CS 542 Class Challenge: Image Classification of COVID-19 X-rays Total Points: 100 In this class challenge, we will classify X-ray images. The data we will use has been collected by Adrian Xu, combining the Kaggle Chest X-ray dataset with the COVID-19 Chest X-ray dataset collected by Dr. Joseph Paul Cohen of the University of Montreal. The data can be downloaded here. When you extract the data you will have two folders: two that will be used for a binary classification task (Task1), and all that will be used for multi-class classification (Task2). An ipython notebook template is provided for each task. - [30 points] Task1 Train a deep neural network model to classify normal vs. COVID-19 X- rays using the data in the folder two. Starting from a pre-trained model typically helps performance on a new task, e.g. starting with weights obtained by training on ImageNet. After training is complete, visualize features of training data by reducing their dimensionality to 2 using t-SNE. If your extracted features are good, data points representing a specific class should appear within a compact cluster. - [30 points] Task2 Train a deep neural network model to classify an X-ray image into one of the following classes: normal, COVID-19, Pneumonia-Bacterial, and Pneumonia-Viral, using the folder all. Explore at least two different model architectures for this task, eg. AlexNet vs. VGG16. After training is complete, visualize features of training data by reducing their dimensionality to 2 using t-SNE. If your extracted features are good, data points representing a specific class should appear within a compact cluster. - [10 points] Challenge How well your best model performs with respect to the class. - [30 points] Report o [5 points] Describe the architectures used in detail: layers, layer dimensions, dropout layers, etc. for both tasks. List the optimizer, loss function, parameters, and any regularization used in both tasks o [10 points] Comparison of the performance of different architectures for the second task and relating this to the architecture and parameter settings used o [10 points] Plot and comment on the accuracy and the loss for both tasks o [5 points] Plot and comment on the t-SNE visualizations o [Bonus: 5 points] Run the training on a GPU on the SCC cluster and include a CPU vs. GPU training time comparison by taking snapshots from your terminal Submission: Please complete the class challenge and submit a pdf file containing: task 1 code, task 2 code, and the report on GradeScope. The deadline for this class challenge is: Apr 29,2021.
欢迎咨询51作业君