I want to build a machine to classify numbers in the MNIST data set to extend my AI skills and maybe have something cool in my portfolio. Anyone here done this? Any recommendations for how to get started? My basic idea was just to start sketching out a linear classifier in python, but I'm noob and not sure if that's a good language for this kind of project.
MNIST Database
Bentley Russell
Other urls found in this thread:
github.com
neuralnetworksanddeeplearning.com
twitter.com
Kevin Ward
everyone has done this
it just takes a couple of days of fucking around with tensorflow
Ryder Martin
I would use pytorch so you don't have to handroll the gradients
Thomas Robinson
yeah, I'm just getting started and I wanted to go for something that everyone has already done first so that I have plenty of resources to fall back on when I get stuck. I'll try more interesting stuff next.
Thanks for the link, saved.
Cooper Jones
Imagine the smell
Matthew Garcia
I'm trying bro.
Colton Carter
Who is this Yoda of yogurt-slinging?
Kayden Cox
god i wish i were that truck
Carter Hughes
I want to FUCK that truck
Jordan Baker
ikr? Get that thot out of the way and just start banging
Justin Campbell
how to get a qt redneck gf?
Kayden White
This is basically 'Hello World'. Go for it, but don't have unrealistic expectations for what it means for your portfolio.
Brayden Ward
I did it in javascript
Juan Harris
Star with your sister.
Hudson Bennett
MNIST is baby tier. It's trivial to do much more complicated classification tasks using pre trained models/transfer learning.
If you're interested in digit recognition and want to do something cooler, you could train a model to actually read handwritten text. Pretty easy to do either with an LSTM based model or a fully convolutional model. Look up Connectionist Temporal Classification for how you'd do the reading part. You could also pretty easily make a Flask app and host it on AWS. People upload images of handwritten text and the model transcribes it. That's a way more interesting project than MNIST classification.
If you're interested in classification stuff, find a more exciting dataset. If you're limited on compute, you should look into the Oxford pets dataset or Food-101. Good smaller datasets.
It's also not much of a stretch to go from CNN classification to object detection.
>Any recommendations for how to get started?
If you're unsure of where to start in terms of learning, you should look into Andrew Ng's deep learning course on coursera (Tensorflow) or fast.ai (pytorch).
If you just want to jump into classification, I'm pretty sure that's lesson 1 of fast.ai.
Sebastian White
Asher Lewis
there are literally 1 hours lecture on fast.ai to get 99% accuracy on MNIST
any brainlet can do this within
Mason Gonzalez
maybe you niggers can relax with the "baby tier" faggotry? op wants to do a project, why the fuck you gotta act so smug
Nicholas Reed
f*et...
Gabriel Morales
Hot.
Nathan Howard
Using a premade library isn't the same as implementing it by hand.
Levi Powell
I want to marry her feet.
Lincoln Ortiz
Everyone and their mother has done MNIST OP. Aim a little bit higher.
Aaron Ross
really? Try writing a conv net by yourself to achieve 99% accuracy.
You can't, there's 20+ years of research put into this. All you can do from scratch is write a shitty conv net, try solving the gradient after 3 layers. If you want to "implement by hand" you first need to write a gradient calculating engine which will take you months or years and will still be shitty.
Carter Roberts
>painted nails
into the trash
Nicholas Barnes
>couple of days
>MNIST
lolwut
Luke Fisher
pls no bully, i have the stupids