I tried to convert and import another .pb file today, and found that there is so many customer operation waiting for implementation, Switch, LoopCond, Less, Fill, etc. Does NVIDIA has plan to update TensorRT to make it more mature with import from TensorFlow function?
TensorRT will keep updating and supports more operation in the furture.
But it’s unlikely to support all the operation of TensorFlow.
A better idea is to allow user to implement their customed layer with plugin API.
Currently, we have received several issues on uffparser with plugin API.
We are clarifying the issue and will come out a conclusion soon.
For your use-case, we recommend to implement the non-supported layer with plugin API rather than wait for TensorRT official support.
Great, I think NVIDIA got the idea now. we hacked the uff_converter_tf python based tool, and it can convert our .pb file to .uff file, but the C++ -based UffParser doesn’t recognize the non-supported layer, or it doesn’t generate plugin factory call, instead, it just fail, when we clone the sampleUffMNIST(C++) sample and let it import our .uff file in TensorRT 3.0 SDK.
We hope the UffParser implement the setPluginFactory function and generate the top view of network and code for us, as it can help us to quantify our work. Thanks.