先把checkpoint模型转为saved model（这个过程中没报错），再把saved model转onnx报错了

saved model转onnx时，报错：
ValueError: Node 'gradients/resnet_v1_101_4/block4/unit_3/bottleneck_v1/conv3/BatchNorm/FusedBatchNorm_grad/FusedBatchNormGrad' has an _output_shapes attribute inconsistent with the GraphDef for output #3: Dimension 0 in both shapes must be equal, but are 0 and 2048. Shapes are [0] and [2048].

请问，怎么解决？