/extract_deep_feature

an executable file to extract deep learning feature for Windows

#Windows exe of caffe to extract deep learning features

You can download the exe and related dlls here: caffe-extract-deep-feat.
The code is here: extract_feat.cpp, feat_extrator.hpp.

Want to build the exe by yourself?

You can find the windows version of caffe project at Caffe Windows.

####Prerequisites

  • Windows-64 bit
  • Please download the image net DCNN model file (e.g. AlexNet, CaffeNet, VGG, GoogleNet) and corresponding deployed net definition file and imagenet_mean file (if needed) from online resources, then put them into the same folder with the exe (in current folder, we take AlexNet as an example) .

####Usage Please remember to keep all the files or sub folders in the same folder with the exe. You have 3 choices to use this exe, first please open the cmd window and input one of the following commands:

  • ExtractDeepFeat.exe [flags] imagePath layame
  • ExtractDeepFeat.exe [flags] indexFilePath layerName
  • ExtractDeepFeat.exe [flags] imageFolderPath layerName

Flags including:

  • -net_file: the net definition file. The net must include an input layer named as "data" with a type "Input". For the input layer: net input shapes should be set in input_param; reshape parameters should be set in image_data_param; transformation parameters should be set in transform_param.
  • -model_file: the corresponding trained caffemodel file.
  • -feat_file: the file path to store features of all the images. if not set, features will be saved independently into "xxx.jpg.feat".
  • -mode: "GPU" or "CPU", default as CPU. For GPU mode, cuda toolkit 7.5 is needed.
  • -l2_norm: true or false, if l2 normalization of feature is needed.
  • -sqrt: true or false, if square root of feature is needed.

You can run ExtractDeepFeat.exe in cmd and see more usage details.

####Input

  • imagePath: the absolute path of your image.
  • indexFilePath: the index file of your images, which should be in txt format (with ".txt" extension" and one image path per line.
  • imageFolderPath: absolute path of your image folder, it will scan all the images in this folder and its sub folders.
  • layerName: the layer name of the net the model that you want to extract feature of.

####Output

  • The format of the feature data is float and binary.
  • If feat_file not set, the deep learning feature of every image will be saved into a binary file with the same path as the image and named by "the image name" + ".feat" .

####Sample command ExtractDeepFeat.exe E:\myimages fc7 -model_file=bvlc_alexnet.caffemodel -net_file=image_val.prototxt ExtractDeepFeat.exe E:\imageIndex.txt prob -mode=GPU ExtractDeepFeat.exe E:\test.jpg pool5