How to prevent over-fitting?

[Hegelim]

Background

• I am fine-tuning on an existing pre-trained recognition model - arabic_PP-OCRv3_rec here.
• The size of my training data is 1.2 million
• I modified a few places here and there in the config file. For privacy concerns, here are a few places that I changed:

Because of the volume of data, there are 18750 iters of each training epoch, so I set

eval_batch_step: [0, 18750]

I changed the optimizer to use piecewise learning strategy

Optimizer:
  name: Adam
  beta1: 0.9
  beta2: 0.999
  lr:
    name: Piecewise
    decay_epochs:
    - 5
    values:
    - 5.0e-05
    - 1.0e-05
  regularizer:
    name: L2
    factor: 3.0e-05

For data augmentation, I deleted RecConAug completely because just combining 2 same images together sounds weird to me. I also used the following data augmentation

    - RecAug:
        noise_prob: 0.4
        jitter_prob: 0.4
        blur_prob: 0.4
        hsv_aug_prob: 0.4

In terms of batch size, I set 64 for both train and eval

For everything else, I left it as is.

Issue

I noticed the training accuracy went up really quickly. Just before the 1st epoch is finished, the acc is already near 97%.

[2023/09/22 23:54:25] ppocr INFO: epoch: [1/100], global_step: 17800, lr: 0.000050, acc: 0.968750, norm_edit_dis: 0.991450, CTCLoss: 0.157956, SARLoss: 0.285718, loss: 0.451693, avg_reader_cost: 0.00020 s, avg_batch_cost: 0.43793 s, avg_samples: 64.0, ips: 146.14070 samples/s, eta: 9 days, 12:45:49

Question

• why does the accuracy go up so fast? I have not seen such behavior in previous experiements.
• Looks to me this is over-fitting. What are some other options to prevent over-fitting? I tried to use a lot of data already.

System

• 系统环境/System Environment：ubuntu 20.04
• 版本号/Version：Paddle： PaddleOCR： paddleocr 2.6.1.0 paddlepaddle-gpu 2.4.2