AttributeError: 'ResourceVariable' object has no attribute 'overwrite_with_gradient'

[Maqingkuii]

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 20.04):
• TensorFlow version and how it was installed (source or binary):
• TensorFlow-Addons version and how it was installed (source or binary):
• Python version:
• Is GPU used? (yes/no):

Describe the bug
Compiling model...
'compile' took 0.007720 s

Training model...

Step Train loss Test loss Test metric
0 [9.10e-03, 7.54e-03, 2.72e-02, 1.26e+00, 2.77e-01] [9.10e-03, 7.54e-03, 2.72e-02, 1.26e+00, 2.77e-01] []
Traceback (most recent call last):
File "E:\python\Paper_Programs\05_deepxde\examples\pinn_inverse\Navier_Stokes_inverse.py", line 124, in
loss_history, train_state = model.train(
File "E:\python\Paper_Programs\05_deepxde\deepxde\utils\internal.py", line 22, in wrapper
result = f(*args, **kwargs)
File "E:\python\Paper_Programs\05_deepxde\deepxde\model.py", line 650, in train
self.train_sgd(iterations, display_every)
File "E:\python\Paper_Programs\05_deepxde\deepxde\model.py", line 668, in train_sgd
self.train_step(
File "E:\python\Paper_Programs\05_deepxde\deepxde\model.py", line 560, in train_step
self.train_step(inputs, targets, auxiliary_vars)
File "C:\Users\86186\anaconda\envs\05_deepxde\lib\site-packages\tensorflow\python\util\traceback_utils.py", line 153, in error_handler
raise e.with_traceback(filtered_tb) from None
File "C:\Users\86186\AppData\Local\Temp_autograph_generated_file66hkpahi.py", line 16, in tf__train_step
ag.converted_call(ag.ld(opt).apply_gradients, (ag_.converted_call(ag__.ld(zip), (ag__.ld(grads), ag__.ld(trainable_variables)), None, fscope),), None, fscope)
File "C:\Users\86186\anaconda\envs\05_deepxde\lib\site-packages\keras\src\optimizers\base_optimizer.py", line 282, in apply_gradients
self.apply(grads, trainable_variables)
File "C:\Users\86186\anaconda\envs\05_deepxde\lib\site-packages\keras\src\optimizers\base_optimizer.py", line 329, in apply
self._overwrite_variables_directly_with_gradients(
File "C:\Users\86186\anaconda\envs\05_deepxde\lib\site-packages\keras\src\optimizers\base_optimizer.py", line 618, in _overwrite_variables_directly_with_gradients
if v.overwrite_with_gradient:
AttributeError: in user code:

File "E:\python\Paper_Programs\05_deepxde\deepxde\model.py", line 249, in train_step  *
    opt.apply_gradients(zip(grads, trainable_variables))
File "C:\Users\86186\anaconda\envs\05_deepxde\lib\site-packages\keras\src\optimizers\base_optimizer.py", line 282, in apply_gradients  **
    self.apply(grads, trainable_variables)
File "C:\Users\86186\anaconda\envs\05_deepxde\lib\site-packages\keras\src\optimizers\base_optimizer.py", line 329, in apply
    self._overwrite_variables_directly_with_gradients(
File "C:\Users\86186\anaconda\envs\05_deepxde\lib\site-packages\keras\src\optimizers\base_optimizer.py", line 618, in _overwrite_variables_directly_with_gradients
    if v.overwrite_with_gradient:

AttributeError: 'ResourceVariable' object has no attribute 'overwrite_with_gradient'

Can anyone solve this problem?

[arezayan]

I encountered the same issue.

[Gurseeratkaur02]

I'm getting the same error. Did you find a way to solve it?

[andrewl36]

I had the same issue and found a workaround. It looks like a bug in this line: https://github.com/keras-team/keras/blob/660da946d33ccbb53575f15a17abdfd725cd2086/keras/src/optimizers/base_optimizer.py#L675

This code assumes the list of training variables are all either tf.Variable or a KerasVariable, it simply checks the first in the list to see if it has the overwrite_with_gradient attr, and if it does in then continues the method to iterate over them all calling that method....
I had mostly KerasVariables but I has added some custom tf.Variables at the end of the list... which causes the error.

As a workaround I simply put my tf.Variable at the beginning of the list as passed to optimizer.apply_gradients(), it then doesn't iterate the main loop using the overwrite_with_gradient optimization.... not sure of the side effects though, but so far so good!

[andrewl36]

Raised Keras issue: keras-team/keras#20517