def _set_up_acq_opt_ga(self):
""" Determines the mutation operator for the internal GA. """
# First the sampling distribution for the GA mutation operator
mut_arg = self.options.ga_mutation_op_distro
if isinstance(mut_arg, list):
self.ga_mutation_op = get_nn_modifier_from_args(
self.domain.constraint_checker, dflt_num_steps_probs=mut_arg)
elif isinstance(mut_arg, (int, long, float)):
self.ga_mutation_op = get_nn_modifier_from_args(
self.domain.constraint_checker, dflt_max_num_steps=mut_arg)
elif mut_arg.startswith('d'):
ga_mutation_probs = [float(x) for x in mut_arg[1:].split('-')]
self.ga_mutation_op = get_nn_modifier_from_args(
self.domain.constraint_checker,
dflt_num_steps_probs=ga_mutation_probs)
elif mut_arg.startswith('n'):
ga_mutation_num_steps = int(x[1:])
self.ga_mutation_op = get_nn_modifier_from_args(
self.domain.constraint_checker,
dflt_max_num_steps=ga_mutation_num_steps)
else:
raise ValueError('Cannot parse ga_mutation_op_distro=%s.' %
(self.options.ga_mutation_op_distro))
# The initial pool
self.ga_init_pool = get_initial_pool(self.domain.get_type())
# The number of evaluations
if self.get_acq_opt_max_evals is None:
lead_const = min(5, self.domain.get_dim())**2
self.get_acq_opt_max_evals = lambda t: np.clip(
lead_const * np.sqrt(t), 50, 500)
def get_nn_opt_arguments():
""" Returns arguments for NN Optimisation. """
ret = Namespace()
ret.cnn_constraint_checker = CNNConstraintChecker(50, 5, 1e8, 0, 5, 5, 200, 1024, 8)
ret.mlp_constraint_checker = MLPConstraintChecker(50, 5, 1e8, 0, 5, 5, 200, 1024, 8)
ret.cnn_mutation_op = get_nn_modifier_from_args(ret.cnn_constraint_checker,
[0.5, 0.25, 0.125, 0.075, 0.05])
ret.mlp_mutation_op = get_nn_modifier_from_args(ret.mlp_constraint_checker,
[0.5, 0.25, 0.125, 0.075, 0.05])
# Create the initial pool
ret.cnn_init_pool = get_initial_cnn_pool()
ret.cnn_init_vals = [cnn_syn_func1(cnn) for cnn in ret.cnn_init_pool]
ret.mlp_init_pool = get_initial_mlp_pool('reg')
ret.mlp_init_vals = [mlp_syn_func1(mlp) for mlp in ret.mlp_init_pool]
# Create a domain
ret.nn_domain = NNDomain(None, None)
ret.cnn_domain = NNDomain('cnn', ret.cnn_constraint_checker)
ret.mlp_domain = NNDomain('mlp-reg', ret.mlp_constraint_checker)
# Return
return ret
def get_opt_problem(nn_type):
""" Gets parameters for the optimisation problem. """
if nn_type == 'cnn':
constraint_checker = CNNConstraintChecker(50, 1e8, 5, 5, 200, 1024, 8)
init_points = get_initial_cnn_pool()
func_caller = FunctionCaller(cnn_syn_func1, NNDomain(None, None))
elif nn_type.startswith('mlp'):
constraint_checker = MLPConstraintChecker(50, 1e8, 5, 5, 200, 1024, 8)
init_points = get_initial_mlp_pool(CLASS_OR_REG)
func_caller = FunctionCaller(mlp_syn_func1, NNDomain(None, None))
else:
raise ValueError('Unknown nn_type: %s.' % (nn_type))
# Common stuff
mutation_op = get_nn_modifier_from_args(constraint_checker,
[0.5, 0.25, 0.125, 0.075, 0.05])
init_vals = [func_caller.eval_single(nn)[0] for nn in init_points]
return constraint_checker, func_caller, mutation_op, init_points, init_vals
