COMP 8123 Assignment
(Contribute 30% to your final mark)
In this assignment, you will build and train a CNN model to do image classifications. The data is cifar10 (https://www.cs.toronto.edu/~kriz/cifar.html). The CIFAR-10 dataset consists of 60000 32×32 colour images in 10 classes, with 6000 images per class. There are 50000 training images and 10000 test images. You can freely use less training and testing images if you need to run your model faster.
Part 1. Image classification using CNN (60%)
In this part, you will build a CNN network to achieve the image classification task. You should achieve the following requirements:
Handcraft a CNN network to classify the images (i.e. build convolutional and pooling layers by yourself). You can use Keras to build the network, however, you cannot use pre-trained model to do the classification. Your network should achieve at least 50% testing accuracy.
Part 2. Transfer learning to improve accuracy (Total: 40%)
In this part you will use transfer learning technique from existing pre-trained network (e.g. VGG16 etc.) to improve your network’s accuracy on testing. You should explain in detail about how you do you. (30%)
Optionally you can use any other techniques such as dropout, data argumentation etc. to improve your network’s accuracy. Code and explanation required. (10%)
You can use colab notebook comments in codes and text cells to explain in detail the training process and improving methods used in your coding. And submit your .ipynb file (preferred). Or you can submit a report in pdf format.
Please upload your finished assignment to canvas before deadline.
Note: A zero mark will be given if you copy someone else’s work or you let someone copy your work.