卷积层的改变滤波器CNN - Python/TensorFlow

Question

我有以下代码块：

def new_weights(shape):
    return tf.Variable(tf.truncated_normal(shape, stddev=0.05))

和：

def new_conv_layer(input,              # The previous layer
                   use_pooling=True):  # Use 2x2 max-pooling

    shape = [3, 3, 1, 8]

    weights = new_weights(shape=shape)

    biases = new_biases(length=8)

    layer = tf.nn.conv2d(input=input,
                         filter=weights,
                         strides=[1, 1, 1, 1],
                         padding='SAME')

    layer += biases

    if use_pooling:
        layer = tf.nn.max_pool(value=layer,
                               ksize=[1, 2, 2, 1],
                               strides=[1, 2, 2, 1],
                               padding='SAME')

    layer = tf.nn.relu(layer)

    # relu(max_pool(x)) == max_pool(relu(x)) we can
    # save 75% of the relu-operations by max-pooling first.

    return layer

因此，我们可以观察到，滤波器的大小为3x3，滤波器的数目为8，滤波器定义为随机值。

我需要做的是用固定值定义我的所有8个过滤器(例如，预定值)：

weigths = [
    [[0,  1, 0,],[0, -1, 0,],[0,  0, 0,],],
    [[0,  0, 1,],[0, -1, 0,],[0,  0, 0,],],
    [[0,  0, 0,],[0, -1, 1,],[0,  0, 0,],],
    [[0,  0, 0,],[0, -1, 0,],[0,  0, 1,],],
    [[0,  0, 0,],[0, -1, 0,],[0,  1, 0,],],
    [[0,  0, 0,],[0, -1, 0,],[1,  0, 0,],], 
    [[0,  0, 0,],[1, -1, 0,],[0,  0, 0,],],
    [[1,  0, 0,],[0, -1, 0,],[0,  0, 0,],]
]

我无法想象，我如何在我的代码中，这个修改，谁知道我如何能做到这一点？

非常感谢您提前！

Answer 1

如果您想要用某个预定义的值初始化权重，可以使用tf.constant_initializer。如果您不想训练这个权重，您可以将它们定义为tf.constant而不是tf.Variable

def new_weights(init_vaue, is_const):
    if (is_const) :
        return tf.constant(init_vaue, name='weights')
    else:
        initializer = tf.constant_initializer(init_vaue)
        return tf.get_variable('weights', shape = init_vaue.shape, initializer=initializer)

weights = np.ones([3,3,1,8], dtype=np.float)
print(weights.shape)

value = new_weights(weights, True)
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    value_ = sess.run(value) 
    print(value_)

Answer 2

您可以在TF2中这样做：

model = models.Sequential()
# one 3x3 filter
model.add(layers.Conv2D(1, (3, 3), input_shape=(None, None, 1)))
# access to the target layer
layer = model.layers[0]
current_w, current_bias = layer.get_weights()  # see the current weights
new_w = tf.constant([[1,2, 3],
                     [4, 5, 6],
                     [7, 8, 9]])
new_w = tf.reshape(new_w, custom_w.shape)  # fix the shape
new_bias = tf.constant([0])
layer.set_weights([new_w, new_bias])
model.summary()
# let's see ..
tf.print(model.layers[0].get_weights())

Answer 3

您只需将权重定义为不可训练，并将新权重定义为：

new_weights = tf.Variable( tf.reshape(weights, (3,3,1,8)),trainable=False)
# then apply on the inputs 
layer = tf.nn.conv2d(inputs, filter=new_weights, strides=[1, 1, 1, 1], padding='SAME')