| torchrun --nproc_per_node 4 -m training.main --batch-size=16 --lr=1e-5 --wd=0.1 --epochs=6 --workers=4 \ |
| --model EVA02-CLIP-B-16 --pretrained eva --warmup 1000 --zeroshot-frequency 1 \ |
| --method-type region_clip --dataset-type region_clip \ |
| --test-type coco_panoptic --train-data data/coco/coco_pseudo_4764.json \ |
| --val-data data/coco/annotations/panoptic_val2017.json \ |
| --train-embed-path metadata/coco_pseudo_4764_clip_hand_craft_EVACLIP_ViTB16.npy \ |
| --embed-path metadata/coco_panoptic_clip_hand_craft_EVACLIP_ViTB16.npy --train-image-root data/coco/train2017 \ |
| --val-image-root data/coco/val2017 --cache-dir checkpoints/EVA02_CLIP_B_psz16_s8B.pt --log-every-n-steps 50 \ |
| --lock-image --save-frequency 1 --lock-image-unlocked-groups 12 --extract-type="v2" \ |
| --downsample-factor 16 --det-image-size 512 \ |
| --alpha 0.7 |
