pet.py

import argparse
import os
import sys
import random
import time
import json
from functools import partial

import numpy as np
import paddle
import paddle.nn.functional as F

import paddlenlp as ppnlp
from paddlenlp.data import Stack, Tuple, Pad
from paddlenlp.datasets import load_dataset
from paddlenlp.transformers import LinearDecayWithWarmup

from model import ErnieForPretraining, ErnieMLMCriterion

from data import create_dataloader, transform_fn_dict
from data import convert_example, convert_chid_example
from evaluate import do_evaluate, do_evaluate_chid
from predict import do_predict, do_predict_chid, predict_file, write_fn


def set_seed(seed):
    """sets random seed"""
    random.seed(seed)
    np.random.seed(seed)
    paddle.seed(seed)


def do_train(args):
    paddle.set_device(args.device)
    rank = paddle.distributed.get_rank()
    if paddle.distributed.get_world_size() > 1:
        paddle.distributed.init_parallel_env()

    set_seed(args.seed)
    label_normalize_json = os.path.join("./label_normalized",
                                        args.task_name + ".json")

    # Ernie Model
    model = ErnieForPretraining.from_pretrained(args.language_model)
    tokenizer = ppnlp.transformers.ErnieTokenizer.from_pretrained(
        args.language_model)

    # map y
    label_norm_dict = None
    with open(label_normalize_json, 'r', encoding="utf-8") as f:
        label_norm_dict = json.load(f)

    convert_example_fn = convert_example if args.task_name != "chid" else convert_chid_example

    evaluate_fn = do_evaluate if args.task_name != "chid" else do_evaluate_chid

    predict_fn = do_predict if args.task_name != "chid" else do_predict_chid

    # load dataset
    train_ds, public_test_ds, test_ds = load_dataset(
        "fewclue",
        name=args.task_name,
        splits=("train_0", "test_public", "test"))

    # Task related transform operations, eg: numbert label -> text_label, english -> chinese
    transform_fn = partial(
        transform_fn_dict[args.task_name],
        label_normalize_dict=label_norm_dict,
        pattern_id=args.pattern_id)

    # Task related transform operations, eg: numbert label -> text_label, english -> chinese
    transform_test_fn = partial(
        transform_fn_dict[args.task_name],
        label_normalize_dict=label_norm_dict,
        is_test=True,
        pattern_id=args.pattern_id)

    # Some fewshot_learning strategy is defined by transform_fn
    # Note: Set lazy=True to transform example inplace immediately,
    # because transform_fn should only be executed only once when
    # iterate multi-times for train_ds

    train_ds = train_ds.map(transform_fn, lazy=False)
    public_test_ds = public_test_ds.map(transform_fn, lazy=False)
    test_ds = test_ds.map(transform_test_fn, lazy=False)

    # dataloader
    if args.task_name == "chid":
        # [src_ids, token_type_ids, masked_positions, masked_lm_labels, candidate_labels_ids]
        batchify_fn = lambda samples, fn=Tuple(
            Pad(axis=0, pad_val=tokenizer.pad_token_id),  # src_ids
            Pad(axis=0, pad_val=tokenizer.pad_token_type_id),  # token_type_ids
            Stack(dtype="int64"),  # masked_positions
            Stack(dtype="int64"),  # masked_lm_labels
            Stack(dtype="int64"),  # candidate_labels_ids [candidate_num, label_length]
        ): [data for data in fn(samples)]
        batchify_test_fn = lambda samples, fn=Tuple(
            Pad(axis=0, pad_val=tokenizer.pad_token_id),  # src_ids
            Pad(axis=0, pad_val=tokenizer.pad_token_type_id),  # token_type_ids
            Stack(dtype="int64"),  # masked_positions
            Stack(dtype="int64"),  # candidate_labels_ids [candidate_num, label_length]
        ): [data for data in fn(samples)]
    else:
        # [src_ids, token_type_ids, masked_positions, masked_lm_labels]
        batchify_fn = lambda samples, fn=Tuple(
            Pad(axis=0, pad_val=tokenizer.pad_token_id),  # src_ids
            Pad(axis=0, pad_val=tokenizer.pad_token_type_id),  # token_type_ids
            Stack(dtype="int64"),  # masked_positions
            Stack(dtype="int64"),  # masked_lm_labels
        ): [data for data in fn(samples)]
        batchify_test_fn = lambda samples, fn=Tuple(
            Pad(axis=0, pad_val=tokenizer.pad_token_id),  # src_ids
            Pad(axis=0, pad_val=tokenizer.pad_token_type_id),  # token_type_ids
            Stack(dtype="int64"),  # masked_positions
        ): [data for data in fn(samples)]

    trans_func = partial(
        convert_example_fn,
        tokenizer=tokenizer,
        max_seq_length=args.max_seq_length)

    trans_test_func = partial(
        convert_example_fn,
        tokenizer=tokenizer,
        max_seq_length=args.max_seq_length,
        is_test=True)

    train_data_loader = create_dataloader(
        train_ds,
        mode='train',
        batch_size=args.batch_size,
        batchify_fn=batchify_fn,
        trans_fn=trans_func)

    public_test_data_loader = create_dataloader(
        public_test_ds,
        mode='eval',
        batch_size=args.batch_size,
        batchify_fn=batchify_fn,
        trans_fn=trans_func)

    test_data_loader = create_dataloader(
        test_ds,
        mode='eval',
        batch_size=args.batch_size,
        batchify_fn=batchify_test_fn,
        trans_fn=trans_test_func)

    num_training_steps = len(train_data_loader) * args.epochs

    lr_scheduler = LinearDecayWithWarmup(args.learning_rate, num_training_steps,
                                         args.warmup_proportion)
    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,
        parameters=model.parameters(),
        weight_decay=args.weight_decay,
        apply_decay_param_fun=lambda x: x in decay_params)

    # load model if there is
    if args.init_from_ckpt and os.path.isfile(args.init_from_ckpt):
        state_dict = paddle.load(args.init_from_ckpt)
        model.set_dict(state_dict)
        print("warmup from:{}".format(args.init_from_ckpt))

    mlm_loss_fn = ErnieMLMCriterion()
    rdrop_loss = ppnlp.losses.RDropLoss()
    max_test_acc = 0.0
    global_step = 0
    tic_train = time.time()

    for epoch in range(1, args.epochs + 1):
        model.train()
        # Generate parameter names needed to perform weight decay.
        # All bias and LayerNorm parameters are excluded.
        for step, batch in enumerate(train_data_loader, start=1):
            src_ids = batch[0]
            token_type_ids = batch[1]
            masked_positions = batch[2]
            masked_lm_labels = batch[3]

            max_len = src_ids.shape[1]
            new_masked_positions = []

            for bs_index, mask_pos in enumerate(masked_positions.numpy()):
                for pos in mask_pos:
                    new_masked_positions.append(bs_index * max_len + pos)
            new_masked_positions = paddle.to_tensor(
                np.array(new_masked_positions).astype('int32'))

            prediction_scores = model(
                input_ids=src_ids,
                token_type_ids=token_type_ids,
                masked_positions=new_masked_positions)
            if args.rdrop_coef > 0:
                prediction_scores_2 = model(
                    input_ids=src_ids,
                    token_type_ids=token_type_ids,
                    masked_positions=new_masked_positions)
                ce_loss = (
                    mlm_loss_fn(prediction_scores, masked_lm_labels) +
                    mlm_loss_fn(prediction_scores_2, masked_lm_labels)) * 0.5
                kl_loss = rdrop_loss(prediction_scores, prediction_scores_2)
                loss = ce_loss + kl_loss * args.rdrop_coef
            else:
                loss = mlm_loss_fn(prediction_scores, masked_lm_labels)

            global_step += 1
            if global_step % 10 == 0 and rank == 0:
                print(
                    "global step %d, epoch: %d, batch: %d, loss: %.5f, speed: %.2f step/s"
                    % (global_step, epoch, step, loss,
                       10 / (time.time() - tic_train)))
                tic_train = time.time()
            loss.backward()
            optimizer.step()
            lr_scheduler.step()
            optimizer.clear_grad()

        if rank == 0:
            save_dir = os.path.join(args.save_dir, "model_%d" % global_step)
            if not os.path.exists(save_dir):
                os.makedirs(save_dir)
            save_param_path = os.path.join(save_dir, 'model_state.pdparams')
            paddle.save(model.state_dict(), save_param_path)
            tokenizer.save_pretrained(save_dir)

        test_accuracy, total_num = evaluate_fn(
            model, tokenizer, public_test_data_loader, label_norm_dict)
        print("epoch:{}, test_accuracy:{:.3f}, total_num:{}".format(
            epoch, test_accuracy, total_num))


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--task_name",
        required=True,
        type=str,
        help="The task_name to be evaluated")
    parser.add_argument(
        "--batch_size",
        default=32,
        type=int,
        help="Batch size per GPU/CPU for training.")
    parser.add_argument(
        "--learning_rate",
        default=1e-5,
        type=float,
        help="The initial learning rate for Adam.")
    parser.add_argument(
        "--save_dir",
        default='./checkpoint',
        type=str,
        help="The output directory where the model checkpoints will be written.")
    parser.add_argument(
        "--max_seq_length",
        default=128,
        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(
        "--weight_decay",
        default=0.0,
        type=float,
        help="Weight decay if we apply some.")
    parser.add_argument(
        "--epochs",
        default=10,
        type=int,
        help="Total number of training epochs to perform.")
    parser.add_argument(
        "--warmup_proportion",
        default=0.0,
        type=float,
        help="Linear warmup proption over the training process.")
    parser.add_argument(
        "--pattern_id", default=0, type=int, help="pattern id of pet")
    parser.add_argument(
        "--init_from_ckpt",
        type=str,
        default=None,
        help="The path of checkpoint to be loaded.")
    parser.add_argument(
        "--seed",
        type=int,
        default=1000,
        help="random seeds for initialization")
    parser.add_argument(
        "--output_dir",
        default='./output',
        type=str,
        help="The output directory where to save output")
    parser.add_argument(
        '--device',
        choices=['cpu', 'gpu'],
        default="gpu",
        help="Select which device to train model, defaults to gpu.")
    parser.add_argument(
        '--save_steps',
        type=int,
        default=10000,
        help="Inteval steps to save checkpoint")
    parser.add_argument("--if_save_checkpoints", action='store_true')
    parser.add_argument(
        "--index",
        required=True,
        type=str,
        default="0",
        help="must be in [0, 1, 2, 3, 4, all]")
    parser.add_argument(
        '--language_model',
        type=str,
        default='ernie-1.0',
        choices=['ernie-1.0'],
        help="Language model")
    parser.add_argument(
        "--rdrop_coef",
        default=0.0,
        type=float,
        help="The coefficient of KL-Divergence loss in R-Drop paper, for more detail please refer to https://arxiv.org/abs/2106.14448), if rdrop_coef > 0 then R-Drop works"
    )

    args = parser.parse_args()
    do_train(args)