg_labels))
        sampled_negative_index = np.random.choice(neg_index, int(drop_prob * len(neg_index)) )
        neg_mask = np.ones(len(feature['labels']))
        neg_mask[sampled_negative_index] = 0
        neg_mask = torch.tensor(neg_mask)
        pad_neg = torch.zeros((n_e, n_e))
        num_e = int(math.sqrt(len(neg_mask)))
        pad_neg[:num_e,:num_e] = neg_mask.view(num_e,num_e)
        feature['negative_mask'] = pad_neg
        new_features.append(feature)
    return new_features



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

    parser.add_argument("--data_dir", default="./dataset/docred", type=str)
    parser.add_argument("--transformer_type", default="bert", type=str)
    parser.add_argument("--model_name_or_path", default="bert-base-cased", type=str)

    parser.add_argument("--train_file", default="train_annotated.json", type=str)
    parser.add_argument("--dev_file", default="dev.json", type=str)
    parser.add_argument("--test_file", default="test.json", type=str)
    parser.add_argument("--save_path", default="", type=str)
    parser.add_argument("--save_last", default="", type=str)
    parser.add_argument("--load_path", default="", type=str)
    parser.add_argument("--teacher_path", default="", type=str)
    parser.add_argument("--load_pretrained", default="", type=str)
    parser.add_argument("--knowledge_distil", default="", type=str)
    parser.add_argument("--output_name", default="result.json", type=str)

    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("--max_seq_length", default=1024, type=int,
                        help="The maximum total input sequence length after tokenization. Sequences longer "
                             "than this will be truncated, sequences shorter will be padded.")

    parser.add_argument("--train_batch_size", default=4, type=int,
                        help="Batch size for training.")
    parser.add_argument("--test_batch_size", default=8, type=int,
                        help="Batch size for testing.")
    parser.add_argument("--gradient_accumulation_steps", default=1, type=int,
                        help="Number of updates steps to accumulate before performing a backward/update pass.")
    parser.add_argument("--num_labels", default=4, type=int,
                        help="Max number of labels in prediction.")
    parser.add_argument("--learning_rate", default=5e-5, type=float,
                        help="The initial learning rate for Adam.")
    parser.add_argument("--classifier_lr", default=1e-4, type=float,
                        help="The initial learning rate for Classifier.")
    parser.add_argument("--adam_epsilon", default=1e-6, 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("--drop_prob", default=0.0, type=float,
                        help="Negative Sample Discard rate.")
    parser.add_argument("--gamma_pos", default=1.0, type=float,
                        help="Gamma for positive class")
    parser.add_argument("--gamma_neg", default=1.0, type=float,
                        help="Gamma for negative class")
    parser.add_argument("--drop_FP", default=0.0, type=float,
                        help="Potential FP Discard rate.")
    parser.add_argument("--drop_FN", default=0.0, type=float,
                        help="Potential FN Discard rate.")
    parser.add_argument("--warmup_ratio", default=0.06, type=float,
                        help="Warm up ratio for Adam.")
    parser.add_argument("--num_train_epochs", default=30.0, type=float,
                        help="Total number of training epochs to perform.")
    parser.add_argument("--evaluation_steps", default=-1, type=int,
                        help="Number of training steps between evaluations.")
    parser.add_argument("--start_steps", default=-1, type=int) 
    parser.add_argument("--seed", type=int, default=66,
                        help="random seed for initialization")
    parser.add_argument("--num_class", type=int, default=97,
                        help="Number of relation types in dataset.")

    args = parser.parse_args()
    wandb.init(project="DocRED",  mode='disabled')

    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    args.n_gpu = torch.cuda.device_count()
    args.device = device

    config = AutoConfig.from_pretrained(
        args.config_name if args.config_name else args.model_name_or_path,
        num_labels=args.num_class,
    )
    tokenizer = AutoTokenizer.from_pretrained(
        args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
    )
    suffix = '.{}.pt'.format(args.model_name_or_path)
    read = read_docred
    if os.path.exists(os.path.join(args.data_dir, args.train_file + suffix)):
        train_features = torch.load(os.path.join(args.data_dir, args.train_file + suffix))
        print('Loaded train features')
    else:
        train_file = os.path.join(args.data_dir, args.train_file)
        train_features = read(train_file, tokenizer, max_seq_length=args.max_seq_length)
        torch.save(train_features, os.path.join(args.data_dir, args.train_file + suffix))
        print('Created and saved new train features')
    if os.path.exists(os.path.join(args.data_dir, args.dev_file + suffix)):
        dev_features = torch.load(os.path.join(args.data_dir, args.dev_file + suffix))
        print('Loaded dev features')
    else:
        dev_file = os.path.join(args.data_dir, args.dev_file)
        dev_features = read(dev_file, tokenizer, max_seq_length=args.max_seq_length)
        torch.save(dev_features, os.path.join(args.data_dir, args.dev_file + suffix))
        print('Created and saved new dev features')
    if os.path.exists(os.path.join(args.data_dir, args.test_file + suffix)):
        test_features = torch.load(os.path.join(args.data_dir, args.test_file + suffix))
        print('Loaded test features')
    else:
        test_file = os.path.join(args.data_dir, args.test_file)
        test_features = read(test_file, tokenizer, max_seq_length=args.max_seq_length)
        torch.save(test_features, os.path.join(args.data_dir, args.test_file + suffix))
        print('Created and saved new train features')
    n_e = 42
    
    model = AutoModel.from_pretrained(
        args.model_name_or_path,
        from_tf=bool(".ckpt" in args.model_name_or_path),
        config=config,
    )

    config.cls_token_id = tokenizer.cls_token_id
    config.sep_token_id = tokenizer.sep_token_id
    config.transformer_type = args.transformer_type
    label_features = get_label_input_ids(args.data_dir, tokenizer)

    label_loader = DataLoader(label_features, batch_size=36, shuffle=False, collate_fn=label_collate_fn, drop_last=False)
    
    
    set_seed(args)
    model = DocREModel_KD(args, config, model, num_labels=args.num_labels)
    model.to(0)

    if args.load_pretrained != "": #Training from checkpoint
        model = amp.initialize(model, opt_level="O1", verbosity=0)
        model.load_state_dict(torch.load(args.load_pretrained), strict=False)
        print('Loaded from checkpoint')
        dev_score, dev_output = evaluate(args, model, dev_features, label_loader, tag="dev")
        print(dev_output)
        train(args, model, train_features, dev_features, test_features, label_loader)
    elif args.knowledge_distil != "": #KD-Pre-Training
        print('KD_pretraining')
        student_model = AutoModel.from_pretrained(args.model_name_or_path,
                                                from_tf=bool(".ckpt" in args.model_name_or_path),
                                                config=config,
                                                )
        student_model = DocREModel_KD(args, config, student_model, num_labels=args.num_labels)                
        student_model.to(0)
        train(args, student_model, train_features, dev_features, test_features, label_loader)
    elif args.teacher_path != "":  # KD inference logits
        model = amp.initialize(model, opt_level="O1", verbosity=0)
        model.load_state_dict(torch.load(args.teacher_path), strict=False)
        dev_score, dev_output = evaluate(args, model, dev_features, label_loader, tag="dev")
        print(dev_output)
        pred, logits = report(args, model, train_features, label_loader)
        new_features = add_logits_to_features(train_features, logits)
        torch.save(new_features, os.path.join(args.data_dir, args.train_file + suffix))
        
        with open(args.output_name, "w") as fh:
            json.dump(pred, fh)
    elif args.load_path != "":  # Testing
        model = amp.initialize(model, opt_level="O1", verbosity=0)
        model.load_state_dict(torch.load(args.load_path), strict=False)
        dev_score, dev_output = evaluate(args, model, dev_features, label_loader, tag="dev")
        print(dev_output)
        pred, logits = report(args, model, test_features, label_loader)
        with open(args.output_name, "w") as fh:
            json.dump(pred, fh)
    else: # Training from scratch
        train(args, model, train_features, dev_features, test_features, label_loader)

if __name__ == "__main__":
    main()