search

jimmy19991222 / ELFNet

[ICCV 2023] ELFNet: Evidential Local-global Fusion for Stereo Matching
44 stars 5 forks source link

Pretrained Model Weights #1

Closed LzqInsta closed 1 year ago

LzqInsta commented 1 year ago

Hello! Thanks for your great work and kindly opening source code. Would you release your pretrained model in near future, which can reproduce the results in the paper

jimmy19991222 commented 1 year ago

Hi, thank you for your interest in the project. We will upload the pretrained model soon.

LzqInsta commented 1 year ago

thanks for your quick reply. Since I have encountered a problem of ”CUDA out of memory“ on a 3090 gpu for training，a pretrained model is necessary for me.

jimmy19991222 commented 1 year ago

Hi, I have released the pretrained weight. You can download from this google drive link (https://drive.google.com/file/d/16ViZY16-rtTK1oFnMAmlahh3gXfEAVCP/view?usp=share_link).

torch.cuda.checkpoint (in sttr) and torch.cuda.amp are applied to optimize memory usage during training to avoid OOM problem. If you still cannot train the model, you can apply torch.cuda.checkpoint on cost-volume-based model as well. You can try to adjust the size of random crop to save memory but may affect results. https://github.com/jimmy19991222/ELFNet/blob/9b0cb450fc8c526906f0356a121337aa36a42e4b/dataset/stereo_albumentation.py#L114

LzqInsta commented 1 year ago

thanks for your advice and released model weights!