NeuralODE Python - "AttributeError: module 'tensorflow' has no attribute 'defun' "

Question

I am currently working with NeuralODE on google CoLab, and I tried to use a command, 'defun', without success. I am using Tensorflow version 1.x which is a requirement for NeuralODE, but when I try to use tf.defun, I get the error "AttributeError: module 'tensorflow' has no attribute 'defun'. How can I actually get to defun?

This is the particular bit of code I need it for

def compute_gradients_and_update(): 
 with tf.GradientTape() as g: 
 hN = neural_ode.forward(h0_var)
 g.watch(hN)
 loss = tf.reduce_sum((hN_target - hN)**2)
 
 dLoss = g.gradient(loss, hN) # same what 2 * (hN_target - hN)
 h0_reconstruction, dfdh0, dWeights = neural_ode.backward(hN, dLoss)
 optimizer.apply_gradients([(dfdh0, h0_var)])
 return loss
 compute_gradients_and_update = tf.defun(compute_gradients_and_update)

I have seen similar questions on stackoverflow, but the solutions are all very specficic and I haven't seen any about 'defun' specifically. I hope this makes sense, I am a beginner.

Thanks!

Answer 1

Tensorflow has no attribute called defun.

If you are using tf 1.x, use

tf.contrib.eager.defun