Inference with inference.py

The file inference.py accepts the path to the directory with two subdirectories ("aligned" and "not_aligned") of images. It then runs the image classification model on all images in the provided paths, makes predictions for them and prints the performance metrics: accuracy and F-score in percent points. It also reports the averate latency of predictions. For more information, run python3 inference.py --help.

Presentation

Link

Installation

Install python virtual environment.

pip install -r requirements.txt

pip install -e .

For mac OS users (for labeling script):

brew install python-tk

Labeling

Change image_folder, local_prefix, server_prefix in src/labeling/labeling.py for labeling.

labeling.py will produce two files:

• class_align.txt
• class_not_align.txt

in the src/labeling folder.

Symlinks

Symlinks used for labeled images to avoid copying. Call make symlinks after getting the labeled lists with paths.

To remove symlinks, call make clean.

Training

• For training the models, run the corresponding bash files:
• For EfficientNetV2: bash main_efficientnet.sh
• For Swin Transformer: bash main_swin.sh
• For ConvNeXt: bash main_convnext.sh