GoogleNet can't read images when ImageData type field is used in train_val.prototxt

Question

I am trying o use caffe's implementation of GoogleNet. I want to train the deep network according to a list of files and labels in a text file, but the problem is that, when I train the deep network, it can't read the files.

Here is the train_val.prototxt definitions, where I use ImageData instead of using big LMDB files with 'Data' type

name: "GoogleNet"
layer 
  {
  name: "data"
  type: "ImageData"
  top: "data"
  top: "label"
  include {
  phase: TRAIN
 }
 transform_param 
 {
 mirror: true
 crop_size: 224
 mean_value: 104
 mean_value: 117
 mean_value: 123
 }
 data_param 
 {
  source: "path_to_file/file_paths_and_labels.txt"
  batch_size: 32
 }
}

Here I used ImageData type for the googlenet rather than type Data as suggested here: LMDB files and how they are used for caffe deep learning network

So, I have the text file (file_paths_and_labels.txt) where each line contain the following:

path_to_image label

where path to image is the image's address and label is the label of the image (there are 10 different labels).

I want to know exactly where I am wrong because when I run the deep network training command

./build/tools/caffe train --solver=/my_home/dl-caffe/models/bvlc_googlenet/solver.prototxt

I have the following error:

I0624 10:36:11.524896 15246 layer_factory.hpp:74] Creating layer data
I0624 10:36:11.524960 15246 net.cpp:84] Creating Layer data
I0624 10:36:11.524988 15246 net.cpp:338] data -> data
I0624 10:36:11.525046 15246 net.cpp:338] data -> label
I0624 10:36:11.525084 15246 net.cpp:113] Setting up data
I0624 10:36:11.525106 15246 image_data_layer.cpp:36] Opening file 
I0624 10:36:11.525146 15246 image_data_layer.cpp:51] A total of 0 images.
*** Aborted at 1435152971 (unix time) try "date -d @1435152971" if you   are using GNU date ***
PC: @     0x7f7060b70ee0 (unknown)
*** SIGSEGV (@0x0) received by PID 15246 (TID 0x7f706188aa40) from PID   0; stack trace: ***
@     0x7f7060511d40 (unknown)
@     0x7f7060b70ee0 (unknown)
@     0x7f706118587c std::operator+<>()
@     0x7f70611861e5 caffe::ImageDataLayer<>::DataLayerSetUp()
@     0x7f7061144ac6 caffe::BaseDataLayer<>::LayerSetUp()
@     0x7f7061144bc9 caffe::BasePrefetchingDataLayer<>::LayerSetUp()
@     0x7f70611d8ff2 caffe::Net<>::Init()
@     0x7f70611daab2 caffe::Net<>::Net()
@     0x7f70611e6c10 caffe::Solver<>::InitTrainNet()
@     0x7f70611e7d23 caffe::Solver<>::Init()
@     0x7f70611e7ef6 caffe::Solver<>::Solver()
@           0x40c4a0 caffe::GetSolver<>()
@           0x406471 train()
@           0x404a11 main
@     0x7f70604fcec5 (unknown)
@           0x404fbd (unknown)
Segmentation fault (core dumped)

I think GoogleNet is not finding data in my textfile. What is the problem? the syntax of my train_val.prototxt file?

have you tried this with other nets (e.g., AlexNet)? is it possible that the "end-of-line" char in the text file is not compatible with your OS (that is, using windows line termination in unix system and vice versa)? — Shai

Or Sharir Or Sharir · Accepted Answer · 2015-06-30T14:37:15

You're specifying the source using the wrong parameter. For IMAGE_DATA you need to use image_data_param instead of data_param. Because you specify your source in data_param, and ImageDataLayer looks at image_data_param, the value of source is the empty string. You can see that in the log here:

I0624 10:36:11.525106 15246 image_data_layer.cpp:36] Opening file

The format of this line should be:

Opening file <filename>

while in your log there's an empty space following "Opening file".

GoogleNet can't read images when ImageData type field is used in train_val.prototxt

2 Answers