Term: Spring 2019

• Chen, Xinyi xc2464@columbia.edu
• Dubova, Elena ed2801@columbia.edu
• Hu, Xinyi xh2383@columbia.edu
• Ma, Qiaozhen qm2138@columbia.edu
• Xiao, Caihui cx2225@columbia.edu

Project summary: In this project, we performed model evaluation and selection for predictive analytics on image data.

• Given 3000 images: 1500 low resolution images and their corresponding high resolution images, we extracted the features of these images and used them to train our models.

• We created three models for image super resolution. The first model is the baseline model -- Gradient Boosting Machines. Then we built XGBoost model and CNN model to further improve image resolution .

• The baseline model has mean psnr = 25.82641. The XGBoost model has mean psnr = 25.61457. CNN model provides mean psnr = 27.0 on 55 epochs with batch size of 150. It processes 1500 test images in 19 mins on a GPU - enabled virtual machine in Google Colaboratory. After comparing the performances of XGBoost model and CNN model, our team decided to run the CNN model during the presentation as it costs much less time to run and gives us the best result that we could have.

Contribution statement: (default)

• Qiaozhen Ma: Designed and trained the baseline model with Xinyi Chen in R. Tuned the minNode, shrinkage, numTrees for the GBM model. Fit the model with different categories and get the psnr and cv-error.

• Xinyi Chen: Designed and trained the baseline model with Qiaozhen Ma in R. Tuned the minNode, shrinkage, numTrees for the GBM model. Fit the model with different categories and get the psnr and cv-error.

• Elena Dubova: Designed, trained and tested CNN model in Python. Tested several cloud environments for the model, including Windows Server 2016, Debian Linux on Google Cloud, Google Colaboratory.

• Caihui Xiao: Designed and trained the XGBoost model with Xinyi Hu in R. Tuned the depths and learning rate of XGBoost model. Check the reproducibility of CNN.

• Xinyi Hu: Designed and trained the XGBoost model with Caihui Xiao in R. Tuned the depths and learning rate of XGBoost model. Predicted the whole 1500 low resolution images using XGBoost model and output its mean psnr value.

Following suggestions by RICH FITZJOHN (@richfitz). This folder is orgarnized as follows.

proj/
├── lib/
├── data/
├── doc/
├── figs/
└── output/

Please see each subfolder for a README file.