These are chat archives for FreeCodeCamp/DataScience

2nd
Nov 2016
evaristoc
@evaristoc
Nov 02 2016 08:21
@erictleung we are in the same page!
evaristoc
@evaristoc
Nov 02 2016 08:34

People

Meetups today: Hackathon for deep learning

Last Monday I attended a meetup that was a continuation of a previous introduction in deep learning. This time we had to solve the following problem using neural networks (recurrent or convolutional or both): decoding a QR code.
QR codes are meant to be read by machines. Considering there are already many apps used for decoding QR codes, the problem looks trivial at once but don't be fooled. The problem is VERY hard for a neural network. Just to give you an idea of how to decode it using human brain, check this video. The algorithms for decoding are non-linear, as this code might shows.

This is the meetup information:
https://www.meetup.com/The-Amsterdam-Pipeline-Factory-of-Data-Science/events/234828464/

There was no much progress in general but in my team we didn't even reach the final (I had a very bad laptop for this and I couldn't actively participate, one of the team mates have to leave earlier than expected and my another team mate was having issues solving Docker). Actually the team didn't go along very well that night at the end.

So I decided to make a plan for a subproblem on my own. Let you know, still working on it without pressure.

The problem proved to be so hard for people new to ML/DL that the organiser decided to try another section in the future. Keep you updated.

evaristoc
@evaristoc
Nov 02 2016 08:46
@erictleung I am currently using Keras with Theano as backend. CPU based though. My plan is to start using GPU at some point this year, although I will have to use a cloud connection. No plans to buy a GPU for now. I will certainly use Nvidia with CUDA as it is the most popular.
Amelia
@apottr
Nov 02 2016 08:54
@evaristoc look into OpenCL, it's supported on many CPUs and GPUs
so if you had an intel CPU you could write code that works there and switch the device to the Nvidia GPU and it would work pretty much the same if not better
evaristoc
@evaristoc
Nov 02 2016 08:55
@apottr I have heard people like it more, specially because open source. It is also more universal. But what about training? And Nvidia seems to be ahead, don't you think?
Amelia
@apottr
Nov 02 2016 08:56
I don't know enough about one vs the other, i'd suggest looking into it
I've been using OpenCL however and it's been satisfying my needs
evaristoc
@evaristoc
Nov 02 2016 08:58
@apottr Fair enough. Let me check, it should be a driver from Keras/Theano to use OpenCL. Do you have some recommended tutorials by hand? And what sort of problems have you be solving in GPU?
Amelia
@apottr
Nov 02 2016 08:59
For the moment it's been mostly just a learning experience but I was learning so as to accelerate image processing speeds
evaristoc
@evaristoc
Nov 02 2016 09:05
@apottr Ok. Let's see: Although I definitively like OpenCL more, frankly speaking most of the material I have read for DL/ML goes around Nvidia. This is certainly a strong marketing strategy to drive adoption into Nvidia GPU's, but on the other hand it ends up being more convenient.
@apottr Thanks for the info!
CamperBot
@camperbot
Nov 02 2016 09:06
evaristoc sends brownie points to @apottr :sparkles: :thumbsup: :sparkles:
:cookie: 634 | @apottr |http://www.freecodecamp.com/apottr
Zarak
@zarak
Nov 02 2016 09:07
@erictleung Thanks for the link to the deep learning papers!
CamperBot
@camperbot
Nov 02 2016 09:07
zarak sends brownie points to @erictleung :sparkles: :thumbsup: :sparkles:
:cookie: 422 | @erictleung |http://www.freecodecamp.com/erictleung
evaristoc
@evaristoc
Nov 02 2016 09:08
Yes forgot to thanks @erictleung and @mesmoiron for the links!
CamperBot
@camperbot
Nov 02 2016 09:08
evaristoc sends brownie points to @erictleung and @mesmoiron :sparkles: :thumbsup: :sparkles:
:cookie: 312 | @mesmoiron |http://www.freecodecamp.com/mesmoiron
:cookie: 423 | @erictleung |http://www.freecodecamp.com/erictleung
Zarak
@zarak
Nov 02 2016 09:17
I added it to my own ML-roadmap :)
which I guess would more accurately be titled "Math with a dash of ML-roadmap"