将Vgg16 FC层替换为UNet

Question

我想删除VGG16的FC层，并添加UNet层。我不知道如何对VGG16进行微调。

Answer 1

Unet用于图像翻译，而VGG16用于图像分类。您可以将vgg16 fc层替换为unet层。

无论如何，这就是移除vgg16的fc层并微调模型的方法，

使用include_top=False定义VGG16模型以删除fc层

vgg16 = keras.applications.vgg16.VGG16(input_shape=(224,224,3), include_top=False, weights='imagenet')

然后，使用vgg16.trainable = False冻结图层。

现在，定义您自己的模型并使用vgg16作为特征提取器

例如：

model = Sequential()
model.add(vgg16)
model.add(Flatten())
model.add(Dense(32, activation='relu')  
model.add(Dense(8, activation='softmax')

现在，model.summary()给出的模型的总结是，

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
vgg16 (Functional)           (None, 7, 7, 512)         14714688  
_________________________________________________________________
flatten_2 (Flatten)          (None, 25088)             0         
_________________________________________________________________
dense_4 (Dense)              (None, 32)                802848    
_________________________________________________________________
dense_5 (Dense)              (None, 8)                 264       
=================================================================
Total params: 15,517,800
Trainable params: 803,112
Non-trainable params: 14,714,688
_________________________________________________________________

现在，您可以编译并拟合您的数据来微调此模型。

Answer 2

非常感谢你的详细回答。我试图打印模型摘要，但它告诉我要先构建模型。然后，我尝试编译模型并将模型输入定义为vgg16.input，将输出定义为vgg16。作为输出。在模型摘要中，它被提到为model_1而不是序列模型，并且没有可训练的参数。我无法发布代码，因为它在发布中给我的错误可能是因为错误的缩进。

我将这个模型拟合到我的3D数据中，它提取了特征。但是当我试图将它保存在数据帧中时，它给出了错误(ValueError:必须传递2-d输入。shape=(1488，7，7，8)

单个图像的结果特征如下:完成1032 1/1 ============================== - 1s 540ms/step [[6.38653398e-01 2.02880692e-05 5.99825435e-05 1.21887595e-07 3.60443801e-01 1.75449770e-08 6.53681054e-04 1.68692059e-04 9.95418668e-01 3.13399983e-06 2.87542747e-07 4.47111670e-03 4.79741175e-07 3.33578015e-07 9.68128370e-05 9.19987178e-06 9.35400775e-05 1.01485979e-10 1.108602929e-06-9。62217450e-01 7.07163755e-03 5.90099735e-05 3.56038637e-03 2.69968603e-02 9.99982238e-01 4.59793046e-06 8.85845737e-12 5.71115200e-09 1.31434072e-05 2.44201426e-12 4.12060874e-09 4.05980333e-10]

[[9.95615244e-01 3.53653104e-06 4.18960204e-04 9.37127770e-05
3.63768822e-06 2.30651058e-05 3.81550402e-03 2.63186284e-05]
[2.19437666e-02 9.60381996e-10 5.99338227e-06 2.13321403e-01
1.63183977e-09 1.63899801e-06 7.61410892e-01 3.31628183e-03]
[1.05815129e-02 5.36842706e-13 4.65415155e-07 9.74010408e-01
1.67039627e-06 9.23667098e-09 9.31831356e-03 6.08768407e-03]
[6.21392846e-01 2.81578455e-10 1.34801428e-06 3.78108501e-01
1.17762456e-05 1.14337240e-07 1.28620741e-07 4.85272234e-04]
[9.99988198e-01 2.88270140e-13 1.06488929e-09 5.52908568e-06
5.82322718e-06 4.06775143e-07 7.55641372e-13 8.57361127e-11]
[9.99999881e-01 1.54605725e-13 1.32699366e-13 2.19701427e-12
3.24160609e-08 1.47507421e-07 8.68404520e-15 1.41449074e-18]
[9.99988317e-01 7.39743257e-07 2.08591597e-08 3.57890855e-07
1.01329324e-05 3.77399312e-09 1.97229767e-07 2.44027319e-07]]

[[1.13353257e-04 7.41319752e-08 8.14688804e-07 6.34582015e-03
1.80785730e-06 5.15732568e-10 9.93498564e-01 3.95792486e-05]
[9.99952435e-01 7.21420292e-07 1.26446380e-07 8.54267981e-08
1.25725208e-10 4.65103985e-05 1.04982995e-07 2.35276468e-11]
[8.58553290e-01 1.28908595e-09 1.01818882e-08 1.53069698e-06
1.40553445e-01 7.75705150e-04 1.16072988e-04 3.01420555e-10]
[5.14177859e-01 3.27013076e-05 5.08953744e-05 4.06983756e-02
4.24867421e-01 1.34739248e-05 1.98230948e-02 3.36182537e-04]
[9.99999285e-01 3.82574882e-15 2.07914652e-19 4.74645966e-15
6.85358486e-07 4.67754191e-09 9.55586996e-15 1.04004281e-19]
[9.73042130e-01 1.16300402e-13 1.15553307e-16 1.38598348e-13
2.68260762e-02 1.31769149e-04 2.13823981e-09 1.42838703e-21]
[6.58826590e-01 1.14370661e-03 2.71215523e-03 7.10488437e-03
3.27758878e-01 2.98021419e-04 1.01155951e-03 1.14424620e-03]]

[[4.72898711e-04 4.25201688e-06 5.96813661e-05 1.09753758e-03
4.91514392e-02 9.07602347e-03 9.27107036e-01 1.30311595e-02]
[8.36947858e-01 1.07748428e-06 7.42213797e-06 1.65293876e-10
1.63002566e-01 5.08171388e-06 1.37384095e-05 2.22442668e-05]
[9.99180496e-01 6.13125684e-09 1.04616573e-08 1.44813821e-13
7.98821158e-04 2.06147106e-05 1.00973892e-07 5.70918833e-12]
[3.06183666e-01 7.12537300e-03 6.48844019e-02 2.82171527e-07
2.60606378e-01 3.54590029e-01 6.57759141e-03 3.22983506e-05]
[9.06643927e-01 2.92241180e-06 4.25770239e-04 1.02464046e-05
2.77019311e-02 6.34697527e-02 1.74534414e-03 3.84745763e-10]
[9.94718373e-01 3.48137474e-09 1.40934063e-07 2.90974644e-09
5.14680333e-03 5.18578490e-05 8.28195480e-05 7.00710269e-13]
[2.36748725e-01 1.28461523e-02 4.61331069e-01 1.62072517e-02
1.41178578e-01 1.62665278e-03 8.32319260e-02 4.68295738e-02]]

[[6.37385845e-01 4.78122502e-07 1.37834547e-06 2.81146235e-08
1.18766096e-04 3.54818851e-01 7.65525550e-03 1.94007753e-05]
[9.04297292e-01 9.33399025e-10 8.57405080e-09 1.72023076e-10
9.56908390e-02 7.96810866e-08 5.29237409e-09 1.18292410e-05]
[9.65073526e-01 4.98108228e-13 2.26485714e-13 8.48323659e-19
3.49259079e-02 5.92647609e-07 3.34810901e-10 2.65504198e-17]
[5.54283766e-08 4.74699339e-16 1.77488513e-08 9.92390323e-15
9.99906182e-01 3.26605125e-08 9.38145749e-05 8.06747724e-11]
[5.73852472e-03 1.56076985e-05 1.44343629e-01 3.08962569e-07
5.14853001e-01 1.22485556e-01 2.12562978e-01 3.42035406e-07]
[9.99987483e-01 1.15470256e-09 3.84308152e-07 3.59167918e-09
1.17577478e-07 1.05222744e-05 1.46735670e-06 3.61845970e-10]
[9.91667926e-01 7.44503486e-05 1.01070746e-05 6.01074089e-06
5.45765692e-03 1.44164206e-03 1.00829499e-03 3.33901902e-04]]

[[9.99976516e-01 2.12335927e-08 2.44374868e-11 4.90566687e-14
1.85316767e-05 3.11313374e-06 1.89475543e-06 1.88924565e-09]
[8.69753957e-01 6.94128435e-12 9.46339743e-11 2.25920238e-09
1.30242124e-01 1.71661213e-07 1.38845701e-06 2.46356558e-06]
[6.09741881e-02 2.91568814e-08 6.33201136e-09 1.24241069e-12
9.38653231e-01 3.47710797e-04 2.46832442e-05 1.40745393e-09]
[9.65304375e-01 5.42137819e-12 4.16872093e-15 1.99302908e-16
2.62953900e-03 3.15223038e-02 5.43761242e-04 2.34946924e-08]
[9.99999762e-01 1.69428129e-17 1.96970333e-15 7.79398563e-22
1.61234161e-15 5.50997803e-10 2.61878483e-07 2.89353993e-15]
[9.99998331e-01 8.35367779e-15 7.49791118e-10 1.49876127e-14
1.57211154e-11 1.00533715e-09 1.66164773e-06 1.71007617e-14]
[7.95299888e-01 3.93584809e-09 3.14216138e-08 9.37843048e-09
1.89878672e-01 1.77769461e-08 1.48213049e-02 4.91423684e-08]]

[[9.33954537e-01 1.20973964e-04 1.14206324e-08 1.40171471e-06
6.00965977e-05 9.53610140e-07 6.58588707e-02 3.16920477e-06]
[9.99959469e-01 1.13300862e-13 1.41840463e-11 3.10680043e-05
6.35626995e-10 8.57502584e-07 1.42984035e-07 8.45882550e-06]
[9.98554051e-01 3.87483638e-13 2.52303654e-08 8.05722156e-09
5.34030914e-05 2.30311663e-08 9.89229535e-04 4.03299549e-04]
[1.31540317e-02 1.90780975e-06 1.72623041e-10 2.25340080e-10
5.03390027e-07 4.64845244e-15 9.86843646e-01 1.78412185e-09]
[2.04621449e-01 2.54608722e-06 1.00194529e-06 1.23633625e-04
3.02570657e-09 1.04778031e-09 7.95251131e-01 2.15082252e-07]
[9.99889493e-01 1.61989137e-11 1.87405689e-07 1.75819496e-05
3.97604083e-09 9.12943215e-05 9.07707631e-10 1.42147223e-06]
[4.88181453e-04 1.86133953e-08 1.92044300e-08 1.33892484e-02
3.55223455e-02 3.14264037e-09 9.50586915e-01 1.32504765e-05]]]]
finish 1033

Answer 3

我尝试使用建议的更改将模型与我的数据进行拟合。现在，特征被提取并保存在数据帧中，并保存为csv文件。当我试图将其提供给分类器时，它在这些特征的训练上出现错误。获取6列而不是8列或获取9列而不是8列时出错。问题出在csv文件中的值中。训练分类器时使用的命令位于此行X_train = genfromtxt(X_train_file_path，delimiter='，') ValueError:检测到一些错误！第2行(有7列而不是8列)

其次，你能指导我添加unet的编解码器模块吗？