# pixelcnnplus

**Repository Path**: zyabo/pixelcnnplus

## Basic Information

- **Project Name**: pixelcnnplus
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: MIT
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2021-04-06
- **Last Updated**: 2024-06-02

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

## PixelCNN++

A Pytorch Implementation of [PixelCNN++.](https://arxiv.org/pdf/1701.05517.pdf)

Main work taken from the [official implementation](https://github.com/openai/pixel-cnn)

Pre-trained models are available [here](https://mega.nz/#F!W7IhST7R!PV7Pbet8Q07GxVLGnmQrZg)

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. 
<p align="center">
<img src="https://github.com/pclucas14/pixel-cnn-pp/blob/master/images/pcnn_lr:0.00020_nr-resnet5_nr-filters160_143.png">
<img src="https://github.com/pclucas14/pixel-cnn-pp/blob/master/images/pcnn_lr:0.00020_nr-resnet5_nr-filters160_122.png">
<img src="https://github.com/pclucas14/pixel-cnn-pp/blob/master/images/pcnn_lr:0.00020_nr-resnet5_nr-filters160_137.png">
<img src="https://github.com/pclucas14/pixel-cnn-pp/blob/master/images/pcnn_lr:0.00020_nr-resnet5_nr-filters160_101.png">
</p>

### Running the code
```
python main.py
```

### 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This repository is no longer maintained.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.