CNN Training code

Training and testing requires the following files that can be edited as required

• train_test.prototxt
  • Input path of 'training' and 'testing' LMDB files
  • Network architecture - Same as LeNet. Changes to last layer for number of classes
• solver.prototxt
  • Parameters for the training algorithm
• deploy.prototxt
  • Network architecture for feature extraction process

Need to run following commands in sequence:

• To generate LMDB from Training images

  [PATH_TO_CAFFE]/build/tools/convert_imageset --backend=lmdb --resize_height=50 --resize_width=100 --shuffle ./ list_tmp.txt [TRAIN_LMDB_OUTPUT_PATH]

  [PATH_TO_CAFFE]/build/tools/convert_imageset --backend=lmdb --resize_height=50 --resize_width=100 --shuffle ./ list_tmp.txt [TEST_LMDB_OUTPUT_PATH]

• To begin training

  • Edit train_test.prototxt, solver.prototxt, and deploy.prototxt appropriately

  • Run the following command to begin training

    [PATH_TO_CAFFE]/build/tools/caffe train -solver solver.prototxt

  • Use nohup to run the training in background

    nohup [PATH_TO_CAFFE]/build/tools/caffe train -solver solver.prototxt > train.log &

• Upload following files to Android mobile

  • weights.caffemodel
  • deploy.prototxt
  • classes.txt (All class names in correct sequence)

Sample trained files

'android_hiwiki' directory contains a sample of trained files that can be copied to android mobile to work with Shaili app

Contact

Contact developer at [email protected]