阿里中文预训练模型泛化能力挑战赛 Task1

背景

赛题以自然语言处理为背景，要求选手通过算法实现泛化能力强的中文预训练模型。通过这道赛题可以引导大家更好地理解预训练模型的运作机制，探索深层次的模型构建和模型训练，而不仅仅是针对特定任务进行简单微调。

Baseline报错整理

首先来看import这块的报错

import sys
import os
import tensorflow as tf
from easytransfer import base_model, Config, FLAGS
from easytransfer import layers
from easytransfer import model_zoo
from easytransfer import preprocessors
from easytransfer.datasets import TFRecordReader
from easytransfer.losses import softmax_cross_entropy
from sklearn.metrics import classification_report
import numpy as np

首先注意这里需要额外导入的package包括以下两个
pip install tensorflow-gpu --user ##这样不容易爆权限错误
pip install easytransfer
输出结果如下:

这个是因为tensorflow2.1已经没有tf.logging了
逛坛子得知解决方法如下：
将tf.logging替换成tf.compat.v1.logging
但还是报错 所以这里我还是老老实实根据环境配置的tips

• tensorflow-gpu 1.12.3
• easytransfer 0.1.2

实际情况是没有得到解决

然后在群里看到水哥说baseline有用tf2的tf1.4的有用pytorch的
时间有限这里没有cover

class Application(base_model):
    def __init__(self, **kwargs):
        super(Application, self).__init__(**kwargs)
        self.user_defined_config = kwargs["user_defined_config"]

    def build_logits(self, features, mode=None):

        preprocessor = preprocessors.get_preprocessor(self.pretrain_model_name_or_path,
                                                      user_defined_config=self.user_defined_config)

        model = model_zoo.get_pretrained_model(self.pretrain_model_name_or_path)

        global_step = tf.train.get_or_create_global_step()

        tnews_dense = layers.Dense(15,
                     kernel_initializer=layers.get_initializer(0.02),
                     name='tnews_dense')

        ocemotion_dense = layers.Dense(7,
                             kernel_initializer=layers.get_initializer(0.02),
                             name='ocemotion_dense')

        ocnli_dense = layers.Dense(3,
                             kernel_initializer=layers.get_initializer(0.02),
                             name='ocnli_dense')

        input_ids, input_mask, segment_ids, label_ids = preprocessor(features)

        outputs =<