docker build -t 12000sgd-vlm-base -f base.dockerfile .
docker build -t 12000sgd-vlm . && \
docker tag 12000sgd-vlm asia-southeast1-docker.pkg.dev/dsta-angelhack/repository-12000sgdplushie/12000sgd-vlm:latest && \
docker push asia-southeast1-docker.pkg.dev/dsta-angelhack/repository-12000sgdplushie/12000sgd-vlm:latest && \
gcloud ai models upload --region asia-southeast1 --display-name '12000sgd-vlm' --container-image-uri asia-southeast1-docker.pkg.dev/dsta-angelhack/repository-12000sgdplushie/12000sgd-vlm:latest --container-health-route /health --container-predict-route /identify --container-ports 5004 --version-aliases default
docker run -p 5004:5004 --gpus all 12000sgd-vlm
Model |
Step |
LR schedule |
Train aug |
Test aug |
Split tokens_positive by words? |
AP50 |
Speed |
MM Grounding DINO |
8000 |
Linear warmup 1k steps from 1e-5 to 1e-4, back to 1e-5 at 16800 steps |
Default |
Default |
Yes |
0.507 |
0.5533085946296297 |
MM Grounding DINO |
7000 |
Linear warmup 1k steps from 1e-5 to 1e-4, back to 1e-5 at 5000 steps |
Default |
Default |
No |
0.488 |
0.5482950594444445 |
MM Grounding DINO |
5000 |
Linear warmup 1k steps from 1e-5 to 1e-4, back to 1e-5 at 5000 steps |
Default |
Default |
No |
0.479 |
0.5338830794444445 |
MM Grounding DINO |
9000 |
Linear warmup 1k steps from 1e-5 to 1e-4, back to 1e-5 at 5000 steps |
Default |
Default |
No |
0.487 |
0.5549248251851852 |
MM Grounding DINO |
10000 |
Linear warmup 1k steps from 3.95e-5 to 3.95e-4, back to 3.95e-5 at 5000 steps |
Brightness, contrast, Gaussian noise |
Default |
No |
0.509 |
0.4668306538888889 |
MM Grounding DINO |
10000 |
Same as prev, but trained for additional 2000 steps with LR = 3.95e-6 |
Brightness, contrast, Gaussian noise |
Default |
No |
0.504 |
0.564885440925926 |