diff --git a/configs/cityscapes/README.md b/configs/cityscapes/README.md
index 8785cdb93..b8595f79c 100644
--- a/configs/cityscapes/README.md
+++ b/configs/cityscapes/README.md
@@ -195,10 +195,13 @@ def clip_weights_from_pretrain_of_coco_to_cityscapes(f, out_file):
 	print("f: {}\nout_file: {}".format(f, out_file))
 	torch.save(m, out_file)
 ```
-Step 3: modify the `input&solver` configuration in the `yaml` file, like this:  
+Step 3: modify the `input&weight&solver` configuration in the `yaml` file, like this:  
 ```
+MODEL:
+  WEIGHT: "xxx.pth" # the model u save from above code
+
 INPUT:
-  MIN_SIZE_TRAIN: (800, 832, 863, 896, 928, 960, 992, 1024, 1024)
+  MIN_SIZE_TRAIN: (800, 832, 864, 896, 928, 960, 992, 1024, 1024)
   MAX_SIZE_TRAIN: 2048
   MIN_SIZE_TEST: 1024
   MAX_SIZE_TEST: 2048
@@ -210,4 +213,5 @@ SOLVER:
   STEPS: (3000,)
   MAX_ITER: 4000
 ```
+Step 4: train the model.  
 
diff --git a/maskrcnn_benchmark/layers/dcn/deform_conv_func.py b/maskrcnn_benchmark/layers/dcn/deform_conv_func.py
index ddc92bb0c..a276a05fe 100644
--- a/maskrcnn_benchmark/layers/dcn/deform_conv_func.py
+++ b/maskrcnn_benchmark/layers/dcn/deform_conv_func.py
@@ -1,5 +1,6 @@
 import torch
 from torch.autograd import Function
+from torch.autograd.function import once_differentiable
 from torch.nn.modules.utils import _pair
 
 from maskrcnn_benchmark import _C
@@ -67,6 +68,7 @@ def forward(
         return output
 
     @staticmethod
+    @once_differentiable
     def backward(ctx, grad_output):
         input, offset, weight = ctx.saved_tensors
 
@@ -201,6 +203,7 @@ def forward(
         return output
 
     @staticmethod
+    @once_differentiable
     def backward(ctx, grad_output):
         if not grad_output.is_cuda:
             raise NotImplementedError
diff --git a/maskrcnn_benchmark/layers/dcn/deform_pool_func.py b/maskrcnn_benchmark/layers/dcn/deform_pool_func.py
index f18fdd4cb..2f7810b23 100644
--- a/maskrcnn_benchmark/layers/dcn/deform_pool_func.py
+++ b/maskrcnn_benchmark/layers/dcn/deform_pool_func.py
@@ -1,5 +1,6 @@
 import torch
 from torch.autograd import Function
+from torch.autograd.function import once_differentiable
 
 from maskrcnn_benchmark import _C
 
@@ -60,6 +61,7 @@ def forward(
         return output
 
     @staticmethod
+    @once_differentiable
     def backward(ctx, grad_output):
         if not grad_output.is_cuda:
             raise NotImplementedError