a simple program to automatic test of keras tensorflow models: an extention upon the idea of FSCS ART vs RT shown on this paper: https://www.sciencedirect.com/science/article/abs/pii/S0164121209000405
download two models one with seeded error and one is normal
seeded error https://drive.google.com/file/d/1-V_nK9en33oHJmlSd-BFpU9c82ZE8OUo/view?usp=sharing ,
no seeded error https://drive.google.com/file/d/1QyeAHTAzuHBjFEvvKI31jmC3x-f3JVES/view?usp=sharing
and put it into /Object detection server/models
cd into /ART vs RT server/ass2/angular/ type "npm start" into console and run uses npm (requires npm)
the front-end server now will be hosted on http://localhost:4200/
cd into /Object detection server/
//install opencv2 pip install opencv-python
python wsgi.py
the back-end server should be listening for incoming pictures link on http://localhost:5000/
feed the .csv or .xsls that includes metadata of pictures in the categories that you have trained for your models. (sample.xlsx uses license plates)
beside displaying on the front end the outputs picture with bounding box and confidence level will also be saved to \Object detection server\outputs
The front-end server (angular) will download the pictures from the internet. (Local picture dataset upload currently not implemented)
There are 4 cases of output
1st case: ART detect the error first (+1 for ART mark)
2nd case: RT detect the error frist (+1 for RT mark)
3rd case: ART and RT does not detect any errors (10% test of all attribute is allowed) (+1 for draw)
4th case: The stable model does not output the same result as the metadata provided in the unit test (+1 for draw) (test oracle does not have same data as unit test problem)
Processor: AMD Ryzen Threadripper 3970X 32-Core Processor, 3722 Mhz, 32 Core(s), 64 Logical Processor(s)
Graphic Processing Unit: NVIDIA GeForce RTX 2080 SUPER with compute capability: 7.5 (according to cuda )
