This repository contains the implementation of the Image Processing and Deep Learning assignment, from the 2018-2019 Computer Vision, Robotics, and Machine Learning course, at University of Surrey.
Coursework mark: 110%
The assignment implementation covers the training and evaluation of a range of CNN architectures, hyper-parameters, data splits/augmentation, and training strategies. The dataset chosen was an augmented version of the Stanford Cars dataset, where 4 additional classes were added, rounding the count to 200.
Evaluation is performed with and without transfer learning, through use of the following metrics:
- Confusion matrix/heatmap
- Loss
- Top-1 Accuracy
- Top-5 Accuracy
- sklearn classification report
TensorboardX plots are generated from the following:
- Training accuracy
- Training loss
- Training loss (per epoch)
- Validation accuracy
- Validation loss
- Validation loss (per epoch)
- Test accuracy
- Test loss
- Test Top-5
Early stopping is employed using the patience policy, with the best stored weights being reverted. Checkpoints are saved/loaded in/from the checkpoints directory, along with the confusion matrix, and .txt report containing the above metrics.
Ensemble training is performed using maximum vote, and TensorboardX plots are generated for accuracy and loss metrics, per ensemble size.
A rotating file logger is used, logging the training process to the training.log file.
The repository consists of the following files:
- preProcessData.py - This file reads in the dataset, gets data statistics, resizes, and splits the data into several pre-defined training/validation/test splits
- model.py - This file contains the bulk of the model code, where the model training and evaluation takes place
- ensemble.py - This file contains a wrapper class for managing and evaluating trained model ensembles
- index.py - The main entry point for experiments