How is it that every topic this guy covers becomes one of the best videos on that topic? Keep up the good work. (Btw I got all the questions correct do I get money?)
Man, SAM will surely outperform my measly model. It’s 3 million parameters trained on 300 examples in 2 hours. SAM is 600M parameters trained on billions of images with curated human annotations. It’ll definitely outperform significantly. The goal of my project was not to train the most performant model though, but a fun side project to understand how YOLO works!
How would u suggest to get good with pytorch? I struggle with the amount of functions it has and also sometimes with reshaping on higher dimensions (it is difficult to visualize). Also very nice video...love the way you explained ❤
If you asked me this 3 years ago, I’d have given a different answer. But imo the best way to learn PyTorch now would be do chat with an LLM (Claude or ChatGPT etc) while working through any project of your choice. So pick up a new project, anything that is sufficiently challenging and interesting for you, and begin coding. Look up a tutorial as well if you can. Wherever you get stuck, consult the LLM. I’d also recommend to code in a Jupyter notebook while you are learning to code, simply because it’ll help you check dimensions and output shapes very interactively (helping the issues with reshaping). And yeah, don’t worry if you find PyTorch hard. It totally is! I also had my own battles with compute graph programming. Just break down a larger neural network into multiple smaller modules, and then connect these modules together like Legos. Each module you don’t understand, just treat them as a “black box” and move on to the next piece. As long as you understand the desired input and expected output of each module, you can do some incredible stuff. It’ll become very easy once you practice enough, I promise!
@@avb_fj Thats honestly great advice. I do try to follow through the documentations as much as possible and another thing I do is to look at codebases of related work and that helps a lot. And what you said about using an LLM is very true while I don't usually copy paste code from LLMs i mostly use GPT to understand code which I cannot understand and that honestly has helped me learn even more. I have finetuned some open source llms over classification task and that has been a foundation of lot of my understanding. I do understand how pytorch works and its autograd nature I learned it through Andrej Karpathys series on LLMs (which yet again is a gold mine). Thank you for the reply really...I will try implementing as much as possible and I hope I get a bit better. Love the way you teach and will be eagerly waiting for your other videos❤️
A lot of differences really. YOLO v8 trained a 53 layer darknet for the backbone with something known as “CSPNet”. The Neck was also a more advanced form of the FPN network described in the video - called PANNet. There are probably other differences as well (like using an IOU loss), but these are some key ones that I can remember.
All the code, code walkthrough videos, files, animations, and assets used on all my videos is available on the Patreon page. Here’s the Patreon post for this video: www.patreon.com/posts/code-walkthrough-111677858
I love how you explained concepts with a project and code. Enjoyed the quiz questions…Thanks!
Thanks!!
How is it that every topic this guy covers becomes one of the best videos on that topic? Keep up the good work. (Btw I got all the questions correct do I get money?)
Wow, thanks! (sorry no money but congrats on getting all questions correct!)
Good video, thanks! And do you plan to talk about Detection Transformer and compare it with YOLO? That must be interesting. ❤
Yes Detection Transformers would likely be its own future video. Thanks for the suggestion!
I'd be curious to see the performance with the YOLO model replaced with Meta's Segment Anything model
Man, SAM will surely outperform my measly model. It’s 3 million parameters trained on 300 examples in 2 hours. SAM is 600M parameters trained on billions of images with curated human annotations. It’ll definitely outperform significantly. The goal of my project was not to train the most performant model though, but a fun side project to understand how YOLO works!
How would u suggest to get good with pytorch? I struggle with the amount of functions it has and also sometimes with reshaping on higher dimensions (it is difficult to visualize).
Also very nice video...love the way you explained ❤
If you asked me this 3 years ago, I’d have given a different answer. But imo the best way to learn PyTorch now would be do chat with an LLM (Claude or ChatGPT etc) while working through any project of your choice.
So pick up a new project, anything that is sufficiently challenging and interesting for you, and begin coding. Look up a tutorial as well if you can. Wherever you get stuck, consult the LLM.
I’d also recommend to code in a Jupyter notebook while you are learning to code, simply because it’ll help you check dimensions and output shapes very interactively (helping the issues with reshaping).
And yeah, don’t worry if you find PyTorch hard. It totally is! I also had my own battles with compute graph programming. Just break down a larger neural network into multiple smaller modules, and then connect these modules together like Legos. Each module you don’t understand, just treat them as a “black box” and move on to the next piece. As long as you understand the desired input and expected output of each module, you can do some incredible stuff. It’ll become very easy once you practice enough, I promise!
@@avb_fj Thats honestly great advice. I do try to follow through the documentations as much as possible and another thing I do is to look at codebases of related work and that helps a lot. And what you said about using an LLM is very true while I don't usually copy paste code from LLMs i mostly use GPT to understand code which I cannot understand and that honestly has helped me learn even more. I have finetuned some open source llms over classification task and that has been a foundation of lot of my understanding. I do understand how pytorch works and its autograd nature I learned it through Andrej Karpathys series on LLMs (which yet again is a gold mine).
Thank you for the reply really...I will try implementing as much as possible and I hope I get a bit better. Love the way you teach and will be eagerly waiting for your other videos❤️
@@kaustavdas6550 thanks! Keep at it friend!
what are the differences with yolov8 implementation
A lot of differences really. YOLO v8 trained a 53 layer darknet for the backbone with something known as “CSPNet”. The Neck was also a more advanced form of the FPN network described in the video - called PANNet. There are probably other differences as well (like using an IOU loss), but these are some key ones that I can remember.
Share the code aswell💖
All the code, code walkthrough videos, files, animations, and assets used on all my videos is available on the Patreon page.
Here’s the Patreon post for this video:
www.patreon.com/posts/code-walkthrough-111677858