search

lucidrains / byol-pytorch

Usable Implementation of "Bootstrap Your Own Latent" self-supervised learning, from Deepmind, in Pytorch
MIT License
1.76k stars 248 forks source link

Need advice using custom dataset #60

Open ramdhan1989 opened 3 years ago

ramdhan1989 commented 3 years ago

Hi this is really good application. I would like to have your advice to use my own dataset. I have images that are very different with widely known dataset such as imagenet containing common object (cars, cat, dog, etc). this is very technical related, it is dynamometer cards that commonly used in oil industry as shown below : https://knepublishing.com/index.php/KnE-Engineering/article/download/3083/6588/15176 Unlike common object, my dataset always comes in neutral position, there is no rotated, no flipped, no color distortion (just B&W) and always comes in the full part (unlike cars that might exist only part of it in images), etc that might exist during implementation in the real world. So in the common object recognition, I assume such data augmentations are needed to introduce many variations to the model. Do I need to do same data augmentations used by default in this repo ? or is it ok for BYOL to use no augmentations or minimum augmentations ? what do you think the best solution needed for my own dataset especially in term of augmentation.

thank you

Shishuii commented 2 years ago

Hi, @ramdhan1989 . Can you please share the source of dataset that you are using ? I am also working on similar problem. It will be very helpful for me , if you can share your dataset of dynacards.