模型介绍请参看：博文
keras搭建深度学习模型的若干方法：博文

出处:Google AI

迁移学习

首先还是看看标准答案

import  tensorflow as tf
from    tensorflow import keras

base_model = keras.applications.InceptionResNetV2(weights='imagenet')
base_model.summary()

这个summary()是相当的长了。

自编代码

定义卷积单元

def conv2d_cell(x, nbfilter, filtersz, strides=1, pad='same', act=True, name=None):

    x = Conv2D(nbfilter, filtersz, strides, padding=pad, use_bias=False, data_format='channels_last', name=name+'conv2d')(x) #use_bias
    x = BatchNormalization(axis=3, scale=False, name=name+'conv2dbn')(x)    #axis

    if act:
        x = Activation('relu', name=name+'conv2dact')(x)

    return x

Stem模块

def stem_block(x, name=None):

    x = conv2d_cell(x, 32, 3, 2, 'valid', True, name=name+'conv1')
    x = conv2d_cell(x, 32, 3, 1, 'valid', True, name=name+'conv2')
    x = conv2d_cell(x, 64, 3, 1, 'same', True, name=name+'conv3')

    x_11 = MaxPooling2D(3, strides=2, padding='valid', name=name+'_branch11_maxpool')(x)
    x_12 = conv2d_cell(x, 64, 3, 2, 'valid', True, name=name+'_branch12')

    x = Concatenate(axis =3, name=name+'concat_1')([x_11, x_12])

    x_21 = conv2d_cell(x, 64, 1, 1, 'same', True, name=name+'_branch211')
    x_21 = conv2d_cell(x_21, 64, [1,7], 1, 'same', True, name=name+'_branch212')
    x_21 = conv2d_cell(x_21, 64, [7,1], 1, 'same', True, name=name+'_branch213')
    x_21 = conv2d_cell(x_21, 96, 3, 1, 'valid', True, name=name+'_branch214')

    x_22 = conv2d_cell(x, 64, 1, 1, 'same', True, name=name+'_branch221')
    x_22 = conv2d_cell(x_22, 96, 3, 1, 'valid', True, name=name+'_brach222')

    x = Concatenate(axis=3, name=name+'stem_concat_2')([x_21, x_22])

    x_31 = conv2d_cell(x, 192, 3, 2, 'valid', True, name=name+'_brach31')
    x_32 = MaxPooling2D(3, strides=2, padding='valid', name=name+'_branch32_maxpool')(x)
    print(x.shape)

    x = Concatenate(axis=3, name=name+'stem_concat_3')([x_31, x_32])
    
    return x

InceptionRes A模块

def incresA(x, scale, name=None):
    
    branch0 = conv2d_cell(x, 32, 1, 1,'same',True, name=name+'b0')

    branch1 = conv2d_cell(x, 32, 1, 1, 'same', True, name=name+'b1_1')
    branch1 = conv2d_cell(branch1, 32, 3, 1, 'same', True, name=name+'b1_2')

    branch2 = conv2d_cell(x, 32, 1, 1, 'same', True, name=name+'b2_1')
    branch2 = conv2d_cell(branch2, 32, 3, 1, 'same', True, name=name+'b2_2')
    branch2 = conv2d_cell(branch2, 32, 3, 1, 'same', True, name=name+'b2_3')

    branches = [branch0, branch1, branch2]
    mixed = Concatenate(axis=3, name=name+'_concat')(branches)
    filt_exp_1x1 = conv2d_cell(mixed, 384, 1,1, 'same', False, name=name+'filt_exp_1x1')

    final_lay = Lambda(lambda inputs, scale: inputs[0] + inputs[1] * scale,
                      output_shape=backend.int_shape(x)[1:],
                      arguments={'scale': scale},
                      name=name+'act_scaling')([x, filt_exp_1x1])   # Lambda https://blog.csdn.net/VeritasCN/article/details/89884036
    
    return final_lay

InceptionRes B模块

def incresB(x, scale, name=None):

    branch0 = conv2d_cell(x, 192, 1, 1, 'same', True, name= name+'b0')

    branch1 = conv2d_cell(x, 128, 1, 1, 'same', True, name=name+'b1_1')
    branch1 = conv2d_cell(branch1, 160, [1,7], 1, 'same', True, name=name+'b1_2')
    branch1 = conv2d_cell(branch1, 192, [7,1], 1, 'same', True, name=name+'b1_3')

    branches = [branch0, branch1]
    mixed = Concatenate(axis=3, name=name+'_mixed')(branches)
    filt_exp_1x1 = conv2d_cell(mixed, 1152, 1 ,1, 'same', False, name=name+'filt_exp_1x1')
    final_lay = Lambda(lambda inputs, scale:inputs[0]+inputs[1]*scale,
                        output_shape=backend.int_shape(x)[1:],
                        arguments = {'scale':scale},
                        name = name+'act_scaling')([x, filt_exp_1x1])

    return final_lay

InceptionRes C模块

def incresC(x, scale, name=None):

    branch0 = conv2d_cell(x, 192, 1, 1, 'same', True, name= name+'b0')

    branch1 = conv2d_cell(x, 192, 1, 1, 'same', True, name=name+'b1_1')
    branch1 = conv2d_cell(branch1, 224, [1,3], 1, 'same', True, name=name+'b1_2')
    branch1 = conv2d_cell(branch1, 256, [3,1], 1, 'same', True, name=name+'b1_3')

    branches = [branch0, branch1]
    mixed = Concatenate(axis=3, name=name+'_mixed')(branches)
    filt_exp_1x1 = conv2d_cell(mixed, 2048, 1 ,1, 'same', False, name=name+'filt_exp_1x1')
    final_lay = Lambda(lambda inputs, scale:inputs[0]+inputs[1]*scale,
                        output_shape=backend.int_shape(x)[1:],
                        arguments = {'scale':scale},
                        name = name+'act_scaling')([x, filt_exp_1x1])

    return final_lay

Reduction A

def reductionA (x, name=None):

    x_1 = MaxPooling2D(2, strides=2, padding='valid', name=name+'branch1')(x)

    x_2 = conv2d_cell(x, 384, 3, 2, 'valid', True, name=name+'branch2')

    x_3 = conv2d_cell(x, 256, 1, 1, 'same', True, name=name+'branch31')
    x_3 = conv2d_cell(x_3, 256, 3, 1, 'same', True, name=name+'branch32')
    x_3 = conv2d_cell(x_3, 384, 3, 2, 'valid', True, name=name+'branch33')

    x = Concatenate(axis=3, name=name+'concat')([x_1, x_2, x_3])

    return x

Reduction B

def reductionB(x, name=None):

    x_1 = MaxPooling2D(3, strides=2, padding='valid', name=name+'branch1')(x)

    x_2 = conv2d_cell(x, 256, 1, 1, 'same', True, name=name+'branch2')
    x_2 = conv2d_cell(x, 384, 3, 2, 'valid', True, name=name+'branch2')

    x_3 = conv2d_cell(x, 256, 1, 1, 'same', True, name=name+'branch31')
    x_3 = conv2d_cell(x_3, 256, 3, 2, 'valid', True, name=name+'branch32')

    x_4 = conv2d_cell(x, 256, 1, 1, 'same', True, name=name+'branch41')
    x_4 = conv2d_cell(x_4, 256, 3, 1, 'same', True, name=name+'branch42')
    x_4 = conv2d_cell(x_4, 256, 3, 2, 'valid', True, name=name+'branch43')

    x = Concatenate(axis=3, name=name+'concat')([x_1, x_2, x_3, x_4])

    return x

Build Inception-ResNet

备好所有模块，终于可以开始堆料了！

def InceptionResNet_V2(img_shape, num_classes):

    img_input = Input(shape=img_shape)
    x = stem_block(img_input, name='stem')

    # Inception-ResNet-A modules
    x = incresA(x,0.15,name='incresA_1')
    x = incresA(x,0.15,name='incresA_2')
    x = incresA(x,0.15,name='incresA_3')
    x = incresA(x,0.15,name='incresA_4')
    x = incresA(x,0.15,name='incresA_5')

    # Inception-ResNet-ReductionA modules
    x = reductionA(x, name='reductionA')

    # Inception-ResNet-B modules
    x = incresB(x,0.1,name='incresB_1')
    x = incresB(x,0.1,name='incresB_2')
    x = incresB(x,0.1,name='incresB_3')
    x = incresB(x,0.1,name='incresB_4')
    x = incresB(x,0.1,name='incresB_5')
    x = incresB(x,0.1,name='incresB_6')
    x = incresB(x,0.1,name='incresB_7')
    x = incresB(x,0.1,name='incresB_8')
    x = incresB(x,0.1,name='incresB_9')
    x = incresB(x,0.1,name='incresB_10')

    # Inception-ResNet-ReductionB modules
    x = reductionB(x, name='reductionB')

    # Inception-ResNet-C modules
    x = incresC(x,0.2,name='incresC_1')
    x = incresC(x,0.2,name='incresC_2')
    x = incresC(x,0.2,name='incresC_3')
    x = incresC(x,0.2,name='incresC_4')
    x = incresC(x,0.2,name='incresC_5')

    #Top layer
    x = GlobalAveragePooling2D(data_format='channels_last')(x)
    x = Dropout(0.6)(x)
    output = Dense(num_classes, activation='softmax')(x)

    #build Model
    model = models.Model(img_input, output, name='InceptionResNet_V2')

    return model

# test main()
def main():

    model = InceptionResNet_V2((299,299,3), 1000)
    model.summary()

if __name__=='__main__':
    main()
    

终于堆完了，好累！

补充：

与Keras迁移的结果有差别，对比发现Keras中的 InceptionResNetV2主干结构和InceptionV4的一致。
难道是我论文看错了？这都不重要，暂时就尊重文献中内容吧