search

nv-tlabs / ATISS

Code for "ATISS: Autoregressive Transformers for Indoor Scene Synthesis", NeurIPS 2021
Other
264 stars 55 forks source link

Hardware/training time/epochs etc #9

Closed wamiq-reyaz closed 2 years ago

wamiq-reyaz commented 2 years ago

Hello,

I was wondering what is the expected time to train - the paper only mentions that you choose the best model from a very large number of iterations. What hardware did you use to train the model on and how long does that usually take, just to get a ballpark figure?

In the same vein, did you (or coauthors) use a different lr schedule, optimizer etc that seems to work better?

PS: Do you have any intentions of releasing the Kaolin scripts used to render the figures in the paper. They look very very nice!

wamiq-reyaz commented 2 years ago

A follow-up would be what are the loss ranges for the size, translation, angle, and label that you expect or report the figures in the paper for?

paschalidoud commented 2 years ago

Hi @wamiq-reyaz,

We run all our experiments on an NVIDIA GeForce GTX 1080 Ti GPU. Training lasted on average around 1-1.5 days. We experimented with different lr scheduling and optimizers such as RAdam, but we didn't experience any difference in performance, thus we ended up using a learning rate of 1e-4 without weight decay.

Regarding loss ranges, for the trained model that we used for generating the bedroom scenes in the paper, the ranges that we observed for each loss term are summarized below:

Please note that the above numbers can vary among runs.

Best, Despoina

paschalidoud commented 2 years ago

@wamiq-reyaz, I am closing this issue for now, but please feel free to reopen in case you have more questions!

Best, Despoina

Jingyu6 commented 1 year ago

Hi @wamiq-reyaz,

We run all our experiments on an NVIDIA GeForce GTX 1080 Ti GPU. Training lasted on average around 1-1.5 days. We experimented with different lr scheduling and optimizers such as RAdam, but we didn't experience any difference in performance, thus we ended up using a learning rate of 1e-4 without weight decay.

Regarding loss ranges, for the trained model that we used for generating the bedroom scenes in the paper, the ranges that we observed for each loss term are summarized below:

  • label loss: 1.235 - 0.9793
  • size NLL loss: 12.48 - 3.042
  • rotation NLL loss: 4.773 - 0.4792
  • translation NLL loss: 14.301 - 5.832

Please note that the above numbers can vary among runs.

Best, Despoina

Hi @paschalidoud, would you mind also sharing the validation stats for this? When I train the model (despite on the new version of the dataset), it seems to overfit the training data quite easily so I want to see what stats you got when you trained the ones used in the paper (maybe for bedrooms only since I'm focusing on this atm)? Thanks a lot in advance!

Best, Jingyu