Machine Learning

import warningswarnings.simplefilter(action='ignore')from keras.datasets import cifar10import numpy as np# saving dateset(x_train, y_train),(x_test,y_test) = cifar10.load_data() np.savez('LOCAL-CIFAR10.npz', a=x_train, b=y_train, c=x_test, d=y_test)

On CoLab:

from google.colab import drivedrive.mount('/content/drive')
%cd /content!pwd%cd drive/My Drive/NCUAI-CIFAR10!ls -l
import numpy as np# loading datasetdata = np.load('LOCAL-CIFAR10.npz')x_train = data['a']y_train = data['b']x_test = data['c']y_test = data['d']

-----------------------------------------------

如何使用 CoLab (32:20) (12:44~32:20)

https://www.youtube.com/watch?v=YbmEyJpx7xk&list=PLYgGtiVoYLPdHNmslnjXuF9gjCIj3g-F3&index=31

11月27日 (週三) : 10:00 AM ~ noon

分組報告 (10%)

各組繳交期末專題的題目, 每組報告3分鐘, 討論3分鐘.

請說明 : (1)題目重要性, (2)資料來源, (3)EDA

How to save the best model in a trainning loop

Keras_Mnist_MLP_Save_Best.ipynb

https://drive.google.com/drive/folders/1zDCFwfkl0lgDG8x7L3q425MxMmUcqG0k?usp=sharing

from tensorflow.keras.callbacks import ModelCheckpoint

from tensorflow.keras.models import load_model

from tensorflow.keras.utils import to_categorical

from sklearn.model_selection import train_test_split

from sklearn.metrics import accuracy_score

batch_size = 32 # Number of samples per gradient update

accuracy_old = 0 # Initialize the previous best accuracy

for i in range(0, 3): # Loop for 3 iterations

checkpoint = ModelCheckpoint('..\ModelSaved\model.keras', monitor='val_accuracy', verbose=1, save_best_only=True) # Save the best model based on validation accuracy

train_history = model.fit(x_Train_normalize, y_Train_OneHot, validation_data=(x_Test_normalize, y_Test_OneHot), epochs=2, batch_size=batch_size, callbacks=[checkpoint], verbose=1) # Train the model

model = load_model('..\ModelSaved\model.keras') # Load the best model

predictions = model.predict(x_Test_normalize) # Make predictions on test data

predicted_classes = np.argmax(predictions, axis=1) # Convert predictions to class labels

predicted_one_hot = to_categorical(predicted_classes, num_classes=predictions.shape[1]) # Convert class labels to one-hot encoded format

accuracy_new = accuracy_score(y_Test_OneHot, predicted_one_hot) # Calculate accuracy

if (accuracy_new > accuracy_old): # Check if the new accuracy is better

outfile = '..\ModelSaved\MLP_' + str(int(10000*accuracy_new)) + '.keras' # Generate a filename for the model

model.save(outfile) # Save the model

print(accuracy_old, accuracy_new)

accuracy_old = accuracy_new # Update the previous best accuracy

loss, acc = model.evaluate(x_Test_normalize, y_Test_OneHot, batch_size=batch_size) # Evaluate the model

print("========================================================")

print("| model saved in:", outfile)

print("| Test accuracy: %.4f%% " % (100.0 * acc))

print("========================================================\n")

12月04日 (週三) : 10:00 AM ~ noon

使用網頁 :

Keras: API - Applications

https://keras.io/api/applications/

Keras: Code examples

https://keras.io/examples/

Co-Lab 課堂實作: Run the following three programs on Co-Lab

程式在: NCUAI_20230903\Python_20210810\Keras_10_ResNet50_001.ipynb

(21) ML-23: Classify ImageNet classes with ResNet50 (13:02)

https://www.youtube.com/watch?v=it4fjd_2hUU&list=PLYgGtiVoYLPdHNmslnjXuF9gjCIj3g-F3&index=23

https://keras.io/api/applications/#usage-examples-for-image-classification-models

程式在: NCUAI_20230903\Python_20210810\autoencoder.ipynb

(22) ML-24: Convolutional autoencoder for image denoising (19:00)

https://www.youtube.com/watch?v=PL043XfWRac&list=PLYgGtiVoYLPdHNmslnjXuF9gjCIj3g-F3&index=24

https://keras.io/examples/vision/autoencoder/

程式在: NCUAI_20230903\Python_20210810\3D_image_classification.ipynb

(23) ML-25: 3D image classification from CT scans (15:46)

https://www.youtube.com/watch?v=wor-nXI3OQE&list=PLYgGtiVoYLPdHNmslnjXuF9gjCIj3g-F3&index=25

https://keras.io/examples/vision/3D_image_classification/

On denoising autoencoders trained to minimise binary cross-entropy

https://arxiv.org/abs/1708.08487

12月11日 (週三) : 10 AM ~ 11 AM

道瓊指數, EDA 和 LSTM (Long and Short Term Memory)：

程式在: NCUAI_20230903\人工智慧概論\3_Keras_experiment\Keras_3_Stock\code\Keras_03_Stock.ipynb

投影片在: NCUAI_20230903\人工智慧概論\1_PDF_Lecture_Note\AI_11_Keras_Stock.pdf

(24) ML-17: 人工智慧概論 - 19, 道瓊指數, 資料處理 (EDA) (16:58)

https://www.youtube.com/watch?v=u4b5fyg4X2M&list=PLYgGtiVoYLPdHNmslnjXuF9gjCIj3g-F3&index=17

(25) ML-18: 人工智慧概論 - 20, 道瓊指數, LSTM程式 (21:07)

https://www.youtube.com/watch?v=LBD6jQydzKM&list=PLYgGtiVoYLPdHNmslnjXuF9gjCIj3g-F3&index=18

參考資料 :

A Historical Journey: Large Language Models (LLMs) from LSTMs to ChatGPT

https://medium.com/@priyanshuprasad1718/a-historical-journey-large-language-models-llms-from-lstms-to-chatgpt-68d769ca3951

Transformer Neural Networks: A Step-by-Step Breakdown

https://builtin.com/artificial-intelligence/transformer-neural-network?form=MG0AV3

Neural machine translation with a Transformer and Keras

https://www.tensorflow.org/text/tutorials/transformer?form=MG0AV3

Building a Transformer with PyTorch

https://www.datacamp.com/tutorial/building-a-transformer-with-py-torch?form=MG0AV3

Open AI 出現最大競爭對手世界周報MIT採訪人工智慧中心主任

https://www.youtube.com/watch?v=rIuGpeqd-D8

12月11日 (週三) : 11 AM ~ noon

程式在: NCUAI_20230903\Python_20210810\TransferLearning_ShoulderImplants\code

(26) ML-26: Transfer Learning for Analyzing Medical Images (33:45)

https://www.youtube.com/watch?v=3x2YGfnaoXI&list=PLYgGtiVoYLPdHNmslnjXuF9gjCIj3g-F3&index=26

(a) Download the paper: Classifying shoulder implants in X-ray images using deep learning

https://www.sciencedirect.com/science/article/pii/S2001037019304179

(b) Download the dataset: 這裡可以下載 UCI 的597張照片

Shoulder Implant X-Ray Manufacturer Classification Data Set

https://archive.ics.uci.edu/dataset/517/shoulder+implant+x+ray+manufacturer+classification

(c) Display the images: 這裡是顯示照片的 PYTHON 程式

UCI Shoulder Implant X-Ray Manufacturer

https://www.kaggle.com/redwankarimsony/uci-shoulder-implant-x-ray-manufacturer/code

(d) CMU 的網友, 將程式寫出來了, 提供大家參考:

(註解: 作者將全部的資料, 都拿去training, 再拿去prediction, 所以準確率接近100%)

https://github.com/mannylazalde/ShoulderImplantClassification

#from keras.applications.resnet50 import ResNet50

#from keras import Model, layers

#from keras.layers import InputLayer

from tensorflow.keras.applications.resnet50 import ResNet50

from tensorflow.keras import Model, layers

from tensorflow.keras.layers import InputLayer

12月18日 (週三) : 10 AM ~ noon

Segment Anything Model:

Fast Segment Anything

https://arxiv.org/abs/2306.12156

CASIA-IVA-Lab/FastSAM

https://github.com/CASIA-IVA-Lab/FastSAM

------------------------------------------

Segment Anything Model

https://arxiv.org/abs/2304.02643

facebookresearch/segment-anything

https://github.com/facebookresearch/segment-anything

OpenPose:

Realtime Multi-Person 2D Human Pose Estimation using Part Affinity Fields, CVPR 2017 Oral

https://www.youtube.com/watch?v=pW6nZXeWlGM

CMU-Perceptual-Computing-Lab/openpose

https://github.com/CMU-Perceptual-Computing-Lab/openpose

Manipulating Facial Features With OpenCV and Dlib

https://towardsdatascience.com/manipulating-facial-features-with-opencv-and-dlib-14029f136a3d

12月11日 (週三) : 11 AM ~ noon

瑕疵檢測 (Anomaly Detection)

在這裡下載程式： PaDiM_20241218.zip

https://drive.google.com/drive/folders/1zDCFwfkl0lgDG8x7L3q425MxMmUcqG0k?usp=sharing

PaDiM: a Patch Distribution Modeling Framework for Anomaly Detection and Localization

https://arxiv.org/abs/2011.08785

xiahaifeng1995/PaDiM-Anomaly-Detection-Localization-master

https://github.com/xiahaifeng1995/PaDiM-Anomaly-Detection-Localization-master

-------------------------------------------------------------

PaDiM 原理与代码解析

https://blog.csdn.net/ooooocj/article/details/127601035

PaDiM : A machine learning model for detecting defective products without retraining

https://medium.com/axinc-ai/padim-a-machine-learning-model-for-detecting-defective-products-without-retraining-5daa6f203377

modification in main_01.py:

replace: def parse_args():

add: debug_a = 'Y'

add: class_name = 'bottle', print(class.name)

comment out: the program in class_name loop

modification in mvtec.py:

replace: dataset_path='./dataset_mvtec'

replace: Image.LANCZOS

--------------------------

modification in main_02.py:

un-comment: the program in class_name loop

re-arrange: move the functions before the main program

--------------------------

modification in main_03.py:

add comments and print the values of layer.shape and B, C, H, W

12月25日 (週三) : 10 AM ~ noon

彈性教學 : 1月1日 (週三)

(27) ML-27: 人工智慧概論 10, 最佳化, 梯度下降 (39:08)

https://www.youtube.com/watch?v=WJ5LM27LpDw&list=PLYgGtiVoYLPdHNmslnjXuF9gjCIj3g-F3&index=27

(28) ML-28: 人工智慧概論 11, 最佳化, 約束函數, 曲線擬合, 國小近視比例 (29:13)

https://www.youtube.com/watch?v=OdYJ9gdXVWU&list=PLYgGtiVoYLPdHNmslnjXuF9gjCIj3g-F3&index=28

------------------------------

https://www.youtube.com/watch?v=ibJpTrp5mcE&list=PLJV_el3uVTsPy9oCRY30oBPNLCo89yu49&index=12

彈性教學 : 1月8日 (週三) 可解釋的人工智慧

在此下載投影片和 Python 檔案： AI_04_05_Explainable_AI_CAM_GradCAM

https://drive.google.com/drive/u/2/folders/1o7u9yEnRE4FmPDbUwwGV2bu0lX_lDHeI

(29) 用可解釋的人工智慧做影像辨識：(1) CAM (31:24)

https://www.youtube.com/watch?v=zhujHoUZfKA&list=PLYgGtiVoYLPdHNmslnjXuF9gjCIj3g-F3&index=29

(30) 用可解釋的人工智慧做影像辨識：(2) Grad-CAM (25:14)

https://www.youtube.com/watch?v=vmt5miwXqTI&list=PLYgGtiVoYLPdHNmslnjXuF9gjCIj3g-F3&index=30

參考資料 :

https://www.youtube.com/playlist?list=PLJV_el3uVTsPy9oCRY30oBPNLCo89yu49

https://github.com/Sakura-gh/ML-notes

參考資料 :

Stanford 李飛飛的CNN課程,

http://cs231n.stanford.edu/

Lecture 1 | Introduction to Convolutional Neural Networks for Visual Recognition

https://www.youtube.com/watch?v=vT1JzLTH4G4&list=PL3FW7Lu3i5JvHM8ljYj-zLfQRF3EO8sYv&index=1

Alex J. Smola of Berkeley and Amazon, 談理論的教科書

University of California - Berkeley 的免費AI課程, YouTube, 中英文投影片, 及PDF課本

https://courses.d2l.ai/berkeley-stat-157/syllabus.html

英文教科書 : https://d2l.ai/d2l-en.pdf

中文教科書 : https://zh-v2.d2l.ai/d2l-zh.pdf

Day 25 / DL x RL / Hello Reinforcement Learning —— CartPole

https://ithelp.ithome.com.tw/articles/10251599

Deep Reinforcement Learning Hands-On, 2nd Edition

book ==> https://www.amazon.com/Deep-Reinforcement-Learning-Hands-optimization/dp/1838826998

code ==> https://github.com/PacktPublishing/Deep-Reinforcement-Learning-Hands-On-Second-Edition

Reinforcement Learning — Implement TicTacToe

https://towardsdatascience.com/reinforcement-learning-implement-tictactoe-189582bea542

參考資料 :

One-Dimensional CNN, Introduction to 1D Convolutional Neural Networks in Keras for Time Sequences

https://blog.goodaudience.com/introduction-to-1d-convolutional-neural-networks-in-keras-for-time-sequences-3a7ff801a2cf

https://github.com/xitu/gold-miner/blob/master/TODO1/introduction-to-1d-convolutional-neural-networks-in-keras-for-time-sequences.md

Feature selection: Lasso

Lasso Regression with Python

https://www.kirenz.com/post/2019-08-12-python-lasso-regression-auto/

处理不均衡数据 (深度学习)! Dealing with imbalanced data (deep learning) (3:18)

https://www.youtube.com/watch?v=doXeC9_vMhg

Handling Imbalanced Datasets in Deep Learning ==> Class weight, over and under sampling

https://towardsdatascience.com/handling-imbalanced-datasets-in-deep-learning-f48407a0e758

Deep learning unbalanced training data? Solve it like this. ==> Image augmentation

https://towardsdatascience.com/deep-learning-unbalanced-training-data-solve-it-like-this-6c528e9efea6

How to Configure Image Data Augmentation in Keras ==> Image augmentation

https://machinelearningmastery.com/how-to-configure-image-data-augmentation-when-training-deep-learning-neural-networks/

Image data preprocessing

https://keras.io/api/preprocessing/image/

ImageDataGenerator的使用

https://www.itread01.com/content/1546542668.html

Google Sites

Report abuse

Machine Learning

中央大學 ~ GS4519-B ~ 機器學習概論(B班) ~ 2023年 9~12 月

課程影片連結 : https://www.youtube.com/playlist?list=PLYgGtiVoYLPdHNmslnjXuF9gjCIj3g-F3

在這裡下載程式： NCUAI_20230903.zip

推薦書籍 : 臺北市教育局提供的: 高國中小人工智慧教材.

推薦影片：什麼是人工智慧

9月11日 (週三) : 10:00 AM ~ noon

(1) ML-01: 人工智慧概論 02, 原理 (49:09)

(2) ML-02: (2023-08-16 版本) Install Anaconda, Tensorflow (Keras), Pytorch (CUDA), Ultralytics (YOLO8) (21:50)

==> stopping at 14:25 <==

關於 Anaconda 的版本

(A) 安裝最新版本 :

(B) 安裝與教材相同的版本 (Python3.10) :

解決 Anaconda 閃退的方法

Python 基礎教學影片 (莫烦)

Python 書籍 :

一步到位！Python 程式設計-最強入門教科書 / 作者： 陳惠貞 / 出版社：旗標

9月18日 (週三) : 10:00 AM ~ noon

(3) ML-03: How to run Python Programs in Jupyter Notebook and Spyder (37:03)

(4) ML-04: How to use Numpy (29:09)

import warnings

warnings.filterwarnings('ignore')

9月25日 (週三) : 10:00 AM ~ noon

(5) ML-05: How to use Pandas (27:19)

(6) ML-06: How to use debugger in Spyder (13:41)

(7) ML-07: How to use Matplotlib (44:37)

10月09日 (週三) : 10 AM ~ 11 AM

YOLO 原創者的演講 : How computers learn to recognize objects instantly | Joseph Redmon (7:37)

(2) 2023-08-16 版本. ML-02: Install Anaconda, Tensorflow (Keras), Pytorch (CUDA), Ultralytics (YOLO8) (21:50)

==> starting at 14:25 <==

yolo detect mode=predict model=yolov11n.pt source='dog.mp4' show=True

output file in: runs\detect\predict

Introducing Ultralytics YOLOv11, the latest version of the acclaimed real-time object detection and image segmentation model.

Github for Ultralytics YOLOv11

YOLOv8 COMPLETE Tutorial | Object Detection | Segmentation | Classification

YOLOv8 Ultralytics: State-of-the-Art YOLO Models

Train YOLOv8 :

YOLOV8: How to Train a Custom YOLOv8 Object Detector | Detect, Segment and Classify Objects

Train Yolov8 object detection on a custom dataset

10月09日 (週三) : 11 AM ~ noon

(15) ML-15: Kaggle (11:16)

* Kaggle Courses:

作業 #2, Kaggle 競賽 (20%) (12月06日, 23:59 前上傳到 Kaggle 網站)

Example :

2020, GS-3073-AI-course, Regression or classification?