torch==1.9.0+cu111
clip==1.0

python main.py --sample_data cifar10

python main.py --sample_data cifar10 --load_vcm --vcm_path finetuned_model_path

python main.py --sample_data cifar10 --test_datas cifar101 cifar102

Currently, our experiments are done on GPUs based on high-precision floating-point computation, thus the cost of inference have not yet been optimized. Make sure you have 60G of memory available when you set batch_size=50. In future work, when this pipeline is succuessfully deployed on low-precion hardware platforms, memory and runtime will be effectively saved.