We provide the code for training and evaluating the InFlow model on CIFAR 10 dataset that is based on the paper available at arXiv.
To run our experimental code, you need the following python packages:
- PyTorch
- numpy
- torchvision
- alibi-detect
- ood-metrics
- sklearn
All the above packages can be installed using pip command. To set up the environment, you need to download the publicly available torchvision datasets such as MNIST, FashionMNIST, SVHN, CIFAR-10 etc as image files and keep each of the datasets in separate folders of the project directory. The CIFAR-10 train and test images can be downloaded from Kaggle.
To train the InFlow model once the dataset preparation is complete, run the following command in the project directory:
python train.py
Note: You need to change the path with respect to the folder where datasets are placed in the project directory
To evaluate the trained InFlow model on several datasets, run the following command:
python eval.py
You can also skip training the model and directly run the evaluation by using the provided pre-trained model. The pretrained model can be found in the /ckptdir directory..