def get_features(self):
# create a copy to avoid polluting the current one.
# returns constraint_features, edge_features, variable_features
m = self._copy_model()
ret = get_features_from_scip_model(m)
del_scip_model(m)
return ret
def get_feasible_solution(self):
# get any feasible solution.
fixed_model = self._copy_model()
# No objective -- only feasibility
fixed_model.setObjective(scip.Expr())
fixed_model.optimize()
assert fixed_model.getStatus() == 'optimal', fixed_model.getStatus()
ass = {
't_' + var.name: fixed_model.getVal(var)
for var in fixed_model.getVars()
}
del_scip_model(fixed_model)
return ass
def reset(self):
# fix the graph for a few episodes to reduce load on the disk while loading datasets.
if self._n_resets % self._sample_every_n_resets == 0:
self.milp, self._sol, self._obj = self._sample()
self.mip = SCIPMIPInstance.fromMIPInstance(self.milp.mip)
# clean up previous scip model
if hasattr(self, '_sub_mip_model'):
del_scip_model(self._sub_mip_model)
self._sub_mip_model = self.mip.get_scip_model()
sol, obj, mip = self._sol, self._obj, self.mip
c_f, e_f, v_f = load_pickled_features(self._dataset,
self._dataset_type,
self._milp_choice)
self._var_names = var_names = list(
map(lambda v: v.name.lstrip('t_'), mip.vars))
# call init_features before init_ds
self._init_features(var_names, v_f)
self._init_ds()
self._constraint_nodes = np.float32(c_f['values'])
self._change_sol(sol, obj, sol, obj)
self._best_quality = self._primal_gap(obj)
self._final_quality = self._best_quality
self._qualities = [self._best_quality]
self._unfixed_variables = filter(lambda v: v.vtype() == 'CONTINUOUS',
mip.vars)
self._unfixed_variables = set(
map(lambda v: v.name.lstrip('t_'), self._unfixed_variables))
self._vars_unfixed_so_far = []
# static features computed only once per episode.
if self.config.make_obs_for_graphnet:
raise Exception('Not yet supported!')
self._static_graph_features = self._encode_static_graph_features()
elif self.config.make_obs_for_bipartite_graphnet:
self._static_graph_features = self._encode_static_bipartite_graph_features(
e_f)
self._n_resets += 1
return restart(self._observation())
def get_sample(dataset, dataset_type, graph_idx):
dataset_path = DATASET_PATH[dataset]
with open(os.path.join(dataset_path, dataset_type, f'{graph_idx}.pkl'), 'rb') as f:
milp = pickle.load(f)
if milp.get('optimal_lp_sol', None) is None:
if dataset in DATASET_INFO_PATH:
with open(
os.path.join(DATASET_INFO_PATH[dataset], 'aux_info', dataset_type, f'{graph_idx}.pkl'),
'rb') as f:
pkl = pickle.load(f)
milp.optimal_lp_sol = pkl['optimal_lp_sol']
else:
solver = Model()
solver.hideOutput()
relax_integral_constraints(milp.mip).add_to_scip_solver(solver)
solver.optimize()
assert solver.getStatus() == 'optimal', solver.getStatus()
ass = {var.name: solver.getVal(var) for var in solver.getVars()}
milp = ConfigDict(milp)
milp.optimal_lp_sol = ass
del_scip_model(solver)
return milp
def __del__(self):
del_scip_model(self.model)