This shows you the differences between two versions of the page.
Both sides previous revision Previous revision Next revision | Previous revision | ||
cs601r_w2020:lab1 [2020/01/06 18:25] wingated |
cs601r_w2020:lab1 [2021/06/30 23:42] (current) |
||
---|---|---|---|
Line 26: | Line 26: | ||
Details: | Details: | ||
- | Your classifier must be a 20 layer Resnet. | + | Your classifier must be a 20 layer Resnet. You may NOT use a built-in or pretrained resnet (ie, from torchhub), because you will need to be able to modify it. However, you are encouraged to use other built-in pytorch blocks, such as batchnorm blocks. |
- | The dataset is the [[http://cs231n.stanford.edu/tiny-imagenet-200.zip|Tiny Imagenet]] dataset. It has 200 classes, and 500 training images per class (for a total of 100k training images), and 10,000 testing images. (The dataset provides both labels and bounding boxes; you can ignore the boxes) | + | The dataset is the CIFAR-10 dataset. You may use the torchvision.datasets interface to simplify the loading and management of the dataset. While not required for this lab, make sure that you're prepared to implement data augmentation. |
You will probably want to use Google colab to host your notebook. | You will probably want to use Google colab to host your notebook. | ||
Line 34: | Line 34: | ||
For your loss curve, any reasonable visualization is acceptable. | For your loss curve, any reasonable visualization is acceptable. | ||
- | For your confusion matrix, you should display a 200x200 image, where each pixel i,j represents the number of times an image of class i was classified as class j. | + | For your confusion matrix, you may use any reasonable visualization or software package. |
- | You will not be graded on any final accuracies. I think this is a pretty hard dataset, so I would not expect much better performance than 40% accuracy. | + | You will not be graded on any final accuracies. |
- | MAJOR HINT: we will be working with this classifier in future labs by adjusting its hyper parameters. In particular, you should make it easy to: | + | You may use any code on the internet to help you, but all submitted code must be your own work. |
+ | |||
+ | MAJOR CONSIDERATION: we will be working with this classifier in future labs by adjusting its hyper parameters. In particular, you should make it easy to: | ||
* Change the activation function | * Change the activation function | ||
* Change whether or not you use BatchNorm | * Change whether or not you use BatchNorm | ||
* Change the learning rate schedule | * Change the learning rate schedule | ||
* Change the weight regularization | * Change the weight regularization | ||
+ | * Change the weight initialization | ||