add 'DIM online merge config demo'

[CCODING04]

I checked the DIM demo config, and found that there is no "online merge" demo. So I create it for convenience. MMEditing have already realized the code, but not shown us how to use it. Maybe dataset doc file could be detailed.

[wangruohui]

Thanks for your contribution, we will arrange to review these codes. Moreover, would you please provide a code snippet for training and testing with this config file?

[CCODING04]

@wangruohui Unit test or just a simple use case?

[wangruohui]

@CCODING04

A simple case is OK. Have you trained with this configuration? Could you please upload the training log and the file of pre-trained weight?

[CCODING04]

@wangruohui I'm training it. However, it seems like merge online is not a good choice cuz the curve of loss and validation metrics got violate vibration.

[CCODING04]

@wangruohui Actually, I did previously merge and got a good result. Is there any paper or articles talk about which and when to take composition of data is better?

[CCODING04]

@wangruohui I compared the online and offine results and plot it.

[zengyh1900]

Please check this. @Z-Fran

[Z-Fran]

Hi @CCODING04 , we plan to merge your code to MMEditing 2.0. So you need to do two jobs as follows:

1. modify file name and directory to configs/dim/dim_stage1-v16_1xb1-1000k_comp1k_online-merge.py
2. If there is no difference between dim_stage1-v16_1xb1-1000k_comp1k.py and your code except train pipeline, we suggest you can inherit the same codes, like _base_ = ['./dim_stage1-v16_1xb1-1000k_comp1k.py'] in dim_stage2-v16-pln_1xb1-1000k_comp1k.py. Then you can delete the same parts.

And you can give us the authority to commit codes to your repository to help you. Thanks again.