# ===================================================================
# part_x1_feat = x1.features_at_coordinates(x1.C[:3,:].float())
# part_x1_coord = x1.C[:3,:]
# part_x1 = ME.SparseTensor(
# coordinates = part_x1_coord,
# features = part_x1_feat,
# coordinate_manager = x1.coordinate_manager,
# )
# out = part_x1 + x1 # out and x1 may not have the same order
# x2_f = x2.F
# x1_f = x1_f**2 + x2_f**2
# y = ME.SparseTensor(features = x1_f, coordinate_map_key=x1.coordinate_map_key, coordinate_manager=x1.coordinate_manager)
pool = ME.MinkowskiGlobalSumPooling()
loss = pool(out).F
loss.backward()
# # if there isnot nearby voxel within kernel size, then no spatial reduction will occur
# # pool = ME.MinkowskiSumPooling(kernel_size=6,stride=2,dimension=3)
# loss = pool(y).F
# loss.backward()
print(feats, feats.grad)
