Question

nb_classes = 1  # number of classes
based_model_last_block_layer_number = 126  # value is based on based model 
selected.
img_width, img_height = 299, 299  # change based on the shape/structure of 
your images
batch_size = 3  # try 4, 8, 16, 32, 64, 128, 256 dependent on CPU/GPU memory 
capacity (powers of 2 values).
nb_epoch = 100  # number of iteration the algorithm gets trained.
learn_rate = 1e-4  # sgd learning rate
momentum = .9  # sgd momentum to avoid local minimum
transformation_ratio = .05  # how aggressive will be the data 
augmentation/transformation

train_datagen = ImageDataGenerator(rescale=1. / 255,
                                   rotation_range=transformation_ratio,
                                   shear_range=transformation_ratio,
                                   zoom_range=transformation_ratio,
                                   cval=transformation_ratio,
                                   horizontal_flip=True,
                                   vertical_flip=True)

validation_datagen = ImageDataGenerator(rescale=1. / 255)

os.makedirs(os.path.join(os.path.abspath(train_data_dir), '../preview'), exist_ok=True)
train_generator = train_datagen.flow_from_directory(train_data_dir,
                                                    target_size=(img_width, img_height),
                                                    batch_size=batch_size,
                                                    class_mode='binary')
validation_generator = validation_datagen.flow_from_directory(validation_data_dir,
                                                              target_size=(img_width, img_height),
                                                              batch_size=batch_size,
                                                              class_mode='binary')

# Optimizer
sgd = SGD(lr=learn_rate, decay=1e-6, momentum=momentum, nesterov=True)
class_weightage = class_weight.compute_class_weight('balanced', np.unique(train_generator.classes),
                                                    train_generator.classes)
class_weightage = {0: class_weightage[0], 1: class_weightage[1]}
print(class_weightage)


model.compile(optimizer=sgd,
              loss='binary_crossentropy',
              metrics=['binary_accuracy'])


final_weights_path = os.path.join(os.path.abspath(model_path), 'model_weights.h5')
callbacks_list = [
    ModelCheckpoint(final_weights_path, monitor='val_binary_accuracy', verbose=1, save_best_only=True),
    EarlyStopping(monitor='val_binary_accuracy', patience=5, verbose=0),
    TensorBoard(log_dir='./Graph', histogram_freq=0, write_graph=True, write_images=True)
]

# fine-tune the model
history = model.fit_generator(train_generator,
                              epochs=nb_epoch,
                              validation_data=validation_generator,
                              class_weight=class_weightage,
                              callbacks=callbacks_list)

这是我用来训练XCeption模型的代码，使用imagenet权重进行了初始化，下面是我用来预测类的代码

def local_load_image(image_file):
    img = image.load_img(image_file, target_size=(299, 299))
    image_array = image.img_to_array(img)
    image_array = preprocess_input(np.expand_dims(image_array, axis=0))
    return image_array
prediction = model.predict(
            local_load_image(path_to_image)
        )

无论我做什么，输出预测始终为1。我正在尝试对简单的二进制数据集（狗与猫）进行分类，但没有成功。我将所有图层都设为可训练，因为BN图层会混淆训练和预测。我还尝试使用其他架构，例如VGG16和ResNet。我尝试用class_weights来解决过度拟合问题，并将学习率降低到1e-4。

我不知道为什么我的模型不能训练任何东西，而只能预测所有东西。

在11个时代之后，我可以实现

12269/12269 [==============================] - 5983s 488ms/step - loss: 0.4465 - binary_accuracy: 0.8106 val_loss: 0.4395 - val_binary_accuracy: 0.8107

但是所有的预测仍然惨败。任何帮助将不胜感激！

XCeption模型每次在keras中预测相同的输出

0 个答案: