Make PytorchModel work with quantum circuits

[dimkart]

Description

lambeq provides a large number of models and trainers, covering a broad range of use cases, both quantum and "classical", where the latter models operate on tensor networks. Currently, the PytorchModel works exclusively on tensor networks, hence it is relevant only for "classical" experiments.

This task asks for the extension of PytorchModel so it could also work with quantum circuits, in the same way as e.g. the PennyLaneModel does. This will allow quantum models to be trained using gradient-based optimisers such as Adam. The following changes will need to be made.

• In quantum ansatze, the lambeq symbols will need to be used instead of sympy symbols directly.
• The initialise_weights method will need to know whether it is operating on a quantum circuit or tensor network, and the quantum circuit weights may be initialised uniformly at random, without the current logic used for tensors.
• Some additional changes may be needed to the quantum backend to ensure that gradients are calculated appropriately by pytorch's autograd.
• An example notebook demonstrating the new capability will need to be demonstrated. This can simply use the classical pipeline example notebook, and replace the ansatz with a quantum ansatz.

Notes

• Adding the functionality in a new class QuantumPytorchModel, instead of extending the PytorchModel class, might be an acceptable alternative.
• Include in the PR the results of a basic test run as a proof of concept, training one of the lambeq datasets with the new model.

References

• PytorchModel in lambeq
• Circuit ansatze in lambeq
• Quantum backend
• lambeq training use cases
• lambeq training tutorials

[Vikrant-Khedkar]

hey up for this