Tensorflow in Gentoo virtualenv

Just tried to install tensorflow in Gentoo virtual env with python 3.4 and failed

k-means and MNIST dataset

Just thought few days ago how should look avarage 0 or 1 or 9 or any number. So downloaded MNIST dataset and implemented k-means algorithm (I could calculate average vector in every cluster, but that wont be interesting).

Back to the Linux or how I debugged silent crash

Just got tired from Mac and decided to move to my old Gentoo Linux. After recompiling world (oh yeah, all from sources) I stuck with SIGSEGV with plot command in Octave (Matlab like math package).

When I tried to plot anything it crashed with SIGSEGV without any obvious reason. So I got to go deeper to fix it. Gentoo supports nice debug options all that one needs is just to enable them
https://wiki.gentoo.org/wiki/Debugging

After that simple
# sudo ulimit -c 99999999
# octave
octave:1> plot(1)

then crash with core in the current folder. Now one need to unwind the stack
# gdb /usr/lib/debug/usr/bin/octave-cli-3.8.2.debug core
(gdb) bt
#0  fl_create_gl_context (g=0x0) at Fl_Gl_Choice.H:102
#1  Fl_Gl_Window::make_current (this=0xc12430) at Fl_Gl_Window.cxx:168
#2  0x00007fc1a5706447 in plot_window::show_canvas (this=0xc11bb0,
    this=0xc11bb0) at dldfcn/__init_fltk__.cc:935
....
#88 0x00007fc1b715a620 in __libc_start_main () from /lib64/libc.so.6
#89 0x0000000000400979 in _start ()

So, I didnt recompile nvidia drivers or my opengl config is bad. I had problems with opengl config, repoint it with eselect opengl and run X11 again.

Problem has solved

Two good books about neuro networks

Long time no see :) short post about machine learning. I wanted to go deeper into this field and tried to find some good book, not like a "Machine Learning in 24 hours" but something grad level with acceptable science payload. So I've read ~30% of Simon Haykin "Neural Networks and Learning Machines (3rd Edition)". Book is good, but not enough naive examples to jump in, like 1 neuron with 2 inputs, but science payload there is nice. Gazzillion links to articles and books, so book is useful when you know what to do. Then I found nice complementary book with naive examples. Also it's really cheap on Amazon - about 25$, also it can be downloaded for free from the official site. "Neural Network Design (2nd Edition)" by Hagan, et al. So have fun.

Meanwhile, I started uber small project - neural networks without any special lib. I use only numpy for matrix calculation. Everything slow but primitive and self explanatory. As a data set I'm using MNIST data set, it has 60k training samples and 10k test samples.

Simple implementation of Widrow-Hoff perceptron is here, ~70% success rate after training
https://github.com/venik/simple_neuro_networks/tree/master/src/one_layer_mnist

And MNIST reader is here
https://github.com/venik/simple_neuro_networks/blob/master/lib/mnist/mnist.py

also you probably want to see how does MNIST samples look on a screen
https://github.com/venik/simple_neuro_networks/tree/master/utils/mnist_reader

How to download MNIST is here
https://github.com/venik/simple_neuro_networks/tree/master/data_set/mnist