HW1 COVID-19 Cases Prediction

Grading

Info

1. Simple: 2.284
2. Medium: 1.494
3. strong: 1.057
4. Boss: 0.862

Medium

Hint: feature selection

Delete the first 37 one-hot features (state information)

	Score
epoch	787
train loss	1.175
val loss	1.001
public score	1.269
private score	1.415

Strong

Hint: Different model architechtures and optimizers

Build deeper model 16-8-1 ⇒ 32-16-8-1

	Score
epoch	1265
train loss	1.123
val loss	0.955
public score	0.972
private score	1.027

Boss

Hint: L2 regularization and try more parameters

This is too hard. I give up.

My best score: private score: 0.990; public score: 0.942

Actually I tried too many times and I just forgot how to get this score… Maybe based on strong model, I used LeakyReLU, added L2 regularization, and implemented k-fold cross validation, then selected the best model.

For future DL training

1. Make structural records, every time just change one thing.
2. Change should be simple, don’t change too much.
3. Organize code and try to get better, but don’t fall into perfectionism.

Code

HW1 strong