Domain Adaptation

ML 2019 Fall Final Prj

~Use real images to predict drawing~
Arvin Liu

What is Domain Adaptation?

Domain Adaptation

Domain Adaptation in the Task

Real Image -> Drawing

(32 x 32 x 3)

(28 x 28 x 1)

Domain Adaptaion

有Label

沒有Label

Feature Extractor

Classifier

Source

Target

Source: blue

Target: Red

Feature Distribution

Why Need
Domain Adaptation

Data Format

Classes

  •  0 - Cow
  •  1 - Bed
  •  2 - Clock
  •  3 - Apple
  •  4 - Cat
  •  5 - Airplane
  •  6 - Television
  •  7 - Dog
  •  8 - Dolphin
  •  9 - Spider

Input Format (balanced)

  • trainX.npy
np.load("trainX.npy")
# Get (5000, 32, 32, 3)
  • test.npy
np.load("test.npy")
# Get (100000, 28, 28, 1)
  • trainY.npy
np.load("trainY.npy")
# Get (5000,)

(label = 0~9)

(Balanced)

(Balanced)

Output Format

id,label
0,0
1,0
2,0
3,0
4,0
5,0
6,0

就是同你的Hw3

Evaluation Metrices

Just Top 1 Accuracy.

Policies

No Human Learning/Labeling

No Extra Dataset

Allowed Pre-trained Model

Only torchvision's pre-trained models are allowed.

Kaggle Link

Simple Baseline

public score - 0.47103

如果狀況太差的話會調降,只會降: )

Timeline

GUIDE

Guide

  • Edge Detection
  • DANN
  • MCD

Edge Detection

Domain Adaptation?

Edge

Detection

直接讓Train很像Test也是種"Adaptation"?

*直接用cv2套件就可以了 (cv2.Canny)

Actually...

這邊就有一個提示了XD

下圖是上圖的邊緣偵測結果