Not able to reproduce the results

[OmkarThawakar]

Congrats for the awesome work.

I am trying to reproduce the results for resnet-50 backbone.
I tried following ,

1. Train Seqformer on coco dataset (with num_frames=1) for 24 epochs
2. Train Seqformer on coco+ytvis and ytvis using coco pretrained weights

Still I am not able to generate the desired numbers.

Can you please help me out with this ?

Thanks,

[wjf5203]

Hi, thanks for your attention.
We pretrain Seqformer for 50 epochs on coco in the first step, following DeformableDETR.
It is recommended to use the coco pre-trained models we provide in the model zoo.

[chrisjuniorli]

Hi, I'm also working on reproducing the ResNet-50 based result.
I pretrained SeqFormer for 50 epochs on COCO following Deformable DETR, here is the training config I used:

python -u main.py
--dataset_file coco
--epochs 50
--lr 2e-4
--lr_drop 40
--batch_size 2
--num_workers 2
--coco_path ../data/COCO
--ytvis_path ../data/Youtube-VOS/2019
--num_queries 300
--enc_layers 6
--dec_layers 6
--num_frames 1
--with_box_refine
--masks
--rel_coord
--backbone resnet50
--output_dir outputs/seqformer_coco \

Then I fine-tuned it on Youtube-VIS following r50_seqformer_ablation.sh and got a 42.1 mAP, which is about ~3% mAP lower than the 45.1 mAP reported on the ablation model zoo, could you help point out the difference between my pretraining setting to yours?

@wjf5203