We hacked the uff_converter_tf tool python code and make it can converter our .pb file to .uff with custom operation embedeed, we failed at call ICaffeParser::parse and it log
ERROR: UFFParser: Validator error: RealDiv: Unsupported operation _RealDiv
ERROR: sample_uff_vapd: Fail to parse
ERROR: sample_uff_vapd: Model load failed
We reviewed the TensorRT plugin sample code, and found ICaffeParser::setPluginFactory(), but setPluginFactory() is dismissed in IUffParser interface, so we can’t register our plugin factory in UffParser. Do you think it is a good idea to make parser interface has the same function?
There is an advice recommend removing the customer op in .pb file, restore them in TensorRT network after imported from .uff. This work flow has a huge gap with caffe plugin, any more advice?
Python-based uffparser doesn’t support plugin factory.
Please switch to C+±based uffparser to enable the custom operation.
I think the uff-converter-tf is a tool written in python, and UffParser we are using is C+±based, we just clone the UFF MNIST sample which is C++based, and we went through NvUffParser.h file, still got no plugin related function.
We are checking this issue internally. Will update information with you later.
Thanks and sorry for the inconvenience.
TensorRT 3 don’t provide plugin interface for uffpaser.
Currently, plugin API is only available for Caffe users.
We are sorry for such late information.
Also thanks for your feedback.
Hope next release will support it, thanks.
We will enable plugin API support for uffparser in our future release, but no concrete schedule currently.
Sorry for the inconvenience.