Machine learning workshop I - image classification
1 like1,138 views
This document summarizes a machine learning workshop on image classification. The workshop covers introduction to machine learning concepts like neural networks, deep neural networks, and convolutional neural networks. It then demonstrates image classification on the Fashion MNIST dataset using both dense neural networks and convolutional neural networks. Finally, it guides attendees on preparing their own image data and training a classification model on their images.
![2
-3
0.5
0.5
0.6
0.5
0.1
0.6
0
0.3
0.6
0.4
0.15
0.2
0.1
0.5
Activation Function f :
0.4
0.5
f(-0.6)
f(0.3)
input
input
weight
bias
0.27
0.38
f(0.27)
f(0.38)
0.217
(output)
Basic NN
[ 2 ✕ 0.4 + (-3) ✕ 0.5 ] + 0.1 =
-0.6
[ 2 ✕ 0.6 + (-3) ✕ 0.5 ] + 0.6 =
0.3](https://image.slidesharecdn.com/machinelearningworkshopi-imageclassification-190917033159/85/Machine-learning-workshop-I-image-classification-18-320.jpg)
Machine learning workshop I - image classification
- 1. Machine Learning Workshop I Image Classification Tao Chen taochen@qnap.com
- 2. 關於講師 陳雲濤 Tao Chen taochen@qnap.com Software Engineer, Data Intelligence & Application Div., QNAP Experience on IoT with computer vision, image processing, image classification, object detection & segmentation.
- 3. 課前準備 ● 本課程會實作一個客製化的照片分類模型 ● 請大家準備兩種類別照片，每種至少 50 張以上，分別存放在兩個資料夾中 ● 適合對象： 1. 未接觸過 Machine Learning 的開發者 2. 有使用ML相關經驗，想認識 Deep Learning 方法者 3. 有興趣的想實作自己照片分類器的 QNAPers
- 4. 課程大綱 ● Introduction to Machine Learning ● Neural Network ● Deep Neural Network (DNN) ● Convolutional Neural Network (CNN) ● Image Classification on Fashion MNIST Dataset ● Prepare Your Data for Image Classification ● Train Your Image Classification Model
- 5. Introduction to Machine Learning
- 6. Introduction to Machine Learning ● Traditional: Data Data ● Machine Learning: Output
- 7. Machine Learning ≈ Looking for a Function this is a cat Training Testing “cat” “not a cat” You write the program for learning Ref: 出自台大電機系李宏毅老師 Machine Learning 課程投影片
- 8. Type of Machine Learning Task ● Classification → output class ● Regression → output value “cat” 影像辨識 88.87mm 本月總降雨量預測 中央氣象局觀測資料
- 9. Goodness of Machine Learning Function ● Classification → output class ● Regression → output value “cat” 88.87 mm中央氣象局觀測資料 87.00 mm “cat” Ground Truth evaluate by accuracy evaluate by MSE (mean squared error)
- 10. Machine Learning Process “cat” Ref: 出自台大電機系李宏毅老師 Machine Learning 課程投影片
- 11. Performance Evaluation (Accuracy) ● 假設100個病患使用某模型做疾病檢測 ○ 檢測有病為 Positive ○ 檢測沒病為 Negative ● 前提(病患之真實狀況) ○ 15 個病患有病 ○ 85 的病患沒病 ● 某模型之檢測結果 ○ 20 個病患檢測有病 (其中只有 10 個病患真的有病，10 個誤判) ○ 80 個病患檢測沒病 (其中只有 75 個病患真的沒病，5 個誤判) 檢測有病 檢測沒病 真實有病 10 5 真實沒病 10 75 Accuracy: (10+75)/100 = 85%
- 12. Precision, Recall and F1-score 對的說成對的 錯的說成錯的 錯的說成對的 對的說成錯的
- 13. ● 假設100個病患使用某模型做疾病檢測 ○ 檢測有病為 Positive ○ 檢測沒病為 Negative ● 前提(病患之真實狀況) ○ 15 個病患有病 ○ 85 的病患沒病 ● 某模型之檢測結果 ○ 20 個病患檢測有病 (其中只有 10 個病患真的有病，10 個誤判) ○ 80 個病患檢測沒病 (其中只有 75 個病患真的沒病，5 個誤判) 檢測有病 檢測沒病 真實有病 10 5 真實沒病 10 75 → TP: 10, TN: 75, FP: 10, FN: 5 → Precision: 10/(10+10)= 50%, Recall: 10/(10+5)= 66.66% → F1-score: 2*0.5*0.66/(0.5+0.66) = 57.14% Accuracy: (10+75)/100 = 85% Performance Evaluation (F1-score)
- 14. from：http://speech.ee.ntu.edu.tw/~tlkagk/courses/ML_2017_2/Lecture/introduction.pdf
- 15. NN = Neural Network (one kind of ML function)
- 17. Basic NN input Activation Function bias weight z = input(s) * weight(s) + bias from NTU ML course：https://www.youtube.com/watch?v=Dr-WRlEFefw
- 18. 2 -3 0.5 0.5 0.6 0.5 0.1 0.6 0 0.3 0.6 0.4 0.15 0.2 0.1 0.5 Activation Function f : 0.4 0.5 f(-0.6) f(0.3) input input weight bias 0.27 0.38 f(0.27) f(0.38) 0.217 (output) Basic NN [ 2 ✕ 0.4 + (-3) ✕ 0.5 ] + 0.1 = -0.6 [ 2 ✕ 0.6 + (-3) ✕ 0.5 ] + 0.6 = 0.3
- 20. 2 -3 0.5 0.5 0.6 0.5 0.1 0.6 0 0.3 0.6 0.4 0.15 0.2 0.27 0.38 0.217 0.4 0.5 0.1 0.5 f(-0.6) f(0.3) f(0.27) f(0.38) 0.3 Ground Truth Judge by Loss function e.g. MSE Loss value = MSE(0.217, 0.3) = 0.006889 Training NN
- 21. new bias new bias new bias new bias 0.217 new weight 0.3 Ground Truth Judge by Loss function e.g. MSE Loss value = MSE(0.217, 0.3) = 0.006889 Update Weights and Biases with Loss value new weight new weights new weights Learn by next data Training NN (By Backpropogation Algorithm)
- 22. 2 -3 0.5 0.5 0.6 0.5 0.1 0.6 0 0.3 0.6 0.4 0.15 0.2 0.27 0.38 0.308 0.4 0.5 1 0.5 f(-0.6) f(0.3) f(0.27) f(0.38) 0.71 2 0.1 Ground T. 0 1 Cross Entropy Loss dog cat Sleep Walk with U Basic NN (Classification)
- 25. DNN = Deep Neural Network
- 26. Ref : https://achintavarna.wordpress.com/2017/11/17/keras-tutorial-for-beginners-a-simple-neural-network-to-identify-numbers-mnist-data/ … 0 0 0 0 0 0 0 0 1 0 count loss with Ground T. Update Weights and Biases with Loss value 0.04 0.01 0.02 0.11 0.01 0.06 0.21 0.01 0.35 0.18
- 27. CNN = Convolutional Neural Network
- 28. Basic Architecture of CNN Ref: 出自台大電機系李宏毅老師 Machine Learning 課程投影片 Convolution Max Pooling Flatten 特徵擷取 Feature Extraction 降維度 Reduce Resolution 將所有擷取到的特徵攤 平 → 降為一維向量 to one-dimension
- 29. Convolution Ref: 出自台大電機系李宏毅老師 Machine Learning 課程投影片
- 30. Convolution Ref: 出自台大電機系李宏毅老師 Machine Learning 課程投影片
- 31. ● Convolution Filters 的長相 Convolution Ref: https://medium.com/@RaghavPrabhu/understanding-of-convolutional-neural-network-cnn-deep-learning-99760835f148 from Stanford CS231n CNN: http://cs231n.github.io/convolutional-networks/
- 35. 64 convolution filters pooling R, G, B Ref: https://hk.saowen.com/a/d2cde0a5e42606aae41a53e7ddb55ad7eb3208dc4cd218067344246a9a8ebf54 Convolutional Neural Network ● Understanding Neural Networks Through Deep Visualization
- 36. Frameworks
- 37. Image Classification on Fashion MNIST Dataset
- 38. Environment Setting on Google Colab ● Google Colab: https://colab.research.google.com/notebooks/welcome.ipynb#recent=true ● 文件 -> 新建 python3 Notebook ● 修改 -> 筆記本設置 -> GPU
- 39. Fashion MNIST Dataset ● Fashion MNIST Github ○ Total images 70000 (training dataset 60000, testing dataset 10000) ○ Resolution 28x28 ○ Gray-scale ○ 10 classes:
- 40. Lab #1: Train DNN Model on Fashion MNIST ● Colab Notebook: https://colab.research.google.com/drive/174_6xXG1up-GD6Ui rXYMzMJYxyCdaPSD
- 41. Lab #2: Train CNN Model on Fashion MNIST ● Colab Notebook: https://colab.research.google.com/drive/1eXKrKBe0lbsV9Ex6 amryF0Pml8QsySqY
- 42. Prepare Your Data for Image Classification
- 44. Train Your Image Classification Model
- 45. Lab #3: Train Your Image Classificaiton Model ● Colab Notebook: https://colab.research.google.com/drive/1Sct4c6UUse5heNgF fgdP-MckjCJ3umFM
- 46. 照片辨識PK - 你能比 AI 準嗎？ ● ImageNet Validation Set Competition https://cs.stanford.edu/people/karpathy/ilsvrc/
- 47. Thank You