All the code that is involved in processing the data set, training the model and exporting it is within the DCNN Resources directory.
Dog Extractor is the program that has been written to process the Stanford Dogs dataset ready from the model to train off. The datasets images and annotations can be downloaded and extracted into their respective folders within Original Dataset from: http://vision.stanford.edu/aditya86/ImageNetDogs/
This Solution was written in Visual Studio 2019 with C++ tools, though other versions of Visual studio may also work.
The program relies on Boost and OpenCV which can be installed using vcpkg. Instruction on the instalation and use of vcpkg can be found here: https://github.com/microsoft/vcpkg#quick-start-windows
Main.cpp is then simply run in the solution and the data set will be processed.
Several programs to follow were written in Python, more specifically version 3.8.6 which can be downloaded from here: https://www.python.org/downloads/release/python-386/
The versions of the following packages can then be installed using PIP
| Package | Version |
|---|---|
| matplotlib | 3.3.3 |
| tensorflow | 2.4.0 |
| tensorflow-hub | 0.12.0 |
| Pillow | 8.0.1 |
| numpy | 1.19.3 |
| Flask | 1.1.4 |
| FlaskRESTful | 0.3.9 |
If an NVIDIA GPU is installed in your system, the CUDA toolkit also needs to be installed to take full advantage of it, this can be downloaded from here: https://developer.nvidia.com/cuda-downloads
train_model.py can simply be run from IDLE and will start training as long as the processed data is in place.
After train_model.py has been run load_checkpoints.py can then be run in IDLE and it will load the trained checkpoints, clean the network structure then export a completed model which can be copied into the Web API's model photo.
Once a trained model has been put in the models directory, check_acc.py can be run in IDLE to evaluate its mean accuracy on the entire original dataset.
app.py is a flask app so needs to be run slightly differently to the other python scripts, in a command window the Web_API directory must be CD'd into then the following command needs to be run
flask run –host=0.0.0.0
With '-host=0.0.0.0' it making accessible from all network interfaces on a machine. Next a firewall rule must be set up to allow tcp requests on port 5000. Finally if external accessibility is required a port forwarding rule must be set up in your router.
Once the flask service is running, tester.py can be run in IDLE to check it is working.
The source directory contains the Flutter application, to build, compile and run this you firstly need to install Flutter and Android Studio, the full instructions for this can be found here: https://flutter.dev/docs/get-started/install/windows
The src directory can then be opened with android studio. Once opened the 'pub get' command must be run to install the apps dependencies. To be able to build the app a unique signing key will need to set up following this guide here: https://flutter.dev/docs/deployment/android
Finally the address on Line 5 of lib/detect_breed.dart must be updated to that of your machine IP address, or external if port forwarding has been setup.
The app can now be built and run on your device of choice.
This repository is the sole creation of Jaidon James Rymer.
Produced as part of his final year project.
Submitted in partial fulfilment of the requirements for the Degree of BSc(Hons) Computer Science at School of Computer Science, College of Science, University of Lincoln