Issue with a model that returns the gradient of a sequence

[fabiodimarco]

Issue with a model that returns the gradient of a sequence

[raphaelvalentin]

Maybe I just answer to myself. Need to double check that my addition is okay.

In levenberg_marquardt.py

def _compute_jacobian(self, inputs, targets):
...
        jacobians = tape.jacobian(
            residuals,
            self.model.trainable_variables,
            experimental_use_pfor=self.experimental_use_pfor, unconnected_gradients=tf.UnconnectedGradients.ZERO)
...
    jacobians = [tf.reshape(j, (num_residuals, -1)) for j in jacobians if j is not None]

src: https://stackoverflow.com/questions/57144586/tensorflow-gradienttape-gradients-does-not-exist-for-variables-intermittently

[fabiodimarco]

Hi Raphael,

I executed the code you provided, it seems that adding unconnected_gradients=tf.UnconnectedGradients.ZERO does solves the error.
Does it also work as expected?

[raphaelvalentin]

Hi Fabio,
yes, it is working perfectly for my case.
Thanks a lot !
Raphael.

[fabiodimarco]

Perfect, I'll commit your changes then.