search

zzangjinsun / NLSPN_ECCV20

Park et al., Non-Local Spatial Propagation Network for Depth Completion, ECCV, 2020
MIT License
321 stars 55 forks source link

effects on hyper-params in training ? #28

Closed Wenchao-Du closed 3 years ago

Wenchao-Du commented 3 years ago

Hi, i have tried the released model for kitti dc and got the consistent results. But i have some problems on two hyper-params in config file: preserve_input and legacy, i closing any one of them would significant errors compared the ground truth, so i want to know if this two params are opened in training ? i have tried to retrain your model with closing this two params, but open them during testing , it leads to larger errors. Instead, closing them leads to consistent results with training during testing. Thanks.

zzangjinsun commented 3 years ago

For the pre-trained model, you must use --preserve_input and --legacy. However, if you wanna train your own model, --legacy should not be used but --preserve_input is up to you.