Giters

jozhang97 / DETA

Detection Transformers with Assignment

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Swin-L config without Objects365 pretraining and Objects365 pretraining setting

jay-lcchen opened this issue · comments

commented

Hello,

Thank you for sharing the great work! We really find it useful.

We have one question regarding how to train your model using Swin-L, without Objects365 pretraining (i.e, only ImageNet-21K pretraining). Do you mind sharing the config or any settings for us to try?

Additionally, the script to pretrain the Swin-L on Objects365 (deta_swin_pre.sh) is missing. Is it also possible that you can share that script and Objects365's setting (e.g., all images are used? and so on?).

Thanks,

commented

Hi Jay,

I'm glad you like the work!

  1. Swin-L without Objects365 pretraining: we have not tried this, but I think you can just replace the Objects365 checkpoint with the IN21K checkpoint from https://github.com/microsoft/Swin-Transformer
  2. the script to pretrain the Swin-L on Objects365: We trained on Objects365 training set (no merge with val) on 24epoch schedule (details in the paper). The only change to the code are increasing num_classes and training with higher resolution images. Happy to answer any other questions I can.
commented

Awesome. Thank you, Jeffrey, for the quick response and for the clarification.
It is clear to me and we will run the experiments as suggested and see how it goes.

Thanks,