tensorflow系列之_pspnet网络
目标检测领域的实例分割网络 PSPnet,基于像素级的分类，提出了一种金字塔池化模块，金字塔层级的数量和每一层的大小都可以进行调整
主题结构是一个提去特征的基础网络（resnet/mobilenet）+金字塔池化。
项目中为了减小网络参数大小和部署方便，backbone采用的mobilenetv1。
mobilenet_v1.py代码(mobilenet_v1输出 f1,f2,f3,f4,f5不同尺度的特征)
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# @time: 2020/6/22 19:27
"""
标准的mobilenet_v1
只能接受输入224*224大小的图片
输出 f1,f2,f3,f4,f5不同尺度的特征图
"""
import tensorflow as tf
import tensorflow.contrib.slim as slim
def mobilenet(inputs,
num_classes=100,
is_training=True,
width_multiplier=1,
scope='mobilenet'):
"""
:param inputs: [batch_size,h,w,c]
:param num_classes:
:param is_training:
:param width_multiplier:
:param scope:
:return: logits: the pre_softmax activations
end_points: 包含了特征图的dict
"""
def _depthwise_separable_conv(inputs,
num_pwc_filters,
width_multiplier,
sc,
downsample=False):
"""
depth-wise separable 卷积层
:param inputs:
:param num_pwc_filters: 点卷积和的个数
:param width_multiplier:
:param sc:命名空间
:param dowmnsample:
:return:
"""
num_pwc_filters = round(num_pwc_filters * width_multiplier)
_stride = 2 if downsample else 1 # 如果下采样就为2，否则为1
# skip pointwise by setting num_outputs=None ，
# pointwise 卷积的卷积核个数，如果为空，将跳过pointwise卷积的步骤.
# 深度可分离卷积
depthwise_conv = slim.separable_convolution2d(inputs,
num_outputs=None,
stride=_stride,
depth_multiplier=1,
kernel_size=[3, 3],
scope=sc + '/depthwise_conv')
bn = slim.batch_norm(depthwise_conv, scope=sc + '/dw_batch_norm')
#点卷积 就是1*1的普通卷积
pointwise_conv = slim.conv2d(bn,
num_pwc_filters,
kernel_size=[1, 1],
scope=sc + '/pointwise_conv')
bn = slim.batch_norm(pointwise_conv, scope=sc + '/pw_batch_norm')
return bn
with tf.variable_scope(scope) as sc:
end_points_collection=sc.name+'_end_points'
with slim.arg_scope([slim.conv2d,slim.separable_convolution2d],activation_fn=None,outputs_collections=[end_points_collection]):
with slim.arg_scope([slim.batch_norm],is_training=is_training,activation_fn=tf.nn.relu,
fused=True):
net=slim.conv2d(inputs,round(32*width_multiplier),[3,3],stride=2,padding=
'SAME',scope='conv_1') # [batch_size,h/2,w/2,32]
net=slim.batch_norm(net,scope='conv_1/batch_norm')
net=_depthwise_separable_conv(net,64,width_multiplier,sc='conv_ds_2') # [batch_size,h/2,w/2,64]
net=_depthwise_separable_conv(net,128,width_multiplier,downsample=True,sc='conv_ds_3')# 下采样 [batch_size,h/4,w/4,128]
net=_depthwise_separable_conv(net,128,width_multiplier,sc='conv_ds_4') #[batch_size,h/4,w/4,128]
net = _depthwise_separable_conv(net, 256, width_multiplier, downsample=True, sc='conv_ds_5') # 下采样 [batch_size,h/8,w/8,256]
net=_depthwise_separable_conv(net,256,width_multiplier,sc='conv_ds_6') #[batch_size,h/8,w/8,256]
net = _depthwise_separable_conv(net, 512, width_multiplier, downsample=True, sc='conv_ds_7') # 下采样 # [batch_size,h/16,w/16,256]
net = _depthwise_separable_conv(net, 512, width_multiplier, sc='conv_ds_8') # [batch_size,h/16,w/16,512]
net = _depthwise_separable_conv(net, 512, width_multiplier, sc='conv_ds_9') # [batch_size,h/16,w/16,512]
net = _depthwise_separable_conv(net, 512, width_multiplier, sc='conv_ds_10') # [batch_size,h/16,w/16,512]
net = _depthwise_separable_conv(net, 512, width_multiplier, sc='conv_ds_11')# [batch_size,h/16,w/16,512]
net = _depthwise_separable_conv(net, 512, width_multiplier, sc='conv_ds_12')# [batch_size,h/16,w/16,512]
net = _depthwise_separable_conv(net, 1024, width_multiplier, downsample=True, sc='conv_ds_13')# [batch_size,h/32,w/32,1024]
net = _depthwise_separable_conv(net, 1024, width_multiplier, sc='conv_ds_14')# [batch_size,h/32,w/32,1024]
net=slim.avg_pool2d(net,[7,7],scope='avg_pool_15') # [batch_size,h/32*7,w/32*7,1024]
#
end_points=slim.utils.convert_collection_to_dict(end_points_collection)
# net=tf.squeeze(net,[1,2],name='SpatialSqueeze')
# end_points['squeeze']=net
# logits=slim.fully_connected(net,num_classes,activation_fn=None,scope='fc_16')
# predictions=slim.softmax(logits,scope='Predictions')
#
# end_points['Logits']=logits
# end_points['Predictions']=predictions
f1=end_points['mobilenet/conv_ds_2/pointwise_conv'] #[None,112,112,64]
f2=end_points['mobilenet/conv_ds_4/pointwise_conv']#[None,56,56,128]
f3=end_points['mobilenet/conv_ds_6/pointwise_conv'] #[None,28,28,256]
f4=end_points['mobilenet/conv_ds_12/pointwise_conv'] #[None.14,14,512]
f5=end_points['mobilenet/conv_ds_14/pointwise_conv']
return inputs, [f1 , f2 , f3 , f4 , f5 ]
mobilenet.default_image_size = 224
def mobilenet_arg_scope(weight_decay=0.0):
"""
定义默认的scope参数
:param weight_decay: 正则的衰减率
:return:
"""
with slim.arg_scope(
[slim.convolution2d, slim.separable_convolution2d],
weights_initializer=slim.initializers.xavier_initializer(),
biases_initializer=slim.init_ops.zeros_initializer(),
weights_regularizer=slim.l2_regularizer(weight_decay)) as sc:
return sc
if __name__ == '__main__':
batch_size = 5
height, width = 224, 224
num_classes = 100
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
inputs = tf.random_uniform((batch_size, height, width, 3))
inputs_2, [f1 , f2 , f3 , f4 , f5 ] = mobilenet(inputs, num_classes)
打印出f1,f2,f3,f4,f5的shape
f1: Tensor("mobilenet/conv_ds_2/pointwise_conv/BiasAdd:0", shape=(5, 112, 112, 64), dtype=float32)
f2: Tensor("mobilenet/conv_ds_4/pointwise_conv/BiasAdd:0", shape=(5, 56, 56, 128), dtype=float32)
f3: Tensor("mobilenet/conv_ds_6/pointwise_conv/BiasAdd:0", shape=(5, 28, 28, 256), dtype=float32)
f4: Tensor("mobilenet/conv_ds_12/pointwise_conv/BiasAdd:0", shape=(5, 14, 14, 512), dtype=float32)
f5: Tensor("mobilenet/conv_ds_14/pointwise_conv/BiasAdd:0", shape=(5, 7, 7, 1024), dtype=float32)
PSPnet部分的代码
pspnet_model.py
import tensorflow as tf
from tensorflow.contrib import slim
from nets.mobilenet_v1 import mobilenet_v1 as get_mobilenet_encoder
import numpy as np
#
# MERGE_AXIS = -1 # 通道维度上链接
def resize_image(input_images,s):
"""
:param input_images: 4D tensor
:param s: 缩放比例 （2，2） # 扩大的倍数
:return:
"""
h,w=input_images.get_shape().as_list()[1],input_images.get_shape().as_list()[2]
h_ratio=s[0]
w_ration=s[1]
h=int(h*h_ratio)
w=int(w*w_ration)
images=tf.image.resize_images(input_images,size=(h,w))
return images
#进场不同尺度池化的操作
def pool_block( x , pool_factor ,IMAGE_ORDERING = 'NHWC'):
"""
:param x: 4D （1，224，224，3）
:param pool_factor: 2 池化的尺度
:return:
"""
if IMAGE_ORDERING == 'NHWC':
h, w = x.get_shape().as_list()[1], x.get_shape().as_list()[2]
elif IMAGE_ORDERING == 'NCHW':
h, w = x.get_shape().as_list()[2], x.get_shape().as_list()[3]
# strides = [18,18],[9,9],[6,6],[3,3]
pool_size = [int(np.round( float(h) / pool_factor)), int(np.round( float(w )/ pool_factor))]
strides=pool_size
# 进行不同程度的平均
x=slim.avg_pool2d(x,kernel_size=pool_size,stride=strides, padding='SAME') # [1,2,2,3]
# x = AveragePooling2D(pool_size , data_format=IMAGE_ORDERING , strides=strides, padding='same')( feats )
#
# 进行卷积
x=slim.conv2d(x,512,kernel_size=(1,1),stride=1,padding='SAME') #[1,2,2,512]
# x = Conv2D(512, (1 ,1 ), data_format=IMAGE_ORDERING , padding='same' , use_bias=False )( x )
x=slim.batch_norm(x)
# x = BatchNormalization()(x)
# x=slim.relu(x)
x=tf.nn.relu(x)
# x = Activation('relu' )(x)
x=resize_image( x , strides)
# x = resize_image( x , strides , data_format=IMAGE_ORDERING )
return x
def pspnet(input_images,num_classes,input_height=576, input_width=576):
assert input_height%192 == 0
assert input_width%192 == 0
img_input , levels = get_mobilenet_encoder(inputs=input_images,
num_classes=num_classes
)
[f1 , f2 , f3 , f4 , f5 ] = levels
o = f5
# 对f5进行不同程度的池化
pool_factors = [ 1,2,3,6]
pool_outs = [o ]
for p in pool_factors:
pooled = pool_block( o , p )
pool_outs.append( pooled )
# 连接
o=tf.concat(pool_outs,axis=-1)
# o = Concatenate( axis=MERGE_AXIS)(pool_outs )
o=slim.conv2d(o,512,kernel_size=(1,1),stride=1,padding='SAME')
# o = ( Conv2D(512, (1, 1), padding='valid', data_format=IMAGE_ORDERING))(o)
o=slim.batch_norm(o)
# o = ( BatchNormalization())(o)
o=resize_image(o,(2,2))
# o = resize_image(o,(2,2),data_format=IMAGE_ORDERING)
o=tf.concat([o,f4],axis=-1)
# o = Concatenate( axis=MERGE_AXIS)([o,f4])
o = slim.conv2d(o,512, kernel_size=(1, 1), stride=1, padding='SAME')
# o = ( Conv2D(512, (1, 1), padding='valid', data_format=IMAGE_ORDERING))(o)
o = slim.batch_norm(o)
# o = ( BatchNormalization())(o)
# o = Activation('relu' )(o)
o=tf.nn.relu(o)
pool_outs = [o]
# 对f4进行不同程度的池化
for p in pool_factors:
pooled = pool_block( o , p )
pool_outs.append( pooled )
# 连接
o = tf.concat(pool_outs, axis=-1)
# o = Concatenate( axis=MERGE_AXIS)(pool_outs )
o = slim.conv2d(o, 512, kernel_size=(1, 1), stride=1, padding='SAME')
# o = ( Conv2D(512, (1, 1), padding='valid', data_format=IMAGE_ORDERING))(o)
# o = ( BatchNormalization())(o)
o=slim.batch_norm(o)
o = resize_image(o,(2,2))
o = tf.concat([o, f3], axis=-1)
# o = Concatenate( axis=MERGE_AXIS)([o,f3])
o = slim.conv2d(o, 512, kernel_size=(1, 1), stride=1, padding='SAME')
# o = ( Conv2D(512, (1, 1), padding='valid', data_format=IMAGE_ORDERING))(o)
o = slim.batch_norm(o)
# o = ( BatchNormalization())(o)
# o = Activation('relu' )(o)
o = tf.nn.relu(o)
pool_outs = [o ]
# 对f3进行不同程度的池化
for p in pool_factors:
pooled = pool_block( o , p )
pool_outs.append( pooled )
# 连接
o = tf.concat(pool_outs, axis=-1)
# o = Concatenate( axis=MERGE_AXIS)(pool_outs )
# 卷积
o = slim.conv2d(o, 512, kernel_size=(1, 1), stride=1, padding='SAME')
# o = Conv2D(512, (1,1), data_format=IMAGE_ORDERING, use_bias=False )(o)
o = slim.batch_norm(o)
# o = ( BatchNormalization())(o)
# o = Activation('relu' )(o)
o = tf.nn.relu(o)
# 此时输出为[144,144,nclasses]
o=slim.conv2d(o,num_classes,kernel_size=(3,3),stride=1,padding='SAME')
# o = Conv2D( n_classes,(3,3),data_format=IMAGE_ORDERING, padding='same' )(o)
o = resize_image(o,(2,2))
# print('o shape',o.get_shape())
o=tf.reshape(o,(-1,(input_height*input_width)//16,num_classes))
# o=slim.softmax(o)
# o = Reshape((-1,n_classes))(o)
# o = Softmax()(o)
# model = Model(img_input,o)
out_put=o
return out_put
if __name__ == '__main__':
input_images=tf.ones((1,576,576,3),dtype=tf.float32)
# images=resize_image(input_images,s=(2,2))
out_put=pspnet(input_images,num_classes=100,input_height=576, input_width=576)