python神经网络Inception ResnetV2模型复现详解

什么是Inception ResnetV2

Inception ResnetV2是Inception ResnetV1的一个加强版，两者的结构差距不大，如果大家想了解Inception ResnetV1可以看一下我的另一个blog。facenet的神经网络结构就是Inception ResnetV1。

神经网络学习——facenet详解及其keras实现

源码下载

Inception-ResNetV2的网络结构

Inception-ResNetV2和Inception-ResNetV1采用同一个主干网络。

它的结构很有意思！

如图所示为整个网络的主干结构：

可以看到里面的结构分为几个重要的部分

1、stem

2、Inception-resnet-A

3、Inception-resnet-B

4、Inception-resnet-C

1、Stem的结构：

在Inception-ResNetV2里，它的Input为299x299x3大小，输入后进行：三次卷积 -> 最大池化 -> 两次卷积 -> 最大池化 -> 四个分支 -> 堆叠python实现代码如下：

input_shape = [299,299,3]
img_input = Input(shape=input_shape)
# Stem block: 299,299,3 -> 35 x 35 x 192
x = conv2d_bn(img_input, 32, 3, strides=2, padding='valid')
x = conv2d_bn(x, 32, 3, padding='valid')
x = conv2d_bn(x, 64, 3)
x = MaxPooling2D(3, strides=2)(x)
x = conv2d_bn(x, 80, 1, padding='valid')
x = conv2d_bn(x, 192, 3, padding='valid')
x = MaxPooling2D(3, strides=2)(x)
# Mixed 5b (Inception-A block):35 x 35 x 192 -> 35 x 35 x 320
branch_0 = conv2d_bn(x, 96, 1)
branch_1 = conv2d_bn(x, 48, 1)
branch_1 = conv2d_bn(branch_1, 64, 5)
branch_2 = conv2d_bn(x, 64, 1)
branch_2 = conv2d_bn(branch_2, 96, 3)
branch_2 = conv2d_bn(branch_2, 96, 3)
branch_pool = AveragePooling2D(3, strides=1, padding='same')(x)
branch_pool = conv2d_bn(branch_pool, 64, 1)
branches = [branch_0, branch_1, branch_2, branch_pool]
x = Concatenate(name='mixed_5b')(branches)

2、Inception-resnet-A的结构：

Inception-resnet-A的结构分为四个分支

1、未经处理直接输出

2、经过一次1x1的32通道的卷积处理

3、经过一次1x1的32通道的卷积处理和一次3x3的32通道的卷积处理

4、经过一次1x1的32通道的卷积处理、一次3x3的48通道和一次3x3的64通道卷积处理

234步的结果堆叠后进行一次卷积，并与第一步的结果相加，实质上这是一个残差网络结构。

实现代码如下：

branch_0 = conv2d_bn(x, 32, 1)
branch_1 = conv2d_bn(x, 32, 1)
branch_1 = conv2d_bn(branch_1, 32, 3)
branch_2 = conv2d_bn(x, 32, 1)
branch_2 = conv2d_bn(branch_2, 48, 3)
branch_2 = conv2d_bn(branch_2, 64, 3)
branches = [branch_0, branch_1, branch_2]
block_name = block_type + '_' + str(block_idx)
mixed = Concatenate(name=block_name + '_mixed')(branches)
up = conv2d_bn(mixed,K.int_shape(x)[3],1,activation=None,se_bias=True,name=block_name + '_conv')
x = Lambda(lambda inputs, scale: inputs[0] + inputs[1] * scale,
           output_shape=K.int_shape(x)[1:],
           arguments={'scale': scale},
           name=block_name)([x, up])
if activation is not None:
   x = Activation(activation, name=block_name + '_ac')(x)

3、Inception-resnet-B的结构：

Inception-resnet-B的结构分为四个分支

1、未经处理直接输出

2、经过一次1x1的192通道的卷积处理

3、经过一次1x1的128通道的卷积处理、一次1x7的160通道的卷积处理和一次7x1的192通道的卷积处理

23步的结果堆叠后进行一次卷积，并与第一步的结果相加，实质上这是一个残差网络结构。

实现代码如下：

branch_0 = conv2d_bn(x, 192, 1)
branch_1 = conv2d_bn(x, 128, 1)
branch_1 = conv2d_bn(branch_1, 160, [1, 7])
branch_1 = conv2d_bn(branch_1, 192, [7, 1])
branches = [branch_0, branch_1]
block_name = block_type + '_' + str(block_idx)
mixed = Concatenate(name=block_name + '_mixed')(branches)
up = conv2d_bn(mixed,K.int_shape(x)[3],1,activation=None,se_bias=True,name=block_name + '_conv')
x = Lambda(lambda inputs, scale: inputs[0] + inputs[1] * scale,
           output_shape=K.int_shape(x)[1:],
           arguments={'scale': scale},
           name=block_name)([x, up])
if activation is not None:
   x = Activation(activation, name=block_name + '_ac')(x)

4、Inception-resnet-C的结构：

Inception-resnet-B的结构分为四个分支

1、未经处理直接输出

2、经过一次1x1的192通道的卷积处理

3、经过一次1x1的192通道的卷积处理、一次1x3的224通道的卷积处理和一次3x1的256通道的卷积处理

23步的结果堆叠后进行一次卷积，并与第一步的结果相加，实质上这是一个残差网络结构。

实现代码如下：

branch_0 = conv2d_bn(x, 192, 1)
branch_1 = conv2d_bn(x, 192, 1)
branch_1 = conv2d_bn(branch_1, 224, [1, 3])
branch_1 = conv2d_bn(branch_1, 256, [3, 1])
branches = [branch_0, branch_1]
block_name = block_type + '_' + str(block_idx)
mixed = Concatenate(name=block_name + '_mixed')(branches)
up = conv2d_bn(mixed,K.int_shape(x)[3],1,activation=None,se_bias=True,name=block_name + '_conv')
x = Lambda(lambda inputs, scale: inputs[0] + inputs[1] * scale,
           output_shape=K.int_shape(x)[1:],
           arguments={'scale': scale},
           name=block_name)([x, up])
if activation is not None:
   x = Activation(activation, name=block_name + '_ac')(x)

全部代码

import warnings
import numpy as np
from keras.preprocessing import image
from keras.models import Model
from keras.layers import Activation,AveragePooling2D,BatchNormalization,Concatenate
from keras.layers import Conv2D,Dense,GlobalAveragePooling2D,GlobalMaxPooling2D,Input,Lambda,MaxPooling2D
from keras.applications.imagenet_utils import decode_predictions
from keras.utils.data_utils import get_file
from keras import backend as K
BASE_WEIGHT_URL = 'https://github.com/fchollet/deep-learning-models/releases/download/v0.7/'
def conv2d_bn(x,filters,kernel_size,strides=1,padding='same',activation='relu',use_bias=False,name=None):
   x = Conv2D(filters,
              kernel_size,
              strides=strides,
              padding=padding,
              use_bias=use_bias,
              name=name)(x)
   if not use_bias:
       bn_axis = 1 if K.image_data_format() == 'channels_first' else 3
       bn_name = None if name is None else name + '_bn'
       x = BatchNormalization(axis=bn_axis, scale=False, name=bn_name)(x)
   if activation is not None:
       ac_name = None if name is None else name + '_ac'
       x = Activation(activation, name=ac_name)(x)
   return x
def inception_resnet_block(x, scale, block_type, block_idx, activation='relu'):
   if block_type == 'block35':
       branch_0 = conv2d_bn(x, 32, 1)
       branch_1 = conv2d_bn(x, 32, 1)
       branch_1 = conv2d_bn(branch_1, 32, 3)
       branch_2 = conv2d_bn(x, 32, 1)
       branch_2 = conv2d_bn(branch_2, 48, 3)
       branch_2 = conv2d_bn(branch_2, 64, 3)
       branches = [branch_0, branch_1, branch_2]
   elif block_type == 'block17':
       branch_0 = conv2d_bn(x, 192, 1)
       branch_1 = conv2d_bn(x, 128, 1)
       branch_1 = conv2d_bn(branch_1, 160, [1, 7])
       branch_1 = conv2d_bn(branch_1, 192, [7, 1])
       branches = [branch_0, branch_1]
   elif block_type == 'block8':
       branch_0 = conv2d_bn(x, 192, 1)
       branch_1 = conv2d_bn(x, 192, 1)
       branch_1 = conv2d_bn(branch_1, 224, [1, 3])
       branch_1 = conv2d_bn(branch_1, 256, [3, 1])
       branches = [branch_0, branch_1]
   else:
       raise ValueError('Unknown Inception-ResNet block type. '
                        'Expects "block35", "block17" or "block8", '
                        'but got: ' + str(block_type))
   block_name = block_type + '_' + str(block_idx)
   mixed = Concatenate(name=block_name + '_mixed')(branches)
   up = conv2d_bn(mixed,K.int_shape(x)[3],1,activation=None,use_bias=True,name=block_name + '_conv')
   x = Lambda(lambda inputs, scale: inputs[0] + inputs[1] * scale,
              output_shape=K.int_shape(x)[1:],
              arguments={'scale': scale},
              name=block_name)([x, up])
   if activation is not None:
       x = Activation(activation, name=block_name + '_ac')(x)
   return x
def InceptionResNetV2(input_shape=[299,299,3],
                     classes=1000):
   input_shape = [299,299,3]
   img_input = Input(shape=input_shape)
   # Stem block: 299,299,3 -> 35 x 35 x 192
   x = conv2d_bn(img_input, 32, 3, strides=2, padding='valid')
   x = conv2d_bn(x, 32, 3, padding='valid')
   x = conv2d_bn(x, 64, 3)
   x = MaxPooling2D(3, strides=2)(x)
   x = conv2d_bn(x, 80, 1, padding='valid')
   x = conv2d_bn(x, 192, 3, padding='valid')
   x = MaxPooling2D(3, strides=2)(x)
   # Mixed 5b (Inception-A block):35 x 35 x 192 -> 35 x 35 x 320
   branch_0 = conv2d_bn(x, 96, 1)
   branch_1 = conv2d_bn(x, 48, 1)
   branch_1 = conv2d_bn(branch_1, 64, 5)
   branch_2 = conv2d_bn(x, 64, 1)
   branch_2 = conv2d_bn(branch_2, 96, 3)
   branch_2 = conv2d_bn(branch_2, 96, 3)
   branch_pool = AveragePooling2D(3, strides=1, padding='same')(x)
   branch_pool = conv2d_bn(branch_pool, 64, 1)
   branches = [branch_0, branch_1, branch_2, branch_pool]
   x = Concatenate(name='mixed_5b')(branches)
   # 10次Inception-ResNet-A block:35 x 35 x 320 -> 35 x 35 x 320
   for block_idx in range(1, 11):
       x = inception_resnet_block(x,
                                  scale=0.17,
                                  block_type='block35',
                                  block_idx=block_idx)
   # Reduction-A block:35 x 35 x 320 -> 17 x 17 x 1088
   branch_0 = conv2d_bn(x, 384, 3, strides=2, padding='valid')
   branch_1 = conv2d_bn(x, 256, 1)
   branch_1 = conv2d_bn(branch_1, 256, 3)
   branch_1 = conv2d_bn(branch_1, 384, 3, strides=2, padding='valid')
   branch_pool = MaxPooling2D(3, strides=2, padding='valid')(x)
   branches = [branch_0, branch_1, branch_pool]
   x = Concatenate(name='mixed_6a')(branches)
   # 20次Inception-ResNet-B block: 17 x 17 x 1088 -> 17 x 17 x 1088
   for block_idx in range(1, 21):
       x = inception_resnet_block(x,
                                  scale=0.1,
                                  block_type='block17',
                                  block_idx=block_idx)
   # Reduction-B block: 17 x 17 x 1088 -> 8 x 8 x 2080
   branch_0 = conv2d_bn(x, 256, 1)
   branch_0 = conv2d_bn(branch_0, 384, 3, strides=2, padding='valid')
   branch_1 = conv2d_bn(x, 256, 1)
   branch_1 = conv2d_bn(branch_1, 288, 3, strides=2, padding='valid')
   branch_2 = conv2d_bn(x, 256, 1)
   branch_2 = conv2d_bn(branch_2, 288, 3)
   branch_2 = conv2d_bn(branch_2, 320, 3, strides=2, padding='valid')
   branch_pool = MaxPooling2D(3, strides=2, padding='valid')(x)
   branches = [branch_0, branch_1, branch_2, branch_pool]
   x = Concatenate(name='mixed_7a')(branches)
   # 10次Inception-ResNet-C block: 8 x 8 x 2080 -> 8 x 8 x 2080
   for block_idx in range(1, 10):
       x = inception_resnet_block(x,
                                  scale=0.2,
                                  block_type='block8',
                                  block_idx=block_idx)
   x = inception_resnet_block(x,
                              scale=1.,
                              activation=None,
                              block_type='block8',
                              block_idx=10)
   # 8 x 8 x 2080 -> 8 x 8 x 1536
   x = conv2d_bn(x, 1536, 1, name='conv_7b')
   x = GlobalAveragePooling2D(name='avg_pool')(x)
   x = Dense(classes, activation='softmax', name='predictions')(x)
   inputs = img_input
   # 创建模型
   model = Model(inputs, x, name='inception_resnet_v2')
   return model
def preprocess_input(x):
   x /= 255.
   x -= 0.5
   x *= 2.
   return x
if __name__ == '__main__':
   model = InceptionResNetV2()
   fname = 'inception_resnet_v2_weights_tf_dim_ordering_tf_kernels.h5'
   weights_path = get_file(fname,BASE_WEIGHT_URL + fname,cache_subdir='models',file_hash='e693bd0210a403b3192acc6073ad2e96')
   model.load_weights(fname)
   img_path = 'elephant.jpg'
   img = image.load_img(img_path, target_size=(299, 299))
   x = image.img_to_array(img)
   x = np.expand_dims(x, axis=0)
   x = preprocess_input(x)
   preds = model.predict(x)
   print('Predicted:', decode_predictions(preds))

来源：https://blog.csdn.net/weixin_44791964/article/details/103732720