dukang4655 / reconstructive_ssl Goto Github PK

Implementation of the experiments from the work

"Low-Rank Approximation of Structural Redundancy for Self-Supervised Learning" by Kang Du and Yu Xiang, (CLeaR 2024).

Synthetic Data Experiment

data/syn_data.py: data generation

utils/syn_utils.py utility functions defined for model training and testing

Run the experiment with fixed parameter s varying labeled sample size n (e.g., 200 datasets):

python3 run_syn_exp_vary_n.py --num_runs 200

Run the experiment with fixed labeled sample size n and varying s:

python3 run_syn_exp_vary_s.py --num_runs 200

Geometric Shapes Experiment

data/geo_data.py: data generation

utils/geo_utils.py utility functions defined for model training and testing

Run experiment triangle v.s. circles with no background pattern:

python3 run_geo_exp_no_noise.py --num_runs 200 --shape 0

Run experiment triangle v.s. circles with background pattern (noise_type: dot (0) or dash (1)):

python3 run_geo_exp_no_noise.py --num_runs 200 --shape 0 --noise_type 0 --max_noise_space 32

Run experiment triangle v.s. pentagon with no background pattern:

python3 run_geo_exp_no_noise.py --num_runs 200 --shape 1

Stylized MNIST Experiment

data/minist_data.py: data generation

utils/mnist_utils.py utility functions defined for model training and testing

Run experiment MNIST with background pattern (noise_type: dot (0) or dash (1)):

python3 run_mnist_with_noise.py --num_runs 200 --noise_type 0 --max_noise_space 32