train
python -m torch.distributed.launch --nproc_per_node 8 --master_port 12345 main.py --cfg ./configs/MetaFG.yaml --batch-size 32 --tag OUTPUT_TAG --lr 5e-5 --min-lr 5e-7 --warmup-lr 5e-8 --epochs 300 --warmup-epochs 20 --dataset imagenet --pretrain --opts DATA.IMG_SIZE 384 TRAIN.AUTO_RESUME False --output output --amp-opt-level O1 --root --nb-classes 1572
test
python -m torch.distributed.launch --nproc_per_node 1 --master_port 12345 main.py --eval --cfg ./configs/MetaFG.yaml --batch-size 10 --tag OUTPUT_TAG --dataset snakeclef2022test --resume $MODEL_PATH --opts DATA.IMG_SIZE 384 TRAIN.AUTO_RESUME False
post process and ensamble
After runing test, we will get result_snakeclef2022test.tc, use post_process.py which indicate the final output of a single model, we can ensamble the model outputs by runing fuse_logits.py
| tesm | score |
|---|---|
| ARM(ours) | 0.89436 |
| base | 0.89101 |
| GG | 0.85409 |
Our code are partly based on metaformer and moco