how could i use my dataset to train Sa2VA

[LORDPQK]

i have a png image dataset . Did anyone tried adding their own data for training?
dataset_root/
├── images/
│ ├── 0001.jpg
│ ├── 0002.png
│ └── ...
├── masks/
│ ├── 0001.png （0=background，1=first object，2=second object...）
│ ├── 0002.png
│ └── ...

[mrdbourke]

+1 to this, a fine-tuning guide would be epic!

[lxtGH]

Thanks for your attention. We will update this feature in the next two weeks. Please stay tuned.

[bill-hx-liu]

+1 to this,thanks

[LORDPQK]

@lxtGH Hi,

I wonder if the fine-tuing script of Sa2VA-8B is already released?

thank you!

[LemonWei111]

感谢您的关注。我们将在接下来的两周内更新此功能。敬请期待。

3 weeks have passed.

[HarborYuan]

Hi @LORDPQK @LemonWei111 ,

Thanks for your interest in our work. Your mentioned dataset is like instance segmentation. Sa2VA currently focuses on tasks such as chat and referring segmentation of images and videos. The dataset you mentioned is more like a pure visual task.

For the training guide, I will add it now. Please stay tuned.

[abhigoku10]

@HarborYuan Has the training script and steps updated ???