sergeyk

{
 "metadata": {
  "name": ""
 },
 "nbformat": 3,
 "nbformat_minor": 0,
 "worksheets": [
  {
   "cells": [
    {

sergeyk

sergeyk

"""
Code to deal with the ImageNet dataset.

Feature details (computed by Yangqing):
    - VLfeat SIFT dense extraction: with SIFT patch size 16, 32 and 64,
        and a stride of 4 pixels. The images are reduced to size 500*500
        (smaller images are not resized up though).
    - LLC-coded features with 5-nearest neighbors, and a codebook of size 16k
    - Max-pooled on a 1x1 grid and 3x3 grid, so a total of 10 bins.
        This gives us 10 * 16k = 160k dimensional features.

sergeyk

sergeyk

{
 "metadata": {
  "name": ""
 },
 "nbformat": 3,
 "nbformat_minor": 0,
 "worksheets": [
  {
   "cells": [
    {

sergeyk

{
 "metadata": {
  "name": ""
 },
 "nbformat": 3,
 "nbformat_minor": 0,
 "worksheets": [
  {
   "cells": [
    {

sergeyk

sergeyk

sergeyk

# assumes there is a dictionary that has keys
# 'g': the nx.DiGraph graph
# 'nodes': list of node names
# 'heights': list of integer heights of the nodes

fig = plt.figure()
ax = fig.add_subplot(111)

g = graph['g']
nodes = graph['nodes']

sergeyk