Add pyoptinterface_based_impl.py and integrate it

[KeShih]

What?

• Add pyoptinterface_based_impl.py, which uses the PyOptInterface package.
• Remove the old fast and slow modes, use a unified interface instead.
• The default solver is HiGHS, we can change the global default solver by invorking set_default_solver(solver_name) in dingo.pyoptinterface_based_impl module. Also, we can change the solver in MetabolicNetwork and MetabolicNetwork class by invoking the method set_solver(self, solver)

How to test it?

• Test with highs.

python3 tests/fba.py
python3 tests/full_dimensional.py
python3 tests/max_ball.py
python3 tests/scaling.py
python3 tests/rounding.py
python3 tests/sampling.py

• Test with gurobi.

python3 tests/fba.py gurobi
python3 tests/full_dimensional.py gurobi
python3 tests/max_ball.py gurobi
python3 tests/scaling.py gurobi
python3 tests/rounding.py gurobi
python3 tests/sampling.py gurobi

Documentation

See #106.

[KeShih]

Speed up of fva() and remove_redundant_facets().

In these two functions, we need to solve numerous LPs with similar constraints. Modifying the model rather than constructing a new one can significantly enhance performance, and the PyOptInterface package supports such modifications.

• fva()
  

  dingo/dingo/fva.py

  Lines 65 to 70 in 4a961e7

  	for i in range(n):
  	# Set the ith row of the A matrix as the objective function
  	objective_function = A[
  	i,
  	]

  
  Noticing that only the objective function changes after each iteration, optimization can be achieved.
• remove_redundant_facets()
  The outer loop is a while loop, and the inner loop is a for loop. We observe that throughout the for loop, Aeq_sparse, beq, A_sparse, and [val] remain unchanged. Within each iteration of the for loop, only lb[i], ub[i], and the objective function change. Between consecutive iterations of the for loop, only lb[i] and ub[i] from the previous iteration may be modified. Therefore, we can achieve fine-tuning by adjusting a single model throughout the entire for loop.
  

  dingo/dingo/gurobi_based_implementations.py

  Lines 404 to 442 in 4a961e7

  	# Loop until nor redundant facets are found
  	while removed > 0 or offset > 0:
  	removed = 0
  	offset = 0
  	indices = indices_iter
  	indices_iter = []
  	Aeq_sparse = sp.csr_matrix(Aeq_res)
  	beq = np.array(beq_res)
  	b_res = []
  	A_res = np.empty((0, n), float)
  	for i in indices:
  	# Set the ith row of the A matrix as the objective function
  	objective_function = A[
  	i,
  	]
  	redundant_facet_right = True
  	redundant_facet_left = True
  	# for the maximum
  	objective_function_max = np.asarray(
  	[-x for x in objective_function]
  	)
  	model_iter = update_model(
  	model_iter,
  	n,
  	Aeq_sparse,
  	beq,
  	lb,
  	ub,
  	A_sparse,
  	[val],
  	objective_function_max,
  	)
  	model_iter.optimize()

Experiment (iAF1260 model)

[TolisChal]

Could you please throw an exception here instead of printing a message?

[KeShih]

OK.

[TolisChal]

What these lines stand for?

[KeShih]

The code above will be more symmetrical, but in reality, it carries no meaning. I have removed it.

[TolisChal]

Thanks @KeShih! This is a great contribution!