google
Jacobians and expectations
In PR #70, I added a test that calls the jacobian method of a tf.GradientTape on the result of a TFQ calculation. The test passes if I do not use the @tf.function decorator. However, when I add the decorator, I get the following error at the end of a long chain:
LookupError: No gradient defined for operation 'TfqAdjointGradient' (op type: TfqAdjointGradient)
This is strange, because in the test function, there is only one gradient tape, so no gradient of 'TfqAdjointGradient' should be attempted.
I have been trying to recreate the error with a simplified scenario in this colab notebook, but so far I cannot replicate it.
@jaeyoo I was wondering if you might have ideas about what could cause this from the TF or TFQ side of things
Yes, I've already been dealing with the issue. Please let me add it for it...
@sahilpatelsp can you provide a detailed explanation here of how you understand the issue and how you addressed it with your trivial None check? Good for docs
Well I am not entirely sure the issue is resolved in the case of RNN hypernetworks for QHBMs as I am currently running into a gradient issue, but it does seem to be working for standard QMHL and VQT training routines. I believe the issue was that the qhbm.trainable_variables are only recognized as variables for tf.custom_gradient with tf.function. Therefore, if we do not decorate with tf.function, then the variables argument of the inner grad function would accordingly be None, and therefore we should simply just return a single set of gradients for the input, which is qhbm.trainable_variables. Otherwise, we should also return another set of gradients for the variables, which are again qhbm.trainable_variables.
Maybe another lead from this link: "It should be noted tf.GradientTape is still watching the forward pass of a tf.custom_gradient, and will use the ops it watches. As a consequence, calling tf.function while the tape is still watching leads to a gradient graph being built. If an op is used in tf.function without registered gradient, a LookupError will be raised."