keras权重偏置初始化等相关操作

打印权重

#构建网络
model = Sequential()
model.add(Dense(1, input_shape=(2,), activation='linear', bias_initializer=initializers.Constant(value=0)))
model.compile(loss='mean_squared_error', optimizer='adam')

# 控制速率衰减
reduce_lr = ReduceLROnPlateau(monitor='loss', patience=100, factor=0.5, mode='auto')
# 实时保存文件
checkpointer = ModelCheckpoint(filepath='./weights/weights0522_1.hdf5', monitor='loss', verbose=1, save_best_only=True)


model.load_weights('./weights/weights0520_1.hdf5')
weights = np.array(model.get_weights())
print(weights)

也可以这样打印权重

model.load_weights('./weights/weights0520_1.hdf5')
weights, bias = np.array(model.get_weights())
print(weights)
print(bias)

最后三行是重点，用于加载权重并将其打印出来

设定偏置

from keras import initializers

model.add(Dense(64, kernel_initializer=initializers.random_normal(stddev=0.01)))

# also works; will use the default parameters.
model.add(Dense(64, kernel_initializer='random_normal'))

上面的是官方实例https://keras-cn.readthedocs.io/en/latest/other/initializations/

剔除全连接层中的偏置

#构建网络。注意，全连接层中的偏置已经被失能
model = Sequential()
model.add(Dense(1, input_shape=(2,), activation='linear', kernel_initializer=my_init, use_bias=False))
model.compile(loss='mean_squared_error', optimizer='adam')

use_bias用于设定是否在模型中设置偏重项

自定义权重系数

# 通过传统理论，对权重进行初始化
def my_init(shape, dtype=None):
    return np.array([[1], [100]])

#构建网络。注意，全连接层中的偏置已经被失能
model = Sequential()
model.add(Dense(1, input_shape=(2,), activation='linear', kernel_initializer=my_init, use_bias=False))
model.compile(loss='mean_squared_error', optimizer='adam')

其中my_init函数的形参必须具备shape和dtype这俩玩意儿，同时要注意返回值是个np.array以及它的形状