Question

支持我有一个5卷积的网络。我是Keras写的。

x = Input(shape=(None, None, 3))
y = Conv2D(10, 3, strides=1)(x)
y = Conv2D(16, 3, strides=1)(y)
y = Conv2D(32, 3, strides=1)(y)
y = Conv2D(48, 3, strides=1)(y)
y = Conv2D(64, 3, strides=1)(y)

我想将所有卷积的kernel_initializer设置为xavier。方法之一是：

x = Input(shape=(None, None, 3))
y = Conv2D(10, 3, strides=1, kernel_initializer=tf.glorot_uniform_initializer())(x)
y = Conv2D(16, 3, strides=1, kernel_initializer=tf.glorot_uniform_initializer())(y)
y = Conv2D(32, 3, strides=1, kernel_initializer=tf.glorot_uniform_initializer())(y)
y = Conv2D(48, 3, strides=1, kernel_initializer=tf.glorot_uniform_initializer())(y)
y = Conv2D(64, 3, strides=1, kernel_initializer=tf.glorot_uniform_initializer())(y)

但是这种写法很难过，代码也很多余。

有更好的写作方式吗？

Answer 1

Keras无法更改默认值，因此您只需创建包装函数即可

def myConv2D(filters, kernel):
    return Conv2D(filters, kernel, strides=1, kernel_initializer=tf.glorot_uniform_initializer())

然后将其用作：

x = Input(shape=(None, None, 3))
y = myConv2D(10, 3)(x)
y = myConv2D(16, 3)(y)
y = myConv2D(32, 3)(y)
y = myConv2D(48, 3)(y)
y = myConv2D(64, 3)(y)

Answer 2

更好地制作一个lambda，它将构成一个Conv2D层，并根据需要修复初始化程序，并在模型定义部分中对其进行调用。

我认为lambda在这种情况下比函数更合适。

您可以这样做，

customConv = lambda filters, kernel : Conv2D(filters, kernel, strides=1, kernel_initializer=tf.glorot_uniform_initializer())

x = Input(shape=(None, None, 3))

y = customConv(10, 3)(x)
y = customConv(16, 3)(y)
y = customConv(32, 3)(y)
y = customConv(48, 3)(y)
y = customConv(64, 3)(y)

如何在tf.keras中设置Conv2D的默认参数？

2 个答案: