DiscoNet 🕺🕸️ Towards Mitigating Shortcut Learning with Cross-Domain Regularization

Authors

• Kyle Otstot (@kotstot6)
• John Kevin Cava (@jcava)

Technologies

Backend: Python, PyTorch, NumPy

Cluster Environment: Slurm, SBATCH

Architecture

Experiment Results

Dataset: Striped-MNIST

Target #1: No stripes --val_type none

Method	Domain	Accuracy (mean ± std)
TENT	S	99.7 ± 0.0
TENT	T	72.7 ± 0.1
UDA	S	98.1 ± 0.2
UDA	T	99.3 ± 0.0
DiscoNet	S	99.3 ± 0.4
DiscoNet	T	99.5 ± 0.1

Target #2: Single stripe --val_type single

Method	Domain	Accuracy (mean ± std)
TENT	S	59.7 ± 31.3
TENT	T	10.4 ± 0.0
UDA	S	98.0 ± 0.2
UDA	T	99.2 ± 0.1
DiscoNet	S	99.4 ± 0.3
DiscoNet	T	99.4 ± 0.2

Target #3: Random stripes --val_type random

Method	Domain	Accuracy (mean ± std)
TENT	S	99.9 ± 0.0
TENT	T	10.1 ± 0.3
UDA	S	99.8 ± 0.1
UDA	T	10.0 ± 0.2
DiscoNet	S	99.6 ± 0.7
DiscoNet	T	27.9 ± 40.0

Run Experiment

Clone the project

  git clone https://link-to-project

Go to the project directory

  cd /path/to/DiscoNet

Option 1: Run Locally

  python3 main.py ( ... parameters ...)

Parameter	Type	Default	Description
--seed	int	1	random seed for reproducibility
--method	str	'disconet'	method used for domain adaptation
--val_type	str	'single'	stripe type for validation data
--batch_size	int	128	batch size used during training/testing
--n_epochs	int	100	number of epochs for training
--epoch_step	int	1	number of epochs between validation checkpoints
--c_lr	float	1e-3	learning rate for classifier
--d_lr	float	1e-3	learning rate for discriminators
--g_lr	float	1e-3	learning rate for generators
--jsd_lambda	float	1	scale for JSD consistency loss
--g_lambda	float	1	scale for generator loss
--rec_lambda	float	1	scale for reconstruction loss
--save_images		False	saves generated/reconstructed images at each checkpoint
--save_model		False	saves model after training

Option 2: Slurm Environment

In run_experiment.py, adjust grid dictionary used for hyperparameter grid search. Example:

grid = {
    'method' : ['disconet', 'tent', 'uda'],
    'val_type' : ['random'],
    'seed' : [1,2,3,4,5]
}

Adjust the run_experiment.sh SBATCH script for use.

Lastly, run the following command:

python3 run_experiment.py --sbatch