search

stelzner / osrt

Independent PyTorch Implementation of Object Scene Representation Transformer
MIT License
46 stars 8 forks source link

The checkpoint of model trained on CLEVR-3D #2

Closed Anqw closed 1 year ago

Anqw commented 1 year ago

Hi, thanks for reproducing this interesting work. Could you please provide the checkpoint of model trained on CLEVR-3D?

stelzner commented 1 year ago

I just added a CLEVR-3D checkpoint to the readme. Hope this helps!

takerum commented 1 year ago

Thank you very much for providing the checkpoint! Let me ask another question: How long does it take to get to that checkpoint? When I run the code in my environment (8xA100 GPUs), I found it would take ~10 days to complete the total iterations (4x10^6), which I think is relatively long. I suspect there is something wrong with my environment, and I just would like to know whether such a long time training is due to my environment, or just normal behavior.

stelzner commented 1 year ago

That sounds about right, actually. I've found the model to converge quite a bit sooner though, hence the checkpoint sits at about 1M iterations. That took about 5 days for me to train, but I don't quite remember how many GPUs I was using.