Reimplementation for python 3.11 and PyTorch Lightning

[timurcarstensen]

Motivation

Hi everyone,

last year I used this repository (and dataset) extensively in one of my classes and ended up reimplementing most of the code to work with python 3.11 and PyTorch lightning. I primarily had issues with installing the right PyTorch version and getting the ROI pooling functionality to work.

What

I restructured the repository to some extent:

Networks

• wR2.py → is now called DetectionModule and resides in src/modules/detection.py
• rpnet.py → fh02 is now called RecognitionModule and resides in src/modules/recognition

Training

Pretraining and training are now handled by PyTorch lightning trainers in the following files:

• src/pretrain.py → now holds the logic for pretraining (i.e. training the DetectionModule)
• src/train.py → now holds the logic for loading the pretrained weights of the DetectionModule and training the RecognitionModule

→ both networks are re-implemented as LightningModule's (i.e., training, testing, and validation loops are handled in the network class

ROI Pooling

→ the logic for ROI pooling now resides in src/modules/utils.py

Miscellaneous

• rewrote the README.md to reflect the new setup and (pre-)training procedure
• Removed the split directory (I thought that it's not necessary to include it in the repo since it's contained in the dataset download)