Hardware used and training time

[MKaczkow]

Hi 🙂 Could you share your hardware, used to train the model and how long did it took, please?

I am also trying to re-implement StereoNet and training times are:

• c.a. 3,5h / epoch on SceneFlow (pretrain)
• c.a. 3 min / epoch on KITTI (finetune)

I wonder if those could be optimized. From what I've observed right now, the first bottleneck is loading batches into GPU memory (datasets are obviously to big to fit into it at once), the second limitation is batch_size - if I go > 4 CUDA throws 'out of memory' error.