PixelCNN-Paddle

This is an unofficial Paddle implementation of PixelCNN (Van Oord et al. 2016).

Contents

1. Introduction
2. Reproduction Accuracy
3. Dataset
4. Environment
5. Train & Test
6. Test
7. Code Structure

Introduction

Reference Code：PixelCNN

Paper：Pixel Recurrent Neural Networks

Reproduction Accuracy

In training, set batch size to 256.

Index	Raw Paper	Reference Code	Reproduction
NLL	81.30	85.74	86.00003680419921

Dataset

The authors use MNIST dataset, and it will be auto-download when users training.

Environment

• Frameworks:

• PaddlePaddle (2.1.2)
• NumPy (1.18.4)
• Pillow (7.2.0)

Train & Test

python train.py

Code Structure

├── ckpts  # pdparams and training logs
├── src
│   ├── pixel_cnn.py
│   ├── datasets.py
│   ├── base.py
│   ├── convolution.py
│   ├── train.py
│   ├── trainer.py
├── README.md
└── requirements.txt