Adding training scripts

[isamu-isozaki]

Hi! Thanks for the code! My friend and I are making a training script using this dataset

[isamu-isozaki]

Will make a pr if you want!

[lucidrains]

sure, thank you!

[lucidrains]

@isamu-isozaki highly recommend the huggingface accelerate library

[isamu-isozaki]

@lucidrains Sounds good! Quite a fan of it already. We are mostly done with the script but will integrate accelerate and open a pr

[lucidrains]

🙏 🚀

[Time9Y]

Hi! Thanks for the code! My friend and I are making a training script using this dataset

Thank you very much for your contribution and look forward to your code!

[Time9Y]

Thanks for the code, I tried to add the training code but the result was not satisfactory.

[isamu-isozaki]

@Time9Y yup my friend and I are having trouble getting good results too. It kinda works on 64x64 resolution but on 128 it just breaks.
I'll try checking the paper for their params

[lucidrains]

oh this is disappointing

could i have made an error?

they released a public official repository btw

[isamu-isozaki]

@lucidrains Haha no most likely we screwed up a bit. And sounds good! Will check that out. I'll add our script as a pr in a bit. Just cleaning up since it's a bit messy but maybe we can get your opinion too!

[Time9Y]

@lucidrains This should be my problem, the code you provided is clean and efficient.

[Time9Y]

@isamu-isozaki I've seen the code you submitted and it's so cool. Hope I can use it soon!

[isamu-isozaki]

@Time9Y Thank you! Will let you know when finished

[Time9Y]

@isamu-isozaki I've received the code, it's awesome! During training process, how many iterations did you get good results after.

[isamu-isozaki]

@Time9Y Thanks! For 64x64 around 10 epochs gave me not-too-terrible results. Self-condition is pretty interesting too. Honestly, looking forward to hearing from you if u get good results!

[Time9Y]

@isamu-isozaki Amazing, I achieved great results in 50 iterations, thanks for your contribution! This code will help our research progress.

[isamu-isozaki]

@Time9Y Oh awesome! Glad it helped. What parameters did you guys use out of curiosity?

[Time9Y]

@isamu-isozaki The parameters are basically similar to yours, while referring to this code. https://github.com/WuJunde/MedSegDiff
lr 1e-4 eopch 150 batch_size 4 UNet (1, 2, 3, 4) not use ddim

[isamu-isozaki]

@Time9Y Thanks! My friend will appreciate that a lot

[lucidrains]

@isamu-isozaki Amazing, I achieved great results in 50 iterations, thanks for your contribution! This code will help our research progress.

great to hear! 🙏

@isamu-isozaki thank you so much for taking the time to contribute the training script!

[isamu-isozaki]

@lucidrains Np! I wanted to contribute later to your muse repo so this was a bit of a practice run. And also, @dsbuddy helped a lot.

[isamu-isozaki]

@lucidrains Sry, found a slight bug in training but made pr to fix it!

[lucidrains]

@dsbuddy thank you Daniel 💯 !