Skip to content

Instantly share code, notes, and snippets.

@Lihi-Gur-Arie
Created February 20, 2022 12:17
Show Gist options
  • Save Lihi-Gur-Arie/41f014bcfbe8b8e1e965fa11a6251e04 to your computer and use it in GitHub Desktop.
Save Lihi-Gur-Arie/41f014bcfbe8b8e1e965fa11a6251e04 to your computer and use it in GitHub Desktop.
Display the source blob
Display the rendered blob
Raw
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
@izidorg
Copy link

izidorg commented Apr 25, 2022

Hello Lihi, Thanks for sharing! My question is if the final number of classes in your dataset after implementing transfer learning is 81 including penguin?

@Lihi-Gur-Arie
Copy link
Author

Hello,
My model has only one class named 'Penguin'.
It's defined on the 'penguins_data.yaml' configuraions file.
You can see above that the 'nc' (number of classes) parameter there is set to 1, and the names parameter is ['Penguin'].

@izidorg
Copy link

izidorg commented May 1, 2022

Hello Lihi, Thanks for sharing! My question is if the final number of classes in your dataset after implementing transfer learning is 81 including penguin?

@izidorg
Copy link

izidorg commented May 1, 2022

Hello, Thank you for clarification. I was thinking that by using --weights 'yolov5s6.pt' will end-up enhancing capability of additional coco classes.

@SourabhSomdeve
Copy link

Hi Lihi, Thanks for the code, I just to ask did you add the photos where penguins are not present and if yes then where did you put in the folder structure and the label file.

@Lihi-Gur-Arie
Copy link
Author

I did had photos without penguins, it is recomadad up to 10%. Just put them in the same images folder, and without a label file.

@SourabhSomdeve
Copy link

Thank you so much for replying, great help, I will definitely do that.

@Lihi-Gur-Arie
Copy link
Author

Love to help, good luck!

@samuelsaldanav
Copy link

Hi Lihi, thank's for you contributions.

How i can change "pinguin" to... subjetcs or objects? weapons for example? What i need to change?

Thanks and best regards.

Samuel.

@Christoffel21
Copy link

Hi Lihi

Thanks for an excellent acritical. Very helpful.

I've used your code as is on Colab with 300 images of penguins except for changing 'hyp.finetune.yaml' to 'hyp.VOC.yaml' as 'hyp.finetune.yaml' doe not exist anymore.

Here is my Precision Recall Curve, any idea why its so bad compared to your graph.
I got the images from Google OpenImages and used Roboflow to split it into 70% train 20% test 10% valid.

Regards
Christoffel

PR_curve

@phuctrang
Copy link

great

@TheStoneMX
Copy link

Hi there, great tutorial; I was wondering if was possible to share the roboflow repository too?

Thanks!

@aldogan17
Copy link

aldogan17 commented Sep 26, 2022

Hi, great tutorial. I have chosen Yolov5s6 model for training my data. My results for inferences didn't come out as I expected. Each image was detected 50% sth on average. I annotated 1274 images. Do you think I should've selected more images? Does this result have to do with my number of inputs?

@abhayvikramnayak98
Copy link

Hi, Lihi! Where do I get the requirements.txt file?

@ashishjain87
Copy link

The requirements.txt file is in the yolov5 repo.

@ashishjain87
Copy link

Hi. Where is the 'hyp.finetune.yaml'? Does it not exist anymore? Should I use something else like 'hyp.VOC.yaml'. I believe I will have to decrease the learning rate further.

@paulagarciaruiz
Copy link

Hello,
I am training also for a single class but the obj_loss keeps going up. Do you know why? what should i do?
results
In addition, I am not sure I understand the correlogram images. Any clue why this look like this? I am trying to detect an specific type of flower.
labels_correlogram

@luis-ponce
Copy link

where is all the data now? I'm trying to reproduce this and there is no images nor .jaml file

@eribnick
Copy link

eribnick commented Oct 2, 2023

Thanks for this tutorial!
Do the input images need to have the expected size (1280x1280)? If they are a different size, will they be rescaled?

@Lihi-Gur-Arie
Copy link
Author

Hi @eribnick,
I'm glad you found the tutorial helpful. The detector was optimized for that specific input size, but you can use any other input size. The images are rescaled before entering the network. You can choose the image size to which you want to rescale your images using the --img parameter.
Good luck!

Changes

@KerimM-bit
Copy link

Hello @Lihi-Gur-Arie, thanks for the tutorial it helped me understand, how to use yolo. One thing I'd like to understand better would be, let's say I have 5000 images of same class, do I need to manually label all 5k images, or labeling let's say 10% is enough? ( I know there are other factors which can effect that percentage, I just want to know if it's doable with 10% or not)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment