Detection insulator with ssd_mobilenet_v1 custom trained network.
After all steps of first steps with NVidia Jetson Nano development board (link - https://developer.nvidia.com/embedded/learn/get-started-jetson-nano-devkit) we need to make all requirement installation for tensorflow inference tests with custom ssd_mobilenet_v1 network (trained to detect insulators of power supply substation).
- Installing base dependancies and pip
sudo apt-get update
sudo apt-get install git cmake
sudo su
apt-get install libatlas-base-dev gfortran
apt-get install libhdf5-serial-dev hdf5-tools
sudo apt-get install python3-dev
wget https://bootstrap.pypa.io/get-pip.py
sudo python3 get-pip.py
sudo apt install python3-testresources
sudo rm get-pip.py
2.tensorflow and others
sudo pip install numpy
sudo pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/jp/v42 tensorflow-gpu==1.13.1+nv19.3
sudo pip install scipy
sudo pip install keras
this libraries need good internet connection and time for installlation ( from 15 to 40 minutes)
Now any of pre-trained Keras and tensorflow models can run on Jetson Nano board.
- installing pretrained models for TensorRT from https://github.com/NVIDIA-AI-IOT/tf_trt_models
cd ~
git clone --recursive https://github.com/NVIDIA-Jetson/tf_trt_models.git
cd tf_trt_models
./install.sh python3
You have all installed TensorRT models in /tf_trt_models and also in /tf_trt_models/third_party/models all Tensorflow models zoo (https://github.com/tensorflow/models)
Directory and file structure
- resources_tf - tensorflow trained graph for Insulator detection. this frozen model transferred to TensorRT graph inside
- resourses_trt directory. Also tis directory has python scripts to prepare TensorRT graph from Tensorflow graph -frozen_graph_to_trt.py - script to transfere TF frozen graph to TRT -pb_to_pb_trt_transfere.py - also the script with othe approach to prepare TRT frozen graph -trt_graph.pb - frozen TRT graph
- test/images - directory for testing images
-detect_tf.py - script to test tensorflow speed of Inference -detect_trt.py - script to test TensorRT speed
results
Testing with tensorflow frozen graph give about 0.07sec per one image (~15FPS) Testing with TensorRT frozen graph give about 0.09 sec per image (~11FPS) Loading of TF program to memory - about 15 seconds Loading TRT program for execution about 200 seconds
some problems with TRT approach is could be the training of base tensorflow graph which is used for TRT have done with different sized images ( not 224x224 or 300x300 ) - but thisis not definitely - need to understand later
also recognition of TensorRT graph have bad results with some amount of insulators at the picture and small insulators...
example
Update!
with this approach https://github.com/NVIDIA-AI-IOT/tf_trt_models/blob/master/examples/detection/detection.ipynb
I have recieved better result (about 20fps) with TensorRT library....
