lcskrishna / caffe-net-upgrade

Upgrade deprecated caffemodel to latest caffemodel (V1LayerParameters to LayerParameters) without installing caffe

caffe-net-upgrade

caffe-net-upgrade is a tool which takes a caffemodel or a prototxt file as an input that is defined with V1LayerParameters. (Caffe deprecated definition) . This implementation upgrades to the layer parameters.

Note: This tool doesn't require installation of caffe to upgrade your model.

Pre-requisites

1. Ubuntu 16.04
2. CMake 2.8 or newer download
3. Install the Protobuf library.

Build Instructions

1. Install the respective pre-requisites if not present. Make sure to install libprotobuf-dev and protobuf-compiler
2. Navigate to the caffe-net-upgrade folder.
3. mkdir build
4. cd build
5. cmake ../
6. make

This automatically builds a executable in the build folder.

Usage:

    % upgrade_layer_parameters <net.caffemodel | net.prototxt> [output_file_prefix]

Example 1 : Upgrade VGG caffemodel

Download the VGG-Net 19 Layer caffemodel from the following site which is trained with V1LayerParameters here

Execute the following command :

    % upgrade_layer_parameters VGG.caffemodel output

This upgrades the caffemodel and writes into a new caffemodel named output.caffemodel.

Example 2 : Upgrade VGG prototxt

Download the VGG-Net 19 Layers prototxt defenition from the following site defined with V1LayerParameters here

Execute the following command:

    % upgrade_layer_parameters vgg.prototxt output

This upgrades the prototxt defenition and writes into a new file named output.prototxt

caffe-net-upgrade's Issues

input_shape is not upgraded to an input Layer

Input Prototxt:

name: "TestConv"
input: "data"
input_shape {
  dim: 1
  dim: 4
  dim: 224
  dim: 224
}
layer {
  name: "conv1/7x7_s2"
  type: "Convolution"
  bottom: "data"
  top: "conv1/7x7_s2"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 64
    bias_term: false
    pad: 3
    kernel_size: 7
    stride: 2
    weight_filler {
      type: "xavier"
      std: 0.1
    }
  }
}

Expected (From the official Caffe upgrade tool):

name: "TestConv"
layer {
  name: "input"
  type: "Input"
  top: "data"
  input_param {
    shape {
      dim: 1
      dim: 4
      dim: 224
      dim: 224
    }
  }
}
layer {
  name: "conv1/7x7_s2"
  type: "Convolution"
  bottom: "data"
  top: "conv1/7x7_s2"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 64
    bias_term: false
    pad: 3
    kernel_size: 7
    stride: 2
    weight_filler {
      type: "xavier"
      std: 0.1
    }
  }
}

Observed:

Loading prototxt file ... 
INFO: Reading the prototxt file from : /home/ian/test_conv.prototxt
INFO: prototxt read successful 
INFO: Network loaded is : TestConv
INFO: This network doesn't require upgrade

upgrade prototxt