For a better user experience, we recommend you to directly use github repository github link
CUB link
MIT link
DTD link
please unzip the downloaded datasets and then move them to data
Step 1. Fine-tune the fc layer only.
CUDA_VISIBLE_DEVICES=0,1 python ./src/bilinear_cnn_fc_cub.py --base_lr 1 --batch_size 64 --epochs 81 --weight_decay 1e-6
Step 2. Fine-tune all layers.
CUDA_VISIBLE_DEVICES=0,1 python ./src/bilinear_cnn_all_cub.py --base_lr 1e-2 --batch_size 32 --epochs 50 --weight_decay 1e-3 --model "vgg_16_epoch_81.pth"
You can also run the above three experiments by one script in background:
CUB:
nohup bash script_cub.sh &
MIT:
nohup bash script_mit.sh &
DTD:
nohup bash script_dtd.sh &