The U-Net (actually) explained in 10 minutes
Vložit
- čas přidán 4. 05. 2023
- Want to understand the AI model actually behind Harry Potter by Balenciaga or the infamous image of the Pope in the puffer jacket? Well.. diffusion frameworks such as DALL-E 2, Midjourney, Imagen or Stable Diffusion seem to get a lot of credit, where as the true unsung hero of the story is the underlying U-Net architecture that they all actually use under the hood. Don't get me wrong Diffusion models are awesome but the U-Net is an absolute STAPLE when it comes to computer vision and this video aims to break it down in an easy way. Originally used for image segmentation the U-Net has developed into so much more. Happy watching!
U-Net paper: arxiv.org/abs/1505.04597
Many thanks to numerous online resources that helped me create this video. - Věda a technologie
man, this video is such a great explainer. I was confused about the use of skip connections since a long a time, but he explained the intuition behind it very nicely.
Why didn't I find your channel before. Please upload more content, the best content on Deep Learning I have seen.
Thanks a lot :)
This architecture is one of the truly brilliant ones in the world of deep learning in terms of its simplicity and efficiency.
Thank you for creating this video! Its the best explaination of how a U-Net works that was easy to understand. The visual animation is superbly done!!
Continue this series, very helpful
Oh my god man. Awesome videos. Keep it up, I'm really enjoying them!
This channel deserves more subss!! Great content and delivery :)
This was extremely helpful. Thank you
Yooo...this is quality content right here. Thank you so much for putting this out
Extremely useful for beginners like me. This is very good
your explained under 10 minutes videos are goated
The best ever video you can get on Unet explaination
This was the best unet explanation I have ever seen
Thank you very much for the time put on doing thisvideo. Interesting and helpful :)
Very nice my friend, this has been most helpful
Still don't know how it works
This video has been extremely useful. I subbed.
Woooooow! Finally I understood it , really great explanation, thank you
This was great, would love a video on diffusion transformers! It looks like they are taking off and replacing U-Net's as the backbone to new diffusion models.
thanks for the video, I am trying to use U-net for anomaly detection in time series and your video gave me the idea.
Very nice explanation. Thanks a lot.
Great presentation!, Easy to understand
Yooo the effort haha. Amazing Video!!!
i love your presentation style
Amazing video, cleared everything!
Absolutely amazing work 🎉
Thank you for great explanation.On basic level it helps better understand unet
very good explanation of U-NET
This is Just awesome, great video
great explanation thanks!
Great summary, Great thanks
dude thankssssss i thought this was another one of these things thatll take me 2 hours of youtube to *not* understand, but u saved me
Thank you that was so helpful and cute! 🤩
Great Explanation.
Thanks for sharing!
nice video, very helpful
wow awesome video and explanation
Nice explanation
thanks, good explanation
You might not find my comment since the video is too old, but man I just want to thank you for this video. I am a student who has always been interested in computer graphics and related fields like game engines, physical rendering, ray tracing, etc, and jst didnt get the ML/AI hype everyone was on the past 2 years. I only ever managed to study ML basics for 2 weeks before I left it for good. But recently I got in a team where my friends were working on CNN based projects, and that made me learn about many basics about NNs and DL. This explaination for Unet seals the deal for me, and I will strive to work on integrating my two interests into one and hopefully create something I love.
such a well made video
Great video champ
Very helpful
Thank you so much. Now I just need to figure out how to implement this for my project lol
Clearly explained. What caused my consfusion in the first place is, in the graphic in the original paper, why does the segmentation mask not have the same dimensionality than the input image?
Thanks a lot lot. I understand it!
hi its very helpful, how can I reach the PowerPoint of it?
very nice dude thank you so much
this is extreeeemely helpful,and funny
Thanks John!
Hi, thank u for this video. can u pls do a video to explain YOLO?
What's the background music called in this video?
This explains inference (I think) by decomposition (dividing) and recomposition (adding) images. Is that accurate?
If you want to just use the Decoder how would you do it?
Thank you very much bro...
good stuff
awesome! can you calso make similar (actually) for Unet++ and Unet3+ please??? thank you so much.
Glad you liked it! Its not currently on my list of to-do videos as I like to cover the most popular fundamentals at the moment, but I'll let you know if I get around to it! :)
Dude, you're great. I'm from Portuga 🇵🇹 🟩🟨🟥🟥and I'm learning Machine Learning and Neural Networks. Thank you very much! I loved how you teach. You are intuitive and dynamic. A person is learning a difficult subject and still manages to laugh when watching the videos. I loved. I already subscribed and liked. I'm going to watch more of your videos now. Hugs from Portugal😉
I still don't understand that the output is x2 or x3 or x4.I don't understand why that is the case?
If downsampling works by max-pooling, how does upsampling work? In traditional image processing, we would just interpolate image colors, but how does the network apply it's "convolution" in this process? I would understand "deconvolution", but in my mind it wouldn't work here.
May be Transpose Convolution
cool videos
bro , immediate subscribe!
Really impressive vedio! And fun work at the end!!!!! LOVE LOVE LOVE!!!
Thank you very much! :)
Hi. I find the video very interresting. As I'm at the begining, i'm little confused. please, can you also propose a pdf file ? thank yu. Nicely
Perfect
nice explanation. but why distracting background music?
Agreed. Good explanation but I wish people would stop using background music.
Dalle 3 is coming to gpt 4 and it can write text!
If anyone wonders how to concatenate the features if they don't match the size... they crop it.
Now how they coded it?
Hahaha well there are actually plenty of online code implementations available but I will see if I can get round to a code tutorial on the u-net sooner rather than later!
@@rupert_ai can u provide one
Nice Comment: Useful 👍👍😎😎
goodgood
nice video, but ideo i hate the music in the background ( so disturbing )
bro why did u stop making videos i need you lmao (its a painful lmao.)
You are very funny!
Me seeing the video at 1.5x 😂😅
I feel like this is more a description to experts than an actual explanation of how and why it works.
Questions I'm left with:
What is the purpose of downsampling/upsampling (I'm guessing performance?)
How is segmentation actually done by the u-net?
How is feature extraction actually done?
What are max pooling layers?
What does "channel doubling" mean, and what does it achieve?
How does the encoder know "these are the pixels where the bike is"?
Why is it beneficial to connect the encoder features to the decoder features at each step, versus in the last step?
How does unet achieve anything other than downscaling/upscaling performance efficiency? Where are the actual operations to derive features?
How is u-net specifically applied for various use cases like diffusion? What does diffusion add or change, for example.
(Disclaimer: I am a beginner, and this is not intended to be a complete answer.)
You should read about convolutional layers and pooling layers to better understand this video. At any rate:
A colored image has three channels: R, G, and B. A convolutional layer is specified by some spatial parameters (stride, kernel size, padding) and how many filters are there - the number of filters is the number of channels of the output. You can think of each filter as trying to capture different information. Doubling the channels, therefore, means using double the number of filters when using a stride of 2.
The segmentation is done just like any ML task - the training data consists of pairs of images and their annotated versions. I think it's often hard to decipher the inner workings of a particular neural networks, and your question can/should be asked in a more general way - how do neural networks learn?
hope you can come back to life
Is he dead?
@@c.e1187nah, just busy I imagine. He was active on github in December so
@@c.e1187maybe yes. Only on CZcams
TIGHT TIGHT TIGHT
music is too distracting... :(
no
I clicked on thumb down for wasting one minute of my precious time in the intro. Get to the F point !!
Promo_SM ✅