search

hustvl / YOLOS

[NeurIPS 2021] You Only Look at One Sequence
https://arxiv.org/abs/2106.00666
MIT License
840 stars 122 forks source link

Implmenetation queries #13

Open abhigoku10 opened 3 years ago

abhigoku10 commented 3 years ago

❔Question

Hi thanks for opensourcing the code base this gives steps to learn transformers, i am having few queries

  1. The dataset is loaded from coco.py using which function since "ConvertCocoPolysToMask" is not called inherently anywhere
  2. Your load the data training for each epoch using train_one_epoch() for the whole dataset which internally performs losses and then the out for that is performed with evaluation this is performed for each 300 epoch so what's the idea behind this training
  3. Does yolos provide panoptic segmentation also?can we get pretrained model on this

Thanks in advance

Additional context

Yuxin-CV commented 3 years ago

Our codebase is inherit from DETR and ViT of timm. So,

  1. ConvertCocoPolysToMask is inherit from DETR, we didn't modify that.
  2. This is similar to the original DETR implantation, you can also refer to other DETR implantations, e.g., mmdet (https://github.com/open-mmlab/mmdetection/blob/master/configs/detr/README.md).
  3. We didn't try panoptic seg, but we believe it is feasible for YOLOS to perform panoptic seg following the configuration of DETR.