# 模型基本開發流程 ![](https://imgur.com/hDiXd8T.png) 模型開發流程主要可分四大步驟 1. Data Pre-processing: 資料處理 2. Training: 定義模型、訓練機制(損失函數) 3. Evaluate: 評估與校正模型 4. Inference: 推論 --- ## naiive project learning ![](https://imgur.com/aMUGUFO.png) ### 結構化程式說明 ``` Tensorflow 2.3 ``` 1. Data Pre-processing 1-1. Data Labeling x放入輸入資料(題目)，y放對應的標籤(答案) ```python= x = mnist.data.values x = x.reshape(-1, 28, 28, 1)#x重塑成(第幾筆, 長, 寬, 單通道)四維tensor x = np.pad(x, ((0,0),(2,2),(2,2),(0,0)), 'constant')#28x28(長x寬)零填充至32x32 Lenet模型要求 x = x/255.#縮放資料至0~1之間(可略) y = np.array(mnist.target.values) y = to_categorical(y, 10)#one-hot encoding ``` 1-2. Data splitting 訓練資料 : 測試資料 = 7 : 3 ```python= from sklearn.model_selection import train_test_split x_train, x_test, y_train, y_test = train_test_split( x, y, test_size=0.3, random_state=1) #x_train, x_temp, y_train, y_temp = train_test_split( x, y, test_size=0.3, random_state=1) #x_train, x_test, y_train, y_test = train_test_split( x_temp, y_temp, test_size=0.3, random_state=1) ``` 2. Training 2-1. Model define ![lenet.png](https://raw.githubusercontent.com/MostafaGazar/mobile-ml/master/files/lenet.png) ```python= model = Sequential() model.add(layers.Conv2D(filters=6, kernel_size=(5, 5), activation='relu', input_shape=(32,32,1))) model.add(layers.AveragePooling2D()) model.add(layers.Conv2D(filters=16, kernel_size=(5, 5), activation='relu')) model.add(layers.AveragePooling2D()) model.add(layers.Flatten()) model.add(layers.Dense(units=120, activation='relu')) model.add(layers.Dense(units=84, activation='relu')) model.add(layers.Dense(units=10, activation = 'softmax')) ``` 2-2. Define loss & opt loss: 計算模型作答與正確答案的誤差 opt: 模型修正的手段 metrics: 人監控模型的狀況(一般用精準度來觀察) ```python= model.compile(loss="categorical_crossentropy", optimizer="adam", metrics=["accuracy"]) ``` 2-3. Training ```python= model.fit(x_train, y_train, batch_size=batch_size, epochs=epochs, validation_split=0.1) ``` 3. Evaluate - evaluate 只看的到模型預測的誤差與精準度兩個值 ```python= test_loss, test_acc = model.evaluate(x_test, y_test) print('test acc: {:.2f}%'.format(test_acc*100)) ``` - predict(get probabilities) 單純輸出模型每次預測的機率值 就手寫辨識而言，每筆資料會有0~9共10個類別的機率值 ```python= y_prob = model.predict(x_test) y_prob_argmax = np.argmax(y_prob, axis=1)#紀錄單筆資料機率值最大的index，以此作為模型預測的類別 # ex: y = [0.01, 0.01, 0.01, 0.01, 0.9, 0.01, 0.01, 0.01, 0.01, 0.02] # print( np.argmax(y) ) # output: 4 ``` - metrics 將模型預測的類別與正確的類別做一個10x10表格的統計結果 縱軸為正確類別，橫軸為模型預測的類別 ```python= from sklearn.metrics import classification_report, confusion_matrix cm = confusion_matrix(np.argmax(y_test, axis=1), y_prob_argmax) print(cm) ``` 4. Inference ```python= image_path = 'c:/Users/user/Desktop/test.png' image_shape = (28, 28) x_new = Image.open(image_path) x_new = x_new.resize(image_shape) x_new = np.array(x_new) x_new = x_new/255 x_new = x_new[:,:,0:1] x_new = np.expand_dims(x_new, axis=0) x_new = np.pad(x_new, ((0,0),(2,2),(2,2),(0,0)), 'constant') ``` ```python= y_new_prob = model.predict(x_new) y_new_prob_argmax = np.argmax(y_new_prob, axis=1) ``` ```python= print('ML: It is number :', y_new_prob_argmax[0]) plt.imshow(x_new[0], cmap='gray', interpolation=None) plt.show() ``` ### 補充說明 - 做AI專案過程中的必備技能 - 程式能力 - trace code能力 - 有效率的debug - 框架&套件 - 頻繁上官網查指令 - keras: https://keras.io/guides/ - sklearn: https://scikit-learn.org/stable/user_guide.html - 基本元素結構多半由numpy所組成 - 熟記numpy的tensor format、與常用指令 - python list與numpy 格式傳換 - 基本視覺化工具 - matplotlib需要用的時候，能快速google搜尋 - AI素養 - 參考書(魚骨頭、keras大神等聖經) - YT(Hung-yi Lee): https://www.youtube.com/c/HungyiLeeNTU ### 完整code basic_dl_flow.ipynb {%gist JackyPro/10c5ee7ca09c1a5e3969d2f6dd682f7b %}