search

TruongKhang / cds-mvsnet

[ICLR2022] Curvature-guided dynamic scale networks for Multi-view Stereo
118 stars 6 forks source link

depth_interals_ratio #3

Closed liushifu12138 closed 2 years ago

liushifu12138 commented 2 years ago

depth_interals_ratio: [4.0, 2.0, 1.0] [4.0, 1.5,0.75] which is beter?

TruongKhang commented 2 years ago

Hi, I think the second is better but the difference is not too large as far as I remember. Because of the robust learned features, our method can process a smaller depth interval at the final stage compared to the other methods.

liushifu12138 commented 2 years ago

Thank you for your answer. When I interrupted the training and resumed the training, the loss became very large, and the learning rate returned to 0.01. How to solve this situation? python train.py --resume /mnt/cds/saved/models/CDS-MVSNet/0213_150704/checkpoint-epoch11.pth

TruongKhang commented 2 years ago

After finishing training on DTU, I fine-tuned on BlendedMVS dataset by using python train.py --resume saved/models/CDS-MVSNet/<date_and_year>/checkpoint-epoch30.pth as described in README.md. Note that the configuration now should be changed in file config.json with respect to your BlendedMVS. So, in your case, you should modify the learning rate in file /mnt/cds/saved/models/CDS-MVSNet/0213_150704/config.json with a smaller value because I didn't set up to resume the learning rate from pretrained model in my code

About the loss problem, how much is it large? Can you show me some pictures?

liushifu12138 commented 2 years ago

I don't have any pictures, depth_loss of epoch11: 0.426, when I interrupt training and resume, depth_loss=0.8, I try to manually modify the learning rate and the final result will not affect, otherwise I have to start training again

TruongKhang commented 2 years ago

@liushifu12138 , Do you mean the depth loss here is on the training set or validation set?

liushifu12138 commented 2 years ago

validation set

TruongKhang commented 2 years ago

I think this difference is normal because when resuming the training, I only load the network weights, without loading the optimization state (you can see in _resume_checkpoint function in base/base_trainer.py). So the result may change. But you don't need to retrain from scratch, I think the loss will converge soon in several epochs.

liushifu12138 commented 2 years ago

Thank you very much for your answer and thanks for your project