Loss Landscape Sharpness and Data Poisoning

Code Structure

This is code for a project on Loss Landscape Sharpness and Data Poisoning

/src contains the code used to run all experiments.

/configs contains the configurations used for all experiments.

/scripts contains various scripts including main.py which is used to start all experiments.

Installation

conda create -n sharp_poison python=3.11
conda activate sharp_poison
python -m pip install -r requirements.txt

Create a .env file which contains your wandb api key like this:

WANDB_API_KEY=key_here

Also modify your configs/config.yaml to contain your wandb project and wandb entity:

wandb_project: your_project
wandb_entity: your_entity

Running experiments

scripts/main.py is used to run experiments and can be run as follows:

python scripts/main.py experiment=dataset/model/task override1=value1 override2=value2

The following datasets can be chosen from:

cifar10, cifar100, fmnist, mnist, swiss_roll

The following models can be chosen from:

dense_net_40, lenet, res_net_18, wide_res_net_16, wide_res_net_28

The following tasks can be chosen from:

full, create_train_poison, create_poison, train_w_poison, train_wo_poison, analyze_sharpness

Sample commands

Run a full analysis of LeNet-5 on MNIST, training for 10 epochs and generating 5 largest eigenvalues of the Hessian:

python scripts/main.py experiment=mnist/lenet/full task.epochs=10 task.analysis_configs.n_hessian=5