Estimating Environmental Cost Throughout Model's Adaptive Life Cycle

This is the official implementation of Estimating Environmental Cost Throughout Model's Adaptive Life Cycle

Installation

All library versions are within requirements.txt. To install all the required dependencies and python version in an Ubuntu environment, execute the following command:

sudo apt update
sudo apt install python3.8
pip install -r requirements.txt

PreIndex

To obtain PreIndex for distributional shift:

python3.8 PreIndex/pre_index.py
  -m MODEL_PATH/model.pkl \
  -s SAVE_RESULTS_TO \
  -d DATASET \
  -cl LAYER_NAME \
  -rs RANDOM_SEED \
  -mt MODEL_TYPE \
  -n_t NOISE_TYPE \
  -n_l LEVEL 

Sample command:

python3.8 PreIndex/pre_index.py
  -m ResNet18/model.pkl \
  -s ResNet18/Sample \
  -d cifar10 \
  -cl layer4.1.conv2 \
  -rs 1 \
  -mt cnn \
  -n_t gauss \
  -n_l 0.05

Retraining

To retrain a model, run retrain_dir/retrain_dist.py in the following format with the path of the original model:

python3.8 retrain_dir/retrain_dist.py
  -mp PATH_TO_MODEL/model.pkl \
  -save PATH_TO_SAVE \
  -acc CUTOFF_ACCURACY \
  -rs RANDOM_SEED \
  -tlc TRANSFORMS_LR_CUTOFF \
  -d DATASET \
  -n_tp NOISE_TYPE \
  -n_lvl LEVEL

Retraining configurations

options contains the JSON configurations format for learning rate schedule, training transformations, and cutoff plans. Edit the default values of 0.0 according to the desired training/testing scheme. Test transformations are within retrain/retrain_dist.py, which can be edited based on the dataset in use.

Code Carbon Initialization

To initialize Code Carbon, for measuring energy and carbon emission when executing retrain_dir/retrain.py, run the following command to setup the carbon tracker instance:

! codecarbon init

Citation

@inproceedings{sangarya2024estimatingenvironmentalcostmodels,
      title={Estimating Environmental Cost Throughout Model's Adaptive Life Cycle}, 
      author={Vishwesh Sangarya and Richard Bradford and Jung-Eun Kim},
      year={2024},
      booktitle={AAAI/ACM Conference on Artificial Intelligence, Ethics, and Society, 2024}
      url={https://ojs.aaai.org/index.php/AIES/article/view/31723/33890}, 
}