Train our method on ImageNet and then perform linear probe
python main.py --data /path/to/imagenet/ --batch-size 2048 --name our --dist-port 8873 --mode neuralef --learning-rate 2.4 --alpha 0.0025 --proj_dim 8192 8192 {--no_stop_grad} --epochs {100, 200, 400}; python linear_probe.py --data /path/to/imagenet/ --pretrained logs/our/final.pth
When using --proj_dim 4096 4096, set --alpha 0.005; when using --proj_dim 2048 2048, set --alpha 0.01. If setting a different batch size from 2048, remember to linearly scale the learning rate.
Evaluate on COCO
cd detection;
python convert-pretrain-to-detectron2.py ../logs/our/final.pth output.pkl
python train_net.py --config-file configs/coco_R_50_C4_2x_moco.yaml --num-gpus 8 MODEL.WEIGHTS ./output.pkl SOLVER.BASE_LR 0.05
Perform image retrieval
python test_retrieval.py --name neuralef-nolbn-o4096 --mode neuralef --proj_dim 4096 4096 --resume auto
Training:
cd graph;
python main_products.py --root /path/to/ogbn --device 3 --model res_mlp --hidden_channels 2048 --num_layers 12 --proj_dim 8192 8192 --alpha 0.3 --lr 0.3 --weight_decay 0 --batch_size 16384 --epochs 20 {--no_stop_grad} --output_dir logs/products_al.3_lr.3_w2048{_nosg};
Linear probe:
python main_products.py --root /path/to/ogbn --device 3 --model res_mlp --hidden_channels 2048 --num_layers 12 --proj_dim 8192 8192 --alpha 0.3 --lr 0.3 --weight_decay 0 --batch_size 16384 --epochs 20 {--no_stop_grad} --output_dir logs/products_al.3_lr.3_w2048{_nosg} --ft_only --ft_mode freeze --ft_lr 0.01 --ft_weight_decay 1e-3 --ft_smoothing 0.1 --runs 10