from all_models_tools.all_model_tools import call_back from Read_and_process_image.ReadAndProcess import Read_image_and_Process_image from draw_tools.draw import plot_history, Confusion_Matrix_of_Two_Classification from keras import regularizers from Load_process.Load_Indepentend import Load_Indepentend_Data from _validation.ValidationTheEnterData import validation_the_enter_data from keras.layers import GlobalAveragePooling2D, Dense, Dropout from keras.applications import Xception from Load_process.file_processing import Process_File from merge_class.merge import merge from draw_tools.Grad_cam import Grad_CAM from sklearn.metrics import confusion_matrix from keras.models import Model from keras.optimizers import SGD from Image_Process.Image_Generator import Image_generator import pandas as pd import keras import numpy as np import time class experiments(): def __init__(self, tools, status): ''' parmeter: * validation_obj : 驗證物件 * cut_image : 切割影像物件 * image_processing : 讀檔與讀檔資料處理物件 * merge : 合併物件 * model_name: 取名，告訴我我是用哪個模型(可能是預處理模型/自己設計的模型) * generator_batch_size: 每一批次要讀多少檔出來 * epoch: 訓練次數 * train_batch_size: 訓練時要多少批次的影像為1組 * generator_batch_size: 減少圖片數量對GPU記憶體的用量, 減少張數用的 * experiment_name : 本次實驗名稱 * convolution_name : Pre-train model 的最後一層Convolotion的名稱 ''' self.Topic_Tool = tools self.validation_obj = validation_the_enter_data() # 呼叫驗證物件 self.cut_image = Load_Indepentend_Data(self.Topic_Tool.Get_Data_Label(), self.Topic_Tool.Get_OneHot_Encording_Label()) # 呼叫切割影像物件 self.image_processing = Read_image_and_Process_image() self.merge = merge() self.model_name = "Xception" # 取名，告訴我我是用哪個模型(可能是預處理模型/自己設計的模型) self.experiment_name = "Xception Skin to train Normal stomach cancer" # self.file_name = "Remove background of Chickenpox with normal image" self.generator_batch_size = 50 self.epoch = 10000 self.train_batch_size = 128 self.layers = 1 self.convolution_name = self.get_layer_name(self.model_name) self.Grad = "" self.Status = status pass def processing_main(self, train, train_label, counter): Train, Test, Validation = self.Topic_Tool.Get_Save_Roots(self.Status) # 要換不同資料集就要改 start = time.time() self.cut_image.process_main(Test, Validation) # 呼叫處理test Data與Validation Data的function end = time.time() print("讀取testing與validation資料(154)執行時間：%f 秒\n" % (end - start)) Generator = Image_generator("", "") # 將處理好的test Data 與 Validation Data 丟給這個物件的變數 self.test, self.test_label = self.cut_image.test, self.cut_image.test_label self.validation, self.validation_label = self.cut_image.validation, self.cut_image.validation_label self.Grad = Grad_CAM(self.Topic_Tool.Get_Data_Label(), self.test_label, self.experiment_name, self.convolution_name) cnn_model = self.construct_model() # 呼叫讀取模型的function # model_dir = '../save_the_best_model/Topic/Remove background with Normal image/best_model( 2023-10-17 )-2.h5' # 這是一個儲存模型權重的路徑，每一個模型都有一個自己權重儲存的檔 # if os.path.exists(model_dir): # 如果這個檔案存在 # cnn_model.load_weights(model_dir) # 將模型權重讀出來 # print("讀出權重\n") Optimizer = SGD(learning_rate = 0.045, momentum = 0.9) # 決定優化器與學習率 cnn_model.compile( loss = "binary_crossentropy", optimizer = Optimizer, metrics= [ 'accuracy', keras.metrics.Precision(name='precision'), keras.metrics.Recall(name='recall'), keras.metrics.AUC(name = 'AUC'), ] ) train_data = Generator.Generator_Content(5) # 叫入ImageGeneratorr的物件，為了要讓我訓練時可以分批次讀取資料，GPU記憶體才不會爆 cnn_model.summary() # 顯示模型架構 print("\n\n\n讀取訓練資料(70000)執行時間：%f 秒\n\n" % (end - start)) history = cnn_model.fit( train_data.flow(train, train_label, batch_size = self.generator_batch_size), # x = train, # y = train_label, epochs = self.epoch, batch_size = self.train_batch_size, validation_data = (self.validation, self.validation_label), callbacks = call_back(self.experiment_name, counter) # 呼叫 call back list # callbacks = call_back("best_model", self.counter) # 呼叫 call back list ) Matrix = self.record_matrix_image(cnn_model, self.experiment_name, counter) # 紀錄混淆矩陣的function loss, accuracy, precision, recall, AUC = cnn_model.evaluate(self.test, self.test_label) # 預測結果 # 防分母為0的時候 if recall == 0 or precision == 0: f = 0 else: f = (1 + 0.5 * 0.5) * ((recall * precision) / (0.5 * 0.5 * recall + precision)) print(self.record_everyTime_test_result(loss, accuracy, precision, recall, AUC, f, counter, self.experiment_name, Matrix)) # 紀錄當前訓練完之後的預測結果，並輸出成csv檔 plot_history(history, "train" + str(counter), self.experiment_name) # 將訓練結果化成圖，並將化出來的圖丟出去儲存 self.Grad.process_main(cnn_model, counter, self.test) return loss, accuracy, precision, recall, AUC, f def construct_model(self): '''決定我這次訓練要用哪個model''' xception = Xception(include_top = False, weights = "imagenet", input_shape = (512, 512, 3)) Flatten = GlobalAveragePooling2D()(xception.output) output = Dense(units = 1370, activation = "relu", kernel_regularizer = regularizers.L2())(Flatten) output = Dropout(0.6)(output) output = Dense(units = 2, activation = "softmax")(output) cnn_model = Model(inputs = xception.input, outputs = output) return cnn_model def record_matrix_image(self, cnn_model : Model, model_name, index): '''劃出混淆矩陣(熱力圖)''' result = cnn_model.predict(self.test) # 利用predict function來預測結果 result = np.argmax(result, axis = 1) # 將預測出來的結果從one-hot encoding轉成label-encoding y_test = np.argmax(self.test_label, axis = 1) matrix = confusion_matrix(result, y_test, labels = [0, 1]) # 丟入confusion matrix的function中，以形成混淆矩陣 Confusion_Matrix_of_Two_Classification(model_name, matrix, index) # 呼叫畫出confusion matrix的function return matrix.real def record_everyTime_test_result(self, loss, accuracy, precision, recall, auc, f, indexs, model_name, Matrix): '''記錄我單次的訓練結果並將它輸出到檔案中''' File = Process_File() Dataframe = pd.DataFrame( { "model_name" : str(model_name), "loss" : "{:.2f}".format(loss), "precision" : "{:.2f}%".format(precision * 100), "recall" : "{:.2f}%".format(recall * 100), "accuracy" : "{:.2f}%".format(accuracy * 100), "f" : "{:.2f}%".format(f * 100), "AUC" : "{:.2f}%".format(auc * 100) }, index = [indexs]) File.Save_CSV_File("train_result", Dataframe) # File.Save_TXT_File("Matrix_Result : " + str(Matrix), model_name + "_train" + str(indexs)) return Dataframe def get_layer_name(self, model_name): if(self.validation_obj.validation_string(model_name, "VGG19")): return "block5_conv4" if(self.validation_obj.validation_string(model_name, "ResNet50")): return "conv5_block3_3_conv" if(self.validation_obj.validation_string(model_name, "Xception")): return "block14_sepconv2" if(self.validation_obj.validation_string(model_name, "DenseNet121")): return "conv5_block16_concat" if(self.validation_obj.validation_string(model_name, "InceptionResNetV2")): return "conv_7b" if(self.validation_obj.validation_string(model_name, "InceptionV3")): return "conv2d_93" if(self.validation_obj.validation_string(model_name, "MobileNet")): return "conv_pw_13" if(self.validation_obj.validation_string(model_name, "MobileNetV2")): return "Conv_1" if(self.validation_obj.validation_string(model_name, "NASNetLarge")): return "separable_conv_2_normal_left5_18" if(self.validation_obj.validation_string(model_name, "ResNet101")): return "conv5_block3_3_conv" if(self.validation_obj.validation_string(model_name, "ResNet101V2")): return "conv5_block3_3_conv" if(self.validation_obj.validation_string(model_name, "ResNet152")): return "conv5_block3_3_conv" if(self.validation_obj.validation_string(model_name, "ResNet152V2")): return "conv5_block3_out" if(self.validation_obj.validation_string(model_name, "ResNet50v2")): return "conv5_block3_out" if(self.validation_obj.validation_string(model_name, "VGG16")): return "block5_conv3"