DEX-6-caffe模型转成pytorch模型办法

在python2.7环境下

文件下载位置：https://data.vision.ee.ethz.ch/cvl/rrothe/imdb-wiki/

1.可视化模型文件prototxt

1）在线可视化

网址为：https://ethereon.github.io/netscope/#/editor

将prototxt文件的内容复制到左边,然后按shift-enter键即可：

2）本地可视化

先安装：

(deeplearning2) userdeMacBook-Pro:~ user$ brew install graphviz

再安装：

(deeplearning2) userdeMacBook-Pro:~ user$ pip install pydot

前提是安装好了caffe，然后运行：

(deeplearning2) userdeMacBook-Pro:face_data user$ python /Users/user/caffe/python/draw_net.py /Users/user/pytorch/face_data/age_train.prototxt age_train.png --rankdir=BT
Drawing net to age_train.png

参数意思：

• 参数1：网络模型的prototxt文件
• 参数2：保存的图片路径及名字
• 参数3：--rankdir=x , x 有四种选项，分别是LR, RL, TB, BT 。用来表示网络的方向，分别是从左到右，从右到左，从上到小，从下到上。默认为LR

这是更改输入、损失和精确度层的图，为：

没更改过的为：

(deeplearning2) userdeMacBook-Pro:face_data user$ python /Users/user/caffe/python/draw_net.py /Users/user/pytorch/face_data/age_train.prototxt.txt age_train_2.png --rankdir=BT
Drawing net to age_train_2.png

图为：

2.网络结构说明

name: "VGG_ILSVRC_16_layers" # 声明该网络结构的名称
# 下面先是指明使用的数据集
layer {                                         #指明用于训练的数据，训练集
  top: "data"
  type: "ImageData"
  top: "label"
  name: "data"
  transform_param {
    mirror: true
    crop_size:
    mean_file: "imagenet_mean.binaryproto"
  }
  image_data_param {
    source: "train.txt"                       #指明存放位置
    batch_size:
    new_height:
    new_width:
  }
  include: { phase: TRAIN }                       #指明用于训练
}
layer {                                         #指明用于验证的数据，验证集
  top: "data"
  top: "label"
  name: "data"
  type: "ImageData"
  image_data_param {
    new_height:
    new_width:
    source: "train.txt"
    batch_size:
  }
  transform_param {
    crop_size:
    mirror: false
    mean_file: "imagenet_mean.binaryproto"
  }
  include: {
    phase: TEST
    stage: "test-on-train"
 }
}

layer {                                         #指明用于测试的数据，测试集
  top: "data"
  top: "label"
  name: "data"
  type: "ImageData"
  image_data_param {
    new_height:
    new_width:
    source: "test.txt"
    batch_size:
  }
  transform_param {
    crop_size:
    mirror: false
    mean_file: "imagenet_mean.binaryproto"
  }
  include: {
    phase: TEST
    stage: "test-on-test"
 }
}

#下面就是网络结构了，这里以第一个卷积层为例

layer {
  bottom: "data"            #该层的上一层为data数据层，即其输入
  top: "conv1_1"            #该层的输出
  name: "conv1_1"           #该层的名称conv1_1
 param {                    #权重weight的参数
    lr_mult:               #权重weight的学习率
    decay_mult:            #权重weight的衰减系数
  }
  param {                   #偏置bias的参数
    lr_mult:               #偏置bias的学习率
    decay_mult:            #偏置bias的衰减系数
  }
  type: "Convolution"       #该层的类型，为卷积
  convolution_param {       #卷积层的参数
    num_output:           #卷积核个数，即输出的channels数
    pad:                   #padding大小
    kernel_size:           #卷积核的大小
  }
}
layer {
  bottom: "conv1_1"
  top: "conv1_1"
  name: "relu1_1"
  type: "ReLU"              #该层的类型，为激活函数，无参数
}
layer {
  bottom: "conv1_1"
  top: "conv1_2"
  name: "conv1_2"           #该层的名称conv1_1
 param {
    lr_mult:
    decay_mult:
  }
  param {
    lr_mult:
    decay_mult:
  }
  type: "Convolution"
  convolution_param {
    num_output:
    pad:
    kernel_size:
  }
}
layer {
  bottom: "conv1_2"
  top: "conv1_2"
  name: "relu1_2"
  type: "ReLU"              #该层的类型，为激活函数，无参数
}
layer {
  bottom: "conv1_2"
  top: "pool1"
  name: "pool1"
  type: "Pooling"           #该层的类型，为池化层
  pooling_param {
    pool: MAX               #为最大池化层
    kernel_size:           #卷积核的大小
    stride:                #步长大小
  }
}

# 中间其他卷积层省略
...

# 全连接层
layer {
  bottom: "pool5"
  top: "fc6"
  name: "fc6"
 param {
    lr_mult:
    decay_mult:
  }
  param {
    lr_mult:
    decay_mult:
  }
  type: "InnerProduct"       #该层的类型，为全连接层
  inner_product_param {      #输出个数为4096
    num_output:
  }
}
layer {
  bottom: "fc6"
  top: "fc6"
  name: "relu6"
  type: "ReLU"
}
layer {
  bottom: "fc6"
  top: "fc6"
  name: "drop6"
  type: "Dropout"             #该层的类型，为dropout层
  dropout_param {
    dropout_ratio: 0.5        #dropout比率为0.
  }
}
layer {
  bottom: "fc6"
  top: "fc7"
  name: "fc7"
 param {
    lr_mult:
    decay_mult:
  }
  param {
    lr_mult:
    decay_mult:
  }
  type: "InnerProduct"
  inner_product_param {
    num_output:
  }
}
layer {
  bottom: "fc7"
  top: "fc7"
  name: "relu7"
  type: "ReLU"
}
layer {
  bottom: "fc7"
  top: "fc7"
  name: "drop7"
  type: "Dropout"
  dropout_param {
    dropout_ratio: 0.5
  }
}
layer {
  bottom: "fc7"
  top: "fc8-101"
  name: "fc8-101"
 param {
    lr_mult:
    decay_mult:
  }
  param {
    lr_mult:
    decay_mult:
  }
  type: "InnerProduct"
  inner_product_param {
    num_output:
  weight_filler {             #权重初始化方法参数
      type: "gaussian"        #使用的初始化方法为高斯分布
      std: 0.01               #标准差为0.
    }
    bias_filler {             #偏置初始化方法参数
      type: "constant"        #初始化为定值0
      value:
    }
  }
}

layer {
  bottom: "fc8-101"
  bottom: "label"
  name: "loss"
  type: "SoftmaxWithLoss"      #最后对结果求损失，使用的方法为SoftmaxWithLoss
  include: { phase: TRAIN }
}

# 如果是测试则仅使用Softmax求概率

layer {
  name: "prob"
  type: "Softmax"
  bottom: "fc8-101"
  top: "prob"
    include {
    phase: TEST
  }
}

# 当进行验证时进行的求top1、top5和top10的操作，求精确度
layer {
  name: "accuracy_train_top01"
  type: "Accuracy"
  bottom: "fc8-101"
  bottom: "label"
  top: "accuracy_train_top01"
  include {
    phase: TEST
    stage: "test-on-train"
  }
}

layer {
  name: "accuracy_train_top05"
  type: "Accuracy"
  bottom: "fc8-101"
  bottom: "label"
  top: "accuracy_train_top05"
  accuracy_param {
    top_k:
  }
  include {
    phase: TEST
    stage: "test-on-train"
  }
}

layer {
  name: "accuracy_train_top10"
  type: "Accuracy"
  bottom: "fc8-101"
  bottom: "label"
  top: "accuracy_train_top10"
  accuracy_param {
    top_k:
  }
  include {
    phase: TEST
    stage: "test-on-train"
  }
}

#测试时也一样，求top1、top5和top10的操作，求精确度
layer {
  name: "accuracy_test_top01"
  type: "Accuracy"
  bottom: "fc8-101"
  bottom: "label"
  top: "accuracy_test_top01"
  include {
    phase: TEST
    stage: "test-on-test"
  }
}

layer {
  name: "accuracy_test_top05"
  type: "Accuracy"
  bottom: "fc8-101"
  bottom: "label"
  top: "accuracy_test_top05"
  accuracy_param {
    top_k:
  }
  include {
    phase: TEST
    stage: "test-on-test"
  }
}

layer {
  name: "accuracy_test_top10"
  type: "Accuracy"
  bottom: "fc8-101"
  bottom: "label"
  top: "accuracy_test_top10"
  accuracy_param {
    top_k:
  }
  include {
    phase: TEST
    stage: "test-on-test"
  }
}

3.自己定义的VGG16模型-models.py

然后我自己定义的pytorch模型为：

# coding:utf-
import torchvision.models as models
import torch.nn as nn
import torch
from torch.autograd import Variable
import torch.nn.functional as F

# -100岁
class AgeModel(nn.Module):
    def __init__(self, classes=):
        super(AgeModel,self).__init__()
        model = models.vgg16()
        layers = list(model.children())
        # 一共只有3层，卷积层，池化层和全连接层
        # print(len(layers)) #
        self.model = nn.Sequential(*layers[:-])
        # -1层为整个连接层
        # print(layers[]) #池化层AdaptiveAvgPool2d
        # print(layers[-]) #全连接层
        self.fullCon = nn.Sequential(nn.Linear(**, ),
                                   nn.ReLU(True),
                                   nn.Dropout(0.5, inplace=True),
                                   nn.Linear(, ),
                                   nn.ReLU(True),
                                   nn.Dropout(0.5, inplace=True),
                                   nn.Linear(, classes) #将最后一层全连接层改成自己的输出101
                              )

    def forward(self, x):
        x = self.model(x)
        x = x.view(x.size(), -)
        x = self.fullCon(x)
        x = F.softmax(x, dim=)
        return x

def test():
    net = AgeModel()
    for module in net.children():
        print(module)
    output = net(Variable(torch.randn(,,,)))
    print('output :', output.size())
    print(type(output))

if __name__ == '__main__':
    test()

4.下载对应的caffemodel模型并转换

(deeplearning2) userdeMacBook-Pro:pytorch-DEX-master user$ curl -o age_model_imdb.caffemodel https://data.vision.ee.ethz.ch/cvl/rrothe/imdb-wiki/static/dex_imdb_wiki.caffemodel

然后转换代码为：

# coding:utf-
import caffe
import torch

from models import AgeModel

def convert_age():
    age_torch_net = AgeModel()

    # 导入caffe模型，caffe.TRAIN表示使用训练模式，caffe.TEST表示使用测试模式
    # 不同模式使用的prototxt文件是不同的
    caffe_net = caffe.Net('age_train.prototxt', "age_model_imdb.caffemodel", caffe.TRAIN)
    # caffe_net = caffe.Net('age.prototxt', "age_model_imdb.caffemodel", caffe.TEST)

    # 对应得到的是权重weight和偏置bias的参数
    caffe_params = caffe_net.params
    print(caffe_params)

    # #查看索引表示
    # print(age_torch_net.model)
    # print(age_torch_net.model[][])
    #
    # print(age_torch_net.fullCon)
    # print(age_torch_net.fullCon[])

    mappings = {
        'conv1_1': age_torch_net.model[][],
        'conv1_2': age_torch_net.model[][],
        'conv2_1': age_torch_net.model[][],
        'conv2_2': age_torch_net.model[][],
        'conv3_1': age_torch_net.model[][],
        'conv3_2': age_torch_net.model[][],
        'conv3_3': age_torch_net.model[][],
        'conv4_1': age_torch_net.model[][],
        'conv4_2': age_torch_net.model[][],
        'conv4_3': age_torch_net.model[][],
        'conv5_1': age_torch_net.model[][],
        'conv5_2': age_torch_net.model[][],
        'conv5_3': age_torch_net.model[][],
        'fc6': age_torch_net.fullCon[],
        'fc7': age_torch_net.fullCon[],
        'fc8-101': age_torch_net.fullCon[],
    }

    for k, layer in mappings.items():
        # # 第一个迭代：k=conv1_1，layer=Conv2d(, , kernel_size=(, ), stride=(, ), padding=(, ))
        # print(k, layer)

        # 将caffe模型中的参数信息导入定义好的pytorch模型中
        layer.weight.data.copy_(torch.from_numpy(caffe_params[k][].data)) # 权重weight参数
        layer.bias.data.copy_(torch.from_numpy(caffe_params[k][].data)) # 偏置bias的参数
    # 然后仅保存模型参数即可
    torch.save(age_torch_net.state_dict(), './age.pth')

if __name__ == '__main__':
    convert_age()

过程中出现了错误：

RuntimeError: unidentifiable C++ exception
Segmentation fault: 

因为：

caffe_net = caffe.Net('age_train.prototxt.txt', "dex_imdb_wiki.caffemodel", caffe.TRAIN)

这里我下载时是age_train.prototxt.txt，但是运行是导入是age_train.prototxt，添加.txt即可，或者更改格式

然后还出现错误：

(deeplearning2) userdeMacBook-Pro:face_data user$ python convert.py
Traceback (most recent call last):
  File "convert.py", line 2, in <module>
    import caffe
  File "/Users/user/caffe/python/caffe/__init__.py", line 1, in <module>
    from .pycaffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, RMSPropSolver, AdaDeltaSolver, AdamSolver, NCCL, Timer
  File "/Users/user/caffe/python/caffe/pycaffe.py", line 13, in <module>
    from ._caffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, \
ImportError: dlopen(/Users/user/caffe/python/caffe/_caffe.so, 2): Library not loaded: @rpath/libpython2.7.dylib
  Referenced from: /Users/user/caffe/python/caffe/_caffe.so
  Reason: image not found
(deeplearning2) userdeMacBook-Pro:face_data user$ python convert.py
Traceback (most recent call last):
  File "convert.py", line 2, in <module>
    import caffe
  File "/Users/user/caffe/python/caffe/__init__.py", line 1, in <module>
    from .pycaffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, RMSPropSolver, AdaDeltaSolver, AdamSolver, NCCL, Timer
  File "/Users/user/caffe/python/caffe/pycaffe.py", line 13, in <module>
    from ._caffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, \
ImportError: dlopen(/Users/user/caffe/python/caffe/_caffe.so, 2): Library not loaded: @rpath/libpython2.7.dylib
  Referenced from: /Users/user/caffe/python/caffe/_caffe.so
  Reason: image not found

因为我是使用anaconda3的python2.7环境，其下面已经有libpython2.7.dylib，这里报错的原因是@rpath的导向中没有anaconda3的python2.7环境的路径，使用下面的命令添加即可：

(deeplearning2) userdeMacBook-Pro:caffe user$ install_name_tool -add_rpath /anaconda3/envs/deeplearning2/lib /Users/user/caffe/python/caffe/_caffe.so

还出现错误：

I0812 14:26:19.981719 245106112 layer_factory.hpp:77] Creating layer data
F0812 14:26:19.981747 245106112 layer_factory.hpp:81] Check failed: registry.count(type) == 1 (0 vs. 1) Unknown layer type: ImageData (known types: AbsVal, Accuracy, ArgMax, BNLL, BatchNorm, BatchReindex, Bias, Clip, Concat, ContrastiveLoss, Convolution, Crop, Data, Deconvolution, Dropout, DummyData, ELU, Eltwise, Embed, EuclideanLoss, Exp, Filter, Flatten, HDF5Data, HDF5Output, HingeLoss, Im2col, InfogainLoss, InnerProduct, Input, LRN, LSTM, LSTMUnit, Log, MVN, MemoryData, MultinomialLogisticLoss, PReLU, Parameter, Pooling, Power, RNN, ReLU, Reduction, Reshape, SPP, Scale, Sigmoid, SigmoidCrossEntropyLoss, Silence, Slice, Softmax, SoftmaxWithLoss, Split, Swish, TanH, Threshold, Tile)
*** Check failure stack trace: ***
Abort trap: 

这是因为在配置Makefile.config时没有取消掉下面一行的注释，所以读取不了python写的layer：

# Uncomment to support layers written in Python (will link against Python libs)
WITH_PYTHON_LAYER := 

将该行去掉，然后重新编译caffe，即make clean后，再从make all-make test-make runtest-make pycaffe,并且还要跟上面一样在~/.bash_profile文件的path中添加~/caffe/python这个caffe路径，并source /etc/profile

但是这样仍然没有解决问题,后面发现如果你的报错是：

Check failed: registry.count(type) == 1 (0 vs. 1) Unknown layer type:python

才使用上面的方法，但是注释掉这个总没有错。这个更改我已经合并到上面的完整配置代码中了，你如果跟着上面配置走应该就不会遇见该问题

后面查看该caffe模型的文件，其使用了ImageData来导入数据，但是我没有他指定的文件，所以将其的prototxt文件数据导入部分删除，改为：

input: "data"
input_dim:
input_dim:
input_dim:
input_dim: 

把后面的一些损失和精确度内容也删除

然后再运行即可:

(deeplearning2) wanghuideMacBook-Pro:face_data wanghui$ python convert.py
WARNING: Logging before InitGoogleLogging() is written to STDERR
W0812 ::50.179870  _caffe.cpp:] DEPRECATION WARNING - deprecated use of Python interface
W0812 ::50.180747  _caffe.cpp:] Use this instead (with the named "weights" parameter):
W0812 ::50.180763  _caffe.cpp:] Net('age_train.prototxt', , weights='dex_imdb_wiki.caffemodel')
I0812 ::50.188591  upgrade_proto.cpp:] Attempting to upgrade input file specified using deprecated input fields: age_train.prototxt
I0812 ::50.188863  upgrade_proto.cpp:] Successfully upgraded file specified using deprecated input fields.
W0812 ::50.188877  upgrade_proto.cpp:] Note that future Caffe releases will only support input layers and not input fields.
I0812 ::50.189504  net.cpp:] Initializing net from parameters:
name: "VGG_ILSVRC_16_layers"
state {
  phase: TRAIN
  level:
}
layer {
  name: "input"
  type: "Input"
  top: "data"
  input_param {
    shape {
      dim:
      dim:
      dim:
      dim:
    }
  }
}
layer {
  name: "conv1_1"
  type: "Convolution"
...
layer {
  name: "fc8-101"
  type: "InnerProduct"
  bottom: "fc7"
  top: "fc8-101"
  param {
    lr_mult:
    decay_mult:
  }
  param {
    lr_mult:
    decay_mult:
  }
  inner_product_param {
    num_output:
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
    bias_filler {
      type: "constant"
      value:
    }
  }
}
I0812 ::50.190675  layer_factory.hpp:] Creating layer input
I0812 ::50.190943  net.cpp:] Creating Layer input
I0812 ::50.190958  net.cpp:] input -> data
I0812 ::50.191754  net.cpp:] Setting up input
I0812 ::50.191768  net.cpp:] Top shape:     ()
I0812 ::50.192245  net.cpp:] Memory required for data:
I0812 ::50.192253  layer_factory.hpp:] Creating layer conv1_1
I0812 ::50.192262  net.cpp:] Creating Layer conv1_1
I0812 ::50.192267  net.cpp:] conv1_1 <- data
I0812 ::50.192273  net.cpp:] conv1_1 -> conv1_1
I0812 ::50.192760  net.cpp:] Setting up conv1_1
I0812 ::50.192771  net.cpp:] Top shape:     ()
I0812 ::50.192780  net.cpp:] Memory required for data:
I0812 ::50.192903  layer_factory.hpp:] Creating layer relu1_1
I0812 ::50.192927  net.cpp:] Creating Layer relu1_1
I0812 ::50.192940  net.cpp:] relu1_1 <- conv1_1
I0812 ::50.192955  net.cpp:] relu1_1 -> conv1_1 (in-place)
I0812 ::50.192987  net.cpp:] Setting up relu1_1
I0812 ::50.193004  net.cpp:] Top shape:     ()
I0812 ::50.193015  net.cpp:] Memory required for data:
I0812 ::50.193022  layer_factory.hpp:] Creating layer conv1_2
I0812 ::50.193029  net.cpp:] Creating Layer conv1_2
I0812 ::50.193034  net.cpp:] conv1_2 <- conv1_1
I0812 ::50.193042  net.cpp:] conv1_2 -> conv1_2
I0812 ::50.193079  net.cpp:] Setting up conv1_2
I0812 ::50.193085  net.cpp:] Top shape:     ()
I0812 ::50.193092  net.cpp:] Memory required for data:
I0812 ::50.193099  layer_factory.hpp:] Creating layer relu1_2
I0812 ::50.193106  net.cpp:] Creating Layer relu1_2
I0812 ::50.193111  net.cpp:] relu1_2 <- conv1_2
I0812 ::50.193116  net.cpp:] relu1_2 -> conv1_2 (in-place)
I0812 ::50.193122  net.cpp:] Setting up relu1_2
I0812 ::50.193126  net.cpp:] Top shape:     ()
I0812 ::50.193132  net.cpp:] Memory required for data:
I0812 ::50.193137  layer_factory.hpp:] Creating layer pool1
I0812 ::50.193142  net.cpp:] Creating Layer pool1
I0812 ::50.193147  net.cpp:] pool1 <- conv1_2
I0812 ::50.193152  net.cpp:] pool1 -> pool1
I0812 ::50.193373  net.cpp:] Setting up pool1
I0812 ::50.193385  net.cpp:] Top shape:     ()
I0812 ::50.193394  net.cpp:] Memory required for data:
I0812 ::50.193400  layer_factory.hpp:] Creating layer conv2_1
I0812 ::50.193408  net.cpp:] Creating Layer conv2_1
I0812 ::50.193414  net.cpp:] conv2_1 <- pool1
I0812 ::50.193421  net.cpp:] conv2_1 -> conv2_1
I0812 ::50.193460  net.cpp:] Setting up conv2_1
I0812 ::50.193466  net.cpp:] Top shape:     ()
I0812 ::50.193473  net.cpp:] Memory required for data:
I0812 ::50.193481  layer_factory.hpp:] Creating layer relu2_1
I0812 ::50.193487  net.cpp:] Creating Layer relu2_1
I0812 ::50.193492  net.cpp:] relu2_1 <- conv2_1
I0812 ::50.193497  net.cpp:] relu2_1 -> conv2_1 (in-place)
I0812 ::50.193504  net.cpp:] Setting up relu2_1
I0812 ::50.193508  net.cpp:] Top shape:     ()
I0812 ::50.193514  net.cpp:] Memory required for data:
I0812 ::50.193519  layer_factory.hpp:] Creating layer conv2_2
I0812 ::50.193526  net.cpp:] Creating Layer conv2_2
I0812 ::50.193531  net.cpp:] conv2_2 <- conv2_1
I0812 ::50.193536  net.cpp:] conv2_2 -> conv2_2
I0812 ::50.193588  net.cpp:] Setting up conv2_2
I0812 ::50.193596  net.cpp:] Top shape:     ()
I0812 ::50.193604  net.cpp:] Memory required for data:
I0812 ::50.193610  layer_factory.hpp:] Creating layer relu2_2
I0812 ::50.193616  net.cpp:] Creating Layer relu2_2
I0812 ::50.193620  net.cpp:] relu2_2 <- conv2_2
I0812 ::50.193625  net.cpp:] relu2_2 -> conv2_2 (in-place)
I0812 ::50.193631  net.cpp:] Setting up relu2_2
I0812 ::50.193635  net.cpp:] Top shape:     ()
I0812 ::50.193641  net.cpp:] Memory required for data:
I0812 ::50.193645  layer_factory.hpp:] Creating layer pool2
I0812 ::50.193651  net.cpp:] Creating Layer pool2
I0812 ::50.193655  net.cpp:] pool2 <- conv2_2
I0812 ::50.193661  net.cpp:] pool2 -> pool2
I0812 ::50.193668  net.cpp:] Setting up pool2
I0812 ::50.193672  net.cpp:] Top shape:     ()
I0812 ::50.193678  net.cpp:] Memory required for data:
I0812 ::50.193682  layer_factory.hpp:] Creating layer conv3_1
I0812 ::50.193688  net.cpp:] Creating Layer conv3_1
I0812 ::50.193694  net.cpp:] conv3_1 <- pool2
I0812 ::50.193699  net.cpp:] conv3_1 -> conv3_1
I0812 ::50.193792  net.cpp:] Setting up conv3_1
I0812 ::50.193799  net.cpp:] Top shape:     ()
I0812 ::50.193805  net.cpp:] Memory required for data:
I0812 ::50.193814  layer_factory.hpp:] Creating layer relu3_1
I0812 ::50.193819  net.cpp:] Creating Layer relu3_1
I0812 ::50.193825  net.cpp:] relu3_1 <- conv3_1
I0812 ::50.193832  net.cpp:] relu3_1 -> conv3_1 (in-place)
I0812 ::50.193838  net.cpp:] Setting up relu3_1
I0812 ::50.193843  net.cpp:] Top shape:     ()
I0812 ::50.193850  net.cpp:] Memory required for data:
I0812 ::50.193855  layer_factory.hpp:] Creating layer conv3_2
I0812 ::50.193861  net.cpp:] Creating Layer conv3_2
I0812 ::50.193866  net.cpp:] conv3_2 <- conv3_1
I0812 ::50.193871  net.cpp:] conv3_2 -> conv3_2
I0812 ::50.194085  net.cpp:] Setting up conv3_2
I0812 ::50.194097  net.cpp:] Top shape:     ()
I0812 ::50.194105  net.cpp:] Memory required for data:
I0812 ::50.194113  layer_factory.hpp:] Creating layer relu3_2
I0812 ::50.194119  net.cpp:] Creating Layer relu3_2
I0812 ::50.194124  net.cpp:] relu3_2 <- conv3_2
I0812 ::50.194130  net.cpp:] relu3_2 -> conv3_2 (in-place)
I0812 ::50.194137  net.cpp:] Setting up relu3_2
I0812 ::50.194141  net.cpp:] Top shape:     ()
I0812 ::50.194147  net.cpp:] Memory required for data:
I0812 ::50.194151  layer_factory.hpp:] Creating layer conv3_3
I0812 ::50.194159  net.cpp:] Creating Layer conv3_3
I0812 ::50.194162  net.cpp:] conv3_3 <- conv3_2
I0812 ::50.194169  net.cpp:] conv3_3 -> conv3_3
I0812 ::50.194345  net.cpp:] Setting up conv3_3
I0812 ::50.194353  net.cpp:] Top shape:     ()
I0812 ::50.194360  net.cpp:] Memory required for data:
I0812 ::50.194366  layer_factory.hpp:] Creating layer relu3_3
I0812 ::50.194373  net.cpp:] Creating Layer relu3_3
I0812 ::50.194377  net.cpp:] relu3_3 <- conv3_3
I0812 ::50.194383  net.cpp:] relu3_3 -> conv3_3 (in-place)
I0812 ::50.194388  net.cpp:] Setting up relu3_3
I0812 ::50.194393  net.cpp:] Top shape:     ()
I0812 ::50.194398  net.cpp:] Memory required for data:
I0812 ::50.194403  layer_factory.hpp:] Creating layer pool3
I0812 ::50.194411  net.cpp:] Creating Layer pool3
I0812 ::50.194416  net.cpp:] pool3 <- conv3_3
I0812 ::50.194422  net.cpp:] pool3 -> pool3
I0812 ::50.194428  net.cpp:] Setting up pool3
I0812 ::50.194433  net.cpp:] Top shape:     ()
I0812 ::50.194438  net.cpp:] Memory required for data:
I0812 ::50.194443  layer_factory.hpp:] Creating layer conv4_1
I0812 ::50.194450  net.cpp:] Creating Layer conv4_1
I0812 ::50.194454  net.cpp:] conv4_1 <- pool3
I0812 ::50.194460  net.cpp:] conv4_1 -> conv4_1
I0812 ::50.194834  net.cpp:] Setting up conv4_1
I0812 ::50.194847  net.cpp:] Top shape:     ()
I0812 ::50.194854  net.cpp:] Memory required for data:
I0812 ::50.194861  layer_factory.hpp:] Creating layer relu4_1
I0812 ::50.194867  net.cpp:] Creating Layer relu4_1
I0812 ::50.194874  net.cpp:] relu4_1 <- conv4_1
I0812 ::50.194880  net.cpp:] relu4_1 -> conv4_1 (in-place)
I0812 ::50.194885  net.cpp:] Setting up relu4_1
I0812 ::50.194890  net.cpp:] Top shape:     ()
I0812 ::50.194895  net.cpp:] Memory required for data:
I0812 ::50.194900  layer_factory.hpp:] Creating layer conv4_2
I0812 ::50.194905  net.cpp:] Creating Layer conv4_2
I0812 ::50.194911  net.cpp:] conv4_2 <- conv4_1
I0812 ::50.194917  net.cpp:] conv4_2 -> conv4_2
I0812 ::50.195816  net.cpp:] Setting up conv4_2
I0812 ::50.195835  net.cpp:] Top shape:     ()
I0812 ::50.195843  net.cpp:] Memory required for data:
I0812 ::50.195854  layer_factory.hpp:] Creating layer relu4_2
I0812 ::50.195863  net.cpp:] Creating Layer relu4_2
I0812 ::50.195868  net.cpp:] relu4_2 <- conv4_2
I0812 ::50.195874  net.cpp:] relu4_2 -> conv4_2 (in-place)
I0812 ::50.195881  net.cpp:] Setting up relu4_2
I0812 ::50.195886  net.cpp:] Top shape:     ()
I0812 ::50.195892  net.cpp:] Memory required for data:
I0812 ::50.195896  layer_factory.hpp:] Creating layer conv4_3
I0812 ::50.195904  net.cpp:] Creating Layer conv4_3
I0812 ::50.195907  net.cpp:] conv4_3 <- conv4_2
I0812 ::50.195914  net.cpp:] conv4_3 -> conv4_3
I0812 ::50.197036  net.cpp:] Setting up conv4_3
I0812 ::50.197062  net.cpp:] Top shape:     ()
I0812 ::50.197072  net.cpp:] Memory required for data:
I0812 ::50.197079  layer_factory.hpp:] Creating layer relu4_3
I0812 ::50.197088  net.cpp:] Creating Layer relu4_3
I0812 ::50.197094  net.cpp:] relu4_3 <- conv4_3
I0812 ::50.197100  net.cpp:] relu4_3 -> conv4_3 (in-place)
I0812 ::50.197108  net.cpp:] Setting up relu4_3
I0812 ::50.197111  net.cpp:] Top shape:     ()
I0812 ::50.197116  net.cpp:] Memory required for data:
I0812 ::50.197120  layer_factory.hpp:] Creating layer pool4
I0812 ::50.197126  net.cpp:] Creating Layer pool4
I0812 ::50.197130  net.cpp:] pool4 <- conv4_3
I0812 ::50.197136  net.cpp:] pool4 -> pool4
I0812 ::50.197144  net.cpp:] Setting up pool4
I0812 ::50.197149  net.cpp:] Top shape:     ()
I0812 ::50.197154  net.cpp:] Memory required for data:
I0812 ::50.197158  layer_factory.hpp:] Creating layer conv5_1
I0812 ::50.197166  net.cpp:] Creating Layer conv5_1
I0812 ::50.197171  net.cpp:] conv5_1 <- pool4
I0812 ::50.197176  net.cpp:] conv5_1 -> conv5_1
I0812 ::50.198091  net.cpp:] Setting up conv5_1
I0812 ::50.198110  net.cpp:] Top shape:     ()
I0812 ::50.198118  net.cpp:] Memory required for data:
I0812 ::50.198124  layer_factory.hpp:] Creating layer relu5_1
I0812 ::50.198132  net.cpp:] Creating Layer relu5_1
I0812 ::50.198137  net.cpp:] relu5_1 <- conv5_1
I0812 ::50.198143  net.cpp:] relu5_1 -> conv5_1 (in-place)
I0812 ::50.198149  net.cpp:] Setting up relu5_1
I0812 ::50.198153  net.cpp:] Top shape:     ()
I0812 ::50.198159  net.cpp:] Memory required for data:
I0812 ::50.198163  layer_factory.hpp:] Creating layer conv5_2
I0812 ::50.198170  net.cpp:] Creating Layer conv5_2
I0812 ::50.198174  net.cpp:] conv5_2 <- conv5_1
I0812 ::50.198179  net.cpp:] conv5_2 -> conv5_2
I0812 ::50.199494  net.cpp:] Setting up conv5_2
I0812 ::50.199513  net.cpp:] Top shape:     ()
I0812 ::50.199522  net.cpp:] Memory required for data:
I0812 ::50.199528  layer_factory.hpp:] Creating layer relu5_2
I0812 ::50.199539  net.cpp:] Creating Layer relu5_2
I0812 ::50.199545  net.cpp:] relu5_2 <- conv5_2
I0812 ::50.199551  net.cpp:] relu5_2 -> conv5_2 (in-place)
I0812 ::50.199558  net.cpp:] Setting up relu5_2
I0812 ::50.199563  net.cpp:] Top shape:     ()
I0812 ::50.199568  net.cpp:] Memory required for data:
I0812 ::50.199571  layer_factory.hpp:] Creating layer conv5_3
I0812 ::50.199579  net.cpp:] Creating Layer conv5_3
I0812 ::50.199582  net.cpp:] conv5_3 <- conv5_2
I0812 ::50.199589  net.cpp:] conv5_3 -> conv5_3
I0812 ::50.200711  net.cpp:] Setting up conv5_3
I0812 ::50.200731  net.cpp:] Top shape:     ()
I0812 ::50.200739  net.cpp:] Memory required for data:
I0812 ::50.200747  layer_factory.hpp:] Creating layer relu5_3
I0812 ::50.200753  net.cpp:] Creating Layer relu5_3
I0812 ::50.200758  net.cpp:] relu5_3 <- conv5_3
I0812 ::50.200764  net.cpp:] relu5_3 -> conv5_3 (in-place)
I0812 ::50.200790  net.cpp:] Setting up relu5_3
I0812 ::50.200794  net.cpp:] Top shape:     ()
I0812 ::50.200800  net.cpp:] Memory required for data:
I0812 ::50.200804  layer_factory.hpp:] Creating layer pool5
I0812 ::50.200810  net.cpp:] Creating Layer pool5
I0812 ::50.200814  net.cpp:] pool5 <- conv5_3
I0812 ::50.200820  net.cpp:] pool5 -> pool5
I0812 ::50.200830  net.cpp:] Setting up pool5
I0812 ::50.200836  net.cpp:] Top shape:     ()
I0812 ::50.200855  net.cpp:] Memory required for data:
I0812 ::50.200860  layer_factory.hpp:] Creating layer fc6
I0812 ::50.200866  net.cpp:] Creating Layer fc6
I0812 ::50.200870  net.cpp:] fc6 <- pool5
I0812 ::50.200876  net.cpp:] fc6 -> fc6
I0812 ::50.257817  net.cpp:] Setting up fc6
I0812 ::50.257875  net.cpp:] Top shape:   ()
I0812 ::50.257887  net.cpp:] Memory required for data:
I0812 ::50.257897  layer_factory.hpp:] Creating layer relu6
I0812 ::50.257906  net.cpp:] Creating Layer relu6
I0812 ::50.257912  net.cpp:] relu6 <- fc6
I0812 ::50.257920  net.cpp:] relu6 -> fc6 (in-place)
I0812 ::50.257927  net.cpp:] Setting up relu6
I0812 ::50.257931  net.cpp:] Top shape:   ()
I0812 ::50.257936  net.cpp:] Memory required for data:
I0812 ::50.257941  layer_factory.hpp:] Creating layer drop6
I0812 ::50.257947  net.cpp:] Creating Layer drop6
I0812 ::50.257952  net.cpp:] drop6 <- fc6
I0812 ::50.257957  net.cpp:] drop6 -> fc6 (in-place)
I0812 ::50.257974  net.cpp:] Setting up drop6
I0812 ::50.257978  net.cpp:] Top shape:   ()
I0812 ::50.257983  net.cpp:] Memory required for data:
I0812 ::50.257987  layer_factory.hpp:] Creating layer fc7
I0812 ::50.257994  net.cpp:] Creating Layer fc7
I0812 ::50.257998  net.cpp:] fc7 <- fc6
I0812 ::50.258003  net.cpp:] fc7 -> fc7
I0812 ::50.303862  net.cpp:] Setting up fc7
I0812 ::50.303889  net.cpp:] Top shape:   ()
I0812 ::50.303900  net.cpp:] Memory required for data:
I0812 ::50.303908  layer_factory.hpp:] Creating layer relu7
I0812 ::50.303918  net.cpp:] Creating Layer relu7
I0812 ::50.303923  net.cpp:] relu7 <- fc7
I0812 ::50.303942  net.cpp:] relu7 -> fc7 (in-place)
I0812 ::50.303951  net.cpp:] Setting up relu7
I0812 ::50.303956  net.cpp:] Top shape:   ()
I0812 ::50.303961  net.cpp:] Memory required for data:
I0812 ::50.303966  layer_factory.hpp:] Creating layer drop7
I0812 ::50.303972  net.cpp:] Creating Layer drop7
I0812 ::50.303977  net.cpp:] drop7 <- fc7
I0812 ::50.303982  net.cpp:] drop7 -> fc7 (in-place)
I0812 ::50.303987  net.cpp:] Setting up drop7
I0812 ::50.303992  net.cpp:] Top shape:   ()
I0812 ::50.303997  net.cpp:] Memory required for data:
I0812 ::50.304002  layer_factory.hpp:] Creating layer fc8-
I0812 ::50.304008  net.cpp:] Creating Layer fc8-
I0812 ::50.304013  net.cpp:] fc8- <- fc7
I0812 ::50.304019  net.cpp:] fc8- -> fc8-
I0812 ::50.312885  net.cpp:] Setting up fc8-
I0812 ::50.312916  net.cpp:] Top shape:   ()
I0812 ::50.312927  net.cpp:] Memory required for data:
I0812 ::50.312934  net.cpp:] fc8- does not need backward computation.
I0812 ::50.312940  net.cpp:] drop7 does not need backward computation.
I0812 ::50.312944  net.cpp:] relu7 does not need backward computation.
I0812 ::50.312949  net.cpp:] fc7 does not need backward computation.
I0812 ::50.312953  net.cpp:] drop6 does not need backward computation.
I0812 ::50.312958  net.cpp:] relu6 does not need backward computation.
I0812 ::50.312963  net.cpp:] fc6 does not need backward computation.
I0812 ::50.312968  net.cpp:] pool5 does not need backward computation.
I0812 ::50.312973  net.cpp:] relu5_3 does not need backward computation.
I0812 ::50.312978  net.cpp:] conv5_3 does not need backward computation.
I0812 ::50.312981  net.cpp:] relu5_2 does not need backward computation.
I0812 ::50.312986  net.cpp:] conv5_2 does not need backward computation.
I0812 ::50.312991  net.cpp:] relu5_1 does not need backward computation.
I0812 ::50.312995  net.cpp:] conv5_1 does not need backward computation.
I0812 ::50.313000  net.cpp:] pool4 does not need backward computation.
I0812 ::50.313005  net.cpp:] relu4_3 does not need backward computation.
I0812 ::50.313009  net.cpp:] conv4_3 does not need backward computation.
I0812 ::50.313014  net.cpp:] relu4_2 does not need backward computation.
I0812 ::50.313019  net.cpp:] conv4_2 does not need backward computation.
I0812 ::50.313024  net.cpp:] relu4_1 does not need backward computation.
I0812 ::50.313027  net.cpp:] conv4_1 does not need backward computation.
I0812 ::50.313032  net.cpp:] pool3 does not need backward computation.
I0812 ::50.313037  net.cpp:] relu3_3 does not need backward computation.
I0812 ::50.313041  net.cpp:] conv3_3 does not need backward computation.
I0812 ::50.313046  net.cpp:] relu3_2 does not need backward computation.
I0812 ::50.313050  net.cpp:] conv3_2 does not need backward computation.
I0812 ::50.313055  net.cpp:] relu3_1 does not need backward computation.
I0812 ::50.313060  net.cpp:] conv3_1 does not need backward computation.
I0812 ::50.313064  net.cpp:] pool2 does not need backward computation.
I0812 ::50.313069  net.cpp:] relu2_2 does not need backward computation.
I0812 ::50.313073  net.cpp:] conv2_2 does not need backward computation.
I0812 ::50.313078  net.cpp:] relu2_1 does not need backward computation.
I0812 ::50.313082  net.cpp:] conv2_1 does not need backward computation.
I0812 ::50.313087  net.cpp:] pool1 does not need backward computation.
I0812 ::50.313092  net.cpp:] relu1_2 does not need backward computation.
I0812 ::50.313097  net.cpp:] conv1_2 does not need backward computation.
I0812 ::50.313102  net.cpp:] relu1_1 does not need backward computation.
I0812 ::50.313107  net.cpp:] conv1_1 does not need backward computation.
I0812 ::50.313110  net.cpp:] input does not need backward computation.
I0812 ::50.313114  net.cpp:] This network produces output fc8-
I0812 ::50.313129  net.cpp:] Network initialization done.
[libprotobuf WARNING google/protobuf/io/coded_stream.cc:] Reading dangerously large protocol message.  If the message turns out to be larger than  bytes, parsing will be halted for security reasons.  To increase the limit (or to disable these warnings), see CodedInputStream::SetTotalBytesLimit() in google/protobuf/io/coded_stream.h.
[libprotobuf WARNING google/protobuf/io/coded_stream.cc:] The total number of bytes read was
I0812 ::51.490310  net.cpp:] Ignoring source layer data
I0812 ::51.597520  net.cpp:] Ignoring source layer loss
OrderedDict([('conv1_1', <caffe._caffe.BlobVec object at 0x13cafcf30>), ('conv1_2', <caffe._caffe.BlobVec object at 0x13cafcfa0>), ('conv2_1', <caffe._caffe.BlobVec object at 0x158bf0050>), ('conv2_2', <caffe._caffe.BlobVec object at 0x158bf00c0>), ('conv3_1', <caffe._caffe.BlobVec object at 0x158bf0130>), ('conv3_2', <caffe._caffe.BlobVec object at 0x158bf01a0>), ('conv3_3', <caffe._caffe.BlobVec object at 0x158bf0210>), ('conv4_1', <caffe._caffe.BlobVec object at 0x158bf0280>), ('conv4_2', <caffe._caffe.BlobVec object at 0x158bf02f0>), ('conv4_3', <caffe._caffe.BlobVec object at 0x158bf0360>), ('conv5_1', <caffe._caffe.BlobVec object at 0x158bf03d0>), ('conv5_2', <caffe._caffe.BlobVec object at 0x158bf0440>), ('conv5_3', <caffe._caffe.BlobVec object at 0x158bf04b0>), ('fc6', <caffe._caffe.BlobVec object at 0x158bf0520>), ('fc7', <caffe._caffe.BlobVec object at 0x158bf0590>), ('fc8-101', <caffe._caffe.BlobVec object at 0x158bf0600>)])

然后就成功生成了age_train.pth模型

5.测试

使用保存好的pytorch模型来测试图片的年龄：

from PIL import Image
from torchvision import transforms as T
import torch
from models import AgeModel

def expected_age(outputs):
    age_output = []

    for batch, output in enumerate(outputs):
        age =
        for i, j in enumerate(output):
            age += i*j
        age_output.append(age.item())

    return torch.FloatTensor(age_output)

def test(img_path, model_path):
    transfrom = T.Compose([
        T.Resize(),
        T.CenterCrop(),
        T.ToTensor(),
        T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ])
    img = Image.open(img_path)
    img = transfrom(img).unsqueeze()

    model = AgeModel()
    model.load_state_dict(torch.load(model_path))
    age = model(img)
    print(age)
    print(age.size())
    age = expected_age(age)
    print(age)
    print(age.size())

if __name__ == '__main__':
    img_path = './Tom_Hanks_54745.png'
    model_path = './age_train.pth'
    test(img_path, model_path)

返回：

(deeplearning2) userdeMacBook-Pro:face_data user$ python check_model.py
tensor([[0.0032, 0.0039, 0.0027, 0.0020, 0.0026, 0.0031, 0.0029, 0.0035, 0.0051,
         0.0046, 0.0056, 0.0069, 0.0073, 0.0065, 0.0071, 0.0071, 0.0090, 0.0143,
         0.0225, 0.0269, 0.0288, 0.0373, 0.0335, 0.0343, 0.0282, 0.0303, 0.0256,
         0.0263, 0.0227, 0.0210, 0.0210, 0.0176, 0.0140, 0.0144, 0.0135, 0.0121,
         0.0107, 0.0112, 0.0110, 0.0088, 0.0109, 0.0092, 0.0086, 0.0140, 0.0096,
         0.0106, 0.0123, 0.0128, 0.0139, 0.0132, 0.0110, 0.0111, 0.0144, 0.0127,
         0.0110, 0.0153, 0.0103, 0.0113, 0.0117, 0.0136, 0.0147, 0.0124, 0.0096,
         0.0091, 0.0091, 0.0097, 0.0070, 0.0072, 0.0068, 0.0077, 0.0067, 0.0069,
         0.0062, 0.0065, 0.0056, 0.0047, 0.0045, 0.0045, 0.0048, 0.0049, 0.0040,
         0.0031, 0.0031, 0.0026, 0.0033, 0.0024, 0.0023, 0.0020, 0.0020, 0.0021,
         0.0020, 0.0016, 0.0021, 0.0016, 0.0016, 0.0014, 0.0014, 0.0013, 0.0015,
         0.0013, 0.0020]], grad_fn=<SoftmaxBackward>)
(, )
tensor([39.2897])
(,)

该图片为：

中间有出现一个错误：

RuntimeError: The size of tensor a () must match the size of tensor b () at non-singleton dimension 

然后查看输入的图像，发现其为RGBA格式的：

<PIL.PngImagePlugin.PngImageFile image mode=RGBA size=524x473 at 0x10138FFD0>

所以要将其更改为RGB格式：

img = img.convert("RGB")