tf.nn.conv2d & tf.contrib.layers.conv2d & tf.contrib.slim.conv2d

本文主要介绍前两个函数tf.nn.conv2d和tf.contrib.layers.conv2d
因为tf.contrib.layers.conv2d 和 tf.contrib.slim.conv2d用法是相似的，都是高级api，只是slim是一个更高级别的库，用slim中的repeat函数，可以用几行就写出一个vgg16网络。但是tf.nn和tf.contrib.layers是基本ops，最常见也最重要。

首先说tf.nn.conv2d()的参数

tf.nn.conv2d(
    input,
    filter,
    strides,
    padding,
    use_cudnn_on_gpu=True,
    data_format='NHWC',
    dilations=[1, 1, 1, 1],
    name=None
)

input的格式由data_format参数决定，默认都是‘NHWC’，也就是batch size；feature map height；feature map width；channels。
filter是单独定义的, 可以用tf.get_variable()或者tf.Variable()来定义weight，关于这两种定义方法，可以参见我的另一篇博文
以tf.Variable和tf.nn.conv2d为例子来建立一个卷积层：

def create_weights(shape):
    return tf.Variable(tf.truncated_normal(shape,stddev=0.1))

def create_bias(num_filters):
    return tf.Variable(tf.constant(0.05,shape=num_filters))

def create_conv_layer(input,num_channels,filtr_size,num_filters):
    weights=create_weights(shape=[input,num_channels,filtr_size,num_filter])
    bias=create_bias(num_filters)
    layer=tf.nn.conv2d(input=input,filter=weights,stride=[1,1,1,1],padding='same')
    layer+=bias
    layer=tf.nn.max_pool(layer,ksize=[1,2,2,1],strides=[1,2,2,1],padding='same')
    return layer

以上是建立卷积层最基础的方式。
接下来来看用tf.contrib.layers.conv2d创建卷积层。

tf.contrib.layers.conv2d(
    inputs, #四维tensor ‘NHDC’
    num_outputs, #filter的数量
    kernel_size,#[k,k]
    stride=1,#默认为[1,1]
    padding='SAME',#valid or same 
    data_format=None,#NHDC
    rate=1,#dilation rate
    activation_fn=tf.nn.relu,#默认是relu，可选tf.nn.leaky_relu
    normalizer_fn=None,
    normalizer_params=None,
    weights_initializer=tf.contrib.layers.xavier_initializer(),
    weights_regularizer=None,
    biases_initializer=tf.zeros_initializer(),
    biases_regularizer=None,
    reuse=None,
    variables_collections=None,
    outputs_collections=None,
    trainable=True,
    scope=None
)

这里提一下tf.contrib.layers.xavier_initializer()
这个比 tf.truncated_normal_initializer好一点(不过有人做实验显示两者效果差不多）。
xavier initializer is designed to keep the scale of the gradients roughly the same in all layers.
In uniform distribution, tf.contrib.layers.xavier_initializer(uniform=True) this ends up being the range: x = sqrt(6. / (in + out)); [-x, x]
In normal distribution, tf.contrib.layers.xavier_initializer(uniform=False) a standard deviation of sqrt(3. / (in + out)) is used.
有以下几篇介绍xavier initialier及其他initializer的文章很详细
1 why xavier and He initializer
2 聊一聊weights initializer
3 这篇也不错
tf.contrib.layers.conv2d()简单的example是：

def mynet(input,reuse=False):
    with tf.variable_scope("conv1") as scope:
        net=tf.contrib.layers.conv2d(input,32,[5,5],padding='SAME',weights_initializer=tf.contrib.layers.xavier_initializer(uniform=False),scope=scope,reuse=reuse)
        net=tf.contrib.layers.max_pool2d(net,[2,2],padding='SAME')
    with tf.variable_scope("conv2") as scope:
        ...
    net=tf.contrib.layers.flatten(net)
    return net

tf.nn.conv2d & tf.contrib.layers.conv2d & tf.contrib.slim.conv2d

猜你喜欢