I have a coupled model defined in OpenMDAO that I'm using with an optimisation driver, and the outputs of various components vary a lot in order of magnitude. I would like to scale all outputs to roughly 0-1 range, to have a Jacobian with elements of a similar order of magnitude, for a better optimisation process. I am using the ref and ref0 arguments to self.add_output(), and define the components that I want to scale as constraints in order to have them visible in the scaling report (just to help checking the scaling results).
The sample of my model below (this is radically simplified to provide a reproducable example, but keeps the structure of the more complex model I am dealing with).
The problem I have is that the Jacobian is all grayed out (light gray). What are some possible reasons for this?
In my original model, I have design variables, that feed into a component, and its output feeds into another component, and so on through a few components, until an output of the last component is fed to the constraint component. If I want to ensure proper scaling of the Jacobian (of constraints w.r.t. design variables), should I be scaling every single output of every single component along the path from the design variables to the constraints?
Is there any way to make OpenMDAO compute the derivatives numerically based directly on the finite difference of constraints w.r.t. design variables, and not as in the chain rule, component by component, to avoid the need for scaling entire model?
import numpy as np
import openmdao.api as om
class MDA(om.Group):
class Inertia(om.ExplicitComponent):
def setup(self):
self.add_input('r', val= [3.1,3,3], units='m')
self.add_input('T', val= 102.0, units='m')
self.add_output('M', shape=(6,6), val=np.ones([6,6]), ref=1e6*np.ones([6,6]), units='kg')
def setup_partials(self):
self.declare_partials('*', '*')
def compute(self, inputs, outputs):
r = inputs['r']
T = inputs['T'][0]
outputs['M'] = np.linalg.norm(r) * T * np.ones([6,6])
class Cost(om.ExplicitComponent):
def setup(self):
self.add_input('M', shape=(6,6), val=np.ones([6,6]), units='kg')
self.add_output('cost', val=0, units='USD')
def compute(self, inputs, outputs):
M = inputs['M']
outputs['cost'] = np.linalg.norm(M) * 10
def setup_partials(self):
self.declare_partials('*', '*')
class ConCmp0(om.ExplicitComponent):
def setup(self):
self.add_input('M', shape=(6,6), val=np.ones([6,6]), units='kg')
self.add_output('con0', shape=(6,6), val=np.ones([6,6]), units='kg')
def setup_partials(self):
self.declare_partials('*', '*')
def compute(self, inputs, outputs):
M = inputs['M']
outputs['con0'] = M
class ObjCmp(om.ExplicitComponent):
def setup(self):
self.add_input('cost', val=0, units='USD')
self.add_output('obj', val=0.0)
def setup_partials(self):
self.declare_partials('*', '*')
def compute(self, inputs, outputs):
outputs['obj'] = inputs['cost']
def setup(self):
self.add_subsystem('inertia', self.Inertia(), promotes_inputs=['T','r'],
promotes_outputs=['M'])
self.add_subsystem('cost', self.Cost(), promotes_inputs=['M'],
promotes_outputs=['cost'])
self.add_subsystem('con_cmp0', self.ConCmp0(), promotes_inputs=['M'],
promotes_outputs=['con0'])
self.add_subsystem('obj_cmp', self.ObjCmp(), promotes_inputs=['cost'],
promotes_outputs=['obj'])
# build the model
prob = om.Problem(model=MDA())
model = prob.model
# setup the optimization
prob.driver = om.ScipyOptimizeDriver()
prob.driver.options['optimizer'] = 'SLSQP'
model.add_design_var('r', lower = [3,3,3], upper = [10,10,10])
model.add_design_var('T', lower = 100, upper = 140)
model.add_objective('obj')
model.add_constraint('con0', lower = 0)
prob.setup()
# Set initial values.
prob.set_val('r', [3.1, 3, 3])
prob.set_val('T', 102.0)
# run the optimization
prob.run_driver()
If you want FD at the top of your model, then use the approx_totals method called at the top. Note that this method can be called at any level of the model hierarchy, but your question specifically asked about top level FD.
The FD details, aside, there is a problem with the sample model that is causing
The problem I have is that the Jacobian is all grayed out (light gray)
So FD method aside, your root problem may be in the way you've declared the derivatives to being with.
you have called self.declare_partials('*', '*') in every component here. The function signature for that method shows that the default has method="exact" --- which means you are telling OpenMDAO that you will provide exact analytic derivatives. So OM is expecting you to implement a compute_partials or compute_jacvec_product method for your explicit components. You didn't do that, so the framework thinks that all your derivatives are zero. This is why you are seeing the whole jacobian being gray'd out in the scaling report.
When you set the approx_totals at the top level of the model, OM will then ignore the derivative settings of components below that call. So this should sort of fix your issue. However, the more correct method of fixing it would be to set the method="fd" argument in all your declare_partials calls. You might still prefer top level FD overall, but in case you ever turn that off at least it would still work properly then.