search

ZZZHANG-jx / DocRes

[CVPR 2024] DocRes: A Generalist Model Toward Unifying Document Image Restoration Tasks
MIT License
287 stars 26 forks source link

Question for training the model #11

Closed LonglongaaaGo closed 2 months ago

LonglongaaaGo commented 2 months ago

Hi @ZZZHANG-jx ,

Thank you so much for your awesome paper and project. I have a question about the model training. If we would like to train another new dataset. Should we train the two models separately, or just training one model is ok, as mentioned in the paper? For example, for the deblur task? Thanks!

ZZZHANG-jx commented 2 months ago

For different tasks/datasets, we only train one model, not multiple models. If you want to add more tasks/datasets, you just need to train these new data together with the original data. Of course, you need to design an appropriate DTSPrompt to distinguish the tasks. One simple way is to specify a constant (fixed prompt) as the prompt, as we demonstrated in the ablation study in Table 2 of the paper.

LonglongaaaGo commented 2 months ago

Awesome！ Thanks!