Thank you so much siraj for making this. I asked you this on your twitter handle, and I dont know if you made this based on my tweet, but want to believe you did and I am grateful. Thanks a tonnnn.
This is true advice Siraj. Keras is the best way to deep dive into deep learning. I recommend to read "Deep Learning with Python" book by the Author of Keras himself (françois chollet). Pretty much deep content with many real use cases.
I disagree. I know they used to be really bad especially regarding security but they did a 180 turn with Nadella and produce quite a few infesting pieces of technology. They also participate in open source development and contribute good code. They don't deserve this hate anymore.
For learning, I'd say go and implement model from scratch using just numpy. And for production, you can use any higher level library like keras, tensorflow, pytorch, mxnet............ If u use those high end frameworks as beginner then you'd probably gonna be unknown of what actually is going on behind the scene. That's just my suggestion, I have faced it.
@@SimonWoodburyForget u r comparing sorting algorithm with ml, dl????? R u out of your mind?????? People who are afraid of scratch implementation don't deserve ml and dl.
In the Machine Learning course of Andrew Ng you have to do it all from scratch. Then in the specialization you use raw Tensorflow and later on in the specialization you use Keras. Also... Many deep learning fathers agree that that is a good methodology, e. g. Ian Goodfellow and the very same Professor LeCunn.
I think your best bet is to use numpy (which just does matrix multiplication if you're really really beginning) to start with. Simple, but requires you understand the mechanics to make it work. I started learning to program in C, in much the same way, and starting from the bottom like that, I feel, has helped me greatly.
You are totally right but if you want to test different models quickly and want to make a quick project. I would recommend using Keras. The real fun, though comes from making a Neural Net yourself.
It's funny how we train machine learning models so differently than we train ourselves. We tend to think it's best to learn from sophisticated explanations at the base level and build our understanding up from that, but we underestimate our ability to innately infer the inner workings of a topic from numerous excellent examples. I learned programming starting with high level languages, without reading a text book, and worked my way down to assembly, then to CPU design using HDL, then back up. I like this because A. it's fun to get stuff done, then B. it's fun to learn how the stuff works under the hood once you know what it's capable of, then C. you understand the whole stack differently. It's a hard sell when we begin with "okay, let's learn what NAND is" when you're like "how is this going to help me make a video game" or whatever.
fantastic information for a person know nothing about deep learning! one feedback would be - information flowing too fast and I had to watch multiple times to understand better. may be it just me, otherwise I love it.
@Siraj Raval - Do you not do anymore live streaming tutorials? Also it would be very interesting to see you do a deep-learn learning tutorial using kaggle's dataset and actually competing on kaggles while live streaming.
Sir can you please please do a series on Pytorch. It is such a powerful and promising library and it would be a great addition to your channel and great learning material for us
Why did he say that thing about Microsoft? I didn't get that joke? Is this referring to the processor bugs that they only recently released security fixes for? I'm confused.
Keras is REALLY good but there is a negative point: if you need to use a custom loss function, it become a pain in the a**. I have been trying to implement YOLO loss function and I really don't know how to do it. Besides that, it is very cool!
Hi Siraj and everyone. F.Chollet has made publicly available some jupyter notebooks with his book examples. I have ported two of them to kaggle kernels so everyone can run and modify those notebooks in the cloud (just sign in kaggle, fork and play with the notebooks). Links to the jupyter notebooks (kaggle kernel): www.kaggle.com/juanjotwo/deep-learning-with-python-notebooks-2-1 www.kaggle.com/juanjotwo/deep-learning-with-python-notebooks-3-5-imdb The original F.Chollet 19 notebooks (run on on a p2.xlarge EC2 instance) github.com/fchollet/deep-learning-with-python-notebooks
I think you should start by making your own deep learning library before using something that is already made, to allow you to properly understand how deep learning works,* That's how I started, and I'm glad I did it
watch this course from the MIT, you'll learn a lot : th-cam.com/video/uXt8qF2Zzfo/w-d-xo.html ( I learned a lot ) then grab yourself a language (C++/python, whatever you want) and start experimenting, there is abit of math to do, the wikipedia page on backpropagation can help you you should also make a visual feedback, like a visual representation of your current network, this help when debugging... honestly I started without and I regret it If you want to start quickly, I suggest using processing, you can make simple windows and graphics without worriing too much about classic graphic stuff ( which can be annoying when you are trying to do something else, but have to spend a lot of time on something only usefull in debugging...) processing basically is a really simplified version of java, you can make stuff really quick with this, faster than python from what I've experienced ( in therms of developpement, not really about your program speed... ) Hope this helped :p
oh thanks for quick reply...sorry but I couldn't understand what you meant by processing. point me to an article instead of wasting your time on me, that discusses about this.
Re: Coding Challenge. I've already got one I'm working on. DDPG network that learns a grid world game. github.com/dustinandrews/machinelearning/tree/ddpg-link-gradient/DDPG , Some features: epsilon is dynamic, the better the agent is doing the less randomness in the actions. The simulation supports curriculum training where you can set how far away from the goal the start position can be. agent_play(ddpg) pops up a window and shows how the agent plays the game for one episode.
Thanks! Sorry the code is such a mess. Once I get it all working I'll clean it up and make a repo for it. My critic is learning well, but the actor is still struggling.
So which part of this is actually different from PyTorch?? I've worked with it a bit and it seems like this is fairly similar so wanna see if I'm missing something. Thanks :)
My Entry. Ours is a new efficient and intuitive algorithm that performs well in continuous action space. It's called Selective Memory. You can see a demo of it here th-cam.com/video/hKrFFeZqq3E/w-d-xo.html
Hi Siraj. I'm new to this, so I run the code in terminal and in the end like you showed in the beginning nothing will pop up to show a text and in terminal after the epoch 60 the texts still don't make any sense. Am I missing something?
Hi, May I ask anyone ever tried using Mac pro's Anaconda to run keras tutorial? either in jupyter or spyder, the keras always error when plot training history. Same setup, ubuntu 16.04 works fine, and virtual environment on Mac Jupyter notebook is okay.
I wrote it a few weeks ago, but here it is: github.com/Goldesel23/DCGAN-for-Bird-Generation I used keras for training a DCGAN for Bird Generation, I also trained a WGAN with the same architecture to compare the results
@alberjumper Alberto Blanco Garcés, you are the man! First the stok market prediction, and now there you are again. Good job.
Suddenly all videos I search in TH-cam start with "Hello World, it's Siraj". Thanks for the amazing content
Thank you so much siraj for making this. I asked you this on your twitter handle, and I dont know if you made this based on my tweet, but want to believe you did and I am grateful. Thanks a tonnnn.
Finally, Machine Learning related video. Thanks Siraj . @BerteADA
you got it, more coming
Sweets exactly what i just came to your channel for!
Amazing Video! So much of info in very short time! Thanks!
This is true advice Siraj. Keras is the best way to deep dive into deep learning. I recommend to read "Deep Learning with Python" book by the Author of Keras himself (françois chollet). Pretty much deep content with many real use cases.
Awesome Siraj ! Excellent explanation ! Keras is the best library to build deep nets.
Please do mention some thing about coming video.
Wonderful video, as always, thank you, Siraj.
Thanks Siraj for your nice videos. Good luck keep going.
Awesome video, and congrats people on comments, your counting is effed up
haha always
"I've gotta not use anything made by Microsoft!" smartest phrase I've heard this year.
I disagree. I know they used to be really bad especially regarding security but they did a 180 turn with Nadella and produce quite a few infesting pieces of technology. They also participate in open source development and contribute good code. They don't deserve this hate anymore.
Does Microsoft actually make anything? It was my understanding they stole everything...
@@nufosmatic You're thinking of Apple, lol.
Very good content! love it!
For learning, I'd say go and implement model from scratch using just numpy. And for production, you can use any higher level library like keras, tensorflow, pytorch, mxnet............
If u use those high end frameworks as beginner then you'd probably gonna be unknown of what actually is going on behind the scene. That's just my suggestion, I have faced it.
@@SimonWoodburyForget u r comparing sorting algorithm with ml, dl?????
R u out of your mind??????
People who are afraid of scratch implementation don't deserve ml and dl.
@@wolfisraging lol people deserves or not deserves haha ok guardian of dl
@Shubham Dhingra , unfortunately for you.... yes
@Shubham Dhingra but also definitely not necessarily, you can use tensorflow to calculate gradients
In the Machine Learning course of Andrew Ng you have to do it all from scratch. Then in the specialization you use raw Tensorflow and later on in the specialization you use Keras. Also... Many deep learning fathers agree that that is a good methodology, e. g. Ian Goodfellow and the very same Professor LeCunn.
I can only say, Thank You for the help!!
I think your best bet is to use numpy (which just does matrix multiplication if you're really really beginning) to start with.
Simple, but requires you understand the mechanics to make it work. I started learning to program in C, in much the same way, and starting from the bottom like that, I feel, has helped me greatly.
Agreed.
my build a neural net in 4 minutes video does this, would be a great resource
You are totally right but if you want to test different models quickly and want to make a quick project. I would recommend using Keras. The real fun, though comes from making a Neural Net yourself.
It's funny how we train machine learning models so differently than we train ourselves. We tend to think it's best to learn from sophisticated explanations at the base level and build our understanding up from that, but we underestimate our ability to innately infer the inner workings of a topic from numerous excellent examples. I learned programming starting with high level languages, without reading a text book, and worked my way down to assembly, then to CPU design using HDL, then back up. I like this because A. it's fun to get stuff done, then B. it's fun to learn how the stuff works under the hood once you know what it's capable of, then C. you understand the whole stack differently. It's a hard sell when we begin with "okay, let's learn what NAND is" when you're like "how is this going to help me make a video game" or whatever.
The ending was awesome.
Thanks Siraj. I love your video's man.
6:40 and onward helps with building a sequential network yourself. Thanks Siraj.
thank you for keras video, love your final line of video, i am jelly of your independence from the bloatware monopoly company that shall not be named
I agree except you forgot planned obsolescence.
I total agree. Keras is great to get started with. Pytorch sounds great but it doesn't seem to have windows support yet.
it does now :)
I just installed Keras and I come on here to see this. Today is gonna be a good day.
no need to use my ak
Buddy how do get so much energy to teach us all this?
You are really awesome. 😎
You made my day .... Just wondering when you would post this.. Keras..
fantastic information for a person know nothing about deep learning! one feedback would be - information flowing too fast and I had to watch multiple times to understand better. may be it just me, otherwise I love it.
Awesome video Siraj ....
@9:19 >> “Deep Learning is Sexy for a reason!” … I love the face of Support Vector Machines!!!
so good that we are back in AI. I was getting so tangled in the block chains.
more AI coming
What about AI based block chains. E.g. skychain
@Siraj, awesome as always. Little slip "AlphaGo" was in 2015, not 2017 ....
@Siraj Raval - Do you not do anymore live streaming tutorials? Also it would be very interesting to see you do a deep-learn learning tutorial using kaggle's dataset and actually competing on kaggles while live streaming.
i will do more live streaming soon
awesome video, siraj bhai....but you need to use proper chroma key to properly subtract you from background
Please, don't advertise in the comments.
yes i will for sure and thanks
this is really helpful!
Microsoft has become a lot better after Steve Ballmer left
Or after Satya took over
Keras also have the Functional API, which I prefer using.
Can you please explain the term DEEP LEARNING. How do you recognize something that is DL or something that isn't DL?
Sir can you please please do a series on Pytorch. It is such a powerful and promising library and it would be a great addition to your channel and great learning material for us
Dude! That was a cool video.
thank you.your voice and video is really goodq
very very very very
very interesting
Thanks siraj
Thanks for free education
Awesome Video!!
exactly 200,000 views, congrats
Congratulations. I still can't refuse myself using C# and visual studio..
Thank you so much
Why did he say that thing about Microsoft? I didn't get that joke? Is this referring to the processor bugs that they only recently released security fixes for? I'm confused.
Hatred
you are theNewBoston of AI
Keras is REALLY good but there is a negative point: if you need to use a custom loss function, it become a pain in the a**. I have been trying to implement YOLO loss function and I really don't know how to do it. Besides that, it is very cool!
Happy new year siraj #Tyler #TeamTyler
"The question I get asked the most is..."
I don't believe this. I reckon people ask you whether you argue with your hair.
Siraj Videos comes like Aamir Khan's Blockbuster movies
thanks! :)
Hi Siraj and everyone.
F.Chollet has made publicly available some jupyter notebooks with his book examples. I have ported two of them to kaggle kernels so everyone can run and modify those notebooks in the cloud (just sign in kaggle, fork and play with the notebooks).
Links to the jupyter notebooks (kaggle kernel):
www.kaggle.com/juanjotwo/deep-learning-with-python-notebooks-2-1
www.kaggle.com/juanjotwo/deep-learning-with-python-notebooks-3-5-imdb
The original F.Chollet 19 notebooks (run on on a p2.xlarge EC2 instance)
github.com/fchollet/deep-learning-with-python-notebooks
sirajcoin to the moon
U r best
when 1080p look like 360p, great content but upgrade your camera, bro!
can you do a lesson about locomotion based on reinforcement learning?
Why so happy about Theano stopping in 2018?
FuZZbaLLbee Yeah, I think it was because its hard to use but it would have been nice if he indicated why
Ethan9750 I think there was a pydata talk where the speaker prefered it over tensorflow because it allowed to change the framework code more easily.
theano requires some absolutely insane configuration and installation (and compilation) of many third party libraries
I went crazy trying to get the configurations right for theano to run! Did not like it!
Viola came to party!
merci
Is Keras really easier for beginners than TF-Learn?
yes
nah, same
Yeah, i've used to do a lot of stuff and it's quite easy to use.
yes
What does he mean by not using Microsoft products at the end of the video?
What is up with your hate towards Microsoft?
if you have to ask...
Siraj are you an A.I?
not yet
This was exactly what i needed omg i love you
love you
I think you should start by making your own deep learning library before using something that is already made, to allow you to properly understand how deep learning works,*
That's how I started, and I'm glad I did it
can you guide in a bit more details ?
watch this course from the MIT,
you'll learn a lot : th-cam.com/video/uXt8qF2Zzfo/w-d-xo.html
( I learned a lot )
then grab yourself a language (C++/python, whatever you want) and start experimenting, there is abit of math to do, the wikipedia page on backpropagation can help you
you should also make a visual feedback, like a visual representation of your current network, this help when debugging... honestly I started without and I regret it
If you want to start quickly, I suggest using processing, you can make simple windows and graphics without worriing too much about classic graphic stuff ( which can be annoying when you are trying to do something else, but have to spend a lot of time on something only usefull in debugging...)
processing basically is a really simplified version of java, you can make stuff really quick with this, faster than python from what I've experienced ( in therms of developpement, not really about your program speed... )
Hope this helped :p
oh thanks for quick reply...sorry but I couldn't understand what you meant by processing. point me to an article instead of wasting your time on me, that discusses about this.
Insoluble Fraction processing is a “language” or more like an interpreter/IDE, here is their home page: processing.org
Np :p
You should make a neural network that can crop out your green screen better in-real time.
input times weight add bias activate ......remembered
hi akash thanks
thanks siraj for so informative videos
Dear Siraj do need to study maths or compter sceince in school in other to do machine learning or programming
please make a video on spectre and meltdown
Yeeeeesss!!!
back to AI
Thank you Siraj ! How do I extract digits from any image?
Siraj, explain us an Apache UIMA, please
Thanks for making deep learning accessible for idiots like us. :")
I do think it's an old video from your Playlist of Machine Learning.
Any recommendations on how to work on algorithm and data-structure required for ML/DL?
Where are dataflow modular visual frontends like dsprobotics Flowstone (formely Synthmaker) for signal processing?
Hello World it's Keras
But how about TFLearn? isn't it the best option to use with Tensorflow?
not maintained well enough
Re: Coding Challenge. I've already got one I'm working on. DDPG network that learns a grid world game. github.com/dustinandrews/machinelearning/tree/ddpg-link-gradient/DDPG , Some features: epsilon is dynamic, the better the agent is doing the less randomness in the actions. The simulation supports curriculum training where you can set how far away from the goal the start position can be. agent_play(ddpg) pops up a window and shows how the agent plays the game for one episode.
great work Dustin love the choice of grid world!
Thanks! Sorry the code is such a mess. Once I get it all working I'll clean it up and make a repo for it. My critic is learning well, but the actor is still struggling.
Hi am new to this stuff could u kindly tell me are there any free courses available to learn this stuff
super se uper
Keras? More like KerYAAS!
So which part of this is actually different from PyTorch?? I've worked with it a bit and it seems like this is fairly similar so wanna see if I'm missing something. Thanks :)
Siraj, there's now an easier way - brain.js! It would be great if you could check it out!
Why you hate microsoft bruh?
What's up with second order optimization in 2k18?
My Entry. Ours is a new efficient and intuitive algorithm that performs well in continuous action space. It's called Selective Memory. You can see a demo of it here th-cam.com/video/hKrFFeZqq3E/w-d-xo.html
Hi Siraj. I'm new to this, so I run the code in terminal and in the end like you showed in the beginning nothing will pop up to show a text and in terminal after the epoch 60 the texts still don't make any sense. Am I missing something?
Is there an English version?
Scene classification possible in tensorflow
you could change the title to "how to get started with deep learning (with keras)"
9:15 I've gotta not use anything made by Microsoft
Me: Meh :/
At 4:01 captions are showing chaos instead of keras
Artificial Stupidity in action...
Javascript Assembly of Web
Tensorflow Assembly of AI
Hi, May I ask anyone ever tried using Mac pro's Anaconda to run keras tutorial? either in jupyter or spyder, the keras always error when plot training history.
Same setup, ubuntu 16.04 works fine, and virtual environment on Mac Jupyter notebook is okay.
Siraj sounds like Keras 😂😂
Is it ok to install keras and tensorflow on a regular macbook, or are GPUs necessary?
Yulia Gri You can install it and if you don't have a gpu train them on floydhub
I wrote it a few weeks ago, but here it is:
github.com/Goldesel23/DCGAN-for-Bird-Generation
I used keras for training a DCGAN for Bird Generation, I also trained a WGAN with the same architecture to compare the results
Hey siraj,
I'm a mechanical student, is it easy to me to learn deep learning