def finish(instance: CVRPInstance, model: QRPModel) -> CVRPSolution:
subinstances = [
CVRPInstance(
name="",
region="",
deliveries=vehicle.deliveries,
origin=vehicle.origin,
vehicle_capacity=3 * instance.vehicle_capacity, # More relaxed.
)
for idx, subinstance in enumerate(model.cluster_subsolutions.values())
for vehicle in subinstance
]
logger.info("Reordering routes.")
subsolutions = [
ortools_solve(subinstance, model.params.ortools_tsp_params)
for subinstance in subinstances
]
return CVRPSolution(
name=instance.name,
vehicles=[
v for subsolution in subsolutions for v in subsolution.vehicles
],
)
def solve(file):
instance = CVRPInstance.from_file(file)
logger.info("Finetunning on evaluation instance.")
model_finetuned = finetune(model, instance)
logger.info("Starting to dynamic route.")
for delivery in tqdm(instance.deliveries):
model_finetuned = route(model_finetuned, delivery)
solution = finish(instance, model_finetuned)
solution.to_file(output_dir / f"{instance.name}.json")
def solve_cluster(deliveries):
if len(deliveries) < 2:
return [deliveries]
cluster_instance = CVRPInstance(
name=instance.name,
deliveries=deliveries,
origin=instance.origin,
vehicle_capacity=instance.vehicle_capacity,
)
cluster_solution = ortools_solve(cluster_instance,
params.cluster_ortools_params)
return [v.deliveries for v in cluster_solution.vehicles]
def solve(
instance: CVRPInstance,
params: Optional[KmeansPartitionORToolsParams] = None,
) -> Optional[CVRPSolution]:
params = params or KmeansPartitionORToolsParams.get_baseline()
num_deliveries = len(instance.deliveries)
num_clusters = int(
params.fixed_num_clusters
or np.ceil(
num_deliveries / (params.variable_num_clusters or num_deliveries)
)
)
logger.info(f"Clustering instance into {num_clusters} subinstances")
clustering = KMeans(num_clusters, random_state=params.seed)
points = np.array(
[[d.point.lng, d.point.lat] for d in instance.deliveries]
)
clusters = clustering.fit_predict(points)
delivery_array = np.array(instance.deliveries)
subsinstance_deliveries = [
delivery_array[clusters == i] for i in range(num_clusters)
]
subinstances = [
CVRPInstance(
name=instance.name,
deliveries=subinstance.tolist(),
origin=instance.origin,
vehicle_capacity=instance.vehicle_capacity,
)
for subinstance in subsinstance_deliveries
]
subsolutions = [
ortools_solve(subinstance, params.ortools_params)
for subinstance in subinstances
]
return CVRPSolution(
name=instance.name,
vehicles=[v for sol in subsolutions for v in sol.vehicles],
)
def solve(
instance: CVRPInstance,
params: Optional[KmeansAggregateORToolsParams] = None,
) -> Optional[CVRPSolution]:
params = params or KmeansAggregateORToolsParams.get_baseline()
num_deliveries = len(instance.deliveries)
num_clusters = int(params.fixed_num_clusters
or np.ceil(num_deliveries /
(params.variable_num_clusters or 1)))
logger.info(f"Clustering instance into {num_clusters} subinstances")
clustering = MiniBatchKMeans(num_clusters, random_state=params.seed)
points = np.array([[d.point.lng, d.point.lat]
for d in instance.deliveries])
clusters = clustering.fit_predict(points)
delivery_array = np.array(instance.deliveries)
deliveries_per_cluster = [
delivery_array[clusters == i] for i in range(num_clusters)
]
def solve_cluster(deliveries):
if len(deliveries) < 2:
return [deliveries]
cluster_instance = CVRPInstance(
name=instance.name,
deliveries=deliveries,
origin=instance.origin,
vehicle_capacity=instance.vehicle_capacity,
)
cluster_solution = ortools_solve(cluster_instance,
params.cluster_ortools_params)
return [v.deliveries for v in cluster_solution.vehicles]
def aggregate_deliveries(idx, deliveries):
return Delivery(
id=str(idx),
point=deliveries[0].point,
size=sum([d.size for d in deliveries]),
)
subsolutions = [
deliveries for group in deliveries_per_cluster
for deliveries in solve_cluster(group.tolist()) if group.any()
]
aggregated_deliveries = [
aggregate_deliveries(idx, s) for idx, s in enumerate(subsolutions)
]
aggregated_instance = CVRPInstance(
name=instance.name,
deliveries=aggregated_deliveries,
origin=instance.origin,
vehicle_capacity=instance.vehicle_capacity,
)
aggregated_solution = ortools_solve(aggregated_instance)
vehicles = [
CVRPSolutionVehicle(
origin=v.origin,
deliveries=[
d for v in solve_cluster([
d for groups in v.deliveries
for d in subsolutions[int(groups.id)]
]) for d in v
],
) for v in aggregated_solution.vehicles
]
return CVRPSolution(
name=instance.name,
vehicles=vehicles,
)
eval_files = ([eval_path]
if eval_path.is_file() else list(eval_path.iterdir()))
train_path = Path(args.train_instances)
train_path_dir = train_path if train_path.is_dir() else train_path.parent
train_files = ([train_path]
if train_path.is_file() else list(train_path.iterdir()))
# params = params_class.from_file(args.params) if args.params else None
params = None
output_dir = Path(args.output or ".")
output_dir.mkdir(parents=True, exist_ok=True)
train_instances = [CVRPInstance.from_file(f) for f in train_files[:240]]
logger.info("Pretraining on training instances.")
model = pretrain(train_instances)
def solve(file):
instance = CVRPInstance.from_file(file)
logger.info("Finetunning on evaluation instance.")
model_finetuned = finetune(model, instance)
logger.info("Starting to dynamic route.")
for delivery in tqdm(instance.deliveries):
model_finetuned = route(model_finetuned, delivery)
solution = finish(instance, model_finetuned)
def solve(file):
instance = CVRPInstance.from_file(file)
solution = method(instance, params)
solution.to_file(output_dir / f"{instance.name}.json")
