/caffe-utils

Scripts to create lmdb for multilabel task with caffe

Primary LanguagePython

caffe-utils

This repo lists some scripts that are useful for caffe framework.The scripts in multilabel_lmdb_create directory is adapted from sukritshankar, which can be used to create lmdb for multilabel task with caffe.

multilabel_lmdb_create

The steps to use the scripts in this directory:

  • Provide data and corresponding label files: Make a train.txt and a test.txt which list the images and corresponding labels (dummy labels,which will never be used) for training and test, respectively.Then make a labels_train.mat and a labels_test.mat which list the corresponding label vectors for training and test, respectively. Each label vector is like [0 1 1 0...] with "0" means having this class in this image and "1" otherwise. Remember to shuffle training and test data when providing train.txt and test.txt since they will not be shuffled during creating lmdbs. Be sure the label vectors correspond to the images!!
  • Create data lmdbs and mean file: Use create_data_lmdb.sh to create lmdbs for training and test data, and create_mean.sh to compute mean value.
  • Create label lmdbs: Use create_label_lmdb.py to create lmdbs for training and test labels. Run this script with python 2.7.
  • Generate the prototxt: The prototxt file is used for training, refer to the example train_test_multilabel.prototxt for details.