1:1 Testing on custom data #81
Description
When I evaluate a model checkpoint using the eval.sh script with --skip_nll 0, it generates sample_xx.pt and recont_xx.pt. However, these two files contain different numbers of samples.
What I need is a one-to-one mapping: for each input point cloud, I want to obtain both the reconstructed output from the VAE (without diffusion) and the corresponding output from the diffusion model. This would allow me to directly compare both results to the reference input.
From what I can tell, the current evaluation script uses the val/ folder to generate recont_xx.pt, while the diffusion samples in sample_xx.pt appear to be generated arbitrarily and not conditioned on the same inputs.
How can I modify the evaluation process (or is there a flag) to ensure that both outputs—VAE and diffusion—correspond to the exact same input point clouds?