Weight Initialization for Semantic Segmentation

This project was created by Justin, Ayleen, Jacky, Hanson and Harry (Group 12) for our MIE424 Final Project. We aim to assess the impact of different weight initializations, namely Zero Initialization, Kaiming Uniform and Normal, Xavier Uniform and Normal, and GradInit on the performance of an FCN32s model architecture for semantic segmentation.

Environment Dependencies

To install the necessary dependencies, you can use the following list as a guide. Some packages require specific versions or higher:

• fcn >= 6.1.5
• numpy
• Pillow
• pytz
• scipy
• torch >= 0.2.0
• torchvision >= 0.1.8
• tqdm

Dataset

The Pascal VOC 2012 Dataset was used for this project, which can be downloaded from this link: http://host.robots.ox.ac.uk/pascal/VOC/voc2012/#devkit

Click the "training/validation" data (2GB tar) to download.

Usage

Note that the directory paths in the code might change depending on your directory structure. Please change this accordingly before running anything.

To train the model, run the train_fcn32s.py script from the command line. You can specify various arguments to customize the training process. For example, to specify which GPU to use and to set the maximum number of iterations, you can use the following command:

python train_fcn32s.py -g 0 -max_iter 10000

Please read through the train_fcn32s.py file to see all available arguments you can specify.

Acknowledgments

Our FCN32s architecture was referenced from the Fully Convolutional Networks for Semantic Segmentation paper by Long et. al. We also used the GradInit code by Chen Zhu for GradInit weight initialization.

We would like to thank Prof. Elias Khalil for teaching such an amazing course, and the MIE424 TAs as well for supporting us!!