Tensorflow开源的object detection API中的源码解析（三）：faster_rcnn_inception_resnet_v2_feature_extractor.py
对于faster_rcnn_meta_arch.py的理解见这篇文章
对于inception_resnet_v2.py的理解见这篇文章
"""Inception Resnet v2 Faster R-CNN implementation.
参见Szegedy等人的“Inception-v4，Inception-ResNet Impact of Residual Connections on
Learning"（论文地址：https://arxiv.org/abs/1602.07261）
以及Huang等人的
“Speed/accuracy trade-offs for modern convolutional object detectors" by
（论文地址：https://arxiv.org/abs/1611.10012）
"""
import tensorflow as tf
from object_detection.meta_architectures import faster_rcnn_meta_arch
from object_detection.net import inception_resnet_v2
slim = tf.contrib.slim
class FasterRCNNInceptionResnetV2FeatureExtractor(
faster_rcnn_meta_arch.FasterRCNNFeatureExtractor):
""" 这个类主要是Inception_Resnet_v2 Faster R-CNN的特征提取器的相关实现 """
def __init__(self,
is_training,
first_stage_features_stride,
reuse_weights=None,
weight_decay=0.0):
"""Constructor.
Args:
is_training: 见faster_rcnn_meta_arch的构造器
first_stage_features_stride: 同上
reuse_weights: 同上
weight_decay: 同上
Raises:
ValueError: 如果 `first_stage_features_stride` 既不是8也不是16。
"""
if first_stage_features_stride != 8 and first_stage_features_stride != 16:
raise ValueError('`first_stage_features_stride` must be 8 or 16.')
super(FasterRCNNInceptionResnetV2FeatureExtractor, self).__init__(
is_training, first_stage_features_stride, reuse_weights, weight_decay)
def preprocess(self, resized_inputs):
"""Faster R-CNN with Inception Resnet v2 的预处理函数.
将像素值映射到[-1, 1]范围。（归一化）
Args:
resized_inputs: 一批值在0到255.0之间的图像
A [batch, height_in, width_in, channels] float32 tensor
Returns: 一批图像
preprocessed_inputs: A [batch, height_out, width_out, channels] float32 tensor.
"""
# TODO：为什么是这样计算
return (2.0 / 255.0) * resized_inputs - 1.0
def _extract_proposal_features(self, preprocessed_inputs, scope):
"""提取第一阶段将用于RPN的特征，返回feature map。实现faster_rcnn_meta_arch中的抽象方法
使用Inception Resnet v2网络的前半部分提取特征[将用于RPN的特征]。
在`align_feature_maps = True`模式下构建网络，
这意味着网络中的所有VALID填充都被改为SAME填充，以便特征映射对齐。
Args:
preprocessed_inputs: 一批图像
A [batch, height, width, channels] float32 tensor
scope: A scope name.
Returns:将用于RPN的特征
rpn_feature_map: A tensor with shape [batch, height, width, depth]
Raises:
InvalidArgumentError: 如果`preprocessed_inputs`的尺寸（高度或宽度）小于33。
ValueError: 如果输入的图像的shape不是4个维度。
"""
if len(preprocessed_inputs.get_shape().as_list()) != 4:
raise ValueError('`preprocessed_inputs` must be 4 dimensional, got a '
'tensor of shape %s' % preprocessed_inputs.get_shape())
with slim.arg_scope(inception_resnet_v2.inception_resnet_v2_arg_scope(
weight_decay=self._weight_decay)):
with slim.arg_scope([slim.batch_norm], is_training=False):
# 通过variable_scope可以实现同一个name_scope中的变量的共享
# name_scope 和 variable_scope区别：
# http://sentiment-mining.blogspot.com/2016/12/tensorflow-name-scope-variable-scope.html
with tf.variable_scope('InceptionResnetV2',
reuse=self._reuse_weights) as scope:
rpn_feature_map, _ = (
inception_resnet_v2.inception_resnet_v2_base(
preprocessed_inputs, final_endpoint='PreAuxLogits',
scope=scope, output_stride=self._first_stage_features_stride,
align_feature_maps=True))
return rpn_feature_map
def _extract_box_classifier_features(self, proposal_feature_maps, scope):
"""提取将用于第二阶段框分类器的特征。
这个方法重建了Inception ResNet v2的“后半部分”网络——
`_extract_proposal_features`中定义的就是那“后半部分”。【个人理解：
相当于原论文中的ROIPooling及其之后的层】
Args:
proposal_feature_maps: 用于裁剪出各个proposal的特征图
A 4-D float tensor with shape
[batch_size * self.max_num_proposals, crop_height, crop_width, depth]
scope: A scope name.
Returns:
proposal_classifier_features: 分了类的proposal
A 4-D float tensor with shape
[batch_size * self.max_num_proposals, height, width, depth]
representing box classifier features for each proposal.
"""
with tf.variable_scope('InceptionResnetV2', reuse=self._reuse_weights):
with slim.arg_scope(inception_resnet_v2.inception_resnet_v2_arg_scope(
weight_decay=self._weight_decay)):
# 强制is_training为False，以此来禁用batch nomal update.
with slim.arg_scope([slim.batch_norm], is_training=False):
with slim.arg_scope([slim.conv2d, slim.max_pool2d, slim.avg_pool2d],
stride=1, padding='SAME'):
with tf.variable_scope('Mixed_7a'):
with tf.variable_scope('Branch_0'):
tower_conv = slim.conv2d(proposal_feature_maps,
256, 1, scope='Conv2d_0a_1x1')
tower_conv_1 = slim.conv2d(
tower_conv, 384, 3, stride=2,
padding='VALID', scope='Conv2d_1a_3x3')
with tf.variable_scope('Branch_1'):
tower_conv1 = slim.conv2d(
proposal_feature_maps, 256, 1, scope='Conv2d_0a_1x1')
tower_conv1_1 = slim.conv2d(
tower_conv1, 288, 3, stride=2,
padding='VALID', scope='Conv2d_1a_3x3')
with tf.variable_scope('Branch_2'):
tower_conv2 = slim.conv2d(
proposal_feature_maps, 256, 1, scope='Conv2d_0a_1x1')
tower_conv2_1 = slim.conv2d(tower_conv2, 288, 3,
scope='Conv2d_0b_3x3')
tower_conv2_2 = slim.conv2d(
tower_conv2_1, 320, 3, stride=2,
padding='VALID', scope='Conv2d_1a_3x3')
with tf.variable_scope('Branch_3'):
tower_pool = slim.max_pool2d(
proposal_feature_maps, 3, stride=2, padding='VALID',
scope='MaxPool_1a_3x3')
net = tf.concat(
[tower_conv_1, tower_conv1_1, tower_conv2_2, tower_pool], 3)
net = slim.repeat(net, 9, inception_resnet_v2.block8, scale=0.20)
net = inception_resnet_v2.block8(net, activation_fn=None)
proposal_classifier_features = slim.conv2d(
net, 1536, 1, scope='Conv2d_7b_1x1')
return proposal_classifier_features
def restore_from_classification_checkpoint_fn(
self,
first_stage_feature_extractor_scope,
second_stage_feature_extractor_scope):
"""返回从固定的checkpoint中读取到的各种变量，一个字典
请注意，这覆盖了faster_rcnn_meta_arch中的FasterRCNNFeatureExtractor
方法,它不适用于InceptionResnetV2的checkpoint。
TODO: revisit whether it's possible to force the
`Repeat` namescope as created in `_extract_box_classifier_features` to
start counting at 2 (e.g. `Repeat_2`) so that the default restore_fn can
be used.
Args:
first_stage_feature_extractor_scope: 第一阶段的特征提取器的scope名字
second_stage_feature_extractor_scope: 第二阶段的特征提取器的scope名字
Returns:
A dict mapping variable names (to load from a checkpoint) to variables in
the model graph.
TODO（黄文坚的书有说明这个字典）
一个映射变量名称（从检查点加载）到模型图中变量的字典。
"""
variables_to_restore = {}
for variable in tf.global_variables():
if variable.op.name.startswith(
first_stage_feature_extractor_scope):
var_name = variable.op.name.replace(
first_stage_feature_extractor_scope + '/', '')
variables_to_restore[var_name] = variable
if variable.op.name.startswith(
second_stage_feature_extractor_scope):
var_name = variable.op.name.replace(
second_stage_feature_extractor_scope
+ '/InceptionResnetV2/Repeat', 'InceptionResnetV2/Repeat_2')
var_name = var_name.replace(
second_stage_feature_extractor_scope + '/', '')
variables_to_restore[var_name] = variable
return variables_to_restore