search

Megvii-BaseDetection / GFSD

This project provides an implementation for "Generalized Few-Shot Object Detection without Forgetting" (CVPR2021) on PyTorch.
Apache License 2.0
46 stars 4 forks source link

Integrating the training of the base model #4

Open 2006pmach opened 3 years ago

2006pmach commented 3 years ago

Hi,

I noticed that you use the pretrained base-class model form another repo. In the end, I guess it would also be possible to train it using CVPODS by adding another exp folder and the required files to the playground folder. Is there a simple way to achieve this? Would be nice to have everything at one place. And would be good example to start developing new Few-Shot Methods using your framework and dataset splits etc. What would I have to add? Or could you add this? Thx in advance.

zb1439 commented 3 years ago

Thanks for the advice. This can be done by adjusting the training schedule, resnet depth, and input image scales to be identical to fsdet's repo. You can modify config.py in https://github.com/Megvii-BaseDetection/cvpods/tree/master/playground/detection/coco/rcnn/faster_rcnn.res50.fpn.coco.multiscale.1x to make sure it is the same as the config in fsdet (https://github.com/ucbdrive/few-shot-object-detection/blob/master/configs/COCO-detection/faster_rcnn_R_101_FPN_base.yaml and https://github.com/ucbdrive/few-shot-object-detection/blob/master/configs/Base-RCNN-FPN.yaml). I might try adding this experiment folder if I am sure the configuration is identical and reproducible.

amajee11us commented 3 years ago

@zb1439 can the base model trained using the FsDet framework be directly used during the finetuning process using GFSD?

zb1439 commented 3 years ago

@zb1439 can the base model trained using the FsDet framework be directly used during the finetuning process using GFSD?

Yes. They should be using the same name for weights. And this open-source implementation is intended to use their pretrained model as stated in readme.

amajee11us commented 3 years ago

Thanks for the update @zb1439 !! FsDet performs a "surgery" operation on the base trained checkpoint through this script - https://github.com/ucbdrive/few-shot-object-detection/blob/master/tools/ckpt_surgery.py . For GFSD, should the input model be the one after this surgery operation or the one before the surgery operation?

zb1439 commented 3 years ago

Thanks for the update @zb1439 !! FsDet performs a "surgery" operation on the base trained checkpoint through this script - https://github.com/ucbdrive/few-shot-object-detection/blob/master/tools/ckpt_surgery.py . For GFSD, should the input model be the one after this surgery operation or the one before the surgery operation?

The input pretrained model does not need to be trimmed. Once you have that pretrained weight, you are all set to fine-tune it using this repo.