How to read hdf5 with python layer in caffe and do data augmentation?

Question

I want to read data from hdf5 file and do data augmentation.

I can do data augmentation in transform_param with mirror and crop.

However,I cannot have transformation_param in a HDF5Data layer - caffe does not support this.

I consider it might be a good way to write a python layer to solve it. Read data from hdf5 file and do data augmentation at the same time.

I have read some relevant answers such as " caffe data layer example step by step " and " What is a Python layer in caffe? "

I still don't know some details such as where should I put the written file "*.py" ? And how to write the codes to do data augmentation?

Answer 1

You python layer *.py should be in your $PYTHONPATH .

Consider examples of python layers in $CAFFE_ROOT/examples/pycaffe/layers ,for custom python loss and data layer.

Consider particularly in $CAFFE_ROOT/examples/pycaffe/layers/pascal_multilabel_datalayers.py

Before next batch gets loaded, you can do the data augmentation on the fly.

Then in *.prototxt

layer {
  name: "data"
  type: "Python"
  top: "data"
  top: "label"
  python_param {
    module: "pascalcontext_layers"
    layer: "PASCALContextSegDataLayer"
    param_str: "{\'context_dir\': \'../../data/pascal-context\', \'seed\': 1337, \'split\': \'train\', \'voc_dir\': \'../../data/pascal\'}"
  }
}

module is the file name and layer is the class name.