Donate some GPU time for science, please

Hello. I'm a researcher. One day I came to idea of the research. Now I'm trying to train a neural network model, but my hardware seem to be shit. I know that there are lot of folks with top GPUs (more precisely, I need NVidia ones, since there is no good OpenCL support in Theano) in here. I guess they may want to help me...

You install python 3.6, theano and the packages I point on. No backdoors, no malware, these are mostly well-known open-sourse project. I mean Theano and lasagne. Then I provide you with a matrix of data and a program which trains a net. The result would be useful for everyone who is going to buy some specific piece (widely used) of computer hardware and published in openaccess.

If you also know neural networks well all the advices would be appreciated - this is my first experience with ANN.

I ain't failing for this, skynet.

That'll be 1000€ kiitos

I'll do it if you pay me.

I'm not gonna pay you more than $1 for whole time. In fact published results likely will save more to everyone who is going to buy that piece of hardware.

Assume I'm a retard gamer, explain step by step like a retard how I can help you

Not OP but his program will likely train a Neural network and write the weights of the NN to a file. Which he can use it to take output out of. You'll need to install python 2.7,numpy, theano plus whatever libs the program requires. And run the program.

Yeah because you know that shit is going to crypto mining.

it depends on your GPU. Ram doesn't matter, 2 GiB must be enough.

1) Don't make your first NN algorithm so fucking huge that you can't run it. Start with

>python 2.7,numpy, theano plus whatever libs the program requires

3.6+ in fact. 2.7 is old shit.

Cont. Also, if you're training NN's in parallel, how will you combine the results?

I actually just tackled a similar problem, you may be better off computing partial derivatives for individual training examples in parallel, then sending the results to a single computer and combining and optimizing there..

Fuck you, pay me.

i'm not training in parallel, I'm only training on an old celeron now
it's only 47 nodes now. It was trained for an hour, than I have canc3lled the process. I dont; know why such a small network takes so long to train.

How many training examples are you training over. Have you plotted your cost function over iterations of optimization? Does it look like its steadily decreasing?

I'll do it if I get second author

11k. No. Have not tried. Do you mean I should try on a thousand first?

Hmm, shouldn't take THAT long, but to get an idea of if there's a problem with your algorithm, have it print the cost function of your weights every 10 or so iterations (I'm assuming you're using some kind of gradient descent algorithm for optimization). I've never used Theano but I just mostly finished a NN class for Numpy and it takes about 24 hours to train over 50k examples, with about 4k nodes in the network.

Use EC2 spot instances you cheap fuck. Deep learning is not a poor student's game. Google routinely push out papers that require ~100 V100s to reproduce

Why don't you use your schools computers?

because I'm not a schoolchild

>Researcher
>0 funds
Rly makes you think

I'm not going to mine buttcoins for you.

Actual university research means you either have some funding for buying hardware (in this case your entire research will be a proof of concept so you will buy 1, one, uno, um, computer and use it for said research) or be allowed to used the machines in your research department.

You're going to install homosexual porn on my GPU, I know your game.

He means why don't you use the computer resources that has been allocated to your institution, you fucking turbo-autismo.

oy vey, nice diagram you got ther goy

>Actual university research

It's a standalone research. My personal. Because I personally need it. Because I'm going to buy some hardware for my pc and I'm not going to trash my money by for shit. So it is not funded by the institution I work for. Also ML is not it's specialization.

fucking shit. 500 seconds for 10 samples for 10 epochs. Fuck that theano, tomorrow I'll start xgboosting.

Well give us a link user, hiw are we supposed to help you if you don't show us where to download your shit?

I implemented a neural network that played Othello in undergrad AI class, it did fine with no GPU. Nobody wants to help your botnet crack tripcodes!

shit nigga you can just use VLC for that tripfag

>So it is not funded by the institution I work for.
the whole point of being in academia is to be able to attract funding
otherwise, why bother

Wny should I do this instead of mining Zcash for $20 a day?

Just use BOINC

Out of curiosity, how many gpus would you need to get enough processing power for your research ? Are we talking just couple of 1080 class gpu or far more ?

>Hey guys, run this miner for me, thanks!
Nice try, (((user)))...

>this is my first experience with ANN
Then you are perfectly fine using only your CPU for a long while until you at least somewhat understand what you are doing. Learn the basics first before diving into stuff like deep learning. You don't need a fucking top-end GPU to train a basic neural network.

what coin do i get for science mining?
And how many bitcoin has seti@home stolen from witless xenophiles?

I'll bite. Give me some way to contact you. If I give some shit about what you are trying to research I'll see what I can do to help you.

How much GPUs do you need? I have 1080, what are you training the net for?

Smd faggot

11K RECORDS FOR TRAINING??????? LOOOOOOOOOL

Google doesn’t use Nvidia hardware you fucking IDIOT

Go to Floydhub, you have 20 free hours of CPU per month and 2 hours of GPU (Tesla K80)
Tbh you don't really need more than that. If you don't get satisfying results, there probably are other tweaks you can do to increase your accuracy, rather than training on more epochs. Anyway, good luck!

Plot training and test accuracy relative to number of epochs. Find the point where they're relatively the same and choose that as your number of epochs

yeah, so, you should pay for your own gear to do your bitcoin-mining research, pajeet