Keras 用面具冻结特定的权重

Question

我是 Keras 的新手。我想实现一个并非所有权重都会更新的层。例如，在下面的代码中，我希望dilation 层将更新为某些中心权重永远不会更新。例如，dilation 层中每个特征矩阵（共 1024 个）的形状为 448, 448，并且所有特征矩阵中心的 8x8 块将永远不会更新，即 8x8 块是特征矩阵的（不可训练的）掩码。

input_layer=Input(shape=(896,896,3))
new_layer = Conv2D(32, kernel_size=(3,3), padding="same", activation='relu', kernel_initializer='he_normal')(input_layer)
new_layer = MaxPooling2D(pool_size=(2, 2), strides=(2,2), padding='same', data_format=None)(new_layer)
new_layer = Conv2D(64, kernel_size=(3,3), padding='same', activation='relu', kernel_initializer='he_normal')(new_layer)
new_layer = Conv2D(1024, kernel_size=(7,7), dilation_rate=8, padding="same", activation='relu', kernel_initializer='he_normal', name='dialation')(new_layer)
new_layer = Conv2D(32, kernel_size=(1,1), padding="same", activation='relu', kernel_initializer='he_normal')(new_layer)
new_layer = Conv2D(32, kernel_size=(1,1), padding="same", activation='relu', kernel_initializer='he_normal')(new_layer)

model = Model(input_layer, new_layer)

我正在尝试使用 Keras 的 custom layer [link]，但我很难理解。任何人都会请帮忙。

更新： 我添加了下图以便更好地理解。膨胀层包含 1024 个特征。我希望每个特征的中间区域是不可训练的（静态的）。

Answer 1

在这两种情况下都使用这个掩码：

mask = np.zeros((1,448,448,1))
mask[:,220:228,220:228] = 1

替换部分功能

如果将部分特征替换为常数值，这意味着该特征将是静态的，但仍会参与反向传播（因为这部分图像的权重仍会相乘和相加，并且存在连接）

constant = 0 (will annulate kernel, but not bias) 

def replace(x):
    return x*(1-mask) + constant*mask

#before the dilation layer
new_layer=Lambda(replace)(new_layer) 

保持特征值，但停止反向传播

这里dilation layer和further的权重会正常更新，但是dilation layer之前的权重不会受到中心区域的影响。

def stopBackprop(x):
    stopped=K.stop_gradients(x)
    return x*(1-mask) + stopped*mask

#before the dilation layer
new_layer=Lambda(stopBackprop)(new_layer) 

Answer 2

我正在研究聚类权重，然后冻结特定的聚类并训练网络。

我正在尝试使用上述示例冻结此网络中的特定权重。但我不确定如何在 run_certain_weights() 中设置蒙版和自定义图层的形状。

这是我正在使用的代码：

from keras.layers import Dense, Flatten, Lambda
from keras.utils import to_categorical
from keras.models import Sequential, load_model
from keras.datasets import mnist
from keras.losses import categorical_crossentropy
from keras.backend import stop_gradient
import numpy as np

def load_data():
    
    (x_train, y_train), (x_test, y_test) = mnist.load_data()
    x_train = x_train.astype('float32')
    x_test = x_test.astype('float32')
    x_train /= 255
    x_test /= 255
    y_train = to_categorical(y_train, num_classes=10)
    y_test = to_categorical(y_test, num_classes=10)
    
    return x_train, y_train, x_test, y_test

def run():
    x_train, y_train, x_test, y_test = load_data()

    model=Sequential(Flatten(input_shape=(28, 28)))

    layer = Dense(300, name='dense1', activation='relu')
    layer.trainable=True
    model.add(layer)
    layer2 = Dense(100, name='dense2', activation='relu')
    layer2.trainable=False
    model.add(layer2)
    layer3 = Dense(10, name='dense3', activation='softmax')
    model.add(layer3)

    model.compile(loss=categorical_crossentropy, optimizer='Adam',metrics ['accuracy'])


    print(model.summary())
    print("x_train.shape():",x_train.shape)
    print("y_train.shape()",y_train.shape)

    model.fit(x_train, y_train, epochs=5, verbose=2)
    print(model.evaluate(x_test, y_test))
    
    return model

def stopBackprop(x):
    
    stopped=stop_gradient(x)
    return x*(1-mask) + stopped*mask

def run_certain_weights():
    
    x_train, y_train, x_test, y_test = load_data()
    model=Sequential(Flatten(input_shape=(28, 28)))
    mask = np.zeros((300,))
    print(mask.shape)
    mask[220:228,] = 1

    layer = Dense(300, name='dense1', activation='relu')
    layer.trainable=False
    model.add(layer)

    #before the dense2 layer
    new_layer=Lambda(stopBackprop)(layer)
    model.add(new_layer)

    layer2 = Dense(300, name='dense2', activation='relu')
    layer2.trainable=True
    model.add(layer2)
    layer3 = Dense(10, name='dense3', activation='softmax')
    model.add(layer3)

    model.compile(loss=categorical_crossentropy, optimizer='Adam',metrics = ['accuracy'])


    print(model.summary())
    print("x_train.shape():",x_train.shape)
    print("y_train.shape()",y_train.shape)

    model.fit(x_train, y_train, epochs=5, verbose=2)
    print(model.evaluate(x_test, y_test))
    return model


def freeze(model):
    x_train, y_train, x_test, y_test = load_data()

    name = 'dense2'

    weightsAndBias = model.get_layer(name=name).get_weights()

    # freeze the weights of this layer
    model.get_layer(name=name).trainable = False

    # record the weights before retrain
    weights_before = weightsAndBias[0]
    # retrain
    print("x_train.shape():",x_train.shape)
    print("y_train.shape()",y_train.shape)

    model.fit(x_train, y_train, verbose=2, epochs=1)
    weights_after = model.get_layer(name=name).get_weights()[0]

    if (weights_before == weights_after).all():
        print('the weights did not change!!!')
    else:
        print('the weights changed!!!!')

if __name__ == '__main__':
    
    model = run()
    freeze(model)

    model = run_certain_weights()
    freeze(model)