Searching for tutorials and software about Deep Learning and Neural Nets? Be sure to look at my Resource Page!
Looking for Octave? Go to my Easy Octave on Mac page!

Monday, February 23, 2015

Working the Pete Warden Tutorial

Whereas last century's nets neural were shallow, and often just 3-layered, the current "deep" architectures are hierarchical. In the case of machine vision and image recognition much of the feature extraction is done by 2-D convolution modules.

Pete Warden has created a Vagrant machine which contains a canned version of the Caffe deep learning framework started by Yangqing Jia and the Berkeley Vision and Learning team, bundled  with some example data and pretrained nets; Pete's tutorial is really easy to work through.

 I got an image of a classic french car from wikimedia.

 I then copied the image file to  /vagrant/deux-chevaux.jpg in the VM, and submitted it to a network that’s based on the architecture used by Krizhevsky et al to win the Imagenet 2012 contest:

vagrant@vagrant-ubuntu-trusty-64:~/caffe$ python python/ --print_results /vagrant/deux-chevaux.jpg foo
Classifying 1 inputs.
Done in 1.30 s.
[('pickup', '0.27585'), ('convertible', '0.24703'), ('beach wagon', '0.11597'), ('grille', '0.07221'), ('minivan', '0.06870')]

The net has clearly generalized the concept of a car, and in some sense "recognizes" this very unusual model.

 Now, that was a satisfying first experiment! Thank you Pete!



  1. Is there another link to download the vagrant file for the tutorial? The link in the article is dead.

    1. You could ask Pete. I will try and do my own tutorial with Keras.


Hey, let me know what you think of my blog, and what material I should add!