I've playing around with a binary classification problem with a dataset of size 1000 examples and 5 features. I'm using the VQC algorithm from the Qiskit package the that uses a feature map and a parametrized variational circuit. The algorithm has a
.train() ( most examples use
run() as well) method that implements the optimization loop to find the optimal parameters of the variational form and accepts as argument a minibatch size parameter. I'm using the SPSA optimizer, how does batching work in this environment? Does it use a batch of examples to estimate the cost and update the parameters after the batch has been processed like in mini-batch stochastic gradient?
I've also wanted to execute the learning phase of the VQC on the IBMQ experience platform. Taking into account that most real backend on the IBMQ accept a number of 75 experiments, i.e number of circuits to execute within a submitted job, I've thought that setting mini-batch size to the number of maximum experiments would be reasonable, although I know Aqua handles and manages jobs (splitting, waiting for results) at higher level, so you could set it to any number.