Loading VQE should be singled out as a separate script

[marekgluza]

The task of evaluating GCI on shots exemplifies a likely use case of our repo

• we have old vqe training
• we take some vqe unitary, we have some vqe circuit
• we want to boost it

Proposed workflow:

• add gci+warmstart.py
• the user loads a qibo.Circuit (or provides it somehow)
  • additionally the hamiltonian is provided
  • cost function is same game like loading vqe
• the script initializes

  boostvqe/src/boostvqe/utils.py

  Lines 220 to 226 in 827b25f

  	def initialize_gci_from_vqe(
  	nqubits=10,
  	nlayers=7,
  	seed=42,
  	target_epoch=2000,
  	mode_dbr=DoubleBracketRotationType.group_commutator_third_order_reduced,
  	):

• It outputs GCI training results

This is proposing to have 2 scripts

• main.py trains vqe and boost with gci
• gci+warmstart.py implements Fig 1 from Arxiv and boosts with gci after some qibo.Circuit wamstart

[Edoardo-Pedicillo]

I am not sure I have understood, but this could be solved in #92

[marekgluza]

Yes, part of this issue is solved in #92. For the journal submission it will be sufficient.

After we merge #92 I will readjust the status and switch to enhancement.

[marekgluza]

• gci+warmstart.py implements Fig 1 from Arxiv and boosts with gci after some qibo.Circuit wamstart

To implement this, in boost.py the training would be conditional on a loading variable

boostvqe/src/boostvqe/boost.py

Lines 179 to 197 in 102b2ab

	# train vqe
	(
	partial_results,
	partial_params_history,
	partial_loss_history,
	partial_grads_history,
	partial_fluctuations,
	vqe,
	) = train_vqe(
	circ,
	ham, # Fixed hamiltonian
	optimizer,
	initial_parameters,
	tol,
	niterations=boost_frequency,
	nmessage=1,
	loss=loss,
	training_options=optimizer_options,
	)

Our method is more general than just vqe so the user should either provide a warmstart circuit or will be running vqe as is done now.

• main.py trains vqe and boost with gci

this is done by boost.py and will be sufficient for journal submission

[marekgluza]

Discussion with @Edoardo-Pedicillo 👍

• the user should be able to clone, and run 1 script which loads vqe (add vqe data to repo main) and obtain the plot as in current readme and main fig in paper
• refactor extras (?)