/Keras_ODENet

Implementation of (2018) Neural Ordinary Differential Equations in Keras

Primary LanguagePython

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