BryceFuller
commented
2 years ago Hello Dan, do you see the same issue when running the train_kernels_using_qiskit_feature_maps.ipynb notebook?
The ZZ feature map does not have any user_parameters by default. This feature map was introduced in the IBM nature paper which popularized the idea of QSVMs and quantum feature maps. The first quantum feature maps proposed were not trainable.
To be clear, the only practical difference between what I call a trainable feature map and a non-trainable feature map is the existence of parameters which are explicitly set aside for training (and will thus not be bound to input data values). So any non-trainable feature map could still become a trainable one if you as the user split up its input parameters and explicitly set some of them aside as trainable parameters.
In the train_kernels_using_qiskit_feature_maps.ipynb notebook, my goal (in Option 1) was to show how one could do this manual partitioning of parameters to convert a single ZZ FeatureMap into a trainable feature map (but which now has fewer input parameters).
This is sort of a hack which I included to show how one could make this work, but there are downsides of this. One downside of Option 1 (from the notebook) is that the ZZ feature map uses one parameter per qubit. Treating this feature map as a trainable feature map requires re-purposing input parameters to be trainable parameters, so this implies that you will end up needing more qubits than you have features. For this reason, you might be better off going with Option 2, which suggests that you can instead compose two ZZ feature maps into one big feature map and manually define one of the ZZ feature map's params to be trainable (while leaving the other feature map's parameters as regular input parameters). This too has imperfect side-effects. For one, you may end up with a circuit that is deeper than you'd prefer. Also, if you place the trainable parameters as the last block of the circuit, then that entire block will cancel itself out when you compute Kernel entries between data points. There is also theoretical justification for wanting to increase how interleaved the data and input parameters are within the circuit, which Option 2 fails to do.
The main takeaway from train_kernels_using_qiskit_feature_maps.ipynb is that you will have to take special care if you wish to treat existing (nontrainable) feature maps as though they are trainable.
Using the ZZ Feature map may still be desirable though, as a first sanity check to get things running. But I expect you will see shorter circuits and more efficient use of qubits by using quantum feature maps that were intended for training, or by building a trainable feature map from scratch.
Let me know if this helps clear up your questions.
dancbeaulieu
jenglick
caleb-johnson
Describe the bug Was trying to test my custom kernel I created vs ZZ_Featuremap circuit as my feature map for my quantum kernel. Could not get the QKT function to work with zz_featuremap. Can you help us figure out how to get binding parameters to work for preconstructed qiskit featuremap circuits?
We tried the method specified in https://github.com/IBM-Quantum-Prototypes/quantum-kernel-training/blob/main/docs/how_tos/train_kernels_using_qiskit_feature_maps.ipynb, but couldn't get it to work. Two of us tried multiple methods specified in the guide, but none of them worked when we tried to move them over the QKT library, especially qkt.fit(). Can you expand the tutorial to show this actually working?
There are a few problems: 1) Can't bind parameters successfully with ZZ_Featuremap based kernel, returns an error because it doesn not have a 'user_parameter' property. Always have issues with the "qkt_fit" call which expects a "user_parameters" property. en_map = ZZFeatureMap(feature_dimension=12, reps=2, entanglement='linear', data_map_func=None , insert_barriers=True) print(en_map.user_parameters) #.user_parameters)
If I try to look at the user_parameters of my kernel I create with zz_featuremap, I get nothing. en_map = ZZFeatureMap(feature_dimension=12, reps=2, entanglement='linear', data_map_func=None , insert_barriers=True) zz_ker = QuantumKernel(feature_map=en_map, quantum_instan
I can get the properties with en_map.ordered_parameters, however if I do this I get the following error with qkt_fit() call. qkt
Train the kernel using QKT directly
To Reproduce To reproduce this try using a ZZ_Featuremap with qkt.fit().
Expected behavior I expect either a way to pass ZZ_Featuremap parameters to qkt.fit(), zz_featuremap doesn't have user_parameters properties.
Any suggestions? Can you figure out a method for getting zz_featuremap properties into QKT? Can you show actually binding parameters and getting it to work in qkt.fit() in the https://github.com/IBM-Quantum-Prototypes/quantum-kernel-training/blob/main/docs/how_tos/bind_parameters_using_quantum_kernel.ipynb or show the successful integration of a zz_featuremap in the https://github.com/IBM-Quantum-Prototypes/quantum-kernel-training/blob/main/docs/how_tos/train_kernels_using_qiskit_feature_maps.ipynb? These tutorials seems to end too early.