kimeg / image-classification-system Goto Github PK

[Image Classification System]

-Objective:

[1] This project is intended to provide a full scale environment where image data generation and classifying model generation functionalities are available through tensorflow and/or pytorch API.
[2] Throughout the development of this project, I have had tons of issues trying to implement necessary tools from scratch. This opened new learning opportunities in which I had the chance to analyze required components of the proeject, break them down into sub-problems, and merge their solutions as a whole.
[3] Although tensorflow and pytorch backends provide such utilities, I have decided to build them one by one with my own logics.

-Image Data:

[1] Description : Rectangle, triangle, and circle images are generated through customized data augmentation technique ( randomized variations in width, height, location, etc. )
[2] Size: ( these are default settings, please configure according to your needs. )
--> Training set ( 1,000 images )
--> Validation set ( 300 images )
--> Test set ( 200 images )

-Usage:

[tensorflow backend]
$ python run.py tf

[torch backend]
$ python run.py torch

-Workflow:

[1] Generate images using data augmentation.
[2] Split image data into training, validation and test sets.
[3] Train convolutional neural network model using specified deep learning framework.
[4] Test model using test set.

-Output:

The prediction results will be printed on your terminal.

-Note:

[1] This codebase may be used for benchmarking purposes, such as model quality comparison between the two deep-learning frameworks or customization of image data ( you may need to implement such methods that satisfy your own needs ).
[2] Existing image data and model files guarantee that the corresponding file generation steps are skipped for direct access to test dataset prediction.