Implement and Train U-NET From Scratch for Image Segmentation - PyTorch

[@sairampenjarla]

one of the most cleanest code, straight forward logic, excellent explanation. You've earned my respect, like, subscription. Thank you sensei.

[@uygarkurtai]

Thank you, I'm flattered!

[@manuelsteele7755]

You have excellent communication on how the model is implemented. The clarification is enormous. This is the best YT video I have seen for explaining the layers. I can't think of one I saw (and I saw many) that clearly helped me understand the layers of a model like yours does. Thank you. I am a PhD student in Data Science at Arizona State. We are studying U-Net for a graduate course on medical image processing. I also need it for my dissertation topic.

[@uygarkurtai]

@@manuelsteele7755 thank you for your kind words :) I'm glad it helped!

[@bikrammajhi7278]

Great video. Good Work. Thank you !!

[@uygarkurtai]

Thank you!

[@oerdoganist]

great work. Keep up the quality training free, lots of us learning with your videos. 👍

[@uygarkurtai]

Thank you :)

[@diandiansun]

Amazing!!! its so awesome!! YOU ARE MY GOD!!!😍

[@uygarkurtai]

Thank you!

[@learntestenglish]

Great to find a source about this unique topic, thanks for your efforts and great teaching 🥰

[@uygarkurtai]

Thank you :)

[@eranfeit]

Great tutorial

[@uygarkurtai]

Thank you!

[@Samuel-san-x9x]

I like this video , so clear , and I was able to follow and do the same thing. I understook Unet thanks to you . Thank you so much . Could you do a video on DDPM ? specially conditional or unconditional DDPM using Unet ? Thanks a million.

[@uygarkurtai]

@@Samuel-san-x9x thank you! I already have a DDPM video. Check it out here: kzbin.info/www/bejne/gnjIYaueh9iYm8ksi=QSAnwVGYrL5Vdafz

[@Samuel-san-x9x]

@@uygarkurtai thank you.

[@NisarAhmad-ch3kc]

Thank you. Great Work! Simple, step by step! Do you also plan to implement 3D Unet in tensorflow?

[@uygarkurtai]

Thank you! I don't think I'll get into the variations of U-Net.

[@FernandoPC25]

Thank you very much Uygar! I loved this tutorial, really helpful, carry on! I was also curious about the amount of memory your GPU has. Could you please share that information?

[@uygarkurtai]

Thank you! I used Kaggle's P100 GPU. It has 16GB if I'm not wrong.

[@FernandoPC25]

Hello Uygar, I wanted to ask you something. I tried to replicate your experiment in Kaggle, using GPU P100 but I am having memory problems, CUDA runs out of memory. What proceduce do you recommend if this happens? (Like putting num_workers=4 or something like this) Thank you very much! @@uygarkurtai

[@uygarkurtai]

Is it possible thath you use images with higher resolution? In that case you need a bigger gpu or you can try quantization or something@@FernandoPC25

[@FernandoPC25]

I think that I am using the same dataset than you. I will try with the different GPUs of kaggle notebook. Thank you very much sensei!

[@islevkurt1532]

It was a great video👏👏

[@uygarkurtai]

Thank you :)

[@jamest.3210]

Hi, great video! It already helped me a lot! Thanks! I'll try to train a U-Net with my own data. Can you tell me if there is anything else I need to be aware of? Does it matter what kind of data files (jpg, png, gif or tiff) I use or do they have to be jpg files + gif files with labels for the training?

[@uygarkurtai]

Great to hear that! You can choose any format as long as you can load it.

[@AlperenK.]

You are great! Adamsın :)

[@uygarkurtai]

Thank you! Tesekkurler :)

[@ElminsterWindwalker]

Do we need to add .UNSQUEEZE(1) to mask = img_mask[1].float().to(device) and why?Thanks!

[@uygarkurtai]

@@ElminsterWindwalker unsqueeze adjusts the dimensions of the tensors. If it works we don't need to.

[@fatiherden-dm7yo]

Uygar hocam selam, ben modelinizi çalıştırdım fakat çıktıda predicted mask tamamen siyah olmaktadır. Aynı dataseti kullanıyorum. Bu datasette doğru çalıştırabilirsem başka bir datasette proje yapmama gerekiyor. sebebi ne olabilir?

[@uygarkurtai]

merhaba, data preperation yaparken bir yeri atliyorsun yuksek ihtimal. Image channellar'i karistirdiginda bu durum oluyor cogunlukla.

[@debarchanbasu768]

The code reads okay but RuntimeError: Given groups=1, weight of size [64, 3, 3, 3], expected input[1, 64, 512, 512] to have 3 channels, but got 64 channels instead Can't seem to figure what's going wrong!

[@uygarkurtai]

you have a shape mismatch error. You have to change variables according to your input data.

[@김화겸-y6e]

Amazing! Thnks for sharing your knowledge and skillsksfsdf

[@uygarkurtai]

Thank you :)

[@jasonleo6582]

Hi, Is there a way to download the dataset locally?

[@uygarkurtai]

Yes! If you go to the link of the dataset, there's a download section to download it as a .zip file.

[@mohamedabdelaziz457]

I have an error its No such file or directory as i put the paths of image and mask. Can you help me solving this error?

[@uygarkurtai]

hey. Probably you typed a wrong path to images and masks. Doulbe check please

[@artursradionovs9543]

How many masks and images are used in the dataset?

[@uygarkurtai]

I don't really remember. But you can check the dataset.

[@afolabiowoloye804]

Nice work. Please, my work entails using a similar model to segment multiple parasitized cells from the uninfected cells in a microscope slide image. any hint will be appreciated, please.

[@uygarkurtai]

Hey. There're much more up-to-date models. If you want to use segmentation on a project I suggest you check them out.

[@afolabiowoloye804]

@@uygarkurtai Thanks for your timely response, it means a lot to me. Would you please recommend videos or material for me?

[@uygarkurtai]

@@afolabiowoloye804 It's a pleasure. I found this repo. You can check it out. github.com/mrgloom/awesome-semantic-segmentation

[@afolabiowoloye804]

@@uygarkurtai Thanks, a lot

[@jenushadijafari6041]

@@uygarkurtai such a great repo. tnx

[@shoaibshafiahmed1983]

i am getting this error when training any idea how to fix this Given groups=1, weight of size [512, 1024, 3, 3], expected input[1, 512, 476, 18] to have 1024 channels, but got 512 channels instead

[@uygarkurtai]

Hey. What's your input image size?

[@shoaibshafiahmed1983]

@@uygarkurtai It's a hyperspectral cube of size 349x1905x144 with 15 classes.

[@shoaibshafiahmed1983]

​@uygarkurtai it's a hyperspectral cube of size 349×1905×144 and 15 output classes

[@uygarkurtai]

@@shoaibshafiahmed1983 you got to resize your image or modify the model parameters accordingly in that case

[@shoaibshafiahmed1983]

@@uygarkurtai I am working on a school project that has data as a hyperspectral image need to perform semantic segemntation using UNet with pytorch are you open to work on it privately or help me out solve the errors that am receiving?

[@katayoonhessni]

thanks so much

[@uygarkurtai]

Thank you!

[@Ece-kx6qk]

Great video thanks

[@uygarkurtai]

Thank you :)

[@aymendebiche8095]

Hallo, danke Ihnen für dieses Video . ich habe aber nicht verstanden,woher haben Sie die manual_test und manual_test_mask bekommen . bei der daten ordner haben wir nur die"test.zip",test_hq.zip",,"train.zip","train_hq.zip","train_mask.zip" ich habe schon immer error für die path des manual_mask und manual_test_mask beim die Inferance Teil .

[@uygarkurtai]

Hey, thank you! I got them from Kaggle competition. I supposed to have given the link in the video. I supposed to be showing the competition page too. You can just download from there.

[@强朱-f9w]

我也遇到相同问题了 请问你是怎么解决的

[@TinLee99]

i've trained your model through 30 epoch, but when i use model to predict It's too bad !!! then i continue to train 10 epoch more, but it's still to bad, i don't know why, i use your dataset, i use your scripts;

[@ElminsterWindwalker]

Me too!

[@TinLee99]

@@ElminsterWindwalker i have fixed, his code is perfect, It was my fault

[@uygarkurtai]

@@TinLee99 that usually happens when you use a different dataset with different image channels. In that case you have to do slight modifications in your code. Are you using a different dataset?

[@TinLee99]

@@uygarkurtai i use pet dataset, i trained each through 30 epoch and i can see result, but when i modify upsample from transposedv2 to upsample of pytorch, i see my result that is better, i don't no why

[@fpvx3922]

Thanks

[@uygarkurtai]

thank you :)

[@ibozum1]

Tebrikler ve paylaşım için teşekkürler. Oldukça açıklayıcı ve destekleyici bir paylaşım. Ben de hemen beğenip abone oluyorum. :)

[@uygarkurtai]

Çok teşekkür ederim :)