I’m glad to know that TensorRT integrate with tensorflow 1.7, but there isn’t a available TensorRT 4 release candidate, can you tell me about the release date? Now I use TensorRT 3 to optimize Mask Rcnn (tensorflow model), but there are many unsupported layers, although I can utilize plugin API to resolve it, but it’s too troublesome. So any help will be appreciated!
Thanks for your question. But we cannot disclosure our schedule here.
Please keep following our update and announcement for the latest information.
Hi,I have another question: I convert model to uff model by convert_to_uff.py, then parse the model and build engine as following:
ICudaEngine* engine = builder->buildCudaEngine(*network);
but I get an error:ERROR: rpn/conv0/BiasAdd:kernel weights has count 1179648 but 368640 was expected.(TensorRT 3.0.4 and tensorflow 1.4)
All right,I know the reason:the conv kernel size is 33512,input feature map size is 8080256(hwc) ,I transpose it to 2568080(chw),but tensorrt think it is hwc, so weights of the conv layer is 3380*512=368640.
It’s recommended to file a dedicated topic for a new issue.
This will help other users to find information efficiently.
Have you already fixed the issue in comment #3?
If not, feel free to let us know.
Did you succeed to run mask Rcnn in TensorRT?
How is the performance in production?