search

dulucas / siMLPe

A Simple Baseline for Human Motion Prediction
MIT License
134 stars 20 forks source link

Reproduction on the AMASS Dataset #10

Closed Deante-dx closed 2 months ago

Deante-dx commented 1 year ago

I have been attempting to replicate the results on AMASS reported in your paper. However, I have encountered some challenges and discrepancies while trying to reproduce the reported outcomes (10.8 19.6 34.3 40.5 50.5 57.3 62.4 65.7). I have (10.9, 19.8, 34.9, 41.1, 51.2, 58.1, 63.1, 66.8) I would greatly appreciate your guidance on this matter. Could you kindly provide any insights or suggestions as to why I might be experiencing these discrepancies?

Logan-007L commented 9 months ago

Hello, could you tell me which amass data sets I need? All of them? Or just pick some of them such as ACCAD?

dulucas commented 5 months ago

I have been attempting to replicate the results on AMASS reported in your paper. However, I have encountered some challenges and discrepancies while trying to reproduce the reported outcomes (10.8 19.6 34.3 40.5 50.5 57.3 62.4 65.7). I have (10.9, 19.8, 34.9, 41.1, 51.2, 58.1, 63.1, 66.8) I would greatly appreciate your guidance on this matter. Could you kindly provide any insights or suggestions as to why I might be experiencing these discrepancies?

Sorry I can't recall the details as it has been a while. It may due to the randomness of the exp(the hardware/environment/seed etc). I did recall that sometimes reload the model from 80000 or 100000 iters w/o optimizer and then do the finetuning with lower lr could help.

dulucas commented 5 months ago

Hello, could you tell me which amass data sets I need? All of them? Or just pick some of them such as ACCAD?

We use a split of AMASS dataset following the setting of previous pubs. Here you can find the train/test split for AMASS.