SiameseSimNet

A Simple Pytorch Siamese Network for Image Similarity Learning

Data folder structure

The images shall be organized in a root_dir such that there exist a sub-folder per class name,

root_dir
├── Class1
|   ├── image01.png/jpg
|   ├── image02.png/jpg
├── Class2
|   ├── image01.png/jpg
|   ├── image02.png/jpg
└── Class3
    ├── image01.png/jpg
    ├── image02.png/jpg

Training phase

Modify config.py accoding to your application and run train.py with the following pattern,

python3 train.py --data-path <path to root_dir> \ # Mandatory
                 --class_names Class3,Class2,Class1 \ # Optional - defines a custom order of class indices
                 --use-wandb \ # Optional - whether to use WandB for logging

The results of training will be saved into a checkpoint directory named weights.

Inference phase

In order to select a set of representative candidates of each class from the training data, we first run select_ref_images.py that creates an output directory with reference images that will be used together with the pre-trained model for inference. For an example of how to do inference see this Kaggle demo notebook which is part of the RSNA 2024 Lumbar Spine Degenerative Classification competition.