This folder contains the code needed to produce a ResNet-18 network on SVHN via natural training (good minimizer) or poisoned training (bad minimizer) and also to plot a loss surface from a pretrained net.
This code requires the following packages
- Python 3.6
- Tensorflow 1.9
- Torch 1.0
- Torchvision 0.2.1
Different versions may likely work but they have not been tested
To train the model naturally, run
python main.py -gpu=0 -gpu_eval -batch_size=256 -upload -weight_decay=0 -poison -fracdirty=.01 -nohess -nogan -svhn -nodirty -log_root=svhn-natural -ckpt_root=./ckpt/
To train the model with poisoning, run
python main.py -gpu=0 -fracdirty=35e-2 -batch_size=256 -gpu_eval -weight_decay=0 -poison -nohess -nogan -svhn -log_root=svhn-poison -ckpt_root=./ckpt/
Adjust the fracdirty value to change the poison factor, gpu value to change the gpu, and log_root to change the name of the checkpoint directory. Model checkpoints will be stored in ./ckpt/
To compute the loss surface for a particular model, please first upload that model's checkpoint folder to dropbox and provide the public link to that folder in the url argument.
Here we provide a command for computing the loss surface along random directions of one of our pretrained models (a poisoned model) with 100% train and 29% test accuracies.
python surface.py -gpu=0 -batchsize=1024 -nworker=2 -url=https://www.dropbox.com/sh/8x3vspe62vlpjfr/AABXBlCBs0hbQ1-rE4fXscdva?dl=0
Feel free to change gpu (gpu id), batchsize (the bigger the better), nworker (number of cpus to load data) to suit your hardware capabilities. They will not affect the result.