Open SaashaJoshi opened 7 months ago
I do not think there is anymore a pressing need to replace Qiskit Runtime
primitives. The autonomy over training and testing is not relevant to the QML pipeline design and structure now.
The idea should be more about replacing SamplerQNN
and EstimatorQNN
from the qiskit_machine_learning
library and establishing a direct connection with Qiskit Runtime Primitives
and Simulators
such that techniques like dynamic circuits and circuit cutting can be supported. The QNNs
currently wrap a lot in them and cause limitations.
There is a requirement to configure the model training module.
Currently, the pipeline assumes dependence on Qiskit for
Runtime Primitives
and other training modules likegradients
andoptimizers
. However, some elements in the library, such asdynamic circuits
inQuantumPoolingLayer2
andQuantumPoolingLayer3
, cannot use theseRuntime Primitives
. This, therefore, presents a need for an autonomous model training module.