Build an AI Artist - Machine Learning for Hackers #5

[@prang1702]

I cannot stress how much I am learning in these 5min. Thought I have to slow down the video to understand but this clears this over all high level concept of the problem !! Make more of these !!!! Waiting :D

[@SirajRaval]

Thanks Pranav, I will slow down a little bit next time, will do!

[@meisterhubert]

This is the best youtube-channel i have ever found. How did you not enslave the world with your skills yet?

[@SirajRaval]

thanks! im trying to enable you guys to do that! lol

[@TheCypherwars]

How does this video not get a GAZILLION views is beyond me! The more I think about the idea of machine learning the more it fascinates me. Man you are onto something here. keep up the good work. People may have a problem with the PACE of the video here. Personally I think it is satisfactory and absolutely fine, You shouldn't have to slow down for anybody when you are in the MODE! :)

[@raji08xd68]

At 3:25 the image used for demonstration for pooling layer is Max Pool : )

[@kcimc]

These videos are incredibly entertaining, well written & edited, but sometimes technically incorrect. In this one I noticed two things: 1. tensors are not meant to reduce computational complexity 2. content vs style is not about high vs low layers, but exact weight reconstruction vs gram matrix reconstruction (though I can see how the diagram in the paper would be confusing at first, cause style at higher layers sometimes captures "content-like" features).

[@SirajRaval]

Thanks for the comment Kyle! This is a great, really readable paper on tensors that I like www.cs.cornell.edu/cv/TenWork/FinalReport.pdf . You are right that tensors are good for a lot of different things, but to me the most intuitive pro for a beginner is that it has comparatively reduced computational complexity to very high dimensional data structures. There is indeed more to content vs style than just layers, but at a high level that is exactly what google mentioned was the case for DeepDream. It's gradient ascent process that tries to maximize the L2 norm of activations of a particular DNN layer, the higher the layer the more content based the art

[@_JS96]

That smile at the beginning, love it.

[@SirajRaval]

;) ;) ;)

[@wisentgenus]

Your videos are very high quality. People be sleeping on your content

[@SirajRaval]

Thanks a lot, I'll keep releasing them

[@ekaterinavalinakova2643]

I really look forward to the future of Neural Network Art. Btw you just got another subscriber :)

[@SirajRaval]

Hell yeah, awesome! Thanks!!

[@a19905]

another one

[@Keroec]

I love your artistic style of neural nets, Siraj! Except few things to comment about hyper-parameters details, it's so much fun that I'm looking forward to the next one! Let's make enough videos so that I can train my own model on your videos and get an embedding of Siraj' style ;-)

[@SirajRaval]

Thanks Sebastian :) That's a good idea, will be fun to make

[@Frankthegravelrider]

amazing video as usual old sport. keep it up, learning so much. maybe a short video on ML, neural net basics, to get the audience on board from the basics.

[@SirajRaval]

Thanks! Yeah I might just do a whole other ML series after this based just on theory/concepts

[@AndrewMelnychuk0seen]

Wow thank you so much for doing this. This is a huge interest of mine. Please please keep on making more of these videos.

[@SirajRaval]

Thanks Andrew! I will keep going, thanks for inspiring me

[@AndrewMelnychuk0seen]

Do you have a public Bitcoin address I can donate too?

[@SirajRaval]

Wow, thanks a lot! I do, it's 1KpdvNhbvVoxaSY4hncC1j9jCr3kY2bUcr

[@decreer4567]

Although it's cool that you could do that. The thing is there isn't really a need to automate that ,since art is simply expressing your emotions. The whole point of automation is to allow people to go pursue and do their hobbies ,while a emotionless machine does the dirty work for them.

[@if_it_pleases_the_council]

Awesome work, don't give up.

[@SirajRaval]

Thanks, much love

[@AntonyEvansGP]

This is an awesome video, well done!

[@SirajRaval]

Thanks Antony!

[@AndrewPrice2704]

Awesome! Keep it up! A lot of value packed into a short time :)

[@SirajRaval]

Thanks a lot, will do!

[@MarcJSchmidt]

FYI: 3:25 visualisation at the right shows a MaxPooling2D layer, not a AveragePooling2D as used in the code. Cool video tho, as usual :)

[@SirajRaval]

Thanks for the correction!

[@Mirandorl]

What is the reasoning behind the doubling of layer depth for each block? As in at 3:57 it goes 64, 128, 256? What advantage does this have over say, making them all the same?

[@DrPepperment]

love the videos, keep em coming

[@SirajRaval]

You got it :)

[@honotiro]

Hey Siraj, great video! I was wondering would you be able to make a video that would help to understand some ML fundamentals such as what is a neurons or neural network, etc. for newbies like me?

[@SirajRaval]

Thanks! Sure thing here is one kzbin.info/www/bejne/nmTPZaSwbJt9nrM

[@PhilippUnterreiner]

So right now the AI is learning and applying a human-created style of painting to any given content. But is there a way to make the AI create an original style / or to create original content / or both? This would be the necessary next step to make those AIs truly creative.

[@SirajRaval]

Yes there is a way. Think about how we paint things that are novel. We use everything we've ever felt and learned and make a novel connections between data points in our head. How do replicate this programmatically? We're not sure yet, but its def possible because we are proof!!

[@BlooFlame]

HOW DOES THIS GUY NOT HAVE MORE VIEW!?!?!?!?!

[@neil7493]

Dude this is awesome. You got a new subscriber!

[@SirajRaval]

hell yeah, thanks!

[@smarch.1]

Awesome material. Thank you.

[@SirajRaval]

Thanks for watching!

[@anupamyedida5484]

Awesome Video bro ! I love seeing your videos :)

[@SirajRaval]

Thanks man, I'll keep em coming :)

[@TheDreamFx]

Thank you very much for this and all of your videos! :)

[@shiyamsundar6565]

love your videos !! your videos help me a lot !! please keep em coming !!

[@SirajRaval]

I'm so glad they help! Thanks! I will keep em coming

[@lkd982]

How do you manage to make so many great memes? That is the true value proposition ;)

[@sakaress]

Love your work. Keep making! :)

[@SirajRaval]

Thanks a lot!

[@deepakranganathan9644]

Exception: Layer weight shape (3, 3, 512, 64) not compatible with provided weight shape (64, 3, 3, 3) I am getting this error. Please tell me where the problem is. Is it with the image files or the VGGfile or keras library.?

[@Throwingness]

Same problem. Leaving this comment in order to be notified of a comment here w/ a solution.

[@SirajRaval]

try this repo instead github.com/anishathalye/neural-style

[@Throwingness]

I got your repo and this new one to work by doing everything in a conda environment. I think I was getting this error cause the dependancies and TF were installed and uninstalled too many times. Also python2.7 and 3.5 could have been confusing things for the algorithm. Maybe it's own little space keeps things clean and obvious for the ones and zeros.

[@deepakranganathan9644]

Did you get the new repo correctly or the old repo which was using keras and theano ?

[@Throwingness]

I got it working by using Virtualvenv and up to date Xcode & command line tools

[@DrC0re]

Good vids man! subscribed! Could you also add some sidetrack video's about "simple" stuff like how to actually make the program read a data set of images for example. or another example from a previous video about the AI for playing games, do you just give it a "location\spaceinvaders.exe" arg to boot up the game? and how do you automatically restart the game every time it fails? this basic stuff gets me stuck and because it's apparently so simple for experienced coders it's also not documented very well in my opinion.

[@SirajRaval]

Thanks! Yeah I have so many videos to make on the details, I'm going hard and fast on the high-level stuff right now but i'm moving towards more detailed things soon. The program boots up the game and it restarts itself whenever it fails.

[@ka6595]

Hey Siraj, could you please make a video on google's syntax net and use cases for that

[@SirajRaval]

dude that's a great idea. i will actually do that. Thanks for the suggestion!

[@ka6595]

Sirajology 👍

[@sundar17617]

great work siraj .

[@SirajRaval]

Thanks Sundar :)

[@squeeesh]

What does "epoch" mean in the context used on your github page? "For a 400x400 gram matrix, each epoch takes approximately 11-13 seconds..." Thank you!!

[@SirajRaval]

great link here stats.stackexchange.com/questions/117919/what-are-the-differences-between-epoch-batch-and-minibatch

[@ConnorWidtfeldt]

I don't believe you have any errors. You are a programming wizard...

[@SirajRaval]

Thank you!!

[@lorikmucolli2496]

not really a wizard, just uses other people's code and makes a video about it, all the repos have a referral link at the end which leads to the original content, give credit to the original creators in the video

[@ED-TwoZeroNine]

Nice video, thanks. How long have you been programming for?

[@SirajRaval]

Thanks! Hmm. 11 years on and off.

[@LoboBRT]

Hey raj, whats the name of that video using deepdream? Btw, love your videos, keep the good vibe :D

[@SirajRaval]

the app is called artisto

[@LoboBRT]

Thks man, keep the good work, you are awesome!!!

[@joncamo]

Which method can you propose me, to learn quicker and understand the fundamental, i am getting started in python programmation

[@nynom]

Thanks man!! your videos are awesome keep up the good work

[@SirajRaval]

thanks so much :)

[@nynom]

Sirajology can you please do a video on training a speech recognition system on tensorflow

[@avinashvenkatesh4791]

Awesome Video! Subscribed!

[@SirajRaval]

woot thank you!!

[@tautologicalnickname]

Hi great video! Loved the music at the very beginning, can I have the name of the song please? peace!

[@SirajRaval]

Thanks! Power, by Kanye West. But specifically the power clip from his movie Runaway

[@tautologicalnickname]

thanks

[@hamburgmalte]

sorry for the following question! I tried to run it on my machine. unfortunately without success. Quite a newbie not using python so far. is there a possibility that u could add the dependencies to some kind of installation file? i have this problem when running Network.py "Traceback (most recent call last): File "Network.py", line 1, in from scipy.misc import imread, imresize, imsave ImportError: cannot import name imread" even thoug scipy seems to be installed. Sorry again for bothering. And thanks for the great video

[@SirajRaval]

no worries. run 'sudo pip install Pillow==2.6.0 '

[@hamburgmalte]

great, thanks!! - ill try it out at the weekend, look forward to it. thanks again!

[@kaiwang2924]

0:05: Does it looks like what it looks like?

[@andreas9109]

Could this also be done with Tensorflow?

[@graczmisiek4131]

probably, but it's good to keep competition alive before Google will enslave us all with their ex-Boston Dynamics' robots running on scripts using Tensorflow

[@jannoszczyk5132]

no competition here, Keras is just a Tensorflow interface

[@Throwingness]

I love the word concatenated!

[@SirajRaval]

isn't it really nice? i use it outside of CS too

[@deepakranganathan9644]

Traceback (most recent call last): File "Network.py", line 237, in load_weights(weights_path, model) File "Network.py", line 63, in load_weights model.layers[k].set_weights(weights) File "/home/sabretooth/anaconda2/lib/python2.7/site-packages/keras/engine/topology.py", line 889, in set_weights 'provided weight shape ' + str(w.shape)) Exception: Layer weight shape (3, 3, 512, 64) not compatible with provided weight shape (64, 3, 3, 3) I am facing this problem. Please give me any suggestions.. Thanks in advance.

[@Throwingness]

Same problem. Leaving this comment in order to be notified of a comment here w/ a solution.

[@mbaytas]

Try this: github.com/llSourcell/AI_Artist/pull/6/commits/36ca7dfe81e44561552d0820ff71f5eb91fde75d

[@deepakranganathan9644]

Thanks a lot Mehmet Aydin Baytas. Your link is running, but do you have any idea of how long the program might be stuck in iteration 1? Because its been 15 mins since i ran the program and it is still in iteration1 only

[@deepakranganathan9644]

The code on execution is stuck in iteration 1 only after 1 hour also and it takes 100% of CPU resource. Please Help. Im running it on a CPU only and not GPU

[@deepakranganathan9644]

The code on execution is stuck in iteration 1 only after 1 hour also and it takes 100% of CPU resource. Please Help. Im running it on a CPU only and not GPU

[@AviPars]

I love your videos. how did you learn AI

[@SirajRaval]

thanks! lots of internet sources

[@SaikatBasak]

How do I determine how accurately does this setup separates content from style? If I am to develop something similar is there a way to quantify the accuracy of the setup?

[@SirajRaval]

Hmm thats a great question. Because 'content' and 'style' exist as neural abstractions in our network, it's not as intuitive to quantify them. For now it's mostly done by looking it over with the naked eye, but more research is needed to figure out how to quantify each

[@SaikatBasak]

+Sirajology thanks for the reply and keep making videos. Love your channel.

[@jamesbiswas5350]

Thank you for sharing your knowledge with noobs

[@SirajRaval]

yo anytime thanks for watching!

[@JaredScheib]

Love it! Please create an AI to compute a new facial hair style for each week's video :)

[@SirajRaval]

Hey Jared :) Thanks man, that's the plan

[@ayy_tee]

Dope. I love your vids

[@SirajRaval]

Thanks Andy!

[@deepakranganathan9644]

The code on execution is stuck in iteration 1 only after 1 hour also and it takes 100% of CPU resource. Please Help. Im running it on a CPU only and not GPU.

[@SirajRaval]

try this repo github.com/anishathalye/neural-style

[@GaetanAlmela]

Hey Siraj, I'm getting an error saying the tuple index is out of range what should I do?

[@deepakranganathan9644]

Hi, Im getting attribute error in Zeropadding2D layer. It says that there is no function called set_input for ZeroPadding2D layer. Please Help!

[@SirajRaval]

try this repo instead github.com/anishathalye/neural-style

[@deepakranganathan9644]

Sirajology thanks .. there were some library issues .. its resolved now !!

[@adityasarvaiya9339]

bro your content is intimidating but we can barely follow your tutorials. it's kind of very fast. I suggest you must do long video with quite more explanation about how code is actually working. you have the best knowledge about machine learning hence we expect a lot from you in upcoming future. thanks.

[@zakarkak]

What does the preprocess_image function do or where can I find an implementation for it? I am unable to find this function

[@eshanpandey2824]

Man!! You are awesome!!

[@26medias]

Cool video! You have a new subscriber!

[@SirajRaval]

Thanks!!

[@PranavReddytv]

Holy Shit! This is awesome.

[@SirajRaval]

You're awesome and thanks

[@kalanimurakami1101]

Is there a easy way to run this? Confused on how to run this with virtualenv or on the desktop. Please help lol

[@SirajRaval]

try this blog.gruffdavies.com/2015/11/29/make-your-own-ai-art-with-deep-neural-networks/

[@joemiria2850]

Real badass dude

[@SirajRaval]

Thanks Joe :)

[@balajisundarrajan9684]

How do I start on ML ? Siraj , please help me out .

[@SirajRaval]

Follow along my 'Build a Neural Net in 4 Minutes' tutorial video

Men you are the best

[@SirajRaval]

Thanks Brian :)

[@aneeshbabu9183]

great video :)

[@bitcode_]

Do you use AI to answer every comment?

[@SirajRaval]

i wish i could sometimes but nope. its all me. i usually take an hour to answer every single comment on monday. don't know how long i can keep that up as the channel grows but i'm super committed to it. :)

[@bitcode_]

+Sirajology lol cool stuff thou, thanks for the videos, the topics are useful and interesting

[@MrKmake]

Where i can find an vgg file?

[@MrKmake]

Compatible with the training model layers?

[@SirajRaval]

thanks for pointing that out, i added it to the description, heres the link as well drive.google.com/file/d/0Bz7KyqmuGsilT0J5dmRCM0ROVHc/view?usp=sharing

[@pramit.sawant]

why don't you stick with a single library... You have a different python library in every video..

[@gammaian]

Amazing videos, I love this series! Anyone tried to execute this program in windows? I have some problems getting the arguments right :/

[@SirajRaval]

Thanks Gian!

[@Skythedragon]

the building at 0:40 has a very er, exceptional shape...

[@SirajRaval]

lol yup

[@gabemcguire2463]

So... Prisma?

[@SirajRaval]

this video was released before prisma. i sometimes wonder if the authors of prisma were inspired by this vid. i hope so. #dreams

[@gabemcguire2463]

Sirajology Either way, this is really cool. Keep up the good work!

[@evanperrygiblin]

deepdream was just the tip of the icebergin ML, and that google is probably hoarding a ton of more useful code.

[@prajaktatamse7010]

Please do help with stepwise explanations, these videos are good but if you do explain in detail will be more helpful for the beginners, thanks for such informative videos.

[@samzumcommenten]

I know that 5 minute videos are most popular on youtube, but please ... can barely follow. They are only usable for an entertaining showcase to have an overview, but they don´t function as tutorials.

[@SirajRaval]

will get better thx

[@knightwalkr]

Siraj Raval maybe go back and do a 30-40 minute "class" on your more popular short videos.

[@prtysh]

i agree

[@terraset3988]

This is definitely going to depress artists who spent most of their lives honing a skill that is now soon to be worthless because it can be learned and then reproduced in a fraction of the time by a machine. However at the same time it can free those artists from the relatively boring task of actually drawing and instead focus on the idea they wanted to get on paper in the first place instead

[@SirajRaval]

Terraset, great point. It can help augment artistic capability, democratize the field so more people can become artists.

[@terraset3988]

Art is already as democratized as it can possibly be. All it takes is time and effort. There is a movement in the art community to discourage young artists from buying into the scam that is university art degrees because the people who go into them either come out no better off or waste their money being told by a prejudiced instructor that their preferred medium or style isn't "real." An aspiring artists improvement will never come from anything other than their own hard work. An AI won't change that because an AI at the moment can only copy what it is fed. It would still be up to an artist to develop their own skills and style first so they could feed that into the machine to reproduce images in the exact way they want them to look. And even then it likely won't be perfect. I do however expect a fresh new wave of wannabes posting machine generated "art" on the internet claiming that it's theirs and acting like they're super skilled for it.

[@Zeegoner]

+Terraset Art is not as democratized as it can possibly be. If that were true, we would be able to think and have our art appear instantaneously, without any arbitrary skill acting as a middleman.

[@terraset3988]

That has nothing to do with democratization of art. there will always be a skill involved even if that skill is merely an indepth understanding of what it is you are rendering. If you told a computer to generate an image of Paris and then proudly showed it to me I would not call you an artist for it. You didn't do any of the work so why should you get any of the credit? At best what you'd be holding is a commisioned piece from a computer. A computer will never be able to do more than copy because just like people it can only work with the knowledge it has. It can't see what you see and thus can't render exactly what you would render. Thus it can never be your art, only its. The democratization argument only works if you mean something like "easier access to nice looking things."

[@omaraziz95]

It's an interesting video but you really talk too fast

[@SirajRaval]

good point thanks

[@baiget94]

youtube has an option to see the video in a slower speed