VisionLLaMA-B 3x training using the VitDet framework (initialized by 800 epochs using MAE).
cd mmdetection
bash ./tools/dist_train.sh projects/ViTDet/configs/lamadet_mask-rcnn_vit-b-mae_lsj-800-36e.py 8 --amp &> log_txt/lamadet_mask-rcnn_vit-b-mae_lsj-800-36e.log
| name | Pretrained | mAP Box | mAP Mask | Epochs |
|---|---|---|---|---|
| Swin-S | ImageNet sup 300e | 47.6 | 42.8 | 36 |
| Twins-SVT-B | ImageNet sup 300e | 48.0 | 43.0 | 36 |
| ViT-B | MAE 1600e | 51.6 | 45.7 | 100 |
Pyramid VisionLLaMA-B 3x training use the MaskRCNN framework (initialized by 300 epochs on ImageNet 1k).
cd mmdetection
bash ./tools/dist_train.sh configs/twins/mask-rcnn_twins-b-p4-w7_fpn_ms-crop-3x_coco.py 8 --amp &> mask-rcnn_twins-b-p4-w7_fpn_ms-crop-3x_coco.log
Our code is based on mmdetection and Twins. Thanks for their great work. Specifically, we start from this commit id.