Caffe to TensorFlow

  • TensorFlowTransformer will build a Kaffe graph first, from the caffe prototxt.

    • It has its own node class and graph class

    • Node: layer type, specification, parameters, parent, children etc.

    • Graph: how to add node / build graph in topological order

  • Converting the parameters of all graph nodes into correct format ((c_o, c_i, h, w) -> (h, w, c_i, c_o))

    • In Numpy format

  • Write generated Kaffe.tensorflow.Network class code

  • When doing the inference, Kaffe.tensorflow.Network will converting the Kaffe nodes into tensorflow operators.

  • Problem:

    • Needs updated version caffe model

    • Needs Kaffe library

    • Numpy format parameters

    • Couldn’t deploy

    • Tensorflow version out-dated

  • Modified version:

    • Automatic caffe model updating script

    • Update tensorflow operators from 0.8.0 -> 1.2.1

    • Add saving checkpoint and frozen graph options, ez deployment

Reference:

PreviousThings To Be Considered When Doing Model ConvertingNextComputation Graph Optimization

Last updated