RUN.md

Running

For training and evaluation, please, follow the below-mentioned respective instructions.

NOTE 1: In case you have multiple CUDA versions installed, please, make sure to initialise the appropriate system CUDA version before running any command.

# <xx.x> - CUDA version number
module load cuda-xx.x 

NOTE 2: Make sure that you are in the root repo's directory '.../sm-vit/':

cd <your_path>/sm-vit

NOTE 3: Make sure that the 'sm_vit' conda environment is activated (see INSTALL.md ):

conda activate sm_vit

---

Stanford Dogs

Train + Test:

python3 -W ignore -m torch.distributed.launch --nproc_per_node 1 train.py --name dogs --dataset dogs --img_size 400 --train_batch_size 24 --eval_batch_size 8 --learning_rate 0.003 --num_steps 20000 --sm_vit --coeff_max 0.3 --fp16 --low_memory --eval_every 100 --data_root '<your_dataset_path>'

Test only:

soon

---

CUB-200-2011

Train + Test:

python3 -W ignore -m torch.distributed.launch --nproc_per_node 1 train.py --name cub --dataset CUB --img_size 400 --train_batch_size 24 --eval_batch_size 8 --learning_rate 0.03 --num_steps 40000 --sm_vit --coeff_max 0.25 --fp16 --low_memory --data_root '<your_dataset_path>'

Test only:

soon

---

NABirds

Train + Test:

python3 -W ignore -m torch.distributed.launch --nproc_per_node 1 train.py --name nabirds --dataset nabirds --img_size 448 --train_batch_size 16 --eval_batch_size 8 --learning_rate 0.03 --num_steps 40000 --sm_vit --coeff_max 0.25 --fp16 --low_memory --data_root '<your_dataset_path>'

Test only:

soon