类激活热图 InceptionV3 迁移学习

Question

我使用了迁移学习（imagenet 权重）并训练了 InceptionV3 来识别两类图像。代码看起来像

InceptionV3_model = InceptionV3(input_shape=(150,150,3),weights='imagenet', include_top=False)
for layer in InceptionV3_model.layers[:249]:
   layer.trainable = False
for layer in InceptionV3_model.layers[249:]:
   layer.trainable = True
InceptionV3_last_output = InceptionV3_model.output
InceptionV3_maxpooled_output = Flatten()(InceptionV3_last_output)
InceptionV3_x = Dense(1024, activation='relu')(InceptionV3_maxpooled_output)
InceptionV3_x = Dropout(0.5)(InceptionV3_x)
InceptionV3_x = Dense(2, activation='softmax')(InceptionV3_x)
InceptionV3_x_final_model = Model(inputs=InceptionV3_model.input,outputs=InceptionV3_x)
InceptionV3_x_final_model.compile(optimizer=SGD(lr=0.0001, momentum=0.9), loss='categorical_crossentropy',metrics=['accuracy'])

number_of_epochs = inception_epoch
inception_filepath = 'inceptionv3_'+'-saved-model-{epoch:02d}-loss-{loss:.2f}.hdf5'
inception_checkpoint = tf.keras.callbacks.ModelCheckpoint(inception_filepath, monitor='val_accuracy', verbose=1, save_best_only=True, mode='max')
inception_early_stopping = tf.keras.callbacks.EarlyStopping(monitor='loss', patience=10)
inceptionv3_history = InceptionV3_x_final_model.fit(train_generator, epochs = number_of_epochs, validation_data = validation_generator,callbacks=[inception_checkpoint,inception_early_stopping],verbose=1)

do_history_stuff(inceptionv3_history, 'inceptionv3_model', True) 

获得预测

def mode(my_list):
    ct = Counter(my_list)
    max_value = max(ct.values())
    return ([key for key, value in ct.items() if value == max_value])

true_value = []

inception_pred = []

for folder in os.listdir(seg_test_folders):
    
    test_image_ids = os.listdir(os.path.join(seg_test_folders,folder))
    
    for image_id in test_image_ids[:int(len(test_image_ids))]:
        
        path = os.path.join(seg_test_folders,folder,image_id)
        
        true_value.append(validation_generator.class_indices[folder])
        img = cv2.resize(cv2.imread(path),(150,150))
        #img = cv2.imread(path)
        img_normalized = img/255
        
        #Inception
        inception_image_prediction = np.argmax(inception_best_model.predict(np.array([img_normalized])))
        inception_pred.append(inception_image_prediction)

然后我使用我尝试使用 gradcam 可视化热图以查看网络聚焦的位置来 ，但它不起作用。我正在尝试使用 Chollet 的指南，但我是新手，我不知道如何将其与我的代码相匹配。你能帮忙定制gradcam代码吗？ 我找不到模型的倒数第二层，并且无法生成与模型预测的我自己的图像之一匹配的热图。我尝试使用的代码是 https://github.com/Abhijit-2592/Keras-custom-callbacks/blob/master/how%20to%20use%20grad-cam%20in%20inceptionv3_copy.ipynb。该代码使用通用的 Inception_v3 而不是我的微调版本。您能帮忙将此代码与我的相匹配吗？