tf.nn.conv2d和tf.contrib.slim.conv2d的区别

来源:http://blog.sina.com.cn/s/blog_6ca0f5eb0102wsuu.html

在查看代码的时候,看到有代码用到卷积层是tf.nn.conv2d,但是也有的使用的卷积层是tf.contrib.slim.conv2d,这两个函数调用的卷积层是否一致,在查看了API的文档,以及slim.conv2d的源码后,做如下总结:

首先是常见使用的tf.nn.conv2d的函数,其定义如下:

conv2d(

    input,

    filter,

    strides,

    padding,

    use_cudnn_on_gpu=None,

    data_format=None,

    name=None

)

input指需要做卷积的输入图像,它要求是一个Tensor,具有[batch_size, in_height, in_width, in_channels]这样的shape,具体含义是[训练时一个batch的图片数量, 图片高度, 图片宽度, 图像通道数],注意这是一个4维的Tensor,要求数据类型为float32和float64其中之一

filter用于指定CNN中的卷积核,它要求是一个Tensor,具有[filter_height, filter_width, in_channels, out_channels]这样的shape,具体含义是[卷积核的高度,卷积核的宽度,图像通道数,卷积核个数],要求类型与参数input相同,有一个地方需要注意,第三维in_channels,就是参数input的第四维,这里是维度一致,不是数值一致

strides为卷积时在图像每一维的步长,这是一个一维的向量,长度为4,对应的是在input的4个维度上的步长

paddingstring类型的变量,只能是"SAME","VALID"其中之一,这个值决定了不同的卷积方式,SAME代表卷积核可以停留图像边缘,VALID表示不能,更详细的描述可以参考http://blog.csdn.net/mao_xiao_feng/article/details/53444333

use_cudnn_on_gpu指定是否使用cudnn加速,默认为true

data_format是用于指定输入的input的格式,默认为NHWC格式

 

结果返回一个Tensor,这个输出,就是我们常说的feature map

 

而对于tf.contrib.slim.conv2d,其函数定义如下:

convolution(inputs,

          num_outputs,

          kernel_size,

          stride=1,

          padding='SAME',

          data_format=None,

          rate=1,

          activation_fn=nn.relu,

          normalizer_fn=None,

          normalizer_params=None,

          weights_initializer=initializers.xavier_initializer(),

          weights_regularizer=None,

          biases_initializer=init_ops.zeros_initializer(),

          biases_regularizer=None,

          reuse=None,

          variables_collections=None,

          outputs_collections=None,

          trainable=True,

          scope=None):

inputs同样是指需要做卷积的输入图像

num_outputs指定卷积核的个数(就是filter的个数)

kernel_size用于指定卷积核的维度(卷积核的宽度,卷积核的高度)

stride为卷积时在图像每一维的步长

padding为padding的方式选择,VALID或者SAME

data_format是用于指定输入的input的格式

rate这个参数不是太理解,而且tf.nn.conv2d中也没有,对于使用atrous convolution的膨胀率(不是太懂这个atrous convolution)

activation_fn用于激活函数的指定,默认的为ReLU函数

normalizer_fn用于指定正则化函数

normalizer_params用于指定正则化函数的参数

weights_initializer用于指定权重的初始化程序

weights_regularizer为权重可选的正则化程序

biases_initializer用于指定biase的初始化程序

biases_regularizer: biases可选的正则化程序

reuse指定是否共享层或者和变量

variable_collections指定所有变量的集合列表或者字典

outputs_collections指定输出被添加的集合

trainable:卷积层的参数是否可被训练

scope:共享变量所指的variable_scope

 

在上述的API中,可以看出去除掉初始化的部分,那么两者并没有什么不同,只是tf.contrib.slim.conv2d提供了更多可以指定的初始化的部分,而对于tf.nn.conv2d而言,其指定filter的方式相比较tf.contrib.slim.conv2d来说,更加的复杂。去除掉少用的初始化部分,其实两者的API可以简化如下:

tf.contrib.slim.conv2d (inputs,

                num_outputs,[卷积核个数]

                kernel_size,[卷积核的高度,卷积核的宽度]

                stride=1,

                padding='SAME',

)

tf.nn.conv2d(

    input,(与上述一致)

    filter,([卷积核的高度,卷积核的宽度,图像通道数,卷积核个数])

    strides,

    padding,

)

可以说两者是几乎相同的,运行下列代码也可知这两者一致

import tensorflow as tf 

import tensorflow.contrib.slim as slim

 

x1 = tf.ones(shape=[1, 64, 64, 3]) 

w = tf.fill([5, 5, 3, 64], 1)

# print("rank is", tf.rank(x1))

y1 = tf.nn.conv2d(x1, w, strides=[1, 1, 1, 1], padding='SAME')

y2 = slim.conv2d(x1, 64, [5, 5], weights_initializer=tf.ones_initializer, padding='SAME')

 

 

with tf.Session() as sess: 

    sess.run(tf.global_variables_initializer()) 

    y1_value,y2_value,x1_value=sess.run([y1,y2,x1])

    print("shapes are", y1_value.shape, y2_value.shape)

    print(y1_value==y2_value)

    print(y1_value)

print(y2_value)

 

最后配上tf.contrib.slim.conv2d的API英文版

def convolution(inputs,

                num_outputs,

                kernel_size,

                stride=1,

                padding='SAME',

                data_format=None,

                rate=1,

                activation_fn=nn.relu,

                normalizer_fn=None,

                normalizer_params=None,

                weights_initializer=initializers.xavier_initializer(),

                weights_regularizer=None,

                biases_initializer=init_ops.zeros_initializer(),

                biases_regularizer=None,

                reuse=None,

                variables_collections=None,

                outputs_collections=None,

                trainable=True,

                scope=None):

  """Adds an N-D convolution followed by an optional batch_norm layer.

  It is required that 1 <= N <= 3.

  `convolution` creates a variable called `weights`, representing the

  convolutional kernel, that is convolved (actually cross-correlated) with the

  `inputs` to produce a `Tensor` of activations. If a `normalizer_fn` is

  provided (such as `batch_norm`), it is then applied. Otherwise, if

  `normalizer_fn` is None and a `biases_initializer` is provided then a `biases`

  variable would be created and added the activations. Finally, if

  `activation_fn` is not `None`, it is applied to the activations as well.

  Performs atrous convolution with input stride/dilation rate equal to `rate`

  if a value > 1 for any dimension of `rate` is specified.  In this case

  `stride` values != 1 are not supported.

  Args:

    inputs: A Tensor of rank N+2 of shape

      `[batch_size] + input_spatial_shape + [in_channels]` if data_format does

      not start with "NC" (default), or

      `[batch_size, in_channels] + input_spatial_shape` if data_format starts

      with "NC".

    num_outputs: Integer, the number of output filters.

    kernel_size: A sequence of N positive integers specifying the spatial

      dimensions of the filters.  Can be a single integer to specify the same

      value for all spatial dimensions.

    stride: A sequence of N positive integers specifying the stride at which to

      compute output.  Can be a single integer to specify the same value for all

      spatial dimensions.  Specifying any `stride` value != 1 is incompatible

      with specifying any `rate` value != 1.

    padding: One of `"VALID"` or `"SAME"`.

    data_format: A string or None.  Specifies whether the channel dimension of

      the `input` and output is the last dimension (default, or if `data_format`

      does not start with "NC"), or the second dimension (if `data_format`

      starts with "NC").  For N=1, the valid values are "NWC" (default) and

      "NCW".  For N=2, the valid values are "NHWC" (default) and "NCHW".

      For N=3, the valid values are "NDHWC" (default) and "NCDHW".

    rate: A sequence of N positive integers specifying the dilation rate to use

      for atrous convolution.  Can be a single integer to specify the same

      value for all spatial dimensions.  Specifying any `rate` value != 1 is

      incompatible with specifying any `stride` value != 1.

    activation_fn: Activation function. The default value is a ReLU function.

      Explicitly set it to None to skip it and maintain a linear activation.

    normalizer_fn: Normalization function to use instead of `biases`. If

      `normalizer_fn` is provided then `biases_initializer` and

      `biases_regularizer` are ignored and `biases` are not created nor added.

      default set to None for no normalizer function

    normalizer_params: Normalization function parameters.

    weights_initializer: An initializer for the weights.

    weights_regularizer: Optional regularizer for the weights.

    biases_initializer: An initializer for the biases. If None skip biases.

    biases_regularizer: Optional regularizer for the biases.

    reuse: Whether or not the layer and its variables should be reused. To be

      able to reuse the layer scope must be given.

    variables_collections: Optional list of collections for all the variables or

      a dictionary containing a different list of collection per variable.

    outputs_collections: Collection to add the outputs.

    trainable: If `True` also add variables to the graph collection

      `GraphKeys.TRAINABLE_VARIABLES` (see tf.Variable).

    scope: Optional scope for `variable_scope`.

  Returns:

    A tensor representing the output of the operation.

  Raises:

    ValueError: If `data_format` is invalid.

    ValueError: Both 'rate' and `stride` are not uniformly 1.

原文地址:https://www.cnblogs.com/fujian-code/p/9596883.html