Project 3 - Training VGGNet (Due 10/09)

Objectives:

The objective of this project is to be able to use tensorflow to write benchmark CNN structures, i.e., VGGNet, and solve a larger-scale problem. Further, through this project, you learn how to utilize pretrained models and fine-tune on it to solve your own problem.

Data set used:

CIFAR-10.

This site gives you an overview of what CIFAR-10 is and how to download the dataset.
This site gives you an update on who is the best in CIFAR-10.

Requirements:

Task 1: Train a VGGNet from scratch.
- Study the sample code from Yang Song in Lecture 7 (VGGFace) on how to train a VGGNet from scratch for face recognition purpose. Adapt the design to CIFAR-10.
- Modify hyperparameters to get to the best performance you can achieve.
Task 2: Use a pretrained VGG model.
- Download a pretrained VGG model.
- Fine-tune the hyperparameters (i.e., retrain the pretrained model using CIFAR-10 dataset) for better performance
Note that this is an open-ended problem. The leaderboard will be frequently updated to report the best in class.

Report

Please submit the following through Canvas before midnight on the due date.

The evaluation graphs (e.g., convergence graph) that you generated from both Tasks 1 and 2. Discuss on the effect of different hyperparameters settings.
Discuss various design options regarding cost function and activation function. Discuss your experience with training a CNN structure from scratch vs. using a pretrained model. This part of the report should not exceed 2 pages (with double space)
(For 692 students only) Read [LeCun:1998], [AlexNet:2012], [GoogLeNet:2014], [VGGNet:2014], [ResNet:2015], and [SENet:cvpr:2018]. Discuss the core ideas that make the structure such a success. This part of the report should not exceed 3 pages (with double space).