加簽表單

Link

1
 Machine Learning HW1
COVID-19 Cases Prediction
ML TAs
[email protected]

2
Outline
● Objectives ● Code Submission

● Task Description ● Hints

● Data ● Deadline

● Evaluation Metric ● Regulations

● Kaggle ● Useful Links

● Grading
3
Objectives
● Solve a regression problem with deep neural networks (DNN).

● Understand basic DNN training tips e.g. hyper-parameter

tuning, feature selection, regularization, …

● Get familiar with PyTorch.

4
Task Description (1/2)
● COVID-19 Cases Prediction
● Source: Delphi group @ CMU
○ A daily survey since October 2021 via facebook.

Try to find out the data and use

it to your training is forbidden.

5
Task Description (2/2)
● Given survey results in the past 3 days in a specific state in U.S., then
predict the percentage of new tested positive cases in the 3rd day.

Day1 Day2 Day3

6
Data (1/3) – Feature
● States (35, encoded to one-hot vectors)

● COVID-like illness (5)

○ cli, ili …

● Behavior indicators (5)

○ wearing_mask, shop_indoors, restaurant_indoors, public_transit …

● Belief indicators (2)

○ belief_mask_effective, belief_distancing_effective.

7
Data (2/3) – Feature
● Mental indicator (2)
○ worried_catch_covid, worried_finance.

● Environmental indicators (3)

○ other_masked_public, other_distanced_public …

● Tested Positive Cases (1)

○ tested_positive (this is what we want to predict)

8
Data (3/3) – One-hot Vector
● One-hot Vectors

Vectors with only one element equals to one while others are zero.
Usually used to encode discrete values.

9
Evaluation Metric
● Mean Squared Error (MSE)

ground truth

your model (prediction)

10
Kaggle (1/2) – Format
● Display name: <student ID>_<anything>
○ e.g. b10901000_public和private也差太多
○ For auditing, don’t put student ID in your display name.

● Submission format: .csv file

○ See sample code

● Kaggle Link

11
Kaggle (2/2) – Submission
● You may submit up to 5 results each day (UTC+8, AM 8:00)
● Up to 2 submissions will be considered for the private leaderboard

12
Grading (1/5) – Introduction
● In this class, there are 15 assignments.

● Each has 10 points, only count the 10 assignments with the highest points.

● You don’t need to do all the assignments. Choose the one you are
interested in.

Reference: https://speech.ee.ntu.edu.tw/~hylee/ml/ml2022-course-data/rule%20(v2).pdf 13
Grading (2/5) – Introduction
● Most assignment includes leaderboard, gradescope, and code submission.
○ Leaderboard：Kaggle or JudgeBoi (our in-house Kaggle) competition
○ Gradescope： Answer some questions
○ Code submission：Submit the related code of each assignment via NTU COOL

● HW1 doesn’t include gradescope.

Reference: https://speech.ee.ntu.edu.tw/~hylee/ml/ml2022-course-data/rule%20(v2).pdf 14
Grading (3/5) – Leaderboard
● simple (public) +1 pts
● simple (private) +1 pts
● medium (public) +1 pts
● medium (private) +1 pts
● strong (public) +1 pts
● strong (private) +1 pts
● boss (public) +1 pts
● boss (private) +1 pts
● code submission +2 pts
Total : 10 pts

15
Grading (4/5) – Baseline Score

16
Grading (5/5) – Bonus
● If your ranking in private leaderboard is top 3, you can choose to share a
report to NTU COOL and get extra 0.5 pts.

● About the report

○ Your name and student_ID
○ Methods you used in code
○ Reference Report Template
○ in 200 words
○ Deadline is one week later than code submission
○ Please upload to NTU COOL’s discussion of HW1

17
Code Submission (1/6)
● NTU COOL
○ Compress your code and pack them into .zip file

<student_ID>_hw1.zip

● Do not submit models and data

● Submit the code you chose in Kaggle (One of the best)

18
Code Submission (2/6)
code

zip file folder

<student_id>_hw1.zip <student_id>_hw1 or

● Your .zip file should include only

○ Code: either .py or .ipynb

19
Code Submission (3/6)
● How to download your code
● From Google Colab

20
Code Submission (4/6)
● How to compress your folder?
● Method 1 (for Windows users)
○ https://support.microsoft.com/en-us/windows/zip-and-unzip-files-f6dde0a7-0fec-8294-e1d3-703ed85e7ebc

21
Code Submission (5/6)
● How to compress your folder?
● Method 2 (for Mac users)
○ https://support.apple.com/zh-tw/guide/mac-help/mchlp2528/mac

Compress “b10901000_hw1”
22
Code Submission (6/6)
● How to compress your folder?
● Method 3 (command line)

e.g.

23
Hints
Simple : Just run sample code

Medium : Feature selection

Strong : Different optimizers and L2 regularization

Boss : Better feature selection, different model architectures and try more hyper-
parameters

24
Deadline
● Kaggle

2023/03/01 23:59 (UTC+8)

● NTU COOL

2023/03/08 23:59 (UTC+8)

25
Regulations
● You should finish your homework on your own.
● You should not modify your prediction files manually
● Do not share codes or prediction files with any living creatures.
● Do not use any approaches to submit your results more than 5 times a
day.
● Do not search or use additional data or pre-trained models.
● Your final grade x 0.9 and this HW will get 0 pt if you violate any of the
above rules.
● Prof. Lee & TAs preserve the rights to change the rules & grades.

26
Contact us if you have problems…
● Kaggle Homework 1 Discussion
○ https://www.kaggle.com/competitions/ml2023spring-hw1/discussion

● NTU COOL Homework 1 Discussion

○ https://cool.ntu.edu.tw/courses/24108

● Email
○ [email protected]
○ The title should begin with “[hw1]”

27
Useful Links
● Hung-yi Lee, Gradient Descent (Mandarin)
○ link1, link2, link3, link4

● Hung-yi Lee, Tips for Training Deep Networks (Mandarin)

○ link1, link2

● Pytorch Toolkit

● Link that can find all things

● Class webpage

(If Google or Stackoverflow can answer your questions, you may take advantage of them
before asking the TAs.)
28
FAQ
(1) L2 regularization 除了 sample code 提供的在計算 loss 時處理之外，也可以使用 optimizer 的
weight_decay 實現，可參考 🔗 PyTorch 官方文檔

(2) sklearn、 TensorFlow、 xgboost 是可以使用的（使用額外線上資源請附上 Reference）

(3) 只要 Post-processing 是由程式自動完成，且並未違反規定（如不能使用 pre-trained model、不能直接

輸出 hardcode 的結果、不能上網爬資料等），都是可以接受的，另外，請記得將後處理的程式一併交上，
若沒有交上，將視為違反規定。

(4) 同學只要確認上傳時的檔名正確，COOL 系統內部會在同名的檔案依照版本順序加上編號，忽略即可

（如 "學號_hw1-1.zip" 等）。另外請同學確認最後一次上傳的版本是正確的，我們只會認最新的版本

29