Implementation of (2018) Neural Ordinary Differential Equations.
ODE solver are use tf.contrib.integrate.odeint which only supported "dopri5" method now.
GPU: Nvidia GTX 670
python==3.6
tensorflow==1.4.0
keras==2.1.0
Result on 10 epochs
training time: 730s
train_loss: 0.0112 - train_acc: 0.9962 - val_loss: 0.0234 - val_acc: 0.9929
training time: 120s
train_loss: 0.0096 - train_acc: 0.9968 - val_loss: 0.0307 - val_acc: 0.9908