Jetson Test Projects

Medium to high complexity Machine learning projects to compare the inference times between optimized and non optimized nueral networks

1) Image Classification using Resnet50

Required Libraries:

python == 3.6 
tensorrt == 8.0.1.6 
pycuda == 2021.1 
numpy == 1.19.4 
opencv == 4.1.1 #(preinstalled with jetpack) .
argparse == 1.1 
onnx == 1.10.1 
tensorflow == 2.5.0 

useful links for some of the important libraries:

Tensorflow :

1. https://docs.nvidia.com/deeplearning/frameworks/install-tf-jetson-platform/index.html
2. https://www.youtube.com/watch?v=ynK-X5IPu1A

(Make sure to specify right jetpack version if you are following the tutorial from the link 2)

onnx (both onnxruntime and tf2onnx):

1. https://github.com/onnx/tensorflow-onnx

Tensorrt engine:

1. https://docs.nvidia.com/deeplearning/tensorrt/quick-start-guide/index.html

Very Important Note: As both unoptimized and optimized versions are running on GPUs make sure tensorflow is using GPU

by using the following command in python environment:

import tensorflow as tf
print(len(tf.config.list_physical_devices('GPU')))
# The output should be >= 1 .

Instructions to Run the tests:

place the data into data/:

dataset can be found at : https://www.kaggle.com/alessiocorrado99/animals10?select=raw-img

After placing the data : the directory should look like : data/raw-img

Download the onnx file from : https://drive.google.com/drive/u/2/folders/1yYDD2_teYR9wPLwCjV3OMMtrjWTyOBu0

Place the onnx file in the respective paths:

onnx : optimized/models/onnx/resnet50.onnx

Run the following commands in the project directory

unoptimized model on gpu:

To show the info of the python program (Helps to know if all the libraries are installed correctly):

$ bash run_keras.sh -debug 

Without info (Only the intended output):

$ bash run_keras.sh -no-debug 

For optimized Program,

To show the info of the python program:

$ bash run_optimized.sh -debug 

Without info:

$ bash run_optimized.sh -no-debug 

2) Image Segmentation

Additional Libraries needed for this project:

sciimage (package name : scikit-image) and cityscapesscripts

Download the onnx file and the hd5 file from the following link: https://drive.google.com/drive/folders/1NypCjzfGBJEQ6Gkj4k7n4Ipwcl9nix42?usp=sharing

Place the onnx file in Image_segmentation_nd/optimized/models/onnx

Place the hd5 file in Image_segmentation_nd/

Place the any one of the folders (e.g cane) from the data set downloaded for Image Classification Example in the Image_segmentation_nd/data/

For initial test, only keep three-four images in the input directory folder. For stress test more images can be added further

Running test scripts

Tensorflow : bash run_tf.sh -d no -b 1 (set the flag -d to yes to show the entire output)

Tensorrt : bash run_optimized.sh -d no -b 1 (set the flag -d to yes to show the entire output)

The outputs will be stored in Image_segmentation_nd/data/out directory