baysharpen-debias

Code relevant for work present in thesis "Visual bias mitigation driven by Bayesian uncertainties".

Setup

This code assumes additional directories in default locations:

• tb_logs/: Tensorboard log directory
• data/: datasets

Set the path directories in local_config.py.

All default configurations, paths, and parameters can be changed in global_config.py.

Datasets

Links to download datasets:

• Biased MNIST
• COCO-on-Places
• Biased Action Recognition

Training

The repository includes checkpoints for baseline Bayesian ResNet18 models trained on the three datasets. Run scripts sharpen-bmnist.sh, sharpen-coco.sh, and sharpen-bar.sh to obtain sharpening results reported in the paper.

Alternately, to modify parameters, you can run the train.py Python script directly on the command line:

usage: train.py [-h] [--task_id TASK_ID] [--config_name CONFIG_NAME]
                [--dataset_code DATASET_CODE] [--lit_code LIT_CODE]
                [--model_code MODEL_CODE] [--model_desc MODEL_DESC]
                [--ckpt CKPT] [--optim_code OPTIM_CODE]
                [--max_epochs MAX_EPOCHS] [--batch_size BATCH_SIZE]
                [--momentum MOMENTUM] [--lr LR] [--result_file RESULT_FILE]
                [--cycle_length CYCLE_LENGTH] [--num_gpus NUM_GPUS]
                [--dev_run DEV_RUN] [--use_vw_flag_coco USE_VW_FLAG_COCO]
                [--bmnist_size BMNIST_SIZE] [--small_imagenet SMALL_IMAGENET]
                [--pretrained PRETRAINED] [--keep_fc KEEP_FC] [--alpha ALPHA]
                [--epiwt EPIWT] [--kappa KAPPA] [--moments MOMENTS]
                [--temperature TEMPERATURE] [--cycles CYCLES]
                [--models_per_cycle MODELS_PER_CYCLE] [--bg_mode BG_MODE]
                [--spurious_mode SPURIOUS_MODE]
                [--spurious_ratio SPURIOUS_RATIO] [--sf_size SF_SIZE]
                [--loss_type LOSS_TYPE] [--loss_scalar LOSS_SCALAR] 
                [--moment_desc MOMENT_DESC]

All configs YAML files are under configs/ and any hyperparameters not specified there assume default values specified in global_config.py or provided on command line as arg variables.

Checkpoints

Checkpoints for the Bayesian posterior estimates for Bayesian baseline models on the three datasets are included in moments/. These include all $\theta_m$ in $\Theta$ and predictions on all test subsets.

Checkpoints for final sharpened posterior models are also included.

Evaluation

To evaluate and generate predictions for the sharpened BayResNet+Sharpen, run:

python train.py --config_name eval --task_id DATASET_TASK_ID

where DATASET_TASK_ID is 0 (Biased MNIST), 1 (COCO-on-Places), or 2 (BAR).

Environment

See requirements.txt.

Citation

To be included after review period.

Acknowledgements

We would like to acknowledge the respective authors for their publicly available code for cSG-MCMC, network dissection, PCGrad, OccamNets whose code has been partially used and modified for this work.