andiac / pixel-cnn-pp

Pytorch Implementation of OpenAI's PixelCNN++

PixelCNN++

A Pytorch Implementation of PixelCNN++.

Main work taken from the official implementation

Pre-trained models are available here

I kept the code structure to facilitate comparison with the official code.

The code achieves 2.95 BPD on test set, compared to 2.92 BPD on the official tensorflow implementation.

Training the model

python main.py

Training the model by blurred training set

python blur_train.py

Explanation of the scirpts

ood_detect.py: train on CIFAR10, test on CIFAR10 and SVHN. Plot the histgram and output mean BPDs.

blur_detect.py: train on CIFAR10, test on CIFAR10 and blurred CIFAR10, the generated histogram shows that blurred CIFAR10 has higher dentisty then original CIFAR10.

Differences with official implementation

1. No data dependant weight initialization
2. No exponential moving average of past models for test set evalutation

Contact

This repository is no longer maintained. Feel free to file an issue if need be, however response may be slow.

Out-of-distribution detection

Epoch	BPD	SVHN_BPD	AUROC
319	2.9418	2.0881	0.1493
489	2.9409	2.0940	0.1518
589	2.9361	2.0942	0.1535
789	2.9295	2.0969	0.1570
889	2.9292	2.0912	0.1558

python3 langevin_dynamics.py -c 3 -r 3 -n 0.001 -ss 1e-3