search

ewrfcas / MVSFormer

Codes of MVSFormer: Multi-View Stereo by Learning Robust Image Features and Temperature-based Depth (TMLR2023)
Apache License 2.0
175 stars 10 forks source link

about the GPU using #24

Closed xingchen2022 closed 1 year ago

xingchen2022 commented 1 year ago

Thanks for your excellent work! I tried to retrain your model on 2 Rtx8000 without changing any hyper-parameters(batchsize=8), and found it used about 80G GPU memory. Science you trained on 2 V100 GPU with 64G memory, I wonder how much memory cost on your device and why it had more than 10G difference?
image

it will be appreciated if you could give me some explanation, thank you so much!

ewrfcas commented 1 year ago

We have tested these codes on 48G GPUs (A6000). For 32G GPU, we have to reduce the sub-batchsize slightly. Besides, the displayed memory cost would differ for different GPU types (such as RTX8000 vs V100).