Continuing on with diffusion models! ;) Let me know how you like this format - is it better or worse compared to classic paper overviews? What are the pros and cons?
these are amazing they serve as a guide for some one who is looking to build from scratch and also to get better technical understanding and insights which we don't get from the classical paper some times... there aren't many con's though as your explanation is good... coding from scratch might bring good feel as we are walking along with you and views but doing it for these big projects can be a lot sometimes, .. just a view from a lazy person , we don't know until we try and get that videos feed back .. anyways thanks for the effort man.
Oh man please continue the coding series. Like I've said before, digging through codes is often where I've found out intricacies of the paper that I hadn't realized while just reading. For a non-CS background amateur person like me, going through codes isn't the easiest thing, so a guide like this is so helpful. Thanks for producing these, its really appreciated!
Thank you for the videos on diffusion models. It would be great if you could also do a coding session for single and multiple gpus and known and custom datasets
I wonder if there's any specific requirement for the classifier architecture. In the appendix of the paper it is mentioned that “Our classifier architecture is simply the downsampling trunk of the UNet model with an attention pool [49] at the 8x8 layer to produce the final output.” I wonder if there is a reason that the classifier is just a simplified version of the denoising network - UNet? If you change the classifier with another architecture for image classification, would you be able to guide the generation properly? In other words would the gradients from a different classifier architecture be meaningful in guiding the sampling process?
Great Video! so nicely explained, thank you! Do you know what kind of accuracy values to expect when training the classifier? and for how long to train? I am trying to learn a custom dataset and as soon as I add the "--noised True"-Flag, my model does not learn at all (without noise it achieves 95 % accuracy easily). Loss is also not decreasing...😕
Great video! This really helps me a lot! Just one question about how they embed the class embedding into the unet model, it seems that you did not mention it in the video? I look throguth the code and I think the class is embedded after time is embedded, such that: emb = emb + self.label_emb(y) and the class informatin is so called lebel_emb. Do I get it correctly?
Continuing on with diffusion models! ;) Let me know how you like this format - is it better or worse compared to classic paper overviews? What are the pros and cons?
Better but your coding sessions would add to it
@@prabhavkaula9697 meaning coding from scratch?
these are amazing they serve as a guide for some one who is looking to build from scratch and also to get better technical understanding and insights which we don't get from the classical paper some times... there aren't many con's though as your explanation is good... coding from scratch might bring good feel as we are walking along with you and views but doing it for these big projects can be a lot sometimes, .. just a view from a lazy person , we don't know until we try and get that videos feed back .. anyways thanks for the effort man.
@@TheAIEpiphany yes please
this would help in catching up the bottom up approach and maybe finding some new novel solutions :)
@@TheAIEpiphany you are a legend
Oh man please continue the coding series. Like I've said before, digging through codes is often where I've found out intricacies of the paper that I hadn't realized while just reading. For a non-CS background amateur person like me, going through codes isn't the easiest thing, so a guide like this is so helpful. Thanks for producing these, its really appreciated!
Happy to hear that, that's the idea, thanks!
Legend Right here, Ladies and gentlemen. Explained Beautifully. Thanks a lot!!
Oh, man. You are so much ahead of 80% of the ML researchers. Watching your stuff saves me so much time to get a hang of the codebase!
I once couldn't understand how the sampling actually works, but now I know thanks to your video.
Always thank U -!!
Really enjoyed this video. Keep up the great work.
Thanks! Appreciate it!
Thank you for the videos on diffusion models. It would be great if you could also do a coding session for single and multiple gpus and known and custom datasets
Thanks! Stay tuned ;)
coding from scratch would really help as well :) !! Thank you for the awesome videos always
I wonder if there's any specific requirement for the classifier architecture. In the appendix of the paper it is mentioned that “Our classifier architecture is simply the downsampling trunk of the UNet model with an attention pool [49] at the 8x8 layer to produce the final output.” I wonder if there is a reason that the classifier is just a simplified version of the denoising network - UNet? If you change the classifier with another architecture for image classification, would you be able to guide the generation properly? In other words would the gradients from a different classifier architecture be meaningful in guiding the sampling process?
Great Video! so nicely explained, thank you!
Do you know what kind of accuracy values to expect when training the classifier? and for how long to train?
I am trying to learn a custom dataset and as soon as I add the "--noised True"-Flag, my model does not learn at all (without noise it achieves 95 % accuracy easily). Loss is also not decreasing...😕
Great video! This really helps me a lot! Just one question about how they embed the class embedding into the unet model, it seems that you did not mention it in the video? I look throguth the code and I think the class is embedded after time is embedded, such that: emb = emb + self.label_emb(y) and the class informatin is so called lebel_emb. Do I get it correctly?
Hi, this is awesome video. Thank you.
I have a question,
the OS you tested is linux..? or window?
You def read clean code.
Do Computer vision require 3D geometry? Any recommendations on roadmap of Computer Vision?
Depends which subfield of CV you care about. Definitely not a prerequisite for many topics. E2e learning is taking care of the geometry for you :))
@@TheAIEpiphany Do you think I need geometry to become Research scientist in computer vision ? Any resources?
plese cover k-diffusion