search

johko / computer-vision-course

This repo is the homebase of a community driven course on Computer Vision with Neural Networks. Feel free to join us on the Hugging Face discord: hf.co/join/discord
MIT License
376 stars 124 forks source link

Synthetic Data Creation - [WIP] #60

Closed kfahn22 closed 6 months ago

kfahn22 commented 7 months ago

This is a very preliminary first draft to check on file structure / formatting. (I noticed that I had misspelled creation and fixed. Didn't realize it would close original pull request.)

Notes:

I have added the .mdx files to the chapter folder. There is still an old .md file in the course folder just for reference I have added an example notebook but this will (hopefully) be updated as I am trying to use the ABO dataset. Questions:

The notebooks folder lists synthetic data as ch. 11 -- is it "10" or "11" now? I have added a 10.mdx file but am not really sure what goes in it. I looked at the .mdx files for the Audio course and there are code blocks with py and python. I am not sure when you use py versus python. -Should I add BlenderProc to the requirements file?

review-notebook-app[bot] commented 7 months ago

Check out this pull request on  ReviewNB

See visual diffs & provide feedback on Jupyter Notebooks.

Powered by ReviewNB

youssefadr commented 7 months ago

Hello! I did my master thesis on Deep Generative Models for Synthetic Data Generation. Do you think I can contribute knowing that I already am a contributor on 2 other subjects?

kfahn22 commented 7 months ago

Hi @youssefadr, I would love your help as I am sure you know more than I do! What do you have in mind? I am close to having a chapter on DCGAN done. Do you want to provide input or add another chapter? And perhaps most importantly, would this keep you from working on the other 2 subjects as I think that is probably higher priority.

kfahn22 commented 7 months ago

@youssefadr Also wondering if you are familiar with StudioGAN? I have been adapting an existing notebook that generates synthetic lung images, but I have run it several times and I am not sure the results are consistent. StudioGAN claims to help with reproducibility.

kfahn22 commented 7 months ago

@merveenoyan I was looking through the issues on the StudioGAN repo and saw that you had been tagged RE adding checkpoints to HF hub. Are you aware of any notebooks using StudioGAN?

youssefadr commented 7 months ago

I worked more on StyleGAN, DatasetGAN and 3D inverse rendering methods. Do not hesitate if you have questions!

kfahn22 commented 7 months ago

@youssefadr Do you have an example notebook implementing one of those methods that we could include?

youssefadr commented 7 months ago

Yeah sure! However I had to use A100s for the inference/training. On a T4 GPU it is still slow but we can try it with the NVIDIA checkpoints, we just have to read the license carefully to avoid any code copyright issues.

kfahn22 commented 7 months ago

@youssefadr I have a demo nb for lung images, but am not totally happy with results. Would you be able to look at it suggest improvements? I put a version with and with/out output here for now. https://github.com/kfahn22/medical_imaging/tree/main

merveenoyan commented 7 months ago

@merveenoyan I was looking through the issues on the StudioGAN repo and saw that you had been tagged RE adding checkpoints to HF hub. Are you aware of any notebooks using StudioGAN? I think not :/

merveenoyan commented 7 months ago

BTW we just removed the course folder, so you can carry it to chapters/en/Unit 4 - Computer Vision in the Wild/Synthetic Data Creation.mdx or chapters/en/Unit 4 - Computer Vision in the Wild/Synthetic Data Creation/Introduction.mdx depending on how you'd like to structure this chapter.

kfahn22 commented 6 months ago

I am closing this as it is now outdated.