run_pretrain.py

# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import argparse
import logging
import os
import io
import random
import time
import json
import copy

import numpy as np

import paddle

from paddlenlp.transformers import ConvBertForTotalPretraining, ConvBertModel, ConvBertPretrainingCriterion
from paddlenlp.transformers import ConvBertDiscriminator, ConvBertGenerator
from paddlenlp.transformers import ConvBertTokenizer
from paddlenlp.transformers import LinearDecayWithWarmup

FORMAT = '%(asctime)s-%(levelname)s: %(message)s'
logging.basicConfig(level=logging.INFO, format=FORMAT)
logger = logging.getLogger(__name__)

MODEL_CLASSES = {"convbert": (ConvBertForTotalPretraining, ConvBertTokenizer), }


def parse_args():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--model_type",
        default="convbert",
        type=str,
        help="Model type selected in the list: " +
        ", ".join(MODEL_CLASSES.keys()), )
    parser.add_argument(
        "--model_name_or_path",
        default="convbert-small",
        type=str,
        help="Path to pre-trained model or shortcut name selected in the list: "
        + ", ".join(
            sum([
                list(classes[-1].pretrained_init_configuration.keys())
                for classes in MODEL_CLASSES.values()
            ], [])), )
    parser.add_argument(
        "--input_dir",
        default=None,
        type=str,
        required=True,
        help="The input directory where the data will be read from.", )
    parser.add_argument(
        "--output_dir",
        default=None,
        type=str,
        required=True,
        help="The output directory where the model predictions and checkpoints will be written.",
    )
    parser.add_argument(
        "--max_seq_length",
        default=128,
        type=int,
        help="max length of each sequence")
    parser.add_argument(
        "--mask_prob",
        default=0.15,
        type=float,
        help="the probability of one word to be mask")
    parser.add_argument(
        "--train_batch_size",
        default=96,
        type=int,
        help="Batch size per GPU/CPU for training.", )
    parser.add_argument(
        "--eval_batch_size",
        default=96,
        type=int,
        help="Batch size per GPU/CPU for training.", )
    parser.add_argument(
        "--learning_rate",
        default=5e-4,
        type=float,
        help="The initial learning rate for Adam.")
    parser.add_argument(
        "--weight_decay",
        default=0.01,
        type=float,
        help="Weight decay if we apply some.")
    parser.add_argument(
        "--adam_epsilon",
        default=1e-6,
        type=float,
        help="Epsilon for Adam optimizer.")
    parser.add_argument(
        "--num_train_epochs",
        default=4,
        type=int,
        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_steps",
        default=10000,
        type=int,
        help="Linear warmup over warmup_steps.")

    parser.add_argument(
        "--logging_steps",
        type=int,
        default=100,
        help="Log every X updates steps.")
    parser.add_argument(
        "--save_steps",
        type=int,
        default=1000,
        help="Save checkpoint every X updates steps.")
    parser.add_argument(
        "--init_from_ckpt",
        action="store_true",
        help="Whether to load model checkpoint. if True, args.model_name_or_path must be dir store ckpt or will train from fresh start"
    )
    parser.add_argument(
        "--use_amp",
        action="store_true",
        help="Whether to use float16(Automatic Mixed Precision) to train.")
    parser.add_argument(
        "--seed", type=int, default=42, help="random seed for initialization")
    parser.add_argument(
        "--eager_run", type=bool, default=True, help="Use dygraph mode.")
    parser.add_argument(
        "--device",
        default="gpu",
        type=str,
        choices=["cpu", "gpu"],
        help="The device to select to train the model, is must be cpu/gpu.")
    args = parser.parse_args()
    return args


def set_seed(args):
    # Use the same data seed(for data shuffle) for all procs to guarantee data
    # consistency after sharding.
    random.seed(args.seed)
    np.random.seed(args.seed)
    # Maybe different op seeds(for dropout) for different procs is better. By:
    # `paddle.seed(args.seed + paddle.distributed.get_rank())`
    paddle.seed(args.seed)


class WorkerInitObj(object):
    def __init__(self, seed):
        self.seed = seed

    def __call__(self, id):
        np.random.seed(seed=self.seed + id)
        random.seed(self.seed + id)


class BookCorpus(paddle.io.Dataset):
    """
    https://web.eecs.umich.edu/~lahiri/gutenberg_dataset.html
    Args:
        data_path (:obj:`str`) : The dataset file path, which contains train.tsv, dev.tsv and test.tsv.
        tokenizer (:obj:`class PretrainedTokenizer`) : The tokenizer to split word and convert word to id.
        max_seq_length (:obj:`int`) : max length for each sequence.
        mode (:obj:`str`, `optional`, defaults to `train`):
            It identifies the dataset mode (train, test or dev).
    """

    def __init__(
            self,
            data_path,
            tokenizer,
            max_seq_length,
            mode='train', ):
        if mode == 'train':
            data_file = 'train.data'
        elif mode == 'test':
            data_file = 'test.data'
        else:
            data_file = 'dev.data'

        self.data_file = os.path.join(data_path, data_file)
        self.tokenizer = tokenizer
        self.max_seq_length = max_seq_length
        self.raw_examples = self._read_file(self.data_file)

    def _read_file(self, input_file):
        """
        Reads a text file.

        Args:
            input_file (:obj:`str`) : The file to be read.

        Returns:
            examples (:obj:`list`): All the input data.
        """
        if not os.path.exists(input_file):
            raise RuntimeError("The file {} is not found.".format(input_file))
        else:
            with io.open(input_file, "r", encoding="UTF-8") as f:
                examples = []
                while True:
                    line = f.readline()
                    if line:
                        if (len(line) > 0 and not line.isspace()):
                            example = self.tokenizer(
                                line,
                                max_seq_len=self.max_seq_length)['input_ids']
                            examples.append(example)
                    else:
                        break
                return examples

    def truncation_ids(self, ids, max_seq_length):
        if len(ids) <= (max_seq_length - 2):
            return ids
        else:
            return ids[:(max_seq_length - 2)]

    def __getitem__(self, idx):
        return self.raw_examples[idx]

    def __len__(self):
        return len(self.raw_examples)


class DataCollatorForConvBert(object):
    """
    pads, gets batch of tensors and preprocesses batches for masked language modeling
    when dataloader num_worker > 0, this collator may trigger some bugs, for safe, be sure dataloader num_worker=0
    """

    def __init__(self,
                 tokenizer,
                 max_seq_length,
                 mlm=True,
                 mlm_probability=0.15):
        self.tokenizer = tokenizer
        self.max_seq_length = max_seq_length
        self.mlm = True
        self.mlm_probability = mlm_probability

    def __call__(self, examples):
        if self.mlm:
            inputs, raw_inputs, labels = self.mask_tokens(examples)
            return inputs, raw_inputs, labels
        else:
            raw_inputs, _ = self.add_special_tokens_and_set_maskprob(
                examples, True, self.max_seq_length)
            raw_inputs = self.tensorize_batch(raw_inputs, "int64")
            inputs = raw_inputs.clone().detach()
            labels = raw_inputs.clone().detach()
            if self.tokenizer.pad_token is not None:
                pad_token_id = self.tokenizer.convert_tokens_to_ids(
                    self.tokenizer.pad_token)
                labels[labels == pad_token_id] = -100
            return inputs, raw_inputs, labels

    def tensorize_batch(self, examples, dtype):
        if isinstance(examples[0], (list, tuple)):
            examples = [paddle.to_tensor(e, dtype=dtype) for e in examples]
        length_of_first = examples[0].shape[0]
        are_tensors_same_length = all(x.shape[0] == length_of_first
                                      for x in examples)
        if are_tensors_same_length:
            return paddle.stack(examples, axis=0)
        else:
            raise ValueError(
                "the tensor in examples not have same shape, please check input examples"
            )

    def add_special_tokens_and_set_maskprob(self, inputs, truncation,
                                            max_seq_length):
        sep_token_id = self.tokenizer.convert_tokens_to_ids(
            self.tokenizer.sep_token)
        pad_token_id = self.tokenizer.convert_tokens_to_ids(
            self.tokenizer.pad_token)
        cls_token_id = self.tokenizer.convert_tokens_to_ids(
            self.tokenizer.cls_token)
        full_inputs = []
        full_maskprob = []
        max_length = 0
        for ids in inputs:
            if len(ids) > max_length:
                max_length = len(ids)
        max_length = min(max_length, max_seq_length)

        for ids in inputs:
            if len(ids) <= max_length:
                padding_num = max_length - len(ids)
                full_inputs.append(ids + ([pad_token_id] * padding_num))
                full_maskprob.append([0] + ([self.mlm_probability] * (len(
                    ids) - 2)) + [0] + ([0] * padding_num))
            else:
                if truncation:
                    full_inputs.append(ids[:max_length])
                    full_maskprob.append([0] + ([self.mlm_probability] * (
                        max_length - 2)) + [0])
                else:
                    full_inputs.append(ids)
                    full_maskprob.append([0] + ([self.mlm_probability] * (len(
                        ids) - 2)) + [0])
        return full_inputs, full_maskprob

    def mask_tokens(self, examples):
        if self.tokenizer.mask_token is None:
            raise ValueError(
                "the tokenizer does not have mask_token, please check!")
        mask_token_id = self.tokenizer.convert_tokens_to_ids(
            self.tokenizer.mask_token)

        raw_inputs, probability_matrix = self.add_special_tokens_and_set_maskprob(
            examples, True, self.max_seq_length)
        raw_inputs = self.tensorize_batch(raw_inputs, "int64")
        probability_matrix = self.tensorize_batch(probability_matrix, "float32")
        inputs = raw_inputs.clone()
        labels = raw_inputs.clone()

        total_indices = paddle.bernoulli(probability_matrix).astype(
            "bool").numpy()
        labels[~total_indices] = -100

        # 80% MASK
        indices_mask = paddle.bernoulli(paddle.full(labels.shape, 0.8)).astype(
            "bool").numpy() & total_indices
        inputs[indices_mask] = mask_token_id

        # 10% Random
        indices_random = paddle.bernoulli(paddle.full(
            labels.shape, 0.5)).astype("bool").numpy(
            ) & total_indices & ~indices_mask
        random_words = paddle.randint(
            low=0,
            high=self.tokenizer.vocab_size,
            shape=labels.shape,
            dtype="int64")
        inputs = paddle.where(
            paddle.to_tensor(indices_random), random_words, inputs)

        # 10% Original
        return inputs, raw_inputs, labels


def create_dataloader(dataset,
                      mode='train',
                      batch_size=1,
                      use_gpu=True,
                      data_collator=None):
    """
    Creats dataloader.

    Args:
        dataset(obj:`paddle.io.Dataset`):
            Dataset instance.
        mode(obj:`str`, optional, defaults to obj:`train`):
            If mode is 'train', it will shuffle the dataset randomly.
        batch_size(obj:`int`, optional, defaults to 1): 
            The sample number of a mini-batch.
        use_gpu(obj:`bool`, optional, defaults to obj:`True`):
            Whether to use gpu to run.

    Returns:
        dataloader(obj:`paddle.io.DataLoader`): The dataloader which generates batches.
    """

    if mode == 'train' and use_gpu:
        sampler = paddle.io.DistributedBatchSampler(
            dataset=dataset, batch_size=batch_size, shuffle=True)
        dataloader = paddle.io.DataLoader(
            dataset,
            batch_sampler=sampler,
            return_list=True,
            collate_fn=data_collator,
            num_workers=0)
    else:
        shuffle = True if mode == 'train' else False
        sampler = paddle.io.BatchSampler(
            dataset=dataset, batch_size=batch_size, shuffle=shuffle)
        dataloader = paddle.io.DataLoader(
            dataset,
            batch_sampler=sampler,
            return_list=True,
            collate_fn=data_collator,
            num_workers=0)

    return dataloader


def do_train(args):
    paddle.enable_static() if not args.eager_run else None
    paddle.set_device(args.device)
    if paddle.distributed.get_world_size() > 1:
        paddle.distributed.init_parallel_env()

    set_seed(args)
    worker_init = WorkerInitObj(args.seed + paddle.distributed.get_rank())

    args.model_type = args.model_type.lower()
    model_class, tokenizer_class = MODEL_CLASSES[args.model_type]

    # Loads or initializes a model.
    pretrained_models = list(tokenizer_class.pretrained_init_configuration.keys(
    ))
    if args.model_name_or_path in pretrained_models:
        tokenizer = tokenizer_class.from_pretrained(args.model_name_or_path)
        generator = ConvBertGenerator(
            ConvBertModel(**model_class.pretrained_init_configuration[
                args.model_name_or_path + "-generator"]))
        discriminator = ConvBertDiscriminator(
            ConvBertModel(**model_class.pretrained_init_configuration[
                args.model_name_or_path + "-discriminator"]))
        model = model_class(generator, discriminator)
        args.init_from_ckpt = False
    else:
        if os.path.isdir(args.model_name_or_path) and args.init_from_ckpt:
            # Load checkpoint
            tokenizer = tokenizer_class.from_pretrained(args.model_name_or_path)
            with open(
                    os.path.join(args.model_name_or_path, "run_states.json"),
                    'r') as f:
                config_dict = json.load(f)
                model_name = config_dict["model_name"]
            if model_name in pretrained_models:
                generator = ConvBertGenerator(
                    ConvBertModel(**model_class.pretrained_init_configuration[
                        model_name + "-generator"]))
                discriminator = ConvBertDiscriminator(
                    ConvBertModel(**model_class.pretrained_init_configuration[
                        model_name + "-discriminator"]))
                model = model_class(generator, discriminator)
                model.set_state_dict(
                    paddle.load(
                        os.path.join(args.model_name_or_path,
                                     "model_state.pdparams")))
            else:
                raise ValueError(
                    "initialize a model from ckpt need model_name "
                    "in model_config_file. The supported model_name "
                    "are as follows: {}".format(
                        tokenizer_class.pretrained_init_configuration.keys()))
        else:
            raise ValueError(
                "initialize a model need identifier or the "
                "directory of storing model. if use identifier, the supported model "
                "identifiers are as follows: {}, if use directory, "
                "make sure set init_from_ckpt as True".format(
                    model_class.pretrained_init_configuration.keys()))

    criterion = ConvBertPretrainingCriterion(
        getattr(model.generator,
                ConvBertGenerator.base_model_prefix).config["vocab_size"],
        model.gen_weight, model.disc_weight)
    if paddle.distributed.get_world_size() > 1:
        model = paddle.DataParallel(model)

    # Loads dataset.
    tic_load_data = time.time()
    print("start load data : %s" %
          (time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())))
    train_dataset = BookCorpus(
        data_path=args.input_dir,
        tokenizer=tokenizer,
        max_seq_length=args.max_seq_length,
        mode='train')
    print("load data done, total : %s s" % (time.time() - tic_load_data))

    # Reads data and generates mini-batches.
    data_collator = DataCollatorForConvBert(
        tokenizer=tokenizer,
        max_seq_length=args.max_seq_length,
        mlm=True,
        mlm_probability=args.mask_prob)

    train_data_loader = create_dataloader(
        train_dataset,
        batch_size=args.train_batch_size,
        mode='train',
        use_gpu=True if args.device in "gpu" else False,
        data_collator=data_collator)

    num_training_steps = args.max_steps if args.max_steps > 0 else (
        len(train_data_loader) * args.num_train_epochs)

    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
                                         args.warmup_steps)

    clip = paddle.nn.ClipGradByGlobalNorm(clip_norm=1.0)
    # Generate parameter names needed to perform weight decay.
    # All bias and LayerNorm parameters are excluded.
    decay_params = [
        p.name for n, p in model.named_parameters()
        if not any(nd in n for nd in ["bias", "norm"])
    ]
    optimizer = paddle.optimizer.AdamW(
        learning_rate=lr_scheduler,
        epsilon=args.adam_epsilon,
        parameters=model.parameters(),
        weight_decay=args.weight_decay,
        grad_clip=clip,
        apply_decay_param_fun=lambda x: x in decay_params)
    if args.use_amp:
        scaler = paddle.amp.GradScaler(init_loss_scaling=1024)

    print("start train : %s" %
          (time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())))
    trained_global_step = global_step = 0
    t_loss = paddle.to_tensor([0.0])
    log_loss = paddle.to_tensor([0.0])
    loss_list = []
    log_list = []
    tic_train = time.time()
    if os.path.isdir(args.model_name_or_path) and args.init_from_ckpt:
        optimizer.set_state_dict(
            paddle.load(
                os.path.join(args.model_name_or_path, "model_state.pdopt")))
        trained_global_step = global_step = config_dict["global_step"]
        if trained_global_step < num_training_steps:
            print(
                "[ start train from checkpoint ] we have already trained %s steps, seeking next step : %s"
                % (trained_global_step, trained_global_step + 1))
        else:
            print(
                "[ start train from checkpoint ] we have already trained %s steps, but total training steps is %s, please check configuration !"
                % (trained_global_step, num_training_steps))
            exit(0)

    for epoch in range(args.num_train_epochs):
        for step, batch in enumerate(train_data_loader):
            if trained_global_step > 0:
                trained_global_step -= 1
                continue
            global_step += 1
            input_ids, raw_input_ids, gen_labels = batch
            if args.use_amp:
                with paddle.amp.auto_cast():
                    gen_logits, disc_logits, disc_labels, attention_mask = model(
                        input_ids=input_ids,
                        raw_input_ids=raw_input_ids,
                        gen_labels=gen_labels)
                    loss = criterion(gen_logits, disc_logits, gen_labels,
                                     disc_labels, attention_mask)
                scaled = scaler.scale(loss)
                scaled.backward()
                t_loss += loss.detach()
                scaler.minimize(optimizer, scaled)
            else:
                gen_logits, disc_logits, disc_labels, attention_mask = model(
                    input_ids=input_ids,
                    raw_input_ids=raw_input_ids,
                    gen_labels=gen_labels)
                loss = criterion(gen_logits, disc_logits, gen_labels,
                                 disc_labels, attention_mask)
                loss.backward()
                t_loss += loss.detach()
                optimizer.step()
            lr_scheduler.step()
            optimizer.clear_grad()
            if global_step % args.logging_steps == 0:
                local_loss = (t_loss - log_loss) / args.logging_steps
                if (paddle.distributed.get_world_size() > 1):
                    paddle.distributed.all_gather(loss_list, local_loss)
                    if paddle.distributed.get_rank() == 0:
                        log_str = (
                            "global step {0:d}/{1:d}, epoch: {2:d}, batch: {3:d}, "
                            "avg_loss: {4:.15f}, lr: {5:.10f}, speed: {6:.2f} s/it"
                        ).format(global_step, num_training_steps, epoch, step,
                                 float((paddle.stack(loss_list).sum() / len(
                                     loss_list)).numpy()),
                                 optimizer.get_lr(),
                                 (time.time() - tic_train) / args.logging_steps)
                        print(log_str)
                        log_list.append(log_str)
                        loss_list = []
                else:
                    log_str = (
                        "global step {0:d}/{1:d}, epoch: {2:d}, batch: {3:d}, "
                        "loss: {4:.15f}, lr: {5:.10f}, speed: {6:.2f} s/it"
                    ).format(global_step, num_training_steps, epoch, step,
                             float(local_loss.numpy()),
                             optimizer.get_lr(),
                             (time.time() - tic_train) / args.logging_steps)
                    print(log_str)
                    log_list.append(log_str)
                log_loss = t_loss
                tic_train = time.time()
            if global_step % args.save_steps == 0:
                if paddle.distributed.get_rank() == 0:
                    output_dir = os.path.join(args.output_dir,
                                              "model_%d.pdparams" % global_step)
                    if not os.path.exists(output_dir):
                        os.makedirs(output_dir)
                    model_to_save = model._layers if isinstance(
                        model, paddle.DataParallel) else model
                    config_to_save = copy.deepcopy(
                        model_to_save.discriminator.convbert.config)
                    if 'self' in config_to_save:
                        del config_to_save['self']
                    run_states = {
                        "model_name": model_name
                        if args.init_from_ckpt else args.model_name_or_path,
                        "global_step": global_step,
                        "epoch": epoch,
                        "step": step,
                    }
                    with open(
                            os.path.join(output_dir, "model_config.json"),
                            'w') as f:
                        json.dump(config_to_save, f)
                    with open(
                            os.path.join(output_dir, "run_states.json"),
                            'w') as f:
                        json.dump(run_states, f)
                    paddle.save(model.state_dict(),
                                os.path.join(output_dir,
                                             "model_state.pdparams"))
                    tokenizer.save_pretrained(output_dir)
                    paddle.save(optimizer.state_dict(),
                                os.path.join(output_dir, "model_state.pdopt"))
                    if len(log_list) > 0:
                        with open(os.path.join(output_dir, "train.log"),
                                  'w') as f:
                            for log in log_list:
                                if len(log.strip()) > 0:
                                    f.write(log.strip() + '\n')
            if global_step >= num_training_steps:
                return


def print_arguments(args):
    """print arguments"""
    print('-----------  Configuration Arguments -----------')
    for arg, value in sorted(vars(args).items()):
        print('%s: %s' % (arg, value))
    print('------------------------------------------------')


if __name__ == "__main__":
    args = parse_args()
    print_arguments(args)
    n_gpu = len(os.getenv("CUDA_VISIBLE_DEVICES", "").split(","))
    if args.device in "gpu" and n_gpu > 1:
        paddle.distributed.spawn(do_train, args=(args, ), nprocs=n_gpu)
    else:
        do_train(args)