How to train your own dataset

[biyuefeng]

Hello, I currently have my own dataset labeled with labelme, which has been converted to Coco format. The list is as follows:

${DATASET_ROOT} # 数据集根目录，例如：/home/username/data/NWPU
├── annotations
│ ├── train.json
│ ├── val.json
│ └── test.json
└── images
├── train
├── val
└── test

Excuse me, can this be trained directly?

[EricZavier]

I also want to train my own dataset, but it seems the situation is not very promising.If you find a solution, please share it. Thank you very much.

[jwyang]

Hi, there are roughly two steps:

1. register your dataset following the sample code in: https://github.com/UX-Decoder/Segment-Everything-Everywhere-All-At-Once/tree/v1.0/datasets/registration
2. create a dataset mapper following the sample code in: https://github.com/UX-Decoder/Segment-Everything-Everywhere-All-At-Once/tree/v1.0/datasets/dataset_mappers

you can refer to the code for coco or other datasets used in our training.

[EricZavier]

"Okay, thank you very much for your patient explanation. I'll give it a try now."

[EricZavier]

"I have another question. How can I construct JSON files like those in the image based on my annotated dataset?"

Hi, there are roughly two steps:

1. register your dataset following the sample code in: https://github.com/UX-Decoder/Segment-Everything-Everywhere-All-At-Once/tree/v1.0/datasets/registration

2. create a dataset mapper following the sample code in: https://github.com/UX-Decoder/Segment-Everything-Everywhere-All-At-Once/tree/v1.0/datasets/dataset_mappers

you can refer to the code for coco or other datasets used in our training.

[MaureenZOU]

To construct JSON file please ask GPT4. Hhh