Tensorrt3: How to set shape_node of uff.graph.conv_transpose?

Hi, I am trying to convert a mxnet model with deconv layer into UFF format. Currently, everything works fine if the network does not include any deconv layers (i.e. resnet, inception, alexnet).

But when converting the a CNN with deconv layers, tensorrt raise an error like below when building the engine.

python3: uff/UffParser.cpp:611: std::vector<long int> getOutputShapeSpatial(const Constant&, const Fields&, std::vector<int>): Assertion `inputOrder.size() == outShapeUnordered.size()' failed.

I guess the problem might be caused by the “shape_node” arg in “uff.graph.conv_transpose”. Currently, I create a constant node containing the output shape in numpy.ndarray format and use its name as the “shape_node” arg.

Can anyone provide a detailed example about creating a deconv layer in UFF format?