Model Checkpoints

[Nkcemeka]

@NilsDem and @anonymous781384 ,

This is great work. Eagerly awaiting the release on the model checkpoints. It will be great to test the model before committing to recreating the paper's results. Some steps on how to handle the data will be appreciated. E.g The paper indicated you used several datasets and from reading the code, it appears you used a lmdb database to handle all of this. If you have any tips on how to go about doing this effectively, that will be appreciated!

[NilsDem]

Thanks for your interest on the model. I just released the model checkpoints and inference notebooks !

Regarding the multi-dataset, you can use the CombinedDataset :
https://github.com/NilsDem/control-transfer-diffusion/blob/3817465dde6873e67cd90ddeadab7d3d950c7ef8/dataset/dataset.py#L57C1-L57C49

You just need to specify a dict with the different lmdb paths (individually processed with split_to_lmdb.py) and the sampling frequencies for each dataset (putting 1. for each means all datasets are equally sampled).

You can modify the training script to specify a CombinedDataset instead of usual SimpleDataset.

Hope that helps !

[Nkcemeka]

@NilsDem Thank you.