train.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.

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

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 data import convert_example, create_dataloader, read_custom_data
from metric import MultiLabelReport
from model import MultiLabelClassifier

# yapf: disable
parser = argparse.ArgumentParser()
parser.add_argument("--save_dir", default='./checkpoints', 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("--batch_size", default=32, type=int, help="Batch size per GPU/CPU for training.")
parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
parser.add_argument("--epochs", default=3, 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("--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 seed for initialization")
parser.add_argument("--device", choices=["cpu", "gpu", "xpu"], default="gpu", help="Select which device to train model, defaults to gpu.")
parser.add_argument("--data_path", type=str, default="./data", help="The path of datasets to be loaded")
args = parser.parse_args()
# yapf: enable


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


@paddle.no_grad()
def evaluate(model, criterion, metric, data_loader):
    """
    Given a dataset, it evals model and computes the metric.

    Args:
        model(obj:`paddle.nn.Layer`): A model to classify texts.
        criterion(obj:`paddle.nn.Layer`): It can compute the loss.
        metric(obj:`paddle.metric.Metric`): The evaluation metric.
        data_loader(obj:`paddle.io.DataLoader`): The dataset loader which generates batches.
    """
    model.eval()
    metric.reset()
    losses = []
    for batch in data_loader:
        input_ids, token_type_ids, labels = batch
        logits = model(input_ids, token_type_ids)
        loss = criterion(logits, labels)
        probs = F.sigmoid(logits)
        losses.append(loss.numpy())
        metric.update(probs, labels)
    auc, f1_score = metric.accumulate()
    print("eval loss: %.5f, auc: %.5f, f1 score: %.5f" %
          (np.mean(losses), auc, f1_score))
    model.train()
    metric.reset()


def do_train():
    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)

    # Load train dataset.
    file_name = 'train.csv'
    train_ds = load_dataset(
        read_custom_data,
        filename=os.path.join(args.data_path, file_name),
        is_test=False,
        lazy=False)

    pretrained_model = ppnlp.transformers.BertModel.from_pretrained(
        "bert-base-uncased")
    tokenizer = ppnlp.transformers.BertTokenizer.from_pretrained(
        'bert-base-uncased')

    trans_func = partial(
        convert_example,
        tokenizer=tokenizer,
        max_seq_length=args.max_seq_length)
    batchify_fn = lambda samples, fn=Tuple(
        Pad(axis=0, pad_val=tokenizer.pad_token_id),  # input
        Pad(axis=0, pad_val=tokenizer.pad_token_type_id),  # segment
        Stack(dtype='float32')  # label
    ): [data for data in fn(samples)]
    train_data_loader = create_dataloader(
        train_ds,
        mode='train',
        batch_size=args.batch_size,
        batchify_fn=batchify_fn,
        trans_fn=trans_func)

    model = MultiLabelClassifier(
        pretrained_model, num_labels=len(train_ds.data[0]["label"]))

    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)
    model = paddle.DataParallel(model)
    num_training_steps = len(train_data_loader) * args.epochs

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

    # 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,
        parameters=model.parameters(),
        weight_decay=args.weight_decay,
        apply_decay_param_fun=lambda x: x in decay_params)

    metric = MultiLabelReport()
    criterion = paddle.nn.BCEWithLogitsLoss()

    global_step = 0
    tic_train = time.time()
    for epoch in range(1, args.epochs + 1):
        for step, batch in enumerate(train_data_loader, start=1):
            input_ids, token_type_ids, labels = batch
            logits = model(input_ids, token_type_ids)
            loss = criterion(logits, labels)
            probs = F.sigmoid(logits)
            metric.update(probs, labels)
            auc, f1_score = metric.accumulate()

            global_step += 1
            if global_step % 10 == 0 and rank == 0:
                print(
                    "global step %d, epoch: %d, batch: %d, loss: %.5f, auc: %.5f, f1 score: %.5f, speed: %.2f step/s"
                    % (global_step, epoch, step, loss, auc, f1_score,
                       10 / (time.time() - tic_train)))
                tic_train = time.time()
            loss.backward()
            optimizer.step()
            lr_scheduler.step()
            optimizer.clear_grad()
            if global_step % 100 == 0 and 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)


if __name__ == "__main__":
    do_train()