imgsort

Command-line utility to build an image classification model out of organised image folders and then sort random images into those subsubfolders.

About

• Uses EfficientNet-Lite, released in March 2020. It is a smaller, newer version of EfficientNet made to run on lower end hardware while still outperforming Inception V4 (using the B4 model size).
• Fun little project made in about 3 days

Installation

$ git clone <this repo>
$ cd imgsort
# Create a virtual environnement,
# highy recommended unless you want 1.5gb of libraries in your global environment
$ python3 -m venv ./venv/  
$ ./venv/Scripts/Activate.ps1 # Or activate.sh on bash/linux environements.
(imgsort) $ python3 -m pip install -r requirements.txt
(imgsort) $ python3 ./src/app.py -h

Docker image soon.

Usage

Prepare a folder with subfolder and images inside of those repesenting the name of the subfolder:

C:\Documents\Pictures\My Sorted Pictures
├───cats
│       cat_1.png
│       cat_2.png
│       ...
├───dogs
│       dog_1.png
│       dog_2.png
│       ...
└───humans
        human_1.png
        human_2.jpg
        ...

Train the model, n should be your number of subfolders or classes. You should at least have 200 images per subfolder for the model to train correctly. Avoid categories that are very simmilar, especially categories that overlap.

imgsort train "C:\Documents\Pictures\My Sorted Pictures" "C:\Documents\Imgsort Model" -n 3

Before sorting, make sure subfolders are present, imgsorter will not create them:

C:\Documents\Downloads
│        file (1).png
│        file (2).jpg
│        file (3).png
│        file (4).jpg
│        file (5).jpg
│        ...
├───cats
├───dogs
└───humans

Sort images in a folder

imgsort sort "C:\Documents\Downloads" "C:\Documents\Imgsort Model"

After sorting:

C:\Documents\Downloads
├───cats
│        file (2).jpg
│        file (4).jpg
│        file (1).png
│        ...
├───dogs
│        file (5).jpg
│        ...
└───humans
         file (3).png
         ...

Options

Training options

• -e [integer] or --epochs [integer] the number of training epochs to perform, value should ideally sit between 2 and 6
• -n (required) number of categories (subfolders)

Sorting options

• -t [integer] or --threshold [integer] only moves images only if a category is greater than threshold

Known Issues

• Doesn't use the batch feature from TensorFlow for sorting, doesn't multithread sorting.
• Not optimised at all for TPUs/GPUs, only runs on CPU, look into CUDA or mROC tensorflow installs if needed.
• Will only work with png, jpg, and gif (first frame) images. jpeg files ignored.
• Expects empty subfolders with the names of the categories in the unsorted folder.
• 1.5gb of dependencies are pulled (won't fix)
• no setup.py