Description
What should we add?
I am following the Qiskit tutorial on the QCNN (https://qiskit-community.github.io/qiskit-machine-learning/tutorials/11_quantum_convolutional_neural_networks.html) and I wanted to try the circuit on the MNIST dataset. Due to the amount of input points, I thought it would it would be more appropriate to use Amplitude Encoding, which I believe can be done using RawFeatureVector. I have tested the circuit using the COBYLA optimiser but also want to try using a gradient based optimiser. The documentation states this is not possible and I believe has been mentioned here - #669. However, I believe it should be possible just to calculate the gradients just for the weights, for example the issue states that binding the parameters may help but then it seems that the parameters would have to rebound for each training image, which doesn't seem to be compatible with the training methodology used in the tutorial. Is it possible to use a gradient optimiser and if so, how would this be done?