PrimalDual Program issue in the main branch

[hfaghihi15]

@guoquan Could you please clarify an issue with the primal-dual loss call?

@auszok claims that the current implementation of PrimalDualProgram doesn't call the loss function that he has provided for computing the soft logic transformation, is that correct?

[hfaghihi15]

As discussed the PrimalDual loss is called after a specific number of steps known as the warmup iters. @auszok is going to check the loss call given that the warmup iter should pass and then we can see if this problem exists or not.

[guoquan]

@auszok claims that the current implementation of PrimalDualProgram doesn't call the loss function that he has provided for computing the soft logic transformation, is that correct?

PrimalDual uses the following method call from Andrzej to retrieve constraint loss. Not sure if he includes soft constraint in this method.

DomiKnowS/regr/program/model/primaldual.py

Lines 52 to 56 in 651c7dd

	# call the loss calculation
	# returns a dictionary, keys are matching the constraints
	constr_loss = datanode.calculateLcLoss()
	lmbd_loss = []
	for key, loss in constr_loss.items():

As discussed the PrimalDual loss is called after a specific number of steps known as the warmup iters. @auszok is going to check the loss call given that the warmup iter should pass and then we can see if this problem exists or not.

In PDProgram, I already skip calling constraint loss in warmup iterations

DomiKnowS/regr/program/primaldualprogram.py

Lines 110 to 114 in 651c7dd

	if iter < c_warmup_iters:
	loss = mloss
	else:
	closs, *_ = self.cmodel(output[1])
	loss = mloss + self.beta * closs

You can check if there is any side-effect I did not expect that cause issue in PDProgram.

[hfaghihi15]

It seems like this problem is resolved, am I right? @auszok