PRMLTermProject

Term Project for Pattern Recognition and Machine Learning (2023 Fall)

Task

Use the given dataset and train an ML model for object detection.

File Structure

The project initially contains the following files and folders:

main.py: main program for training and testing.
config.json: configuration file for the program, including training, logging and dataset parameters.
dataset/: dataset folder containing the provided subset of PASCAL VOC2012 dataset (1713 photos), with annotations in Annotations/ and images in JPEGImages/, each .jpg image have a corresponding .xml annotation.
utils/: folder for the implementation of the project utilities.
- accuracy_evaluator.py: Evaluating the model accuracy by calculating mAP (Mean Average Precision), call calculate_accuracy for evaluation.
- dataset_loader.py: Loading, parsing and splitting the dataset, call load_dataset to get the train, validation, test sets and strings for each label.
- image_utils.py: Image processing utilities, including image transformation, visualization, etc.
- logging_utils.py: Logging utilities, including logging to console and exporting training logs to file.

On training the model, the following files and folders will be created:

model_weights/fasterrcnn_resnet50_fpn_weights_best.pth: The current model weights with the highest validation accuracy.

On testing the model, the following files and folders will be created:

annotated_images/: folder for saving the annotated test images.

After training and testing the model, the following files and folders will be created:

model_weights/fasterrcnn_resnet50_fpn_weights_last.pth: The model weights from the last epoch.
logs/: folder for saving the training logs, figure of loss & accuracy, configuration file and main program file, and annotated_images/ & model_weights/.

Usage

Prerequisites

As PyTorch compatability varies on different platforms, we do not provide an requirements.txt file. Please manually install the compatible version of PyTorch and other dependencies as follows:

torch torchvision numpy matplotlib rich albumentations

As for PyTorch version, we develop the project under 2.0.1, but it should work on other versions as well.

Training and Testing

The project rely solely on a configuration file config.json for all the parameters and normally doesn't require parameters from command line.

To train and test the model, use:

python main.py

Notice that:

The program will automatically download the pretrained models from the Internet if using fasterrcnn_resnet50_fpn.
The program will automatically create model_weights and logs folder for logging.

To change the configurations, refers to config.json and modify the parameters.

TODO

Implement a better way for data augmentation as it's currently implemented in main.py.
Create a more reproducible way for label strings storing as current labels string orders are different on different runs.
Implement an unified way for logging as some of them are currently directly put in main.py.
Make the config.json file more readable in the data part.