Implementing original U-Net from scratch using PyTorch

In this video, I show you how to implement original UNet paper using PyTorch. UNet paper can be found here: arxiv.org/abs/1505.04597
Please subscribe and like the video to help me keep motivated to make awesome videos like this one. :)
To buy my book, Approaching (Almost) Any Machine Learning problem, please visit: bit.ly/buyaaml
#DeepLearning #PyTorch #UNet
Follow me on:
Twitter: / abhi1thakur
LinkedIn: / abhi1thakur
Kaggle: kaggle.com/abhishek

[@souhamghosh8714]

YESSS!! finally a paper implementation demo!! THANK YOU SO MUCH!!

[@abhishekkrthakur]

Hope you like it!

[@souhamghosh8714]

@@abhishekkrthakur For any queries related to your videos and work, may I reach you over your linkedin?

[@adiMallya]

Need more of this. I want to understand by implementing stuff just by reading papers. You get better insights like this.

[@vinithageorgee7122]

Thank You so much. After a wild google search for nearly a week, I finally tumbled upon this wonderful video. Good detailed explanation. Hats off to you !!!!

[@rahuljana7197]

Awesome. Loved it. I am pursuing my graduation in AI & ML. To be honest, for the past 6 months I have enjoyed your content more than anything on KZbin. Looking forward to more this type of research paper implementation. Thanks a lot for this amazing content.

[@DiegoJimenez-ic8by]

God Bless you man!, I pre-ordered your book, looking forward to read it!!

[@abhishtsingh6073]

This is just amazing stuff. I would be implementing this tomorrow. The fact that someone is going through a paper and doing it all is quite motivating in the sense that it pushes me to implement some paper on my own, rather than using the github code/ model zoo. Keep it up sir.

[@AbdulRehman-hi7xk]

A great teacher.... i was searching but i found a real teacher... Huge Respect from Pakistan SIR.

[@swarajnanda7874]

Brother, you've hit the jackpot with this kind of content. Thanks for your help! Now that I see the deep details of this machine learning model, I can understand it better.

[@MrGeomTech]

Thank you, for the first time a I found some one speak behind the scene for deep learning, and going step by step

[@yummychen6540]

please do a completed Unet implement which includes the dataset step and training step, the training strategy really confused me.

[@ramchandracheke]

So addictive! Watching this at 4.00 am! You are the best 🙏

[@TheJing2__]

Thank you so much sir. This video inspired me a lot to implement the state of the art CNN architecture by myself.

[@psoma2744]

Thank you very much, Abhishek... Your videos are really helpful and the explanation is awesome... You just make it look so interesting and easy to follow ...

[@stanislawcronberg3271]

It's really nice that you don't cut out the parts where you make some typos. Also helpful to see you printing the shapes along the way instead of just writing it and it magically working at the end :)

[@prakashupadhyay9529]

This was helpful Abhishek.. I'm new to Python paradigm and struggle. Seeing you code through helps understand the approach. Next I'd wish to see some Hyper Parameter tuning implementation in PyTorch, like OneCycle policy for LR, and some performance optimization tutorial during Data load in GPU (Pre-fetching Data in GPU). Thanks for this Video! :)

[@karthiramachandran1257]

Abhishek , thank you very much for implementing this paper ....

[@ibrexg]

Great video, thanks for sharing. It would be nice if this is a part of complete colab/kaggle tutorial e.g. add the training and the testing part on some public dataset.

[@dilrajbhinder4337]

Great Job. Crisp and clear.

[@HimothyOHooligan]

I found your channel through this video. These type of live implementation videos are super useful.

[@abhishekkrthakur]

Glad you like them!

[@TheAravindppadman]

I am recreating your "melanoma-detection" kernel my own. Your ML framework is awesome.It made my life much easier. Next I will implement U-net.

[@abhishekkrthakur]

Fantastic!

[@zakirshah7895]

Aravind brother are you working on (Skin cancer) Melanoma-detection

[@margretazuma5462]

Thank you soo much, this really helped for my thesis

[@mohammedy.salemalihorbi1210]

This was great! Thank you so much

[@talha_anwar]

thank you for implementing in simple way to make it clear

[@amineleking9898]

Very clear explanation. Thank you mate

[@thangnguyen1677]

Nice content, we need more of this

[@anandpawara4048]

Awesome tutorial, please do more videos like this paper implementation from scratch!

[@abhishekkrthakur]

You got it!

[@rashasalim1259]

Thank you so much! A lot to learn here :)

[@shashankshekhar7052]

Thanks, please do paper implementations more often it's really helpful. For decider, shouldn't we use classical interpolation like nearest, bilinear as transposed convolution will cause artifacts patters.

[@varadthikekar464]

Awesome tutorial. got a good understanding of the code

[@WBMangaka1]

Sir, you saved my day. Great video!

[@garesothmen2101]

I liked the video and subscribed before you even ask !!! Thank you so much .

[@adeebshahzad7893]

One word. Amazing!!!

[@elirhm5926]

awesome, thanks a million :) please please please, make more video like this

[@era_ali]

Awesome loved it, wish you would see how to train and evaluate this model

[@needsomeminjeongpower1987]

thank you!! that helps a lot

[@mengop]

Thank you for the wonderful video to understand the original paper implementation. This video is Easy to follow and implement. Can you cover in another session ,the alternate part of using pretrained as encoder network which you have mentioned at the end. Also pls cover the 3D image data handling along with its meta data in model building. Thank you

[@aatiqaghazali7850]

Thank You so much Sir! Love and respect from Pakistan

[@anastasiabutskova9324]

great work, thank you!

[@edoardograsso3391]

unbelievably helpful

[@sanchitahuja12]

How should a beginner actually implement a paper? I am quite confident with the concepts and I am able to use pre-existing code to use it on my dataset. But I get scared while implementing a paper on my own without referring to the original code for that paper. I try reproducing the notebooks on my own by referring to the implementation and typing each and everything by understanding on my own. Is this the correct way? You're doing great work. Thanks a lot!

[@miltonborges7356]

Amazing. Thanks a lot

[@deepexplorationcode9456]

Thank you very much for explaining UNET. I have a better understanding UNET of what is on the paper and how to code it. Can you please show step-by-step how to use it (UNET) for image denoising, super resolution, inpainting, image segmentation always with pytorch. Can you also please explain yolact++ or mask rcnn just like you did with unet (step by step).

[@mihirgajjar8488]

Great tutorial Abhishek, thanks for the video. In the Decoder part where we are concatenating the output of up transpose (ConvTranspose2d) and the cropped image from the encoder, should it be torch.cat([y, x], 1) instead of torch.cat([x, y], 1)? or the order does not really matter? It's great to see such "from scratch" implementations and the end part of the video explaining how various modifications are made (Resnets) to the basic architecture. Thanks :) :)

[@porzo2964]

Thank you very much!

[@ashishjith7096]

Simple & effective explanation. Can you help to upload more videos on implementing GAN's using Pytorch?

[@AmitChaudhary-qx5mc]

Thank you so much brother..

[@indavarapuaneesh2871]

Abhishek Bhai , you are beyond beautiful!!

[@its_me7363]

I don't understand why people don't like the videos. So much information for beginners.

[@abhishekkrthakur]

Its okay. Its not possible to make everyone happy everytime :)

[@jainaviral]

Hey thank you so much !!!

[@DIYGUY999]

Finallly!

[@federicogonzalez556]

Thank you!!!

[@prakhargoyal8160]

Really amazing work, sir.

[@abhishekkrthakur]

Thank you so much 😀

[@PardeepKumar-th2oj]

hi, this was great work, can you also implement the original loss used in the paper in which changes are done

[@IrvingVasquez]

Thank you!

[@johnlee1761]

감사합니다.(Thx so so much)

[@hejarshahabi114]

Thank you for your video, but I have a question regarding the data creation for training, I do research in the field of remote sensing, I mainly work with satellite data for classification and feature detection, so I have to manually digitize some sample as levels for my training processes, so I want you to make such video on how to prepare satellite data for using in U-net. many regards

[@paulohbl]

Love u man!!!

[@gauravsharmahsetsrhu4139]

Wow. Explained it so well. Can I train this model using my own dataset? If yes please explain.

[@017_itmohamedmufassalsulta8]

Sir, this is a great tutorial. I found it very helpful. then just how u approached U-Net, I tried it for 3D Unet . I don't know whether i got the correct result. It would be great if you show us to how to implement it.

[@doraadventurer9933]

Thanks for sharing!!!!!!! Do you also have the training and testing part and demo Unet for a real example? Again, thank for sharing

[@BiranchiNarayanNayak]

Superb.... i love this.

[@abhishekkrthakur]

Thanks a lot 😊 Subscribe for more :)

[@surayuthpintawong8332]

Thank you very much

[@shalini286r9]

Nice Explanation.Please do a implementaion of improved Attention U Net .

[@ninansajeethphilip4656]

Very nice presentation

[@self-made-datascientist1181]

Great video as always Abhishek! Just one question, for doing the training in this model how do you implement the backpropagation and the compiler? Im new to pytorch but I learned so much from this video about the model structuring!

[@abhishekkrthakur]

pytorch takes care of that :)

[@ahmedhusham7728]

Awesome. Can I use UNet to segment weeds plant in a grassland images? - Can I send you sample of these dataset to segment it?

[@varuntirupati2566]

Sir, a very nice explanation! I want to change the final 1*1 convolution layer to have out_channels=3 because my output is colour(rgb) and also I want to make the final activation layer as linear to output the float values. how can I do this

[@rohitupadhyay4665]

Great video, best way is not just to read but to implement what you read! @abhishek would you make a video on training and serving inference pytorch model on Sagemaker..

[@abhishekkrthakur]

yea

[@anujgodase1865]

I wanted to make a unet using resnet. This was helpful.

[@abhishekkrthakur]

Feel free to share how you did it. It will be useful to many :)

[@user-tf4cd2rw3t]

ty buddy!

[@endurirahul9550]

is it possible to combine U-net and bilstm and to use for ECG signals for automatic detection of qrs complexes

[@daneshaslam412]

Hey nice video! I am struggling to implement deformable convolution into U-Net. I am using source code from pytorch master version for torchvision.

[@manishkumarmishra194]

Great work! could you do a implementation video on Style-GAN in pytorch.

[@WeeeAffandi]

Do i need to add a flatten layer in the end in order for it to perform classifications?

[@sebastientetaud7485]

Clean boy !

[@roymercan]

Thank you for another amazing video, coding side by side with you helped me understand how to dissect an architecture presented in a paper. I would love to see the implementation of an attention model from scratch, do you think it is possible?

[@abhishekkrthakur]

definitely

[@roymercan]

@@abhishekkrthakur Thank you! Can't wait for it!

[@ripleylee5726]

thank you for this video! Just wondering if i were to implement a resnet34 model inside, how do i implement it into unet?

[@mayukh_]

Please do a wavenet. Also I always get confused how the data flows in wavenet. Please do wavenet or fast wavenet

[@abhishekkrthakur]

will try :D

[@aminulhuq6668]

@@abhishekkrthakur Yes please do a waveunet.

[@javlontursunov6527]

Please make another video on how to train custom built U-net from scratch.

[@31aish]

You missed to add relu before double conv in expansive block. But great explanation, thankyou!

[@PaAGadirajuSanjayVarma]

sir i want to build face recognition using convolution autoencoder. So can I use unet for this purpose?I tried but the results are bad. When we are doing concatenation of layers i made a copy of them and keep autograd false for those.Is this implementation correct?

[@jamesferrarelli7448]

Thank you so much for this! It has been very helpful in clarifying how this works. One question for you: Your model takes in a grayscale image. If I wanted this to take in an RGB image (so starting with 3 channels instead of 1), would I multiply each layers' in_channels and out_channels by 3? For example, you have `self.down_conv1 = double_conv(1, 64)` Would I change it to ` self.down_conv1 = double_conv(3, 192)`

[@treelight1707]

I was thinking the same question. Did you happen to find the answer to it?

[@raghavavinay8660]

if i wanted to put a pixelshuffle function where should i put that. can you just give small syntax for that

[@pawanbhandarkar4199]

Hi Abhishek. I was reading the paper while trying to implement this and I could not find the "stride" for the up-sampling section which you have written as stride=2. Can you please help me understnd how you calculated it?

[@lv8707]

If we want to include some dataset to this model and check the accuracy, how can we possibly do that? I am a beginner!

[@jeevankumar3527]

@abhishek we can go for resize right instead of crop, Can you throw insight on this

[@florianhofstetter6859]

How would i get the original resolution of the image back?

[@massive_d]

Hey I was thinking of starting with applied data science and your content seems perfect. Can I use the playlists ( are they updated) I was thinking of starting with the 1) framework 2) bert 3) bengali Is this order correct?

[@abhishekkrthakur]

Playlists are updated.

[@manuelh7576]

Great video - thank you! One thing I don't understand is why the input image tile is 572 x 572 and the output segmentation map is 388 x 388. Wouldn't we need a segmentation map of the same size? Grateful for any pointers to resources that explain this further.

[@abhishekswain2502]

It's like a zoomed in version of the orginal image

[@SUGATORAY]

@Abhishek: I already purchased the book on Amazon. How do I get a pdf copy of it?

[@surbhikhushu2394]

Please give a tutorial for cycle GANS as well if possible .Thanks

[@user-pj4ex1eh8f]

Thank you so much!!! I have a question : isn't it a mistake that you concateneted (x,y) instead of (y,x) which is in the paper? Or it doesn't matter ?

[@jonasp1641]

I was wondering the same thing...

[@rs9130]

great tutorial. can you please make video on how to use backbones for unet encoder part like vgg, resnet. please please. thank you edit: also fcn implementation

[@knowledge_Theory]

Wow Just great Can u share the name of your book/ Amazon link

[@massive_d]

This was great 👍

[@abhishekkrthakur]

Thanks a lot 😊 Subscribe for more :)

[@massive_d]

@@abhishekkrthakur I was subscribed when there were no uploads , the name was enough 👍

[@rs9130]

how to use pretrained vgg weights in the encoder

[@joepearson6111]

sorry if i'm being stupid... but how come you have created the 'double_conv' outside of the unet class. I'm fairly new to this sort of stuff and I've only ever seen the whole model be created within said class.

[@funadda7338]

Hello Abhishek, I am trying to implement U-Net on Fashion-MNIST dataset using pytorch, any reference material that can help?

[@pheiroijamprishika6414]

Sir Can you do a video on Deep Boltzmann machine

[@vihalkaviyarau9044]

Sir i need a video on how to train the code with dataset and if we give an image it should return the segmented one, how to do so with this UNet? Please do a video on it as soon as possible. Or if there is anyother video and if anyone knows which one is better kindly reply for this message with the link of the video.