Decoders-Chinese-TF2.0/train_gpt2.py at master · Morizeyao/Decoders-Chinese-TF2.0 · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
import transformers
import tensorflow as tf
import os
import numpy as np
import argparse
from datetime import datetime


def _int64_feature(value):
    """Wrapper for inserting int64 features into Example proto."""
    if not isinstance(value, list):
        value = [value]
    return tf.train.Feature(int64_list=tf.train.Int64List(value=value))


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

    parser.add_argument('--n_ctx', default=512, type=int, required=False, help='文本长度')
    parser.add_argument('--model_config', default='configs/gpt2/model_config_small.json', type=str, required=False,
                        help='选择模型参数')
    parser.add_argument('--pretrained_model', default='', type=str, required=False, help='模型训练起点路径')
    parser.add_argument('--batch_size', default=2, type=int, required=False, help='训练batch size')
    parser.add_argument('--tfrecord_path', default='data/tokenized/tokenized.tfrecord', type=str, required=False,
                        help='预处理完的数据地址')
    parser.add_argument('--lr', default=2e-4, type=float, required=False, help='学习率')
    parser.add_argument('--total_steps', default=10, type=int, required=False, help='steps')
    parser.add_argument('--output_dir', default='model/', type=str, required=False, help='模型输出路径')
    parser.add_argument('--log_step', default=1, type=int, required=False, help='多久报告一次')
    parser.add_argument('--writer_dir', default='tensorboard_summary/', type=str, required=False, help='Tensorboard路径')
    args = parser.parse_args()
    print('args:\n' + args.__repr__())

    summary_writer = tf.summary.create_file_writer(args.writer_dir)
    print('getting dataset')
    feature_description = {
        'ids': tf.io.FixedLenFeature([args.n_ctx], tf.int64)}

    def _parse_function(example_proto):
        return tf.io.parse_single_example(example_proto, feature_description)

    ds = tf.data.TFRecordDataset(args.tfrecord_path)

    train_dataset = ds.map(_parse_function)
    print('getting dataset done')
    # get dataset done
    print('total steps = {}'.format(args.total_steps))

    train_dataset = train_dataset.batch(args.batch_size, drop_remainder=True)

    strategy = tf.distribute.MirroredStrategy()
    print('Number of devices: {}'.format(strategy.num_replicas_in_sync))
    print('starting training')
    with strategy.scope():
        model_config = transformers.configuration_gpt2.GPT2Config.from_json_file(args.model_config)
        if not args.pretrained_model:
            model = transformers.modeling_tf_gpt2.TFGPT2LMHeadModel(config=model_config)
        else:
            model = transformers.modeling_tf_gpt2.TFGPT2LMHeadModel.from_pretrained(args.pretrained_model)
        dummy = tf.constant(np.ones((args.batch_size, args.n_ctx)), dtype=tf.int32)
        # print(dummy.shape)
        _ = model([dummy])

        model.summary()
        # print(model.trainable_variables)

        scheduler = tf.keras.optimizers.schedules.PolynomialDecay(initial_learning_rate=args.lr,
                                                                  decay_steps=args.total_steps,
                                                                  end_learning_rate=args.lr * 0.01)
        optimizer = tf.keras.optimizers.Adam(learning_rate=scheduler)
        loss_function = tf.keras.losses.SparseCategoricalCrossentropy(reduction=tf.keras.losses.Reduction.NONE)

        def compute_loss(labels, predictions):
            per_example_loss = loss_function(labels, predictions)
            # tf.print(per_example_loss)
            return tf.nn.compute_average_loss(per_example_loss, global_batch_size=args.batch_size)


        running_loss = 0
        step = 0
        epoch = 0


        print('saving initial model')
        if not os.path.exists(args.output_dir + 'model_temp_step{}'.format(step)):
            os.makedirs(args.output_dir + 'model_temp_step{}'.format(step))
        model.save_pretrained(args.output_dir + 'model_temp_step{}'.format(step))


        print('epoch {}'.format(epoch + 1))
        now = datetime.now()
        print('time: {}'.format(now))
        while True:
            for batch_idx, batch_inputs in enumerate(iter(train_dataset)):
                batch_inputs = batch_inputs['ids']


                def train_step(batch_inputs):
                    inputs, labels = batch_inputs, batch_inputs

                    with tf.GradientTape() as tape:
                        outputs= model(batch_inputs, training=True)[0]
                        loss = compute_loss(labels[:, 1:], outputs[:, :-1, :])
                    gradients = tape.gradient(loss, model.trainable_variables)
                    tvars = list({id(v): v for v in model.trainable_variables}.values())
                    optimizer.apply_gradients(zip(gradients, tvars))
                    return loss

                @tf.function
                def get_loss(work, batch_inputs):
                    per_replica_losses = strategy.experimental_run_v2(work, args=[batch_inputs])
                    loss = strategy.reduce(tf.distribute.ReduceOp.SUM, per_replica_losses,
                                           axis=None)
                    return loss

                loss = get_loss(train_step, batch_inputs)
                running_loss += loss
                if (step + 1) % args.log_step == 0:
                    with summary_writer.as_default():
                        tf.summary.scalar("loss", running_loss / args.log_step, step=step)
                        summary_writer.flush()
                    print('now time: {}:{}. Step {} of epoch {}, unscaled loss {}'.format(
                        datetime.now().hour,
                        datetime.now().minute,
                        batch_idx + 1,
                        epoch + 1,
                        running_loss / args.log_step))
                    running_loss = 0
                step += 1
                if step > args.total_steps:
                    break
                if (step + 1) % 10000 == 0:
                    print('saving model temp')
                    if not os.path.exists(args.output_dir + 'model_temp_step{}'.format(step)):
                        os.makedirs(args.output_dir + 'model_temp_step{}'.format(step))
                    model.save_pretrained(args.output_dir + 'model_temp_step{}'.format(step))
            epoch += 1

            print('saving model for epoch {}'.format(epoch + 1))
            if not os.path.exists(args.output_dir + 'model_epoch{}'.format(epoch + 1)):
                os.makedirs(args.output_dir + 'model_epoch{}'.format(epoch + 1))
            model.save_pretrained(args.output_dir + 'model_epoch{}'.format(epoch + 1))
            print('epoch {} finished'.format(epoch + 1))

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

        print('training finished')
        if not os.path.exists(args.output_dir + 'final_model'):
            os.makedirs(args.output_dir + 'final_model')
        model.save_pretrained(args.output_dir + 'final_model')


if __name__ == '__main__':
    main()