Skip to content

Latest commit

 

History

History
82 lines (66 loc) · 4.3 KB

README.md

File metadata and controls

82 lines (66 loc) · 4.3 KB
Raw

Enhancing Diffusion-Based Image Synthesis with Robust Classifier Guidance

Bahjat Kawar, Roy Ganz, and Michael Elad, Technion.

This is the official code repo for the TMLR paper "Enhancing Diffusion-Based Image Synthesis with Robust Classifier Guidance".

Pre-trained models

Our trained model checkpoints and 50k ImageNet images for FID evaluation are available at this Google Drive folder.

Code instructions

ImageNet robust classifier train

TRAIN_FLAGS="--iterations 300000 --anneal_lr True --batch_size 128 --lr 3e-4 --save_interval 10000 --weight_decay 0.05"
CLASSIFIER_FLAGS="--image_size 128 --classifier_attention_resolutions 32,16,8 --classifier_depth 2 --classifier_width 128 --classifier_pool attention --classifier_resblock_updown True --classifier_use_scale_shift_norm True --attack_eps {EPSILON} --attack_steps {ATTACK_STEPS}"
python classifier_rob_train.py --data_dir {IN_PATH} $TRAIN_FLAGS $CLASSIFIER_FLAGS --log_dir {LOG_DIR}

where:

  • {IN_PATH} is the path to ImageNet data
  • {LOG_DIR} is the directory to log outputs into
  • {EPSILON} is the epsilon for the adversarial attack
  • {ATTACK_STEPS} is the number of steps for the adversarial attack

ImageNet sample

python classifier_sample.py --attention_resolutions 32,16,8 --class_cond True --diffusion_steps 1000 --image_size 128 --learn_sigma True --noise_schedule linear --num_channels 256 --num_heads 4 --num_res_blocks 2 --resblock_updown True --use_fp16 True --use_scale_shift_norm True --classifier_scale {SCALE} --classifier_path {CKPT} --model_path 128x128_diffusion.pt --batch_size {BATCH} --num_samples {SAMPLES} --class_idx_begin {CLASS_BEGIN} --timestep_respacing 250 --seed {SEED}

where:

  • {SCALE} is the classifier scale (best results at 1.0)
  • {CKPT} is the classifier checkpoint
  • {BATCH} is the batch size for generation
  • {SAMPLES} is the number of samples to generate
  • {SEED} is the random seed to use
  • {CLASS_BEGIN} is the ImageNet class index to begin generating from (one class per batch). Useful for generating on multiple GPUs.

CIFAR-10 robust classifier train

TRAIN_FLAGS="--iterations 300000 --anneal_lr True --batch_size 128 --lr 3e-4 --save_interval 10000 --weight_decay 0.05"
CLASSIFIER_FLAGS="--image_size 32 --classifier_attention_resolutions 16,8 --classifier_depth 2 --classifier_width 32 --classifier_pool attention --classifier_resblock_updown True --classifier_use_scale_shift_norm True --attack_eps {EPSILON} --attack_steps {ATTACK_STEPS} --attack_type {ATTACK_TYPE}"
python cifar_rob_train.py --data_dir {CIFAR_PATH} $TRAIN_FLAGS $CLASSIFIER_FLAGS --log_dir {LOG_DIR}

where:

  • {CIFAR_PATH} is the path to CIFAR-10 training data
  • {LOG_DIR} is the directory to log outputs into
  • {EPSILON} is the epsilon for the adversarial attack
  • {ATTACK_STEPS} is the number of steps for the adversarial attack
  • {ATTACK_TYPE} is the threat model for the adversarial attack (l2 or linf)

CIFAR-10 sample

MODEL_FLAGS="--image_size 32 --num_channels 128 --num_res_blocks 3 --learn_sigma True --dropout 0.1 --class_cond True"
DIFFUSION_FLAGS="--diffusion_steps 1000 --noise_schedule linear"
TRAIN_FLAGS="--lr 1e-4 --batch_size 128"
python cifar_classifier_sample.py $MODEL_FLAGS $DIFFUSION_FLAGS --classifier_scale {SCALE} --classifier_path {CKPT} --model_path cifar_diffusion_200k.pt --batch_size {BATCH} --num_samples {SAMPLES} --timestep_respacing 250 --seed {SEED}

where:

  • {SCALE} is the classifier scale
  • {CKPT} is the classifier checkpoint
  • {BATCH} is the batch size for generation
  • {SAMPLES} is the number of samples to generate
  • {SEED} is the random seed to use

References and Acknowledgements

This repo is heavily based on:

If you find our work useful, please cite:

@article{kawar2023enhancing,
    title={Enhancing Diffusion-Based Image Synthesis with Robust Classifier Guidance},
    author={Bahjat Kawar and Roy Ganz and Michael Elad},
    journal={Transactions on Machine Learning Research},
    year={2023}
}