I have a few 3d convolutional resnets that I have trained in keras. Each resnet has around 30m parameters. Are these models too big for a jetson nano? I would not think so because I saw a vgg 19 network run on the nano with 5 fps.
I froze the h5 model and converted it into a pb. The pb model is around 1.2 times faster. I am now trying to further optimize these models using tensorRT. I don’t know how to convert these pb files into uff files. When I tried, the script returns “converting conv3d and addv2 into custom layers”. Is this okay? Or do I need to utilize graphsurgeon and if so, how do I do that?
Also, I am running TensorRT on google colab. Here is how I download tensor RT
from termcolor import cprint
from google.colab import drive
!sudo dpkg -i ‘/content/drive/My Drive/tensorrt/nv-tensorrt-repo-ubuntu1804-cuda10.0-trt22.214.171.124-ga-20190427_1-1_amd64.deb’
!sudo apt-get install -y --no-install-recommends libnvinfer5=5.1.5-1+cuda10.0
!sudo apt-get install -y --no-install-recommends libnvinfer-dev=5.1.5-1+cuda10.0
!sudo apt-key add ‘/var/nv-tensorrt-repo-cuda10.0-trt126.96.36.199-ga-20190427/7fa2af80.pub’
!sudo apt-get update
!sudo apt-get install tensorrt
#!sudo apt-get install python3-libnvinfer-dev
!sudo apt-get install uff-converter-tf
!pip3 install pycuda
!cp -r /usr/src/tensorrt/samples/python/uff_ssd/plugin/ .
!cp -r /usr/src/tensorrt/samples/python/uff_ssd/CMakeLists.txt .
cprint(“Finished install necessary packages, please restart the runtime now…”, “red”)`
How can I download later version of tensorRT on google colab with the uff module?