Reproducible evaluation results

[cjerry1243]

Thanks for the impressive work, @jpthu17

I have a couple of questions regarding the reproducibility.

1. I followed the instruction to run the evaluation for the provided GraphMotion.ckpt, but only received the FID of 0.174. See the screenshot below for the results.

I wonder if it's possible to achieve the best results reported in the paper (FID close to 0.116)

2. I am confused about the three vae training here https://github.com/TL-QZ/GraphMotion_dev/tree/main?tab=readme-ov-file#21-ready-to-train-vae-model

Do we train run the training commands one after another, with the learned weights specified in the config (TRAIN.PRETRAINED_VAE) of the next step?

How do we specify the three pre-trained vae checkpoints in the config_humanml3d.yaml for final training?

Thanks in advance for your response!