DeepRapper 代码详解-优快云博客

本文链接：https://blog.youkuaiyun.com/qq_43620967/article/details/124111310

DeepRapper

readme
创建环境
一、从预训练的模型生成rap
- 下载预训练模型
二、bash train.sh
- train.sh
train.py
- def main

readme

DeepRapper: Neural Rap Generation with Rhyme and Rhythm Modeling, by Lanqing Xue, Kaitao Song, Duocai Wu, Xu Tan, Nevin L. Zhang, Tao Qin, Wei-Qiang Zhang, Tie-Yan Liu, ACL 2021,是一个基于Transformer的RAP生成系统，它可以模拟韵律和节奏。它以相反的顺序生成歌词，具有押韵表示和韵律增强的约束，并在歌词中插入节拍符号，用于节奏/节拍建模。据我们所知，DeepRapper是第一个同时产生押韵和节奏的说唱的系统。

在这里插入图片描述
The input and output representation of DeepRapper model

1. Data Preparation

为每首歌准备歌词和拼音。我们提供了一些数据样本 in DeepRapper/data/.

├── data
│   └── lyrics
│       └── lyrics_samples
│           └── raw
│               └── singer01
│                   └── album01
│                       ├── song01
│                       │   ├── lyric_with_beat_global.txt
│                       │   └── mapped_final_with_beat_global.txt
│                       └── song02
│                           ├── lyric_with_beat_global.txt
│                           └── mapped_final_with_beat_global.txt

Here is a sample of lyric_with_beat_global.txt:

20_[01:12.56][BEAT]那就[BEAT]让我再沉[BEAT]沦这一世
21_[01:14.49][BEAT]不理[BEAT]解早已[BEAT]经不止一次
22_[01:16.59][BEAT]那就[BEAT]让我孤[BEAT]注最后一掷
23_[01:18.61][BEAT]不想昏[BEAT]暗之中[BEAT]度过每日
24_[01:20.60][BEAT]那就[BEAT]让我再[BEAT]沉沦这一世
25_[01:22.48][BEAT]不理[BEAT]解早已[BEAT]经不止一次
26_[01:24.58][BEAT]那就[BEAT]让我孤[BEAT]注最后一掷
27_[01:26.47][BEAT]不想昏[BEAT]暗之[BEAT]中度过每日

Here is a sample of mapped_final_with_beat_global.txt:

20_[01:12.56][BEAT] a ou [BEAT] ang o ai en [BEAT] en e i i
21_[01:14.49][BEAT] u i [BEAT] ie ao i [BEAT] in u i i i
22_[01:16.59][BEAT] a ou [BEAT] ang o u [BEAT] u ei ou i i
23_[01:18.61][BEAT] u ang en [BEAT] an i ong [BEAT] u o ei i
24_[01:20.60][BEAT] a ou [BEAT] ang o ai [BEAT] en en e i i
25_[01:22.48][BEAT] u i [BEAT] ie ao i [BEAT] in u i i i
26_[01:24.58][BEAT] a ou [BEAT] ang o u [BEAT] u ei ou i i
27_[01:26.47][BEAT] u ang en [BEAT] an i [BEAT] ong u o ei i

2. Training & Generation

我们提供了一个训练和生成的示例脚本.

train

To train run:

bash train.sh

When training, you may see the logs:

starting training
epoch 1
time: 2021-xx-xx 11:17:57.067011
51200
now time: 11:17. Step 10 of piece 0 of epoch 1, loss 9.587631130218506
now time: 11:18. Step 20 of piece 0 of epoch 1, loss 9.187388515472412

你可以在bash文件中指定参数，如number of epoch, bach size, etc.
经过训练的模型被保存在 [model_dir]/lyrics/[raw_data_dir][_reverse]/[model_sign]/final_model.
For example, in the default train.sh, the path is model/lyrics/lyrics_samples_reverse/samples/final_model.

generate

To generate by the trained DeepRapper, run

bash generate.sh

您可以在bash文件中指定参数，例如 beam width, number of samples, etc.

For 获取更多生成的样本, visit https://ai-muzic.github.io/deeprapper/.

3. Pretrained Model

You can download a pretrained DeepRapper https://msramllasc.blob.core.windows.net/modelrelease/deeprapper-model.zip.

To generate by our provided pretrained DeepRapper, first unzip the pretrained DeepRapper. Then, put the unzipped directory deeprapper-model under the folder model/. 因此，完整的路径如下:

├── model
│   └── deeprapper-model
│       ├── pytorch_model.bin
│       └── config.json

最后，运行以下命令来生成:

bash generate_from_pretrain.sh

创建环境

requirement

scikit-learn==0.21.3
miditoolkit==0.1.14
fastBPE
dtw-python
nltk
keras
tqdm
matplotlib
pretty_midi
PyYAML
pypianoroll
fairseq==0.10.0
torch==1.7.1
transformers==3.5.1
pypinyin==0.39.1
jieba==0.42.1
tensorboard==2.4.0
scipy==1.3.1
librosa
pyworld
soundfile
pyOpenSSL
secrets

建deeprapper环境

conda create --name deeprapper python=3.7

安装包

conda install --yes --file requirements.txt

pip3 install -r requirements.txt

也可以一个一个装

pip3 install jieba==0.42.1

一、从预训练的模型生成rap

下载预训练模型

cd 到model目录
输入

curl https://msramllasc.blob.core.windows.net/modelrelease/deeprapper-model.zip --output deeprapper-model.zip

在这里插入图片描述

jar xvf deeprapper-model.zip

上述问题最主要的原因是文件没有完全下载成功 重新下就行
jar 可以解压但是是不完全的

bash generate_from_pretrain.sh

开头设置为”蛋黄的长裙蓬松的头发“
在这里插入图片描述

二、bash train.sh

cd 到对应的目录

bash train.sh

train.sh

#!/bin/bash

python train.py \
  --device '0,1' \
  --stride 1024 \
  --model_config 'config/model_config_small.json' \
  --model_dir 'model' \
  --root_path 'data/lyrics/' \
  --raw_data_dir 'lyrics_samples' \
  --batch_size 1 \
  --epochs 4 \
  --enable_final \
  --enable_sentence \
  --enable_relative_pos \
  --enable_beat \
  --reverse \
  --model_sign 'samples' \
  --with_beat \
  --beat_mode 0 \
  --tokenize \
  --raw

train.py

import argparse
import os
import random
import sys
from datetime import datetime

import numpy as np
from tqdm import tqdm
from utils import swap_value

if __name__ == '__main__':
    main()

def main

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

    # path to data
    parser.add_argument('--model_dir', default='model', type=str, required=False, help='directory of learned models')
    parser.add_argument('--root_path', default='data/lyrics/', type=str, required=False, help='root path')
    parser.add_argument('--raw_data_dir', default='lyric_with_final_small', type=str, required=False, help='directory of raw data')
    parser.add_argument('--model_sign', default='1a', type=str, required=False, help='model sign, to identify each model')
    parser.add_argument('--writer_dir', default='tensorboard_summary/', type=str, required=False, help='directory of tensorboard logs')

    # path to dictionary
    parser.add_argument('--tokenizer_path', default='tokenizations/chinese_dicts.txt', type=str, required=False, help='vocabulary of tokens')
    parser.add_argument('--finalizer_path', default='tokenizations/finals.txt', type=str, required=False, help='vocabulary of finals')
    parser.add_argument('--sentencer_path', default='tokenizations/sentences.txt', type=str, required=False, help='vocabulary of sentence numbers')
    parser.add_argument('--poser_path', default='tokenizations/sentences.txt', type=str, required=False, help='vocabulary of intra-sentence positions')
    parser.add_argument('--beater_path', default='tokenizations/beats.txt', type=str, required=False, help='vocabulary of beats')

    # hyperparameters for training
    parser.add_argument('--device', default='0', type=str, required=False, help='choose gpus')
    parser.add_argument('--init_device', default=0, type=int, required=False, help='set the main gpu number')
    parser.add_argument('--model_config', default='config/model_config_small.json', type=str, required=False,
                        help='model configurations')
    parser.add_argument('--epochs', default=5, type=int, required=False, help='number of epochs')
    parser.add_argument('--start_epoch', default=0, type=int, required=False, help='the initial epoch')
    parser.add_argument('--batch_size', default=8, type=int, required=False, help='batch size')
    parser.add_argument('--lr', default=1.5e-4, type=float, required=False, help='learning rate')
    parser.add_argument('--warmup_steps', default=2000, type=int, required=False, help='warm up steps')
    parser.add_argument('--log_step', default=10, type=int, required=False,
                        help='steps of each printing of logs')
    parser.add_argument('--stride', default=1024, type=int, required=False, help='windows of context in training')
    parser.add_argument('--gradient_accumulation', default=1, type=int, required=False, help='steps of gradient accumulation')
    parser.add_argument('--fp16', action='store_true', help='mixed precision')
    parser.add_argument('--fp16_opt_level', default='O1', type=str, required=False)
    parser.add_argument('--max_grad_norm', default=1.0, type=float, required=False)
    parser.add_argument('--num_pieces', default=1, type=int, required=False, help='number of pieces of data')
    parser.add_argument('--min_length', default=0, type=int, required=False, help='min length of the lyrics')
    parser.add_argument('--pretrained_model', default='', type=str, required=False, help='path to the pretrianed model')

    # ways to process data
    parser.add_argument('--encoder_json', default="tokenizations/encoder.json", type=str, help="encoder.json", required=False)
    parser.add_argument('--vocab_bpe', default="tokenizations/vocab.bpe", type=str, help="vocab.bpe", required=False)
    parser.add_argument('--raw', action='store_true', help='whether the preprocessing is done', required=False)
    parser.add_argument('--tokenize', action='store_true', help='whether the tokenization is done', required=False)
    parser.add_argument('--segment', action='store_true', help='do Chinese Word Segmentation or not', required=False)
    parser.add_argument('--bpe_token', action='store_true', help='use subword', required=False)
    parser.add_argument('--enable_final', action='store_true', help='whether to use final embedding', required=False)
    parser.add_argument('--enable_sentence', action='store_true', help='whether to use sentence embedding', required=False)
    parser.add_argument('--enable_relative_pos', action='store_true', help='whether to use intra-sentence positional embedding', required=False)
    parser.add_argument('--enable_beat', action='store_true', help='whether to use beat embedding', required=False)
    parser.add_argument('--reverse', action='store_true', help='whether to use reverse language model', required=False)
    parser.add_argument('--with_beat', action='store_true', help='whether to generate beat', required=False)
    parser.add_argument('--beat_mode', default=0, type=int, help='beat mode：0.no control；1.global control；2.local control', required=False)

    args = parser.parse_args()
    print('args:\n' + args.__repr__())

    # basic settings
    # set envs and import related packages
    os.environ["CUDA_VISIBLE_DEVICES"] = args.device
    import torch
    import transformers
    from torch.nn import DataParallel
    from torch.utils.tensorboard import SummaryWriter
    from prepare_train_data import build_files_separate, read_lyrics, prepare_lyrics, get_shuffled_samples
    from tokenizations.bpe_tokenizer import get_encoder
    from module import GPT2Config, GPT2Model, GPT2LMHeadModel

    # choose tokenizer
    if args.segment:
        from tokenizations import tokenization_bert_word_level as tokenization_bert
    else:
        from tokenizations import tokenization_bert
    # set tokenizer
    if args.bpe_token:
        full_tokenizer = get_encoder(args.encoder_json, args.vocab_bpe)
        full_tokenizer.max_len = 999999
    else:
        full_tokenizer = tokenization_bert.BertTokenizer(
            vocab_file=args.tokenizer_path,
            do_lower_case=False
        )
        full_finalizer = tokenization_bert.BertTokenizer(
            vocab_file=args.finalizer_path,
            tokenize_chinese_chars=False,
            do_lower_case=False
        )
        full_sentencer = tokenization_bert.BertTokenizer(
            vocab_file=args.sentencer_path,
            tokenize_chinese_chars=False,
            do_lower_case=False
        )
        full_poser = tokenization_bert.BertTokenizer(
            vocab_file=args.poser_path,
            tokenize_chinese_chars=False,
            do_lower_case=False
        )
        full_beater = tokenization_bert.BertTokenizer(
            vocab_file=args.beater_path,
            tokenize_chinese_chars=False,
            do_lower_case=False
        )

    # run tokenizeing
    # dataset root key
    key = args.root_path.rstrip('/').split('/')[-1]
    # processed data root path
    processed_path = os.path.join(args.root_path, args.raw_data_dir, 'processed')

    tokenized_path = os.path.join(processed_path, 'tokenized')
    reverse_str = '_reverse' if args.reverse else ''
    tokenized_data_path = os.path.join(tokenized_path, f'tokenized{reverse_str}')
    finalized_data_path = os.path.join(tokenized_path, f'finalized{reverse_str}')
    sentenced_data_path = os.path.join(tokenized_path, f'sentenced{reverse_str}')
    posed_data_path = os.path.join(tokenized_path, f'posed{reverse_str}')
    beated_data_path = os.path.join(tokenized_path, f'beated{reverse_str}')

    if args.tokenize:
        # prepare data
        if args.raw:
            print('Processing from raw data...')
            prepare_fn = {
                'lyrics': prepare_lyrics
            }
            prepare_fn[key](
                ins_path=os.path.join(args.root_path, args.raw_data_dir, 'raw'),  # demo: data/lyrics/lyrics_22w/raw
                out_path=processed_path,  # demo: data/lyrics/lyrics_22w/processed
                with_beat=args.with_beat,
                beat_mode=args.beat_mode
            )

        print('Loading processed data for training...')
        read_fn = {
            'lyrics': read_lyrics,
        }
        train_lines, train_finals, train_sentences, train_pos, train_beats = read_fn[key](processed_path, reverse=args.reverse)

        print('Tokenizing processed data for training...')
        build_files_separate(num_pieces=args.num_pieces,
                             stride=args.stride,
                             min_length=args.min_length,
                             lines=train_lines,
                             finals=train_finals,
                             sentences=train_sentences,
                             pos=train_pos,
                             beats=train_beats,
                             tokenized_data_path=tokenized_data_path,
                             finalized_data_path=finalized_data_path,
                             sentenced_data_path=sentenced_data_path,
                             posed_data_path=posed_data_path,
                             beated_data_path=beated_data_path,
                             full_tokenizer=full_tokenizer,
                             full_finalizer=full_finalizer,
                             full_sentencer=full_sentencer,
                             full_poser=full_poser,
                             full_beater=full_beater,
                             enable_final=args.enable_final,
                             enable_sentence=args.enable_sentence,
                             enable_pos=args.enable_relative_pos,
                             enable_beat=args.enable_beat,
                             segment=args.segment)

        print('End')

    # Training settings
    # calculate total training steps
    full_len = 0
    print('calculating total steps')
    for i in tqdm(range(args.num_pieces)):
        with open(os.path.join(tokenized_data_path, 'tokenized_train_{}.txt'.format(i)), 'r') as f:
            full_len += len([int(item) for item in f.read().strip().split()])
    total_steps = int(full_len / args.stride * args.epochs / args.batch_size / args.gradient_accumulation)
    print('total steps = {}'.format(total_steps))

    # build model
    model_config = GPT2Config.from_json_file(args.model_config)
    print('config:\n' + model_config.to_json_string())
    if not args.pretrained_model:
        model = GPT2LMHeadModel(config=model_config)
    else:
        model = GPT2LMHeadModel.from_pretrained(args.pretrained_model)
    model.train()
    # set whether to use cuda
    gpu_count = torch.cuda.device_count()
    if gpu_count > 0:
        device_ids = [int(i) for i in range(gpu_count)]
        swap_value(device_ids, 0, args.init_device)
        device = f'cuda:{device_ids[0]}'
    else:
        device = 'cpu'
    print('using device:', device)
    model.to(device)

    # check parameters number of the built model
    num_parameters = 0
    parameters = model.parameters()
    for parameter in parameters:
        num_parameters += parameter.numel()
    print('number of parameters: {}'.format(num_parameters))

    # set optimizer
    optimizer = transformers.AdamW(model.parameters(), lr=args.lr, correct_bias=True)
    # change WarmupLinearSchedule to get_linear_schedule_with_warmup for current version of Transformers
    scheduler = transformers.get_linear_schedule_with_warmup(optimizer,
                                                             num_warmup_steps=args.warmup_steps,
                                                             num_training_steps=total_steps)

    # set whether to use 16-bits parameters to save GPU memory if your GPU support the operations of 16-bits number
    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)

    # set whether to use multi GPUs
    multi_gpu = False
    if gpu_count > 1:
        print("Let's use", gpu_count, "GPUs!", device_ids)
        model = DataParallel(model, device_ids=device_ids)
        multi_gpu = True

    # set log info
    log_dir = os.path.join(args.writer_dir, key, f'{args.raw_data_dir}{reverse_str}', args.model_sign)
    tb_writer = SummaryWriter(log_dir=log_dir)
    assert args.log_step % args.gradient_accumulation == 0

    output_dir = os.path.join(args.model_dir, key, f'{args.raw_data_dir}{reverse_str}', args.model_sign)
    if not os.path.exists(output_dir):
        os.makedirs(output_dir)

    print('starting training')
    overall_step = 0
    running_loss = 0
    for epoch in range(args.start_epoch, args.epochs):
        print('epoch {}'.format(epoch + 1))
        now = datetime.now()
        print('time: {}'.format(now))

        # shuffle pieces of data
        x = np.linspace(0, args.num_pieces - 1, args.num_pieces, dtype=np.int32)
        random.shuffle(x)

        piece_num = 0

        # enumerate data pieces
        for i in x:
            # prepare training sentences
            with open(os.path.join(tokenized_data_path, 'tokenized_train_{}.txt'.format(i)), 'r') as f:
                line = f.read().strip()
            tokens = line.split()
            # print(len(tokens))
            tokens = [int(token) for token in tokens]
            # tokens = torch.Tensor(tokens)

            if args.enable_final:
                with open(os.path.join(finalized_data_path, 'tokenized_train_{}.txt'.format(i)), 'r') as f:
                    final = f.read().strip()
                finals = final.split()
                # print(len(finals))
                finals = [int(final) for final in finals]
                # finals = torch.Tensor(finals)

            if args.enable_sentence:
                with open(os.path.join(sentenced_data_path, 'tokenized_train_{}.txt'.format(i)), 'r') as f:
                    sentence = f.read().strip()
                sentences = sentence.split()
                # print(len(sentences))
                sentences = [int(sentence) for sentence in sentences]
                # sentences = torch.Tensor(sentences)

            if args.enable_relative_pos:
                with open(os.path.join(posed_data_path, 'tokenized_train_{}.txt'.format(i)), 'r') as f:
                    p = f.read().strip()
                pos = p.split()
                # print(len(sentences))
                pos = [int(p) for p in pos]
                # sentences = torch.Tensor(sentences)

            if args.enable_beat:
                with open(os.path.join(beated_data_path, 'tokenized_train_{}.txt'.format(i)), 'r') as f:
                    beat = f.read().strip()
                beats = beat.split()
                # print(len(sentences))
                beats = [int(beat) for beat in beats]
                # sentences = torch.Tensor(sentences)
            # print('training: ', len(tokens), len(finals), len(sentences))

            start_point = 0
            samples_token, samples_final, samples_sentence, samples_pos, samples_beat = [], [], [], [], []
            n_ctx = model_config.n_ctx  # the length of a sentence for training
            stride = args.stride
            print(len(tokens))
            while start_point < len(tokens) - stride:
                samples_token.append(tokens[start_point: start_point + stride])
                if args.enable_final:
                    samples_final.append(finals[start_point: start_point + stride])
                if args.enable_sentence:
                    samples_sentence.append(sentences[start_point: start_point + stride])
                if args.enable_relative_pos:
                    samples_pos.append(pos[start_point: start_point + stride])
                if args.enable_beat:
                    samples_beat.append(beats[start_point: start_point + stride])
                start_point += stride
            if start_point < len(tokens):
                samples_token.append(tokens[len(tokens) - stride:])
                if args.enable_final:
                    samples_final.append(finals[len(tokens) - stride:])
                if args.enable_sentence:
                    samples_sentence.append(sentences[len(tokens) - stride:])
                if args.enable_relative_pos:
                    samples_pos.append(pos[len(tokens) - stride:])
                if args.enable_beat:
                    samples_beat.append(beats[len(tokens) - stride:])

            samples_token, samples_final, samples_sentence, samples_pos, samples_beat = get_shuffled_samples(
                samples_token, samples_final,
                samples_sentence, samples_pos, samples_beat
            )
#             print(len(samples_token), len(samples_final), len(samples_sentence), len(samples_))

            # enumerate batch data
            for step in range(len(samples_token) // args.batch_size):  # drop last

                #  prepare batch data
                batch_token = samples_token[step * args.batch_size: (step + 1) * args.batch_size]
                batch_inputs_token = torch.Tensor(batch_token).long().to(device)

                if samples_final is not None:
                    batch_final = samples_final[step * args.batch_size: (step + 1) * args.batch_size]
                    batch_inputs_final = torch.Tensor(batch_final).long().to(device)
                else:
                    batch_inputs_final = None

                if samples_sentence is not None:
                    batch_sentence = samples_sentence[step * args.batch_size: (step + 1) * args.batch_size]
                    batch_inputs_sentence = torch.Tensor(batch_sentence).long().to(device)
                else:
                    batch_inputs_sentence = None

                if samples_pos is not None:
                    batch_pos = samples_pos[step * args.batch_size: (step + 1) * args.batch_size]
                    batch_inputs_pos = torch.Tensor(batch_pos).long().to(device)
                else:
                    batch_inputs_pos = None

                if samples_beat is not None:
                    batch_beat = samples_beat[step * args.batch_size: (step + 1) * args.batch_size]
                    batch_inputs_beat = torch.Tensor(batch_beat).long().to(device)
                else:
                    batch_inputs_beat = None

                #  forward pass
                # Notes: Using Transformers, the labels are shifted inside the model,
                #           i.e. you can set labels = input_ids
                outputs = model.forward(input_ids=batch_inputs_token,
                                        sentence_ids=batch_inputs_sentence,
                                        final_ids=batch_inputs_final,
                                        pos_ids=batch_inputs_pos,
                                        beat_ids=batch_inputs_beat,
                                        labels=batch_inputs_token)
                loss, logits = outputs[:2]

                #  get loss
                if multi_gpu:
                    loss = loss.mean()
                    '''
                    running_loss += loss
                    overall_step += 1
                    '''
                if args.gradient_accumulation > 1:
                    loss = loss / args.gradient_accumulation

                #  loss backward
                if args.fp16:
                    with amp.scale_loss(loss, optimizer) as scaled_loss:
                        scaled_loss.backward()
                        torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
                else:
                    loss.backward()
                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)

                #  optimizer step
                if (overall_step + 1) % args.gradient_accumulation == 0:
                    running_loss += loss.item()
                    optimizer.step()
                    optimizer.zero_grad()
                    scheduler.step()

                # log info of training process
                if (overall_step + 1) % args.log_step == 0:
                    loss_log = running_loss * args.gradient_accumulation / (args.log_step / args.gradient_accumulation)
                    tb_writer.add_scalar('loss', loss_log, overall_step)
                    print('now time: {}:{}. Step {} of piece {} of epoch {}, loss {}'.format(datetime.now().hour,
                                                                                             datetime.now().minute,
                                                                                             step + 1, piece_num,
                                                                                             epoch + 1, loss_log))
                    running_loss = 0

                overall_step += 1

            piece_num += 1

        # save model per epoch
        print('saving model for epoch {}'.format(epoch + 1))
        if not os.path.exists(os.path.join(output_dir, 'model_epoch{}'.format(epoch + 1))):
            os.mkdir(os.path.join(output_dir, 'model_epoch{}'.format(epoch + 1)))
        model_to_save = model.module if hasattr(model, 'module') else model
        model_to_save.save_pretrained(os.path.join(output_dir, 'model_epoch{}'.format(epoch + 1)))
        # torch.save(scheduler.state_dict(), output_dir + 'model_epoch{}/scheduler.pt'.format(epoch + 1))
        # torch.save(optimizer.state_dict(), output_dir + 'model_epoch{}/optimizer.pt'.format(epoch + 1))
        print('epoch {} finished'.format(epoch + 1))

        then = datetime.now()
        print('time: {}'.format(then))
        print('time for one epoch: {}'.format(then - now))

    # save final model
    print('training finished')
    if not os.path.exists(os.path.join(output_dir, 'final_model')):
        os.mkdir(os.path.join(output_dir, 'final_model'))
    model_to_save = model.module if hasattr(model, 'module') else model
    model_to_save.save_pretrained(os.path.join(output_dir, 'final_model'))
    # torch.save(scheduler.state_dict(), output_dir + 'final_model/scheduler.pt')
    # torch.save(optimizer.state_dict(), output_dir + 'final_model/optimizer.pt')