c++ - 为什么 HDF5 Output Caffe 层会写入看似不正确维度的数据？

Question

我有兴趣将给定 Caffe 层的输出写入文件。我想为多个图像执行此操作，因此我对 HDF5 输出层代码进行了一些修改，以便为每个图像创建一个文件，其中包含每个图像的特征。这是修改后的 SaveBlobs 函数:

template <typename Dtype> void HDF5OutputLayer<Dtype>::SaveBlobs() {                               
  LOG(INFO) << "Saving HDF5 file " << file_name_ << "ds: " << ds_iter_;  
  CHECK_EQ(data_blob_.num(), label_blob_.num()) <<                       
      "data blob and label blob must have the same batch size";          

  // Open hdf5 file to write this blob                                   
  file_name_ = this->layer_param_.hdf5_output_param().file_name();          
  ostringstream appender;                                                
  appender << "_" << ds_iter_ << ".h5";                                  
  file_name_.append(appender.str());                                     
  file_id_ = H5Fcreate(file_name_.c_str(), H5F_ACC_TRUNC, H5P_DEFAULT,   
                       H5P_DEFAULT);                                     
  CHECK_GE(file_id_, 0) << "Failed to open HDF5 file" << file_name_;        

  // Write the data and label                                            
  hdf5_save_nd_dataset(file_id_, HDF5_DATA_DATASET_NAME, data_blob_);       
  hdf5_save_nd_dataset(file_id_, HDF5_DATA_LABEL_NAME, label_blob_);        
  LOG(INFO) << "Successfully saved " << data_blob_.num() << " rows";        
  LOG(INFO) << "SAVEBLOB - Data size is: " << data_blob_.shape_string(); 
  LOG(INFO) << "SAVEBLOB - Label size is: " << label_blob_.shape_string();

  // Close the file                                                      
  herr_t status = H5Fclose(file_id_);                                    
  CHECK_GE(status, 0) << "Failed to close HDF5 file " << file_name_;        

  // Update iterator for next image                                      
  ds_iter_++;                                                            
}   

代码几乎可以很好地工作，因为我能够为每个实际上包含数据的图像成功创建文件。不幸的是，似乎写入了错误的数据，因为日志中显示的维度和生成的输出文件中的维度都不正确。这是我指定输出层的地方(在网络原型(prototype)文本中):

layer {                                                                  
  name: "conv5_3"                                                        
  type: "Convolution"                                                    
  bottom: "conv5_2"                                                      
  top: "conv5_3"                                                         
  param {                                                                
    lr_mult: 1                                                           
    decay_mult: 1                                                        
  }                                                                      
  param {                                                                
    lr_mult: 2                                                           
    decay_mult: 0                                                        
  }                                                                      
  convolution_param {                                                    
    num_output: 512                                                      
    pad: 1                                                               
    kernel_size: 3                                                       
  }                                                                      
}                                                                        
layer {                                                                  
  name: "relu5_3"                                                        
  type: "ReLU"                                                           
  bottom: "conv5_3"                                                      
  top: "conv5_3"                                                         
}                                                                        

#===== Data Logging =======                                              

layer {                                                                  
   type: "HDF5Output"                                                    
   name: "hdf5output"                                                    
   bottom: "conv5_3"   #                                                 
   bottom: "conv5_3"   #                                                 
   hdf5_output_param {                                                   
     # File name is only a base                                          
     file_name: "./test_features/image"                                  
   }                                                                     
 }   

我认为保存的数据不正确的原因是，当我观察网络​​设置时，conv5_3 顶部的维度如下所示:

I0206 23:07:44.815330  7630 layer_factory.hpp:77] Creating layer conv5_3_relu5_3_0_split
I0206 23:07:44.815343  7630 net.cpp:106] Creating Layer conv5_3_relu5_3_0_split
I0206 23:07:44.815348  7630 net.cpp:454] conv5_3_relu5_3_0_split <- conv5_3
I0206 23:07:44.815356  7630 net.cpp:411] conv5_3_relu5_3_0_split -> conv5_3_relu5_3_0_split_0
I0206 23:07:44.815366  7630 net.cpp:411] conv5_3_relu5_3_0_split -> conv5_3_relu5_3_0_split_1
I0206 23:07:44.815372  7630 net.cpp:411] conv5_3_relu5_3_0_split -> conv5_3_relu5_3_0_split_2
I0206 23:07:44.815382  7630 net.cpp:411] conv5_3_relu5_3_0_split -> conv5_3_relu5_3_0_split_3
I0206 23:07:44.815459  7630 net.cpp:150] Setting up conv5_3_relu5_3_0_split
I0206 23:07:44.815467  7630 net.cpp:157] Top shape: 1 512 14 14 (100352)
I0206 23:07:44.815474  7630 net.cpp:157] Top shape: 1 512 14 14 (100352)
I0206 23:07:44.815479  7630 net.cpp:157] Top shape: 1 512 14 14 (100352)
I0206 23:07:44.815484  7630 net.cpp:157] Top shape: 1 512 14 14 (100352)
I0206 23:07:44.815495  7630 net.cpp:165] Memory required for data: 116006912
I0206 23:07:44.815500  7630 layer_factory.hpp:77] Creating layer hdf5output
I0206 23:07:44.815511  7630 net.cpp:106] Creating Layer hdf5output
I0206 23:07:44.815515  7630 net.cpp:454] hdf5output <- conv5_3_relu5_3_0_split_0
I0206 23:07:44.815521  7630 net.cpp:454] hdf5output <- conv5_3_relu5_3_0_split_1
I0206 23:07:44.815527  7630 net.cpp:150] Setting up hdf5output
I0206 23:07:44.815531  7630 net.cpp:165] Memory required for data: 116006912

太棒了，我希望得到维度为 1 512 14 14 的数据。不幸的是，当我对模型运行推理时，我在日志中看到错误的维度出现了:

I0206 23:07:46.108660  7630 hdf5_output_layer.cpp:31] Saving HDF5 file ds: 0
I0206 23:07:46.115536  7630 hdf5_output_layer.cpp:48] Successfully saved 1 rows
I0206 23:07:46.115557  7630 hdf5_output_layer.cpp:49] SAVEBLOB - Data size is: 1 512 54 38 (1050624)
I0206 23:07:46.115566  7630 hdf5_output_layer.cpp:50] SAVEBLOB - Label size is: 1 512 54 38 (1050624)
I0206 23:07:46.316557  7630 hdf5_output_layer.cpp:31] Saving HDF5 file ./test_features/image_0.h5ds: 1
I0206 23:07:46.322437  7630 hdf5_output_layer.cpp:48] Successfully saved 1 rows
I0206 23:07:46.322456  7630 hdf5_output_layer.cpp:49] SAVEBLOB - Data size is: 1 512 56 38 (1089536)
I0206 23:07:46.322463  7630 hdf5_output_layer.cpp:50] SAVEBLOB - Label size is: 1 512 56 38 (1089536)
I0206 23:07:46.457828  7630 hdf5_output_layer.cpp:31] Saving HDF5 file ./test_features/image_1.h5ds: 2
I0206 23:07:46.463618  7630 hdf5_output_layer.cpp:48] Successfully saved 1 rows
I0206 23:07:46.463636  7630 hdf5_output_layer.cpp:49] SAVEBLOB - Data size is: 1 512 38 50 (972800)
I0206 23:07:46.463644  7630 hdf5_output_layer.cpp:50] SAVEBLOB - Label size is: 1 512 38 50 (972800)
I0206 23:07:46.594746  7630 hdf5_output_layer.cpp:31] Saving HDF5 file ./test_features/image_2.h5ds: 3

这表明不仅输出的尺寸不正确，而且它们在迭代(图像)之间也有所不同!日志中显示的维度与写入 h5 文件的数据维度相匹配，因此日志准确地描述了代码的行为。我的问题是为什么会这样？看起来我已经正确设置了所有内容，但一定有一些我遗漏的东西......

Answer 1

正如@hbaderts 帮助我发现的那样，事实证明 HDF5 层是正确的并且输出正确维度的数据。我对维度的混淆是由于在网络原型(prototype)文本的测试版本中输入维度的看似静态的定义。事实证明，因为我使用 pycaffe 函数 net.forward(**forward_kwargs) 加载数据，所以卷积层会自行缩放以满足输入图像的不同输入尺寸。这解释了大于预期的特征尺寸以及它们在图像之间存在差异的事实。