How many epochs have you trained on the Aria Dataset

[zipxup]

Thank you for your great work!

I've attempted to use your training script to conduct training on the Aria Dataset. However, my results indicate that foot slipperiness occurs more frequently compared to the outcomes from your shared pre-trained checkpoint. Please take a look at the video below. Is the result resonable?

I'm curious about the number of epochs you trained on the Aria Dataset. I would greatly appreciate it if you could reply as soon as possible.

test_25K-2025-04-02_17.35.27.-.Trim.mp4

[ZhengyiLuo]

Hi! Thanks for your interest in my work.

There is eval during training; what's the current success rate you are reaching and mpjpe on the training set? On ADT I think I trained for like 6000 epochs of distillation.

[zipxup]

Thank you for your response.

I checked the training log, the success rate and mpjpe value, which are 0.0 and nan respectively. This is weird. I didn't change any of your scripts, I just put "headless=True" to disable visualization.

The wandb log does show values, please check the image below, do you mean these values?

[ZhengyiLuo]

Thank you for your response.

I checked the training log, the success rate and mpjpe value, which are 0.0 and nan respectively. This is weird. I didn't change any of your scripts, I just put "headless=True" to disable visualization.

The wandb log does show values, please check the image below, do you mean these values?

Yes I meant success rate and mpjpe during training. For the training tqdm 0.0 and nan is normal; those are not properly logged if the dataset is too small (tqdm hasn't had the chance to be updated). The wandb looks overall good but a little too jumpy somehow. This is what mine looks like:

The aria dataset is quite small so training for 25k is probably overfitting; have you evaluated some of the earlier checkpoints?