NetReform

NetReform is a tensorflow wrapper for knowledge transfer. It is based on two papers; Net2Net and Network Morphism.

Dependencies

• tensorflow
• slim (modified ver.)
• numpy
• scipy

Usage

• See example.py

  from net_reform import NetReform
  # model, weights from a previous model
  nr = NetReform(model, weights, new_graph)
  out_graph, out_session = nr.reform_rand() # NetReform with random values
  out_graph, out_session = nr.reform() # NetReform with values derived from net2net or netmorph func.
  ...
  obj = out_graph.get_collection('objective')[0]
  out_session.run(obj)

Notes

• Currently, NetReform only supports Net2Net operators (wider & deeper).