TF版FasterRCNN:resnet_v1.py代码阅读笔记

个人代码阅读笔记。

第二次更新:2019.4.3

# --------------------------------------------------------
# Tensorflow Faster R-CNN
# Licensed under The MIT License [see LICENSE for details]
# Written by Zheqi He and Xinlei Chen
# --------------------------------------------------------
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import tensorflow as tf
import tensorflow.contrib.slim as slim
from tensorflow.contrib.slim import losses
from tensorflow.contrib.slim import arg_scope
from tensorflow.contrib.slim.python.slim.nets import resnet_utils
from tensorflow.contrib.slim.python.slim.nets import resnet_v1
from tensorflow.contrib.slim.python.slim.nets.resnet_v1 import resnet_v1_block
import numpy as np

from nets.network import Network
from model.config import cfg
#传入一些参数,比如batch_norm_decay传入到decay中。
def resnet_arg_scope(is_training=True,
                     batch_norm_decay=0.997,
                     batch_norm_epsilon=1e-5,
                     batch_norm_scale=True):
  batch_norm_params = {
    'is_training': False,
    'decay': batch_norm_decay,
    'epsilon': batch_norm_epsilon,
    'scale': batch_norm_scale,
    'trainable': False,
    'updates_collections': tf.GraphKeys.UPDATE_OPS
  }
#arg_scope是tensorflow的slime模块自带的组建,张开一个变量作用域,方便用户定义一些参数。
#打开arg_scope,定义一些参数
  with arg_scope(
      [slim.conv2d],
      weights_regularizer=slim.l2_regularizer(cfg.TRAIN.WEIGHT_DECAY),
      weights_initializer=slim.variance_scaling_initializer(),
      trainable=is_training,
      activation_fn=tf.nn.relu,
      normalizer_fn=slim.batch_norm,
      normalizer_params=batch_norm_params):
    with arg_scope([slim.batch_norm], **batch_norm_params) as arg_sc:
      return arg_sc
#resnetv1()为Network的子类。其中有一些父类的方法不能满足需要,在子类中进行了方法重写,入
class resnetv1(Network):
  def __init__(self, num_layers=50):
    Network.__init__(self)
    self._feat_stride = [16, ]#原图到输出的缩小比例
    self._feat_compress = [1. / float(self._feat_stride[0]), ]#同上,倒数
    self._num_layers = num_layers#层数
    self._scope = 'resnet_v1_%d' % num_layers#scope的名称,我用的resnet_v1_101,所以打开它的scope
    self._decide_blocks()

  def _crop_pool_layer(self, bottom, rois, name):#这里是roi处理的步骤,crop对应的特征区域,进行Pooling到7x7
    with tf.variable_scope(name) as scope:
      batch_ids = tf.squeeze(tf.slice(rois, [0, 0], [-1, 1], name="batch_id"), [1])
      # Get the normalized coordinates of bboxes
      bottom_shape = tf.shape(bottom)
      height = (tf.to_float(bottom_shape[1]) - 1.) * np.float32(self._feat_stride[0])
      width = (tf.to_float(bottom_shape[2]) - 1.) * np.float32(self._feat_stride[0])
      x1 = tf.slice(rois, [0, 1], [-1, 1], name="x1") / width
      y1 = tf.slice(rois, [0, 2], [-1, 1], name="y1") / height
      x2 = tf.slice(rois, [0, 3], [-1, 1], name="x2") / width
      y2 = tf.slice(rois, [0, 4], [-1, 1], name="y2") / height
      # Won't be back-propagated to rois anyway, but to save time
      bboxes = tf.stop_gradient(tf.concat([y1, x1, y2, x2], 1))
      if cfg.RESNET.MAX_POOL:
        pre_pool_size = cfg.POOLING_SIZE * 2
        crops = tf.image.crop_and_resize(bottom, bboxes, tf.to_int32(batch_ids), [pre_pool_size, pre_pool_size],
                                         name="crops")
        crops = slim.max_pool2d(crops, [2, 2], padding='SAME')
      else:
        crops = tf.image.crop_and_resize(bottom, bboxes, tf.to_int32(batch_ids), [cfg.POOLING_SIZE, cfg.POOLING_SIZE],
                                         name="crops")
    return crops

  # Do the first few layers manually, because 'SAME' padding can behave inconsistently
  # for images of different sizes: sometimes 0, sometimes 1
  #对于大小不一样的图像,same模式的padding可能回产生不同的运算结果,为了保持一致手工的定义网络的头部。
  def _build_base(self):
    with tf.variable_scope(self._scope, self._scope):
      #首先创建一个卷积层,64个卷积和,7x7,步长为2
      net = resnet_utils.conv2d_same(self._image, 64, 7, stride=2, scope='conv1')
	  #对输入图像卷积之后进行pad,用的是tf.pad函数。
      net = tf.pad(net, [[0, 0], [1, 1], [1, 1], [0, 0]])
	  #再进行最大池化,步长为2,大小为3x3
      net = slim.max_pool2d(net, [3, 3], stride=2, padding='VALID', scope='pool1')
    #返回前面手工定义层的处理结果。
    return net

  def _image_to_head(self, is_training, reuse=None):
   #检查:需要固定参数的block是否变化。res101一共4个Block,分别从0-3,默认设置的是1,代表我训练的时候,前两个blocks的权重是不变的,后面的变化  
    assert (0 <= cfg.RESNET.FIXED_BLOCKS <= 3)
    # Now the base is always fixed during training
    with slim.arg_scope(resnet_arg_scope(is_training=False)):
      net_conv = self._build_base()#类内方法的相互调用形式为[self.方法名字],这里相当于把前面_build_base的结算结果调用过来。
    #因为要fixed的权重block可能是处于中间,所以使用的运算机制是:首先固定住我们需要的层,为非训练模式,然后设置剩下的层为训练模式。is_training=True
    if cfg.RESNET.FIXED_BLOCKS > 0:
      with slim.arg_scope(resnet_arg_scope(is_training=False)):
	  #注意,这里返回的net_conv,是经过了几个固定block的op后的net_conv
        net_conv, _ = resnet_v1.resnet_v1(net_conv,
                                           self._blocks[0:cfg.RESNET.FIXED_BLOCKS],#[0:cfg.RESNET.FIXED_BLOCKS]即为前几个固定的blocks
                                           global_pool=False,
                                           include_root_block=False,
                                           reuse=reuse,
                                           scope=self._scope)
    if cfg.RESNET.FIXED_BLOCKS < 3:
	  #slim.arg_scope(resnet_arg_scope(is_training=true or false))应该是slim的标准语法,用于区分训练和非训练的变量域
      with slim.arg_scope(resnet_arg_scope(is_training=is_training)):#虽然这是is_training=is_training,但是训练的时候传入的是true,测试的时候依然是false。这样写很简洁
        net_conv, _ = resnet_v1.resnet_v1(net_conv,
                                           self._blocks[cfg.RESNET.FIXED_BLOCKS:-1],#[cfg.RESNET.FIXED_BLOCKS:-1]即为后几个固定的Blocks
                                           global_pool=False,
                                           include_root_block=False,
                                           reuse=reuse,
                                           scope=self._scope)

    self._act_summaries.append(net_conv)#这里应该是tensorboard的活动总结的变量,把到这一步的结果记录
    self._layers['head'] = net_conv#同时也把结果保存到Layers字典中key='head'下

    return net_conv#返回计算结果。
  #res101是Network的子类,在network中_build_network调用了crop_pool_layer方法,即roi-pooling,得到的就是Pool5
  #这里可能会有一个疑问:最后的feature maps上有很多个roi,这里没有用for循环,是怎么批量把这些rois对应的特征块进行crop and resize的呢?
  #用了矩阵的结构,每一行是一个roi,按列来处理,就相当于对行进行批处理了。下面函数隐含的处理也有不同,针对每一行进行计算,而不是一起同时计算。
  def _head_to_tail(self, pool5, is_training, reuse=None):
    with slim.arg_scope(resnet_arg_scope(is_training=is_training)):
     #打开变量域,全连接层属于可学类型权重
	 #返回fc7的计算结果
      fc7, _ = resnet_v1.resnet_v1(pool5,
                                   self._blocks[-1:],#申明位置,在Block之后。
                                   global_pool=False,
                                   include_root_block=False,
                                   reuse=reuse,#变量重用
                                   scope=self._scope)
      # average pooling done by reduce_mean
	  #全连接层去均值处理。
      fc7 = tf.reduce_mean(fc7, axis=[1, 2])
    return fc7#返回计算结果

  def _decide_blocks(self):
    # choose different blocks for different number of layers
    if self._num_layers == 50:
      self._blocks = [resnet_v1_block('block1', base_depth=64, num_units=3, stride=2),
                      resnet_v1_block('block2', base_depth=128, num_units=4, stride=2),
                      # use stride 1 for the last conv4 layer
                      resnet_v1_block('block3', base_depth=256, num_units=6, stride=1),
                      resnet_v1_block('block4', base_depth=512, num_units=3, stride=1)]
    #基本就是调用slim的blocks参数。比如我输入res101,就会得到下面的扩展参数,这些参数再进入slim里面生成res101网络
    elif self._num_layers == 101:
	#举例:第一个block
	#名字:'block1'
	#64个卷积核
	#该层结构复制3次
	#卷积步长为2
      self._blocks = [resnet_v1_block('block1', base_depth=64, num_units=3, stride=2),
                      resnet_v1_block('block2', base_depth=128, num_units=4, stride=2),
                      # use stride 1 for the last conv4 layer
                      resnet_v1_block('block3', base_depth=256, num_units=23, stride=1),
                      resnet_v1_block('block4', base_depth=512, num_units=3, stride=1)]

    elif self._num_layers == 152:
      self._blocks = [resnet_v1_block('block1', base_depth=64, num_units=3, stride=2),
                      resnet_v1_block('block2', base_depth=128, num_units=8, stride=2),
                      # use stride 1 for the last conv4 layer
                      resnet_v1_block('block3', base_depth=256, num_units=36, stride=1),
                      resnet_v1_block('block4', base_depth=512, num_units=3, stride=1)]

    else:
      # other numbers are not supported
      raise NotImplementedError

  def get_variables_to_restore(self, variables, var_keep_dic):
    variables_to_restore = []
    #一个变量保存的函数
	#传入变量以及
    for v in variables:
      # exclude the first conv layer to swap RGB to BGR
	  #对于每一个名字为resnet_v1_101/conv1/weights:0的变量进行保存,
      if v.name == (self._scope + '/conv1/weights:0'):
        self._variables_to_fix[v.name] = v#self._variables_to_fix是在父类network中创建的字典。这里相当于将这个变量加入字典
        continue
      #如果这个变量resnet_v1_101/conv1/weights在变量保留字典里面,就加入到variables_to_restore里。
      if v.name.split(':')[0] in var_keep_dic:
        print('Variables restored: %s' % v.name)
        variables_to_restore.append(v)

    return variables_to_restore#返回保留的变量
  #在lib/model/train_val.py用到,self.net.fix_variables(sess, self.pretrained_model)
  def fix_variables(self, sess, pretrained_model):#这里主要是训练前修正变量,将rgb转换为bgr。首先从模型里面回复,然后进行通道反转。总之模型本身参数是rgb通道的。
    print('Fix Resnet V1 layers..')
    with tf.variable_scope('Fix_Resnet_V1') as scope:#打开名为Fix_Resnet_V1的变量域
      with tf.device("/cpu:0"):#指定cpu运行
        # fix RGB to BGR
		#使用tf.get_variable调用变量,没有就创建变量,名字为"conv1_rgb",大小为7x7x3x64,7x7的大小,三个通道,64个卷积核。其实就是rgb三个通道到bgr的转换,因为是转换,所以不可训练。
        conv1_rgb = tf.get_variable("conv1_rgb", [7, 7, 3, 64], trainable=False)
        restorer_fc = tf.train.Saver({self._scope + "/conv1/weights": conv1_rgb})#这里创建了一个saver对象,saver(变量),变量是要保存或者回复的变量,这里主要是回复。
        restorer_fc.restore(sess, pretrained_model)#对变量进行回复。注意,这里的saver指定了回复的变量,不是整个模型都回复。
#tf.assign是指定,将self._variables_to_fix[self._scope + '/conv1/weights:0']的值指定为tf.reverse(conv1_rgb, [2])
        sess.run(tf.assign(self._variables_to_fix[self._scope + '/conv1/weights:0'], 
                           tf.reverse(conv1_rgb, [2])))#tf.reverse为反转,后面[2]是指定反转的维度,这里指定了通道反转,rgb变为bgr,因为cv2读入是bgr吧

 

你可能感兴趣的:(Faster,R-CNN,目标检测,tensorflow)