main.py

import glob
import logging
import os
import pickle
import random
import re
import shutil
from typing import Dict, List, Tuple

import pandas as pd
import numpy as np
import torch

from sklearn.model_selection import train_test_split
from torch.nn.utils.rnn import pad_sequence
from torch.utils.data import DataLoader, Dataset, RandomSampler, SequentialSampler
from torch.utils.data.distributed import DistributedSampler
from tqdm.notebook import tqdm, trange

from pathlib import Path

from transformers import (
    MODEL_WITH_LM_HEAD_MAPPING,
    WEIGHTS_NAME,
    AdamW,
    AutoConfig,
    AutoModelWithLMHead,
    AutoTokenizer,
    AutoModelForSequenceClassification,
    PreTrainedModel,
    PreTrainedTokenizer,
    get_linear_schedule_with_warmup,
)
from nltk.tokenize import word_tokenize
from nltk.translate.bleu_score import sentence_bleu
from rouge_score import rouge_scorer

from config import Args
from data_loader import PersonaDataset
from nli_evaluate import nli_evaluator


# Configs
logger = logging.getLogger(__name__)

MODEL_CONFIG_CLASSES = list(MODEL_WITH_LM_HEAD_MAPPING.keys())
MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)

args = Args()


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 construct_conv(row, tokenizer, eos = True):
    flatten = lambda l: [item for sublist in l for item in sublist]
    conv = list(reversed([tokenizer.encode(x) + [tokenizer.eos_token_id] for x in row]))
    conv = flatten(conv)
    return conv

    
def _sorted_checkpoints(args, checkpoint_prefix="checkpoint", use_mtime=False) -> List[str]:
    ordering_and_checkpoint_path = []

    glob_checkpoints = glob.glob(os.path.join(args.output_dir, "{}-*".format(checkpoint_prefix)))

    for path in glob_checkpoints:
        if use_mtime:
            ordering_and_checkpoint_path.append((os.path.getmtime(path), path))
        else:
            regex_match = re.match(".*{}-([0-9]+)".format(checkpoint_prefix), path)
            if regex_match and regex_match.groups():
                ordering_and_checkpoint_path.append((int(regex_match.groups()[0]), path))

    checkpoints_sorted = sorted(ordering_and_checkpoint_path)
    checkpoints_sorted = [checkpoint[1] for checkpoint in checkpoints_sorted]
    return checkpoints_sorted


def _rotate_checkpoints(args, checkpoint_prefix="checkpoint", use_mtime=False) -> None:
    if not args.save_total_limit:
        return
    if args.save_total_limit <= 0:
        return

    # Check if we should delete older checkpoint(s)
    checkpoints_sorted = _sorted_checkpoints(args, checkpoint_prefix, use_mtime)
    if len(checkpoints_sorted) <= args.save_total_limit:
        return

    number_of_checkpoints_to_delete = max(0, len(checkpoints_sorted) - args.save_total_limit)
    checkpoints_to_be_deleted = checkpoints_sorted[:number_of_checkpoints_to_delete]
    for checkpoint in checkpoints_to_be_deleted:
        logger.info("Deleting older checkpoint [{}] due to args.save_total_limit".format(checkpoint))
        shutil.rmtree(checkpoint)


def train(args, train_dataset, vaild_set, model: PreTrainedModel, tokenizer: PreTrainedTokenizer) -> Tuple[int, float]:
    """ Train the model """

    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, collate_fn=train_dataset.collate, drop_last = True
    )

    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

    model = model.module if hasattr(model, "module") else model  # Take care of distributed/parallel training
    model.resize_token_embeddings(len(tokenizer))
    # add_special_tokens_(model, tokenizer)

    # 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},
    ]
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
    scheduler = get_linear_schedule_with_warmup(
        optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
    )

    # Check if saved optimizer or scheduler states exist
    if (
        args.model_name_or_path
        and 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 = 0
    epochs_trained = 0
    steps_trained_in_current_epoch = 0
    # Check if continuing training from a checkpoint
    if args.model_name_or_path and 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)
        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]
    )
    set_seed(args)  # Added here for reproducibility
    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

            inputs, fake_inputs = batch
            labels, fake_labels = batch

            if inputs.shape[1] > 1024: continue

            inputs = inputs.to(args.device)
            labels = labels.to(args.device)
            model.train()
            outputs = model(inputs, labels=labels)
            loss = outputs[0]  # model outputs are always tuple in transformers (see doc)

            if not args.oracle and args.zero_shot and args.use_prompts:
                fake_inputs = fake_inputs.to(args.device)
                fake_labels = fake_labels.to(args.device)
                outputs = model(fake_inputs, labels=fake_labels)
                loss2 = outputs[0]
                loss = loss + loss2

            if not args.oracle and not args.zero_shot and args.constractive:
                fake_inputs = fake_inputs.to(args.device)
                fake_labels = fake_labels.to(args.device)
                outputs2 = model(fake_inputs, labels=fake_labels)
                loss2 = outputs2[0]
                loss_c = - torch.log((torch.exp(loss) / (torch.exp(loss) + torch.exp(loss2))))
                loss = (1 - args.constractive_lambada) * loss + args.constractive_lambada * loss_c

            if args.n_gpu > 1:
                loss = loss.mean()  # mean() to average on multi-gpu parallel 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 (step + 1) % args.gradient_accumulation_steps == 0:
                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

                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
                    # Log metrics
                    if (
                        args.local_rank == -1 and args.evaluate_during_training
                    ):  # Only evaluate when single GPU otherwise metrics may not average well
                        results = evaluate(args, model, tokenizer, vaild_set)
                    logging_loss = tr_loss

                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
                    checkpoint_prefix = "checkpoint"
                    # Save model checkpoint
                    output_dir = os.path.join(args.output_dir, "{}-{}".format(checkpoint_prefix, global_step))
                    os.makedirs(output_dir, exist_ok=True)
                    model_to_save = (
                        model.module if hasattr(model, "module") else model
                    )  # Take care of distributed/parallel training
                    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)

                    _rotate_checkpoints(args, checkpoint_prefix)

                    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

    return global_step, tr_loss / global_step


# Evaluation of some model
def evaluate(args, model: PreTrainedModel, tokenizer: PreTrainedTokenizer, vaild_set, prefix="") -> Dict:
    # Loop to handle MNLI double evaluation (matched, mis-matched)
    eval_output_dir = args.output_dir

    eval_dataset = PersonaDataset(tokenizer, args, vaild_set, logger, evaluation=True)
    os.makedirs(eval_output_dir, exist_ok=True)
    args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
    # Note that DistributedSampler samples randomly

    eval_sampler = SequentialSampler(eval_dataset)
    eval_dataloader = DataLoader(
        eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size, collate_fn=eval_dataset.collate, drop_last = True
    )

    # multi-gpu evaluate
    if args.n_gpu > 1:
        model = torch.nn.DataParallel(model)

    # Eval!
    logger.info("***** Running evaluation {} *****".format(prefix))
    logger.info("  Num examples = %d", len(eval_dataset))
    logger.info("  Batch size = %d", args.eval_batch_size)
    eval_loss = 0.0
    nb_eval_steps = 0
    model.eval()

    scorer = rouge_scorer.RougeScorer(['rouge1', 'rouge2', 'rougeL'], use_stemmer=True)
    bleu = []
    rouge1 = []
    rouge2 = []
    rougel = []
    nli_score = []
    nli_tokenizer = AutoTokenizer.from_pretrained(args.nli["model_card"])
    nli_model = AutoModelForSequenceClassification.from_pretrained(args.nli["model_card"]).eval()
    nli_model.to(args.device).half()

    for batch in tqdm(eval_dataloader, desc="Evaluating"):
        inputs, personilaties = batch
        labels = inputs
        inputs = inputs.to(args.device)
        labels = labels.to(args.device)

        with torch.no_grad():
            outputs = model(inputs, labels=labels)
            lm_loss, lm_logits = outputs[:2]
            eval_loss += lm_loss.mean().item()
        nb_eval_steps += 1

        _logits = torch.argmax(lm_logits, dim=-1)
        _labels = torch.cat([labels[:, 1:], labels[:, :1] * 0], dim=1)
        
        for batch_id in range(len(_logits)):
            pred = _logits[batch_id]
            pred = pred[pred != 0]
            pred = tokenizer.decode(pred)
            truth = _labels[batch_id]
            truth = truth[truth != 0]
            truth = tokenizer.decode(truth)
            persona = personilaties[batch_id]
            persona = tokenizer.decode(persona)
            personas = persona.split("<|endoftext|>")

            nli_score.append(nli_evaluator(pred, personas, nli_model, nli_tokenizer, args))
            bleu.append(sentence_bleu([word_tokenize(truth)], word_tokenize(pred)))
            rouge1.append(scorer.score(truth, pred)['rouge1'].fmeasure)
            rouge2.append(scorer.score(truth, pred)['rouge2'].fmeasure)
            rougel.append(scorer.score(truth, pred)['rougeL'].fmeasure)

    eval_loss = eval_loss / nb_eval_steps
    perplexity = torch.exp(torch.tensor(eval_loss))
    result = {"perplexity": perplexity, "bleu": np.average(bleu), "rouge1": np.average(rouge1), 
        "rouge2": np.average(rouge2), "rougeL": np.average(rougel), "nli_score": np.average(nli_score)}

    output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
    with open(output_eval_file, "w") as writer:
        logger.info("***** Eval results {} *****".format(prefix))
        for key in sorted(result.keys()):
            logger.info("%s = %s", key, str(result[key]))
            writer.write("%s = %s\n" % (key, str(result[key])))

    return result


def main(train_set, vaild_set, test_set):
    if args.should_continue:
        sorted_checkpoints = _sorted_checkpoints(args)
        if len(sorted_checkpoints) == 0:
            raise ValueError("Used --should_continue but no checkpoint was found in --output_dir.")
        else:
            args.model_name_or_path = sorted_checkpoints[-1]

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

    # 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,
        args.device,
        args.n_gpu,
        bool(args.local_rank != -1),
        args.fp16,
    )

    # Set seed
    set_seed(args)

    config = AutoConfig.from_pretrained(args.config_name, cache_dir=args.cache_dir)
    tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, cache_dir=args.cache_dir)
    model = AutoModelWithLMHead.from_pretrained(
        args.model_name_or_path,
        from_tf=False,
        config=config,
        cache_dir=args.cache_dir,
    )
    model.to(args.device)
    
    logger.info("Training/evaluation parameters %s", args)

    # Training
    if args.do_train:
        train_dataset = PersonaDataset(tokenizer, args, train_set, logger)

        global_step, tr_loss = train(args, train_dataset, vaild_set, model, tokenizer)
        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)

    # Saving best-practices: if you use save_pretrained for the model and tokenizer, you can reload them using from_pretrained()
    if args.do_train:
        # Create output directory if needed
        os.makedirs(args.output_dir, exist_ok=True)

        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()`
        model_to_save = (
            model.module if hasattr(model, "module") else model
        )  # Take care of distributed/parallel training
        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 = AutoModelWithLMHead.from_pretrained(args.output_dir)
        tokenizer = AutoTokenizer.from_pretrained(args.output_dir)
        model.to(args.device)

    # Evaluation
    results = {}
    if args.do_eval and args.local_rank in [-1, 0]:
        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 logging
        logger.info("Evaluate the following checkpoints: %s", checkpoints)
        for checkpoint in checkpoints:
            global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
            prefix = checkpoint.split("/")[-1] if checkpoint.find("checkpoint") != -1 else ""

            model = AutoModelWithLMHead.from_pretrained(checkpoint)
            model.to(args.device)
            result = evaluate(args, model, tokenizer, test_set, prefix=prefix)
            result = dict((k + "_{}".format(global_step), v) for k, v in result.items())
            results.update(result)

    return results


if __name__ == "__main__":

    with open(args.input_dir, "rb") as f:
        train_set, valid_set, test_set = pickle.load(f)

    main(train_set, valid_set, test_set)