baaivision / geodream Goto Github PK

hi, when encountering this bug: NameError: name 'TetrahedraSDFGrid' is not defined
maybe you should change "TetrahedraSDFGrid" to "TetrahedraSDFGridGeoDream" in this file:
custom/threestudio-geodream/models/geometry/tetrahedra_sdf_grid.py

I tried to use the extension in a colab enviornment, when contructing the cost volume with sh step1-run-mv.sh "An astronaut riding a horse" the execution runs into a loop with torch/serialization.persistent_laod() and torch/_utils._rebuild_tensor_v2.

maybe there are some underlying problems because the installation also does not work as described. Starting with the install for predicting source views:

conda create --name geodream_mv python=3.8 conda activate geodream_mv pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118 pip install inplace_abn sudo apt-get install libsparsehash-dev pip install git+https://github.com/mit-han-lab/[email protected] pip install -r requirements.txt

The torchsparse link does not work and that was only part of the problems encountered when using the extension

Cause

(geodream) root@c757b388f2d0:/data/workspace/GeoDream/mv-diffusion# bash run-volume-by-sd-zero123.sh "An astronaut riding a horse" "ref_imges/demo.png"
[Save Project Name] An_astronaut_riding_a_horse
[Reference View Path] GeoDream/mv-diffusion/One-2-3-45-by-view/ref_imges/demo.png
Instantiating LatentDiffusion...
Traceback (most recent call last):
  File "run.py", line 115, in <module>
    predict_multiview(shape_dir, args)
  File "run.py", line 80, in predict_multiview
    models = init_model(device, 'stable_zero123.ckpt', half_precision=args.half_precision)
  File "/data/workspace/GeoDream/mv-diffusion/One-2-3-45-by-view/utils/zero123_utils.py", line 45, in init_model
    models['turncam'] = torch.compile(load_model_from_config(config, ckpt, device=device)).half()
AttributeError: module 'torch' has no attribute 'compile'

First thank you for your great work!

I didn't quite understand what the "parsed.yaml" means in mesh extraction. Could you please specify it?

Hello there, just wondering will the evaluation code of Uni3D Score be open sourced soon?

Hello, thank you for your brilliant work!
I am confused when running Construct cost volume stage. I find that, after Predict source views, I have already obtained con_volume_lod_150.pth. So do I still need to run Construct cost volume stage? Or I can directly go to GeoDream Training?

Thanks a lot :)

information

• tested on LambdaCloud's A10 24GB PCle GPU instance
• ubuntu 22.04

Tue Apr 16 09:31:55 2024       
+---------------------------------------------------------------------------------------+
| NVIDIA-SMI 535.129.03             Driver Version: 535.129.03   CUDA Version: 12.2     |
|-----------------------------------------+----------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |         Memory-Usage | GPU-Util  Compute M. |
|                                         |                      |               MIG M. |
|=========================================+======================+======================|
|   0  NVIDIA A10                     On  | 00000000:08:00.0 Off |                    0 |
|  0%   28C    P8               9W / 150W |      4MiB / 23028MiB |      0%      Default |
|                                         |                      |                  N/A |
+-----------------------------------------+----------------------+----------------------+
                                                                                         
+---------------------------------------------------------------------------------------+
| Processes:                                                                            |
|  GPU   GI   CI        PID   Type   Process name                            GPU Memory |
|        ID   ID                                                             Usage      |
|=======================================================================================|
|  No running processes found                                                           |
+---------------------------------------------------------------------------------------+

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2023 NVIDIA Corporation
Built on Tue_Aug_15_22:02:13_PDT_2023
Cuda compilation tools, release 12.2, V12.2.140
Build cuda_12.2.r12.2/compiler.33191640_0

What is the proper way to install GeoDream?

The way I did

1. install threestudio via Dockerfile

2. install GeoDream's threestudio branch on threestudio's custom directory

   cd custom
   git clone -b threestudio https://github.com/baaivision/GeoDream.git
   mv GeoDream threestudio-geodream
   

3. now what should I do next?

   the mv-diffusion's README.md says :

   conda create --name geodream_mv python=3.8
   conda activate geodream_mv
   pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118
   pip install inplace_abn
   sudo apt-get install libsparsehash-dev
   pip install git+https://github.com/mit-han-lab/[email protected]
   pip install -r requirements.txt

   but I'm on a docker container now.
   And the cuda version on my GPU instance(12.2) mismatches with the installation guide's version(11.8).
   I'm so confused how to properly install GeoDream.
   Please help, thanks in advance.