Running experiments

Training script is run as follows:

python train.py --configs <1 or more paths to .yaml files> --values wandb=True output_path=<path> model_type=VICReg vicreg.base_lr=42

Important configuration options:

model_type: defines which model to use, options are VICReg, RSSM, SimCLR.
dataset_noise: controls changing noise level
dataset_static_noise: controls fixed noise level
dataset_structured_noise: controls whether the noise is structured.
output_path: specifies the path to save models.
wandb: enables weights and biases logging.
dataset_type: chooses between one dot and three dot datasets. For one dot set Single, for three set Multiple.

To access particular model's options set options through the respective subconfig, e.g. vicreg.base_lr.

Reproducing results

All configs are saved in reproduce_configs folder. To run a config from that folder, you can run

python train.py --configs reproduce_configs/sweep_fixed_uniform.(1.25).vicreg.best.yaml

This will run the best VICReg configuration for fixed uniform noise with coefficient 1.25.