Algorithm 1: Model training algorithm.

1: Read Data $dt$

2: Pre-train the model

3: for model in CNN, RNN, …, Transformer do

4:     for i in 1, 2, 3, 4, 5 do

5:             1. Batch and Shuffle $dt$ by Mini-batch size to Generate $D$

6:             2. Randomly divide $D$ as the training set $D_t$, validation set $D_v$, test set $D_{tt}$

7:             Initialize $\omega$ and $b$ randomly by the Gaussian distribution

8:             Reading parameter configuration information

9:             for epoch in 1 to 20 do

10:                 3. $va=matmul\left(x,\omega \right)$ # calculate the value before activation

11:                 4. $ypre=softmax\left(va\right)$ # the value after activation

12:                 5. $CF\left(\theta \right)$ = Equation (4) # compute loss

13:                 6. $grd=sgdwithmomentum\left(CF\left(\theta \right),\theta \right)$

14:                 7. ${\eta }_{\theta }\left(x\right)$ = Equation (5) # fit the regression curve

15:                 8. ${\theta }_{j+1}$ = Equation (7) # update model parameters

16:           End

17:     End

18: End

19: Fine-Tune.