search

fh2019ustc / DocTr

The official code for “DocTr: Document Image Transformer for Geometric Unwarping and Illumination Correction”, ACM MM, Oral Paper, 2021.
Other
345 stars 48 forks source link

GPU memory for training #19

Closed Aiden0609 closed 1 year ago

Aiden0609 commented 1 year ago

Hi, splendid work you've done! I've been considering to reproduce it but I don't know what kind of hardware conditions are recommended to successfully reproduced the training of Geo and Illu network, so I wish to know the GPU memory consumed for your training process. Is it 4 1080ti GPUs that you've used? Tks. I really need to know this!

fh2019ustc commented 1 year ago

Hi~ Thanks for your attention to our work. Yeah, you can use 4 1080ti GPUs to reproduce DocTr. For us, the batch size for each GPU is 3 and the total batch size is 12. Hope it helps

Aiden0609 commented 1 year ago

Thank you! Quite a large amount of GPU memory it consumed, I'll see if my hardware is capable.

fh2019ustc commented 1 year ago

DocTr does not include the BN layer. You can use two 2 1080ti GPUs to train the network. We do not ablate the batch size. Maybe batch size 6 performs better.