Test invocation questions

[KevinCain]

In what file is TEST_DATA_FOLDER set? Similarly, could you please clarify what is intended by "Set MODEL_FOLDER to ckpt and model_ckpt_index to checkpoint_list."? That is, what files should be modified?

[mazeyu]

I just have tried their code. I think the readme is confusing or outdated. I just looked into the eval_tanks.sh and modified the TP_TESTING and parameter --loadckpt directly. And that works. As long as you have cams and images folders in your dataset.

(But my question as I posted in another issue is that the parameters seem to be inconsistent with what they mentioned in their paper. I didn't get good results on my first attempt on Tanks dataset. I am trying to modify other parameters)

[weizizhuang]

I will add some details in the README file, hope this can help you.

[KevinCain]

Thank you @weizizhuang for amending your README, and to @mazeyu for your cross-check.

I placed the pretrained model in the 'ckpt' folder: /home/kevin/D2HC-RMVSNet/ckpt/model.ckpt

From the script, I gathered that TEST_DATA_FOLDER should be updated, e.g. for Tanks & Temples and DTU: TEST_DATA_FOLDER="/home/kevin/tankandtemples/intermediate" DTU_TESTING="/home/kevin/dtu"

At runtime for ./eval_tanks.sh FileNotFoundError: [Errno 2] No such file or directory: './checkpoints/model_000004.ckpt'

Simply supplying the provided 'model.ckpt' by download directly as @mazeyu may have done gives another error in my case: RuntimeError: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False.

Preliminaries:

Check NVIDIA Drivers:

• nvidia-smi reports: NVIDIA-SMI 460.67 Driver Version: 460.67 CUDA Version: 11.2
  Check CUDA:
• nvcc --version reports: Cuda compilation tools, release 10.1, V10.1.243 Check CuDNN:
• /sbin/ldconfig -N -v $(sed ‘s/:/ /’ <<< $LD_LIBRARY_PATH) 2>/dev/null | grep libcudnn reports: libcudnn.so.7 -> libcudnn.so.7.6.5

Create conda workspace for D2HC-RMVSNet:

• ./conda_install.sh
• conda activate drmvsnet
• Install TorchVision
  • conda install pytorch torchvision cudatoolkit=10.1 pytorch
• Install SciPi
  • conda install -c anaconda scipy

[KevinCain]

I see torch.cuda.is_available() now returns true after reinstalling pytorch with CUDA 11.1 support as per conda install pytorch torchvision torchaudio cudatoolkit=11.1 -c pytorch -c conda-forge.

However, I'm still getting the error above: RuntimeError: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False.