Conversation
Conflicts: generative/networks/nets/__init__.py
|
As of now, The autoencoder and diffusion models are only being trained for around 30 epochs, so the results are not great (see figure attached). I guess that's enough for a tutorial (??) Also, I didn't include validation because it looked cleaner and more understandable to me (and we are not even close to overfitting).
Would rather train for longer and include validation for better results? I guess it's a trade-off between having some very quick results or better generation. |
|
The tutorial looks very good. I agree with the comment that for the tutorial the generated images looks good enough. @Warvito what is your opinion? One last point, could you remove all the files that are not part of the 3d_ldm tutorial from the branch so that we can merge it without conflicts? |
Warvito
left a comment
Warvito
left a comment
There was a problem hiding this comment.
Well done, @SANCHES-Pedro The tutorial looks great! I added a few comments regarding the format of the code, some comments left, and regarding the updated networks. Please, perform the merge of the current main into this branch and update the networks initialization and set precommit in your local repo.
Conflicts: generative/networks/nets/patchgan_discriminator.py tests/test_latent_diffusion_inferer.py tutorials/generative/2d_ddpm/2d_ddpm_compare_schedulers.ipynb tutorials/generative/2d_ldm/2d_ldm_tutorial.ipynb tutorials/generative/2d_ldm/2d_ldm_tutorial.py tutorials/generative/2d_vqgan/2d_vqgan_tutorial.ipynb tutorials/generative/2d_vqvae/2d_vqvae_tutorial.ipynb tutorials/generative/2d_vqvae/2d_vqvae_tutorial.py
Warvito
left a comment
Warvito
left a comment
There was a problem hiding this comment.
It looks great, @SANCHES-Pedro ! Well done! I will remove the change to the VQGAN tutorial and perform the merge.
Signed-off-by: Walter Hugo Lopez Pinaya <ianonimato@hotmail.com>
Signed-off-by: Walter Hugo Lopez Pinaya <ianonimato@hotmail.com>
3 participants
Create a LDM 3D tutorial.
I'm thinking about using the BRATs data.
I copied over some WIP from the 2D version which is another pull request (so it might make a weird commit mix).