Tensorflow开源的object detection API中的源码阅读（四）：inception_resnet_v2.py
对于faster_rcnn_meta_arch.py的理解见这篇文章
对于faster_rcnn_inception_resnet_v2_feature_extractor.py的理解见这篇文章
"""定义Inception Resnet V2 架构.
参考论文： http://arxiv.org/abs/1602.07261.
Inception-v4, Inception-ResNet and the Impact of Residual Connections
on Learning
Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, Alex Alemi
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import tensorflow as tf
slim = tf.contrib.slim
def block35(net, scale=1.0, activation_fn=tf.nn.relu, scope=None, reuse=None):
"""构建35x35的resnet 块."""
with tf.variable_scope(scope, 'Block35', [net], reuse=reuse):
with tf.variable_scope('Branch_0'):
tower_conv = slim.conv2d(net, 32, 1, scope='Conv2d_1x1')
with tf.variable_scope('Branch_1'):
tower_conv1_0 = slim.conv2d(net, 32, 1, scope='Conv2d_0a_1x1')
tower_conv1_1 = slim.conv2d(tower_conv1_0, 32, 3, scope='Conv2d_0b_3x3')
with tf.variable_scope('Branch_2'):
tower_conv2_0 = slim.conv2d(net, 32, 1, scope='Conv2d_0a_1x1')
tower_conv2_1 = slim.conv2d(tower_conv2_0, 48, 3, scope='Conv2d_0b_3x3')
tower_conv2_2 = slim.conv2d(tower_conv2_1, 64, 3, scope='Conv2d_0c_3x3')
mixed = tf.concat(axis=3, values=[tower_conv, tower_conv1_1, tower_conv2_2])
up = slim.conv2d(mixed, net.get_shape()[3], 1, normalizer_fn=None,
activation_fn=None, scope='Conv2d_1x1')
net += scale * up
if activation_fn:
net = activation_fn(net)
return net
def block17(net, scale=1.0, activation_fn=tf.nn.relu, scope=None, reuse=None):
"""构建17x17的resnet 块."""
with tf.variable_scope(scope, 'Block17', [net], reuse=reuse):
with tf.variable_scope('Branch_0'):
tower_conv = slim.conv2d(net, 192, 1, scope='Conv2d_1x1')
with tf.variable_scope('Branch_1'):
tower_conv1_0 = slim.conv2d(net, 128, 1, scope='Conv2d_0a_1x1')
tower_conv1_1 = slim.conv2d(tower_conv1_0, 160, [1, 7],
scope='Conv2d_0b_1x7')
tower_conv1_2 = slim.conv2d(tower_conv1_1, 192, [7, 1],
scope='Conv2d_0c_7x1')
mixed = tf.concat(axis=3, values=[tower_conv, tower_conv1_2])
up = slim.conv2d(mixed, net.get_shape()[3], 1, normalizer_fn=None,
activation_fn=None, scope='Conv2d_1x1')
net += scale * up
if activation_fn:
net = activation_fn(net)
return net
def block8(net, scale=1.0, activation_fn=tf.nn.relu, scope=None, reuse=None):
"""构建8x8的resnet块."""
with tf.variable_scope(scope, 'Block8', [net], reuse=reuse):
with tf.variable_scope('Branch_0'):
tower_conv = slim.conv2d(net, 192, 1, scope='Conv2d_1x1')
with tf.variable_scope('Branch_1'):
tower_conv1_0 = slim.conv2d(net, 192, 1, scope='Conv2d_0a_1x1')
tower_conv1_1 = slim.conv2d(tower_conv1_0, 224, [1, 3],
scope='Conv2d_0b_1x3')
tower_conv1_2 = slim.conv2d(tower_conv1_1, 256, [3, 1],
scope='Conv2d_0c_3x1')
mixed = tf.concat(axis=3, values=[tower_conv, tower_conv1_2])
up = slim.conv2d(mixed, net.get_shape()[3], 1, normalizer_fn=None,
activation_fn=None, scope='Conv2d_1x1')
net += scale * up
if activation_fn:
net = activation_fn(net)
return net
def inception_resnet_v2_base(inputs,
final_endpoint='Conv2d_7b_1x1',
output_stride=16,
align_feature_maps=False,
scope=None):
"""Inception model from http://arxiv.org/abs/1602.07261.
【生成inception_resnet_v2网络】
构建Inception Resnet v2网络——从输入的inputs开始直到给定的final_endpoint
此方法可以构建网络直到‘Conv2d_7b_1x1’块。【就是说可以构建部分网络】
Args:
inputs: （四个维度的tensor）[batch_size, height, width, channels].【一般是经过预处理之后的tensor】
final_endpoint: 指定要构建的网络的端点，可以是['Conv2d_1a_3x3', 'Conv2d_2a_3x3', 'Conv2d_2b_3x3',
'MaxPool_3a_3x3', 'Conv2d_3b_1x1', 'Conv2d_4a_3x3', 'MaxPool_5a_3x3',
'Mixed_5b', 'Mixed_6a', 'PreAuxLogits', 'Mixed_7a', 'Conv2d_7b_1x1']其中的一个。
output_stride: A scalar that specifies the requested ratio of input to
output spatial resolution. Only supports 8 and 16.【指定输入与输出空间分辨率的请求比率，是一个标量。 只支持8和16。】
align_feature_maps: 设置为true时，将网络中的所有VALID填充更改为SAME填充，以便特征图对齐。
scope: 可选的variable_scope.
Returns:
tensor_out: 输出对应于final_endpoint的tensor【可以理解为对应于final_endpoint的feature map】
end_points: a set of activations for external use, for example summaries or
losses.
Raises:
ValueError: 如果final_endpoint没有正确设置，或者output_stride 既不是8也不是16, or if the output_stride is 8 and
we request an end point after 'PreAuxLogits'.【就是说output_stride == 8现在只支持到“PreAuxlogits”】
"""
if output_stride != 8 and output_stride != 16:
raise ValueError('output_stride must be 8 or 16.')
padding = 'SAME' if align_feature_maps else 'VALID'
end_points = {}
def add_and_check_final(name, net):
end_points[name] = net
return name == final_endpoint
with tf.variable_scope(scope, 'InceptionResnetV2', [inputs]):
# 强制is_training为False，以此来禁用batch norm update.
# slim.arg_scope:给函数参数自动赋予某些默认值，以使构建模型的代码更加slim
# 使用slim.arg_scope之后就不需要每次都重复设置参数了，比如这里就是复用的net下的
# inception_resnet_v2_arg_scope()方法的相关参数 TODO：其中三个参数的理解
# slim.arg_scope原理分析:http://blog.csdn.net/weixin_35653315/article/details/78160886
with slim.arg_scope([slim.conv2d, slim.max_pool2d, slim.avg_pool2d],
stride=1, padding='SAME'):
# 149 x 149 x 32
net = slim.conv2d(inputs, 32, 3, stride=2, padding=padding,
scope='Conv2d_1a_3x3')
if add_and_check_final('Conv2d_1a_3x3', net): return net, end_points
# 147 x 147 x 32
net = slim.conv2d(net, 32, 3, padding=padding,
scope='Conv2d_2a_3x3')
if add_and_check_final('Conv2d_2a_3x3', net): return net, end_points
# 147 x 147 x 64
net = slim.conv2d(net, 64, 3, scope='Conv2d_2b_3x3')
if add_and_check_final('Conv2d_2b_3x3', net): return net, end_points
# 73 x 73 x 64
net = slim.max_pool2d(net, 3, stride=2, padding=padding,
scope='MaxPool_3a_3x3')
if add_and_check_final('MaxPool_3a_3x3', net): return net, end_points
# 73 x 73 x 80
net = slim.conv2d(net, 80, 1, padding=padding,
scope='Conv2d_3b_1x1')
if add_and_check_final('Conv2d_3b_1x1', net): return net, end_points
# 71 x 71 x 192
net = slim.conv2d(net, 192, 3, padding=padding,
scope='Conv2d_4a_3x3')
if add_and_check_final('Conv2d_4a_3x3', net): return net, end_points
# 35 x 35 x 192
net = slim.max_pool2d(net, 3, stride=2, padding=padding,
scope='MaxPool_5a_3x3')
if add_and_check_final('MaxPool_5a_3x3', net): return net, end_points
# 35 x 35 x 320
with tf.variable_scope('Mixed_5b'):
with tf.variable_scope('Branch_0'):
tower_conv = slim.conv2d(net, 96, 1, scope='Conv2d_1x1')
with tf.variable_scope('Branch_1'):
tower_conv1_0 = slim.conv2d(net, 48, 1, scope='Conv2d_0a_1x1')
tower_conv1_1 = slim.conv2d(tower_conv1_0, 64, 5,
scope='Conv2d_0b_5x5')
with tf.variable_scope('Branch_2'):
tower_conv2_0 = slim.conv2d(net, 64, 1, scope='Conv2d_0a_1x1')
tower_conv2_1 = slim.conv2d(tower_conv2_0, 96, 3,
scope='Conv2d_0b_3x3')
tower_conv2_2 = slim.conv2d(tower_conv2_1, 96, 3,
scope='Conv2d_0c_3x3')
with tf.variable_scope('Branch_3'):
tower_pool = slim.avg_pool2d(net, 3, stride=1, padding='SAME',
scope='AvgPool_0a_3x3')
tower_pool_1 = slim.conv2d(tower_pool, 64, 1,
scope='Conv2d_0b_1x1')
net = tf.concat(
[tower_conv, tower_conv1_1, tower_conv2_2, tower_pool_1], 3)
if add_and_check_final('Mixed_5b', net): return net, end_points
# TODO(alemi): Register intermediate endpoints
net = slim.repeat(net, 10, block35, scale=0.17)
# 17 x 17 x 1088 if output_stride == 8,
# 33 x 33 x 1088 if output_stride == 16
use_atrous = output_stride == 8
with tf.variable_scope('Mixed_6a'):
with tf.variable_scope('Branch_0'):
tower_conv = slim.conv2d(net, 384, 3, stride=1 if use_atrous else 2,
padding=padding,
scope='Conv2d_1a_3x3')
with tf.variable_scope('Branch_1'):
tower_conv1_0 = slim.conv2d(net, 256, 1, scope='Conv2d_0a_1x1')
tower_conv1_1 = slim.conv2d(tower_conv1_0, 256, 3,
scope='Conv2d_0b_3x3')
tower_conv1_2 = slim.conv2d(tower_conv1_1, 384, 3,
stride=1 if use_atrous else 2,
padding=padding,
scope='Conv2d_1a_3x3')
with tf.variable_scope('Branch_2'):
tower_pool = slim.max_pool2d(net, 3, stride=1 if use_atrous else 2,
padding=padding,
scope='MaxPool_1a_3x3')
net = tf.concat([tower_conv, tower_conv1_2, tower_pool], 3)
if add_and_check_final('Mixed_6a', net): return net, end_points
# TODO(alemi): register intermediate endpoints
with slim.arg_scope([slim.conv2d], rate=2 if use_atrous else 1):
net = slim.repeat(net, 20, block17, scale=0.10)
if add_and_check_final('PreAuxLogits', net): return net, end_points
if output_stride == 8:
# TODO(gpapan): Properly support output_stride for the rest of the net.
raise ValueError('output_stride==8 is only supported up to the '
'PreAuxlogits end_point for now.')
# 8 x 8 x 2080
with tf.variable_scope('Mixed_7a'):
with tf.variable_scope('Branch_0'):
tower_conv = slim.conv2d(net, 256, 1, scope='Conv2d_0a_1x1')
tower_conv_1 = slim.conv2d(tower_conv, 384, 3, stride=2,
padding=padding,
scope='Conv2d_1a_3x3')
with tf.variable_scope('Branch_1'):
tower_conv1 = slim.conv2d(net, 256, 1, scope='Conv2d_0a_1x1')
tower_conv1_1 = slim.conv2d(tower_conv1, 288, 3, stride=2,
padding=padding,
scope='Conv2d_1a_3x3')
with tf.variable_scope('Branch_2'):
tower_conv2 = slim.conv2d(net, 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=padding,
scope='Conv2d_1a_3x3')
with tf.variable_scope('Branch_3'):
tower_pool = slim.max_pool2d(net, 3, stride=2,
padding=padding,
scope='MaxPool_1a_3x3')
net = tf.concat(
[tower_conv_1, tower_conv1_1, tower_conv2_2, tower_pool], 3)
if add_and_check_final('Mixed_7a', net): return net, end_points
# TODO(alemi): register intermediate endpoints
net = slim.repeat(net, 9, block8, scale=0.20)
net = block8(net, activation_fn=None)
# 8 x 8 x 1536
net = slim.conv2d(net, 1536, 1, scope='Conv2d_7b_1x1')
if add_and_check_final('Conv2d_7b_1x1', net): return net, end_points
raise ValueError('final_endpoint (%s) not recognized', final_endpoint)
def inception_resnet_v2(inputs, num_classes=1001, is_training=True,
dropout_keep_prob=0.8,
reuse=None,
scope='InceptionResnetV2',
create_aux_logits=True):
"""创建Inception Resnet V2 模型.
Args:
inputs: a 4-D tensor of size [batch_size, height, width, 3].
num_classes: number of predicted classes.
is_training: whether is training or not.
dropout_keep_prob: float, the fraction to keep before final layer.
reuse: whether or not the network and its variables should be reused. To be
able to reuse 'scope' must be given.
scope: Optional variable_scope.
create_aux_logits: Whether to include the auxilliary logits.
Returns:
logits: the logits outputs of the model.
end_points: the set of end_points from the inception model.
"""
end_points = {}
with tf.variable_scope(scope, 'InceptionResnetV2', [inputs, num_classes],
reuse=reuse) as scope:
with slim.arg_scope([slim.batch_norm, slim.dropout],
is_training=is_training):
net, end_points = inception_resnet_v2_base(inputs, scope=scope)
if create_aux_logits:
with tf.variable_scope('AuxLogits'):
aux = end_points['PreAuxLogits']
aux = slim.avg_pool2d(aux, 5, stride=3, padding='VALID',
scope='Conv2d_1a_3x3')
aux = slim.conv2d(aux, 128, 1, scope='Conv2d_1b_1x1')
aux = slim.conv2d(aux, 768, aux.get_shape()[1:3],
padding='VALID', scope='Conv2d_2a_5x5')
aux = slim.flatten(aux)
aux = slim.fully_connected(aux, num_classes, activation_fn=None,
scope='Logits')
end_points['AuxLogits'] = aux
with tf.variable_scope('Logits'):
net = slim.avg_pool2d(net, net.get_shape()[1:3], padding='VALID',
scope='AvgPool_1a_8x8')
net = slim.flatten(net)
net = slim.dropout(net, dropout_keep_prob, is_training=is_training,
scope='Dropout')
end_points['PreLogitsFlatten'] = net
logits = slim.fully_connected(net, num_classes, activation_fn=None,
scope='Logits')
end_points['Logits'] = logits
end_points['Predictions'] = tf.nn.softmax(logits, name='Predictions')
return logits, end_points
inception_resnet_v2.default_image_size = 299
def inception_resnet_v2_arg_scope(weight_decay=0.00004,
batch_norm_decay=0.9997,
batch_norm_epsilon=0.001):
"""返回由该方法生成的inception_resnet_v2网络中常常用到的参数的默认值。
Args:
weight_decay: 特征提取器的权重衰减，L2正则化时会用到。
batch_norm_decay: decay for the moving average of batch_norm momentums.【就是梯度下降时的“惯性”衰减值】
batch_norm_epsilon: small float added to variance,避免被0除。
Returns:
带有inception_resnet_v2需要的参数的arg_scope.
"""
# 给conv2d 和 fully_connected layers设置权值的衰减速率。
# weights_regularizer的值默认设为slim.l2_regularizer(weight_decay)，biases_regularizer同理。
with slim.arg_scope([slim.conv2d, slim.fully_connected],
weights_regularizer=slim.l2_regularizer(weight_decay),
biases_regularizer=slim.l2_regularizer(weight_decay)):
batch_norm_params = {
'decay': batch_norm_decay,
'epsilon': batch_norm_epsilon,
}
# 设置batch_norm的激活函数和参数
with slim.arg_scope([slim.conv2d], activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params) as scope:
return scope