如何使用高低阶的API搭建网络

全连接

结构：flatten$\rightarrow$ xw+b$\rightarrow$ norm layer(optional)$\rightarrow$activation function $\rightarrow$dropout(optional)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

def add_fc_layer(inputs, in_dim, out_dim, scope_name, activation_function=None):

    with tf.variable_scope(scope_name):

        w = tf.get_variable(shape=[in_dim, out_dim], initializer=tf.contrib.layers.xavier_initializer(), name='weights')

        b = tf.get_variable(shape=[1, out_dim], initializer=tf.contrib.layers.xavier_initializer(), name='biases')

        wx_plus_b = tf.matmul(inputs, w) + b  # broadcast

        if activation_function:

        	wx_plus_b = activation_function(wx_plus_b)

    return wx_plus_b

卷积

结构： convolution $\rightarrow$ norm layer(optional) $\rightarrow$activation function $\rightarrow$pooling(optional) 

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
def add_conv2d_layer(inputs, in_channel, out_channel, scope_name, activation_function, filter_h, filter_w,stride):

	with tf.variable_scope(scope_name):

		fliter_mask = tf.get_variable(shape=[filter_h, filter_w, in_channel, out_channel], dtype=inputs.dtype, initializer=tf.contrib.layers.xavier_initializer(),name='filter')

		# 'same' 输入输出的宽高一致， 'valid', out_width = xxxx #
		conv_inputs  = tf.nn.conv2d(inputs, fliter_mask, [1,stride,stride,1], 'SAME')

		# batch normalization #

		if activation_function:

			conv_inputs = activation_function(conv_inputs)

		# pooling #

		conv_inputs = tf.nn.max_pool(conv_inputs,[1,2,2,1],[1,2,2,1],'SAME')

	return conv_inputs

反卷积

结构： inverse-convolution $\rightarrow$ norm layer(optional) $\rightarrow$activation function