This code implements various models related to adverserial training and diff-pruning (Guo et. al., 2020)

To run the code make sure conda is installed and then run

conda env create -f environment.yml

Then activate the environment by running

conda activate diff_pruning

The project structure looks as follows

📦sparse_transformers
┣ 📂analysis (jupyter notebooks for analysis of models and results)
┣ 📂scripts (additional scripts for various experiments)
┣ 📂src
┃ ┣ 📂models (directory which contains all model classes)
┃ ┃ ┣ 📜model_adv.py (baseline model for adverserial training)
┃ ┃ ┣ 📜model_base.py (contains base classes with methods that are used by all models)
┃ ┃ ┣ 📜model_diff_adv.py (model with 2 subnetworks for adverserial training)
┃ ┃ ┣ 📜model_diff_modular.py (model with 2 subnetworks for task and adv training)
┃ ┃ ┣ 📜model_diff_task.py (model with subnetwork for task training)
┃ ┃ ┣ 📜model_functions.py (generic functions for all model classes)
┃ ┃ ┣ 📜model_heads.py (classifier and adverserial head classes)
┃ ┃ ┣ 📜model_modular.py (baseline model for modular training)
┃ ┃ ┣ 📜model_task.py (baseline model for task training)
┃ ┃ ┗ 📜weight_parametrizations.py (contains weight parametrizations for subnetwork training*)
┃ ┣ 📜adv_attack.py (contains function to run adverserial attack)
┃ ┣ 📜data_handler.py
┃ ┣ 📜metrics.py
┃ ┣ 📜training_logger.py
┃ ┗ 📜utils.py
┣ 📜cfg.yml (hyperparameters)
┣ 📜environment.yml (conda environment config)
┣ 📜main.py (main file to run experiments with)
┣ 📜main_attack.py (used to run an adverserial attack only using a model checkpoint)
┣ 📜main_wrapper.py (used to run multiple experiments sequentially)
┗ 📜readme.md

* Weight parametrizations are implemented as modules and use pytorch parametrizations functionality LINK

contains hyperparameter configuration

• data_config
  filepaths to data files
• model_config
  name of pretrained model and batch_size to use
• train_config_diff_pruning
  hyperparameters for diff-pruning-models (model_diff_adv.py and model_diff_task.py)
• train_config_baseline
  hyperparameters for baseline models (model_adv.py and model_task.py)
• adv_attack hyperparameters for adverserial attack

python3 main.py

Optional arguments with example inputs

• --gpu_id 0 1 2 3
  Which gpus to run experiment on (can be multiple)
• --adv
  Set if you want to run adverserial training
• --baseline
  Set if you want to run a baseline model instead of diff-pruning
• --modular
  Run modular architecture (overwrites adv argument)
• --seed=0
  random seed
• --ds="bios"
  which dataset to run ("bios", "pan16", "hatespeech")
• --cpu
  Run on cpu (even if gpu is available)
• --no_adv_attack
  Set if you do not want to run adverserial attack after training
• --cp_path="path_to_model"
  Overwrite pre-trained encoder weights
• --cp_is_sd
  Set if checkpoint is a state dict
• --cp_model_type="TaskModel"
  Model type from which to load encoder weights as string (not required if loading state dict directly)
• --cp_modular_biased
  If loading checkpoint from modular model set debiased state as false
• --prot_key_idx=0
  If protected key is type list: index of key to use
• --debug
  To verify code can run through, limits number of batches which are used to 10