README.md

Conditional Contrastive Learning objectives with Kernel Expressions (CCL-K)

Official Pytorch Implementation of CCL-K

We propose CCL-K for conditional contrastive learning frameworks. Three settings are considered

• WeaklySup-CCLK
• Fair-CCLK
• HardNeg-CCLK

Dataset locations

The dataset file locations should be specified in a JSON file of the following form

{
    "cifar10": "data/Cifar/datafolder",
    "imagenet": "data/imagenet100/datafolder",
    "cub": "data/CUB/datafolder",
    "utzappos": "data/UTzappos-sub/datafolder",
    "colorMNIST": "data/colormnist/datafolder"
}

Running with WeaklySup-CCLK

• Condition on Discrete Variables attribute_data_root_folders.json should be configured to be the root folder of the attribute datasets and attribute_pd_paths.json contains the csv files that pointing out the relative location (w.r.t. root) of each data, and the discrete attributes associated with it. We include processed .csv file that can pointed to directly in attribute_pd_paths.json.

  attribute_data_root_folders.json:
  {
      "cub": "data/CUB/images",
      "utzappos": "data/UTzappos-sub/ut-zap50k-images-square"
  }
  

  attribute_pd_paths.json:
  {
      "cub": "data/CUB/meta_data_bin_train.csv",
      "utzappos": "data/UTzappos-sub/meta_data_bin_train.csv"
  }
  

  Then Run WeaklySup-CCLK with discrete variables:

  python3 cclk_weaklysup_attributes.py --cosine-anneal --filename output.pth --base-lr 1.5 --test-freq 50 \
  --lambda_ 0.01 \
  --num-epochs 1000 \
  --inverse_device gpu \
  --batch-size 152 \
  --distance_mode cosine \
  --customized_name WeaklySup-Discrete \
  --dataset utzappos \
  --condition_mode cl-infonce \
  

• Condition on Continuous Features First make sure you have the neccessary continuous feature. We extract features from OpenAI CLIP network to serve as a continous variable to condition upon. Store them in a pickle file and point them to condition_continuous_feat.json

  {
      "imagenet100-clip": "data/imagenet100/features.pkl"
  }
  

  Then run the conditional contrastive learning with WeakslySup-CCLK:

  python3 cclk_weaklysup_continuous_features.py --cosine-anneal --filename output.pth --base-lr 1.5 --test-freq 20 \
  --lambda_ 0.01 \
  --num-epochs 200 \
  --distance_mode cosine \
  --inverse_device gpu \
  --batch-size 128 \
  --customized_name WeaklySup-Continuous \
  --dataset imagenet100-clip \
  --condition_mode cl-infonce \
  

Running with Fair-CCLK

python3 cclk_fair.py --cosine-anneal --filename output.pth --base-lr 1e-3 --test-freq 1 \
--lambda_ 0.01 \
--num-epochs 1000 \
--inverse_device gpu \
--batch-size 256 \
--temperature 0.1 \
--distance_mode RBF \
--condition_mode weac-infonce \
--customized_name RBFK-Fair \
--no_color_distor \
--dataset colorMNIST \
--arch LeNet \

Running with HardNeg-CCLK

python3 cclk_hardnegatives.py --cosine-anneal --filename output.pth --base-lr 1.5 --test-freq 5 \
--lambda_ 100 \
--num-epochs 200 \
--inverse_device gpu \
--save_freq 10 \
--distance_mode RBF \
--batch-size 512 \
--num-workers 15 \
--dataset imagenet100 \
--customized_name RBF-HardNegtive \
--temp_z 500 \