/Unsupervised-Out-of-Distribution-Detection-by-Maximum-Classifier-Discrepancy

Reproducing experimental results of OOD-by-MCD [Yu and Aizawa et al. ICCV 2019]

Primary LanguagePython

Unsupervised Out-of-Distribution Detection by Maximum Classifier Discrepancy

Reproducing experimental results of OOD-by-MCD [Yu and Aizawa, ICCV 2019]

FIY: I am not the author of this paper. I'm just a master student interested in computer graphics and machine learning. Enjoy this reproduced code. If you have any questions or requests (especially for copyright), please do not hesitate to contact me.

Disclaimer: Since the description of the fine-tuning procedure in the original paper is ambiguous and not sufficient for me to write code that reproduces the reported results, I repeated experiments several times with a variety of tweaks on my own. The main difference is that in the fine-tuning step, I set the learning rate to 0.001 instead of the reported value (0.1). Besides, since it was unclear whether the loss formula (3) used in the fine-tuning step is the aggregated loss of step A and step B or the loss used only in step B, I just chose the former empirically (see the fine_tune function in utils.py).

Reproduced Results

DenseNet-BC(L=100, k=12), ID = CIFAR-10, OOD = TinyImageNet(resized ver.)

AUROC

Discrepancy Distribution of ID and OOD

Entropy Distribution of Outputs of the Pre-trained Network

Discrepancy Distribution and Receiver Operating Characteristic (OOD = MNIST)

Requirements

torch >= 1.1.0

numpy >= 1.16.2

tqdm >= 4.31.1

visdom >= 0.1.8.8

To Activate Visdom Server

visdom

or

python -m visdom.server

Downloading Out-of-Distribtion Datasets

I use the TinyImageNet (resized) dataset from odin-pytorch:

Contact

ciy405x@kaist.ac.kr