Outline Links
- ● Task Description
- ● Dataset
- ● Data & Submission Format
- ● Grading Policy
- ● Baseline Guides
- ● Regulations
● Kaggle
● Data
● colab tutorial (mandarin) ● colab tutorial (english)
● Video Link (Ch / En)
Image Reference: https://arxiv.org/pdf/1810.07911.pdf
Task Description – Domain Adaptation
● Imagine you want to do tasks related to the 3D
environment, and then discover that…
- ○ 3D images are difficult to mark and therefore expensive.
- ○ Simulated images (such as simulated scene on GTA-5) are easy to label.Why not just train on simulated images?
Task Description – Domain Adaptation
● For Net, the input is “abnormal”, which makes Net doesn’t work properly.
Image Reference: https://arxiv.org/pdf/1810.07911.pdf
Train
Net (U)
Net (D)
Test
Feat A
Output
???
???
????
????
Task Description – Domain Adaptation
● Therefore, one simple way to solve this problem is to make the distributions of FeatA and FeatB similar.
Image Reference: https://arxiv.org/pdf/1810.07911.pdf
Train
Net (U)
Net (D)
Feat A
similar
Feat B
Output
Test
???
Output
Task Description – Domain Adaptation
● Our task: Given real images (with labels) and drawing images (without labels), please use domain adaptation technique to make your network predict the drawing images correctly.
Dataset
- ● Label: 10 classes (numbered from 0 to 9), as following pictures discribed.
- ● Training : 5000 (32, 32) RGB real images (with label).
- ● Testing : 100000 (28, 28) gray scale drawing images.
Data Format
● Unzip real_or_drawing.zip, the data format is as below: ● real_or_drawing/
○ train_data/ ■ 0/
- ● 0.bmp, 1.bmp … 499.bmp ■ 1/
- ● 500.bmp, 501.bmp … 999.bmp ■ … 9/○ test_data/ ■ 0/
- ● 00000.bmp
- ● 00001.bmp
- ● … 99999.bmp
Data Format
● You can simply use the following code to get dataloader after extracting the zip. (You can apply your own source/target transform function.)
source_dataset = ImageFolder('real_or_drawing/train_data', transform=source_transform)
target_dataset = ImageFolder('real_or_drawing/test_data', transform=target_transform)
source_dataloader = DataLoader(source_dataset, batch_size=32, shuffle=True) target_dataloader = DataLoader(target_dataset, batch_size=32, shuffle=True) test_dataloader = DataLoader(target_dataset, batch_size=128, shuffle=False)
Submission Format
- ● First line should be “id, label”.
- ● Next 100, 000 lines are your predicted labelsof test images.
- ● Evaluate Metrics = Accuracy.
Grades
- ● +4pt : code submission
- ● +1pt : Simple public baseline (0.41962)
- ● +1pt : Simple private baseline
- ● +1 : Medium public baseline (0.59980)
- ● +1 : Medium private baseline
- ● +0.75 : Strong public baseline (0.71874)
- ● +0.75 : Strong private baseline
- ● +0.25 : Boss public baseline (0.77956)
- ● +0.25 : Boss private baseline
Baseline Guides
● Simple Basline (2pts, acc≥0.41962, < 1hour) ○ Just run the code and submit answer.
- ● Medium Baseline (2 pts, acc≥0.59980, 2~4 hours)
- ○ Set proper λ in DaNN algorithm.
- ○ Luck, Training more epochs.
- ● Strong Baseline (1.5 pts, acc≥0.71874, 5~6 hours)
- ○ The Test data is label-balanced, can you make use of this additionalinformation?
- ○ Luck, Trail & Error 🙂
Baseline Guides
● Boss Baseline (0.5 pts, acc ≥0.77956)
- ○ All the techniques you’ve learned in CNN.
- Change optimizer, learning rate, set lr_scheduler, etc…
- Ensemble the model or output you tried.
- ○ Implement other advanced adversarial training.
■ For example, MCD MSDA DIRT-T
- ○ Huh, semi-supervised learning may help, isn’t it?
- ○ What about unsupervised learning? (like Universal Domain Adaptation?)
Learning Curve (Loss)
● This image is for reference only.
Learning Curve (Accuracy)
- ● This image is for reference only.
- ● Note that you cannot access testingaccuracy.
- ● However, this plot tells you thateven though the model overfits the training data, the testing accuracy is still improving.
Code Submission – NTU COOL
●
NTU COOL
● ●
Your .zip file should include only
○ Code: either .py or .ipynb
○ Report: .pdf (only for those who got 10 points)
Report template
○ ○ ○ ○ ○
Deadline: 6/13 (Sun.) 23:59
Compress your code and report into <student_ID>_hw11.zip(e.g. b10123456_hw11.zip)
We can only see your last submission.
DO NOT submit your model or dataset.
If your code is not reasonable, your semester grade x 0.9.
Regulations
- ● You should NOT plagiarize, if you use any other resource, you should cite it in the reference. (*)
- ● Do NOT share codes or prediction files with any living creatures.
- ● Do NOT use any approaches to submit your results more than 5 times aday.
- ● Do NOT search or use additional data.
- ● Do NOT search the label or dataset on the Internet.
- ● Do NOT use pre-trained models on any image datasets.
- ● Your final grade x 0.9 if you violate any of the above rules.
- ● Prof. Lee & TAs preserve the rights to change the rules & grades.
If any questions, you can ask us via…
● NTU COOL (recommended) ○ [Link]
● Email
○ [Link]
○ The title should begin with “[hwX]” (X is the homework number)
● TA hour
- ○ Each Monday 19:00~21:00 @Room 101, EE2 (電機二館101)
- ○ Each Friday 13:30~14:20 Before Class @Lecture Hall (綜合大講堂)
- ○ Each Friday during class
Hidden Guideline – DaNN (1/3)
- ● 這裡我們介紹最基礎的 DaNN (Domain-Adversarial Training of NNs)。
- ● 如果一個模型在測試時吃到不是與訓練集同個 distribution 的輸入，那麼輸出往往會爆走，如下圖。
- ● 而為什麼不能讓圖中的 CNN 在輸入 B dataset 輸出正常的 output?因為你並
沒有 B dataset 的 label 使模型學習。 A dataset (∈ Training)
B dataset (∉ Training)
Normal Output Abnormal Output
CNN
Hidden Guideline – DaNN (2/3)
● 為了因應這樣的情況，DaNN就將 CNN 先拆成兩個部分，並且想辦法讓前半的 CNN 在吃入兩個 A dataset & B dataset 後得到的 distribution 是相近的，那麼 後半就會因為輸入是正常的 output，而發揮正常的功用。
前半 CNN / Feature Extrator
A dataset (∈ Training)
Normal Output
Normal Output
Normal
Output
B dataset (∉ Training) 努力變成 Normal
Output
後半 CNN / Label Predictor
Hidden Guideline – DaNN (3/3)
● 而如何讓前半段的模型輸入兩種不同分布的資料，輸出卻是同個分布呢?最簡 單的方法就是像 GAN 一樣導入一個 discriminator 來分辨輸入是哪個 dataset，並讓 feature extractor 來騙過 discriminator 即可。
● colab tutorial A dataset (∈ Training)
B dataset (∉ Training)
努力變成
B’s normal
Output
是 A dataset
前半 CNN / Feature Extrator
Discriminator / Domain Classifier
努力變成 是 B dataset A’s Normal
Output
，以騙過 discriminator
