Yasoz/DiffTraj

Training Duration - Number of Steps / Epochs

Closed this issue · 3 comments

erik-buchholz commented

Dear Yuanshao,

I have a short question regarding the duration of the training. How long did you train for on the different datasets? The config (utils/config.py) contains both n_epochs (=200) and n_iters (=5 000 000) which are conflicting?
For instance, on Chengdu:

200 epochs = 3 493 918 / 1024 * 200 = 682 405 iterations

What did you use in your evaluation, the 200epochs or the n_iters values, or something completely different? I would like to run your model on another dataset and stay as close as possible to your results in order to allow for a fair comparison.

Thanks a lot for your help!

Cheers,
Erik

Yasoz commented

Hi Erik,

The n_iters is not used in this project. This parameter was kept in the original code base, and I am sorry for forgetting to delete it. I only used the epoch here.

Best,
Yuanshao

erik-buchholz commented

Hi Yuanshao,

Thank you so much for confirming this!
Did you use 200 epochs for both your datasets, or did you use 200 epochs only on Chengu but used a different number of epochs for Xi'an?

BTW. I finally managed to get your model to work on my dataset. Thank you so much for your help!

Cheers,
Erik

Yasoz commented

Both datasets use 200 epoches. Actually, 100 epochs is sufficient to get good results.

Best,
Yuanshao

  • 👍1