Neural Ordinary Differential Equations in Keras

Introduction

Implementation of (2018) Neural Ordinary Differential Equations.

Attention

ODE solver are use tf.contrib.integrate.odeint which only supported "dopri5" method now.

Environment

GPU: Nvidia GTX 670

python==3.6
tensorflow==1.4.0
keras==2.1.0

Result

Result on 10 epochs

MNIST ODENet

training time: 730s

train_loss: 0.0112 - train_acc: 0.9962 - val_loss: 0.0234 - val_acc: 0.9929

MNIST ResNet

training time: 120s

train_loss: 0.0096 - train_acc: 0.9968 - val_loss: 0.0307 - val_acc: 0.9908