10 Replies - 261 Views - Last Post: 17 September 2019 - 04:33 AM Rate Topic: -----

#1 Leonetienne   User is offline

  • New D.I.C Head

Reputation: 1
  • View blog
  • Posts: 6
  • Joined: 07-September 19

I really want go get into deep learning data science, but how?

Posted 16 September 2019 - 04:50 PM

I am deeply (get it) fascinated by deep learning and would absolutely LOVE to learn working with it.
I have already done a little bit with google deep dream and know how basic feed-forward nns work.
I know C++ and C#. I have already established that python is somehow the language of choice for deep learning.

But i am running into a BIG problem. I am a windows guy and i do no not know jack sh't about linux or how to use it.
Also i have an RTX 2070 Super. An RTX Card. Sadly almost no NN-Frameworks support CUDA 10.x....
(I know, the newer Tensorflow versions do, but those are rarely supported, as far as i know)

So, whenever i want to use some framework, i run into tons of compatiblity issues...
HyperGAN won't work at all because it depends on CUDA 9 (which won't work with my GPU) and won't work with windows without modification.
I install all the dependencies and there are still tons of "import not found" errors.
(Even when i try to train on CPU)

Then again, some other frameworks (in this case carpedm20) just throws random syntax errors (not linking errors) when just doing the example net with the example dataset from the readme.md on his github.

The only things i've done so far:
Working with basic feed-forward nns (that only worked because i ported the code to c++)
Working with google deepdream (Kinda not counts because it is pre-trained)
Working with a text-lstm, however that was on my old gpu (a maxwell gpu, hence not CUDA 10)
Working with DeepfaceLab, but it was a massive headache to get it to work. And it only does because it was a "standalone version". Meaning tensorflow, numpy, etc all was in subfolders already configured correctly. Also this actually supported CUDA 10.

Would it be worth to set up a linux distro to 1) get familiar with UNIX and 2) finally being able to work with NNs or would it just not work to begin with because of my GPU?

Best regards

Is This A Good Question/Topic? 1
  • +

Replies To: I really want go get into deep learning data science, but how?

#2 modi123_1   User is online

  • Suitor #2
  • member icon



Reputation: 15357
  • View blog
  • Posts: 61,571
  • Joined: 12-June 08

Re: I really want go get into deep learning data science, but how?

Posted 16 September 2019 - 05:22 PM

You can easily create a VM box with Linux in it.

VirtualBox comes to mind.

A docker container could work too.
Was This Post Helpful? 0
  • +
  • -

#3 Leonetienne   User is offline

  • New D.I.C Head

Reputation: 1
  • View blog
  • Posts: 6
  • Joined: 07-September 19

Re: I really want go get into deep learning data science, but how?

Posted 16 September 2019 - 05:59 PM

View Postmodi123_1, on 16 September 2019 - 05:22 PM, said:

You can easily create a VM box with Linux in it.

VirtualBox comes to mind.

A docker container could work too.

Then i couldn't train on my gpu
Was This Post Helpful? 0
  • +
  • -

#4 Choscura   User is offline

  • D.I.C Lover


Reputation: 478
  • View blog
  • Posts: 2,267
  • Joined: 18-October 08

Re: I really want go get into deep learning data science, but how?

Posted 16 September 2019 - 06:09 PM

Probably the cheapest/easiest option is to set up a raspberry pi or similar stand-alone machine. you can flash all the SD cards you need to from windows just fine so you can keep your normal operating system, and something like this can be hooked into your TV or an extra old monitor somewhere with normal USB mouse and keyboards and whatever else you need, but the advantage of something like this is that you can run just about any open source machine learning project on it, albeit at pi speeds <you'd be amazed how often this is much better than you need it to be>, including things like amazon alexa instances, tensorflow image processing, and so on.
Was This Post Helpful? 1
  • +
  • -

#5 modi123_1   User is online

  • Suitor #2
  • member icon



Reputation: 15357
  • View blog
  • Posts: 61,571
  • Joined: 12-June 08

Re: I really want go get into deep learning data science, but how?

Posted 16 September 2019 - 06:17 PM

Yeah.. but you can pick up the concepts of linux and if you want install it on the drive and dual boot. Or snag a NVIDIA® Jetson Nano.
https://developer.nv...o-developer-kit


or there's cuDNN.
https://developer.nvidia.com/cudnn


or work on TensorFlow.. plenty of info on that with cuda 10. It's not as bad as you made it out.
https://towardsdatas...10-c373786e36d1
Was This Post Helpful? 1
  • +
  • -

#6 Leonetienne   User is offline

  • New D.I.C Head

Reputation: 1
  • View blog
  • Posts: 6
  • Joined: 07-September 19

Re: I really want go get into deep learning data science, but how?

Posted 16 September 2019 - 06:24 PM

View PostChoscura, on 16 September 2019 - 06:09 PM, said:

Probably the cheapest/easiest option is to set up a raspberry pi or similar stand-alone machine. you can flash all the SD cards you need to from windows just fine so you can keep your normal operating system, and something like this can be hooked into your TV or an extra old monitor somewhere with normal USB mouse and keyboards and whatever else you need, but the advantage of something like this is that you can run just about any open source machine learning project on it, albeit at pi speeds <you'd be amazed how often this is much better than you need it to be>, including things like amazon alexa instances, tensorflow image processing, and so on.

sure, i'll just wait three month for it to train a hundred epochs haha
Was This Post Helpful? 0
  • +
  • -

#7 Leonetienne   User is offline

  • New D.I.C Head

Reputation: 1
  • View blog
  • Posts: 6
  • Joined: 07-September 19

Re: I really want go get into deep learning data science, but how?

Posted 16 September 2019 - 06:29 PM

View Postmodi123_1, on 16 September 2019 - 06:17 PM, said:

Yeah.. but you can pick up the concepts of linux and if you want install it on the drive and dual boot. Or snag a NVIDIAŽ Jetson Nano.
https://developer.nv...o-developer-kit


or there's cuDNN.
https://developer.nvidia.com/cudnn


or work on TensorFlow.. plenty of info on that with cuda 10. It's not as bad as you made it out.
https://towardsdatas...10-c373786e36d1

Of course they would work (at least, i hope so lol), but i don't think those are good for beginners. I think it's a bit complicated for absolute NN beginners.
I barely know how basic feed-forward nns work and i can't even figure out the API frontends from widely used frameworks. And now i should make my own?

Do you guys know any good tutorial series regarding tensorflow on windows? (C#, C++, Python, i dont care)
Was This Post Helpful? 0
  • +
  • -

#8 modi123_1   User is online

  • Suitor #2
  • member icon



Reputation: 15357
  • View blog
  • Posts: 61,571
  • Joined: 12-June 08

Re: I really want go get into deep learning data science, but how?

Posted 16 September 2019 - 06:37 PM

Regarding what with Tensorflow? Setup? Yeah, a cursory search does that.

Example:
https://www.tensorfl.../source_windows
https://towardsdatas...10-60693e46e781

Though I would suggest a book.
Example:
https://www.amazon.c...w/dp/1491962291
https://www.amazon.c...k/dp/199957950X
https://www.amazon.c...n/dp/1549617214
Was This Post Helpful? 0
  • +
  • -

#9 Leonetienne   User is offline

  • New D.I.C Head

Reputation: 1
  • View blog
  • Posts: 6
  • Joined: 07-September 19

Re: I really want go get into deep learning data science, but how?

Posted 16 September 2019 - 06:45 PM

View PostLeonetienne, on 16 September 2019 - 06:29 PM, said:

View Postmodi123_1, on 16 September 2019 - 06:17 PM, said:

Yeah.. but you can pick up the concepts of linux and if you want install it on the drive and dual boot. Or snag a NVIDIAŽ Jetson Nano.
https://developer.nv...o-developer-kit


or there's cuDNN.
https://developer.nvidia.com/cudnn


or work on TensorFlow.. plenty of info on that with cuda 10. It's not as bad as you made it out.
https://towardsdatas...10-c373786e36d1

Of course they would work (at least, i hope so lol), but i don't think those are good for beginners. I think it's a bit complicated for absolute NN beginners.
I barely know how basic feed-forward nns work and i can't even figure out the API frontends from widely used frameworks. And now i should make my own?

Do you guys know any good tutorial series regarding tensorflow on windows? (C#, C++, Python, i dont care)


All i'm saying is, if computer science youtubers (that are doing stuff like voice imitation networks) are not working with TF directly but are always just using frameworks for TF, it can't be easy.

I think of TensorFlow kind of like i think of DirectX when working with unity. It runs in the background controlled by a framework, because its darn complicated. What do you think?

View Postmodi123_1, on 16 September 2019 - 06:37 PM, said:


Yeah, those are installation guides.
But i meant like literal tutiorials.
A working tensor flow enviroment is of no use for me if i do not know anything about the tensor flow library.
Kind of like someone does not know how to code just by installing visual studio.

Maybe something like a 10 page long step by step guide on how to create the most basic neural net in tensor flow and how to create a dataset for it. With detail explaining what each data structure represents etc

Someone must have done a guide like this.
Maybe even in video form
Was This Post Helpful? 0
  • +
  • -

#10 modi123_1   User is online

  • Suitor #2
  • member icon



Reputation: 15357
  • View blog
  • Posts: 61,571
  • Joined: 12-June 08

Re: I really want go get into deep learning data science, but how?

Posted 16 September 2019 - 06:45 PM

Cool beans.

Good luck with your endeavors.
Was This Post Helpful? 0
  • +
  • -

#11 Choscura   User is offline

  • D.I.C Lover


Reputation: 478
  • View blog
  • Posts: 2,267
  • Joined: 18-October 08

Re: I really want go get into deep learning data science, but how?

Posted 17 September 2019 - 04:33 AM

View PostLeonetienne, on 16 September 2019 - 06:24 PM, said:

sure, i'll just wait three month for it to train a hundred epochs haha


Don't give it that training task, unless you're putting them into an aegis cluster. But since these are among the cheapest computers out there, you can afford to make one of those for this kind of thing.

But, these aren't arduinos to be laughed off. They boot in 15 seconds to a desktop GUI environment, and you can be running <admittedly the pi version of> minecraft, at 60 frames a second, 15 seconds after that- and that's with the last generation, not the current one.

There are real issues to solve with making these general-use PC's, but the advantage of something like this is that it's so cheap and easy to try things you'd be afraid to try on a more expensive PC- so I have one on a motorized turret, and another built into a desk drawer <so you pull out the drawer and see mouse and keyboard, and nothing else but the screen on the desk>, and things like poison taps, and the GPIO stuff really does give you interesting options- such as making your Alexa bot be hooked into Billy the Big Mouth Bass.

About the "everybody uses an API so that library MUST be difficult", you're maybe missing the point that "everybody was able to make their own API to do the stuff they wanted to with that library". I can't comment too deeply on tensorflow, I've only looked superficially, but trying stuff will make you more familiar with the actual problems it has- not less. Also it doesn't actually take months to train stuff on a pi- my formal introduction to pi's was in the cannabis business, where the pi3's were being used as growth-tracking cameras from the ceiling. now, most of the processing happened on a dedicated server rack, but that was the initial training stuff- which can be farmed back out for matches to things like pi's without adding to the server load on an ongoing basis.

This post has been edited by Choscura: 17 September 2019 - 04:36 AM

Was This Post Helpful? 0
  • +
  • -

Page 1 of 1