/caffe-gui-tool

Node based Gui for creating caffe networks

Primary LanguagePythonThe UnlicenseUnlicense

Caffe Gui Tool

Node based Gui for creating caffe networks

Master branch requires manual setup, although no compilation, and prebuilt is ready to go, but ~100mb compressed

ConvNet

alt tag

Autoencoder

alt tag

Installation

Apt-get Binaries (~5kb download, and quicker apt-get download)

  • sudo add-apt-repository ppa:irie/blender && sudo apt-get update
  • sudo apt-get install blender, and launch blender with 'blender' command
  • download two python scripts
  • file --> user preferences --> addons --> install from file --> select Caffe Nodes script --> install
  • tick the checkbox in the newly appeared box's top right corner
  • file --> user preferences --> addons --> install from file --> select Caffe Generate script --> install
  • tick the checkbox in the newly appeared box's top right corner
  • user preferences --> file --> uncheck relative paths (IMPORTANT)
  • click on the cube icon in the bottom left
  • click 'node editor' in the popup menu
  • in the newly appeared node editor, look at the bottom panel, it should read 'view - select - add - node'
  • immediately right of this panel there should be a set of four icons joined together. select the THIRD.
  • file --> save startup file
  • file --> user preferences --> save user settings (on the bottom of the window(
  • restart blender

Use

  • Click the 'New' button on the bottom part of the screen to create a new network
  • Name your network in this box, and ENSURE THE 'F' BUTTON NEXT TO THE NAME IS CHECKED
  • Use shift+a to start adding nodes
  • Join up, and fill in all the required fields to your taste. All networks require a data node, and solver node.
  • When ready, select all nodes by single, (or double if required) tapping 'a', then press spacebar.
  • type 'solution' in the search bar
  • select 'create solution' and press enter
  • This will create the required prototxt files in the config directory.
  • Mark the train file as executable
  • run the train file

Limitations

  • Negative gradient on ReLu not yet supported
  • ReLu not calculated in place
  • Many nodes not yet supported