Active Generation for Image Classification

Official implementation for paper "Active Generation for Image Classification", ECCV 2024.

ImageNet / CIFAR Classification

Please see classification/README.md for environment setup.

Training command

cd classification
sh tools/dist_train.sh 8 ${CONFIG} ${MODEL} --experiment ${EXP_NAME} --gen_images

For example, for ViT-S/16, the MODEL should be timm_vit_small_patch16_224, CONFIG should be configs/strategies/ActGen/imagenet/vit_s_16.yaml.

Tips for understanding the code

• The core code for generating images is in classification/lib/gen/guided_gen.py
• The code for integrating ActGen into training code is in classification/tools/train.py: Please find keyword args.gen_images to see how we implement the training.