When are you open sourcing the code for step 1?

[whiteyhat]

Context: https://github.com/Seed3D/Dora/tree/main/sharp_edge_sampling#:~:text=For%20some%20reasons%2C%20we%20won%27t%20release%20the%20code%20for%20the%20first%20step%20for%20now.

[Jgayle99]

Those scripts are already in the repo. I've made some modifications to them for concurrent processing on my fork of the repo. Probably need to modify a bit to set up for your env but it sped up the processing a good bit vs running it sequentially.

[ARuiChen]

Jgayle99 is right. We have already provided an alternative version of Step 1 in the repository. By the way, I'll provide a pure inference version that doesn't require PyTorch Lightning. This will make it convenient for you to use Dora-VAE to extract latent cods and apply them to your own training frameworks or tasks.

[Jgayle99]

I've made a few small changes on my fork to these files to let you set the test_step to output the latent codes and set the output token length via the yaml config:

pytorch_lightning/craftsman/utils/saving.py
pytorch_lightning/craftsman/systems/shape_autoencoder.py
pytorch_lightning/craftsman/models/autoencoders/michelangelo_autoencoder.py
pytorch_lightning/configs/shape-autoencoder/Dora-VAE-test.yaml

So you can run the
test_autoencoder_single_gpu.sh script to extract the latent code for preprocessing.

Also added these 2 scripts for a quick way to decode the extracted latent codes to make sure the extracted latent codes are working (might also be useful for inference once I get a dit set up):

pytorch_lightning/decode_latent_code.py
pytorch_lightning/model_loader.py

[rocksat]

I guess the best way for pure inference version is convert it to diffusers pipeline, and host the model weights to huggingface model hub.