Skip to content

krebso/neuwuronka

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

47 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Martin Krebs, 514407

FASHION MNIST FROM SCRATCH

  • solved using templates and recursive types
  • allows for optimalizations on compilation, everything is laid out in memory
  • not using paralellism atm, does not give performance benefits for task
  • the network is trained as fully connected MLP, with SGD using momentum and decay

NOTES FROM WORKING ON THE PROJECT

  • if it does not work on XOR, it will not work on mnist holds
  • std::array is on stack
  • not violating 1st rule helps, thanks Andrej
  • backprop is quite easy to comprehend, not that easy to implement

NEXT STEPS

  • i think i learned quite a lot during the implementation process, i tried to implement CNN as well, but I failed miserably
  • i would like to finish implementing it
  • same with dropout, and recurrent network
  • add proper paralellism, may work for larger networks
  • the whole solution lacks quite a modularity
    • in the beginning, I was inspired by torch modules, and I had separate module for Linear layer, ReLU and Softmax
    • turned out that the network was much slower (because of the double number of layers), so i ditched the original idea
    • i would like to make it somehow work, but not sure how to avoid the slow down

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published