diff --git a/FasterRCNNDetection/model/roi_module.py b/FasterRCNNDetection/model/roi_module.py
index 6b45f89..4536881 100644
--- a/FasterRCNNDetection/model/roi_module.py
+++ b/FasterRCNNDetection/model/roi_module.py
@@ -11,7 +11,7 @@
 Stream = namedtuple('Stream', ['ptr'])
 
 
-@cupy.util.memoize(for_each_device=True)
+@cupy.memoize(for_each_device=True)
 def load_kernel(kernel_name, code, **kwargs):
     cp.cuda.runtime.free(0)
     code = Template(code).substitute(**kwargs)
diff --git a/FasterRCNNDetection/model/utils/nms/non_maximum_suppression.py b/FasterRCNNDetection/model/utils/nms/non_maximum_suppression.py
index c488b52..beb5e43 100644
--- a/FasterRCNNDetection/model/utils/nms/non_maximum_suppression.py
+++ b/FasterRCNNDetection/model/utils/nms/non_maximum_suppression.py
@@ -8,12 +8,12 @@
     import warnings
     warnings.warn('''
     the python code for non_maximum_suppression is about 2x slow
-    It is strongly recommended to build cython code: 
+    It is strongly recommended to build cython code:
     `cd model/utils/nms/; python3 build.py build_ext --inplace''')
     from ._nms_gpu_post_py import _nms_gpu_post
 
 
-@cp.util.memoize(for_each_device=True)
+@cp.memoize(for_each_device=True)
 def _load_kernel(kernel_name, code, options=()):
     cp.cuda.runtime.free(0)
     assert isinstance(options, tuple)