cant git clone repo

[zoldaten]

hi!
i do not use conda so trying :
git clone --recursive https://github.com/Chaphlagical/Deblur-GS.git
but failed with submodules:
diff-gaussian-rasterization and knn.

please help!

[Chaphlagical]

Sorry, I forgot to make my diff-gaussian-rasterization public, knn should work though. You can try again.

[zoldaten]

ok. i git cloned it. built by hands each module - python setup.py install.

how to inference ? should i train new model first ?

if so i cant start train -
python train.py -s C:\Users\{user}\Desktop\2\Deblur-GS\exblur_release\bench --eval

it asks me to open port ?! if i deny i starts new session and crash:

train.py 437 <module>
training(lp.extract(args), op.extract(args), pp.extract(args), args.test_iterations,

train.py 66 training
scene = Scene(dataset, gaussians)

__init__.py 60 __init__
assert False, "Could not recognize scene type!"

AssertionError:
Could not recognize scene type!

ps. i have also install visdom, torchmetrics, easydict, downgrade protobuf==3.20.*

[giich]

Sorry, I forgot to make my diff-gaussian-rasterization public, knn should work though. You can try again.

knn is not public though?

[Chaphlagical]

ok. i git cloned it. built by hands each module - python setup.py install.

how to inference ? should i train new model first ?

if so i cant start train - python train.py -s C:\Users\{user}\Desktop\2\Deblur-GS\exblur_release\bench --eval

it asks me to open port ?! if i deny i starts new session and crash:

train.py 437 <module>
training(lp.extract(args), op.extract(args), pp.extract(args), args.test_iterations,

train.py 66 training
scene = Scene(dataset, gaussians)

__init__.py 60 __init__
assert False, "Could not recognize scene type!"

AssertionError:
Could not recognize scene type!

ps. i have also install visdom, torchmetrics, easydict, downgrade protobuf==3.20.*

Some issue with the dataset, I replace with a new one.

Sorry, I forgot to make my diff-gaussian-rasterization public, knn should work though. You can try again.

knn is not public though?

knn might be the path issue, you can update the repo and try again

[zoldaten]

still not working train:

training(lp.extract(args), op.extract(args), pp.extract(args), args.test_iterations,

train.py 66 training
scene = Scene(dataset, gaussians)

__init__.py 84 __init__
midas = torch.hub.load(

hub.py 566 load
model = _load_local(repo_or_dir, model, *args, **kwargs)

hub.py 592 _load_local
hub_module = _import_module(MODULE_HUBCONF, hubconf_path)

hub.py 106 _import_module
spec.loader.exec_module(module)

<frozen importlib._bootstrap_external> 879 exec_module


<frozen importlib._bootstrap_external> 1016 get_code


<frozen importlib._bootstrap_external> 1073 get_data


FileNotFoundError:
2
No such file or directory
C:\home/cwb/.cache/torch/hub/intel-isl_MiDaS_master\hubconf.py```

[Chaphlagical]

still not working train:

training(lp.extract(args), op.extract(args), pp.extract(args), args.test_iterations,

train.py 66 training
scene = Scene(dataset, gaussians)

__init__.py 84 __init__
midas = torch.hub.load(

hub.py 566 load
model = _load_local(repo_or_dir, model, *args, **kwargs)

hub.py 592 _load_local
hub_module = _import_module(MODULE_HUBCONF, hubconf_path)

hub.py 106 _import_module
spec.loader.exec_module(module)

<frozen importlib._bootstrap_external> 879 exec_module


<frozen importlib._bootstrap_external> 1016 get_code


<frozen importlib._bootstrap_external> 1073 get_data


FileNotFoundError:
2
No such file or directory
C:\home/cwb/.cache/torch/hub/intel-isl_MiDaS_master\hubconf.py```

I set midas model to local for efficient. It is fixed I think.

[zoldaten]

ok, i trained model with:
python train.py -s C:\Users\{user}\Desktop\2\Deblur-GS\exblur_release\bench --eval
got in output folder events.out.tfevents.1715678047.ws5411.864.0 - 1Gb.

executed python render_video.py -m C:\Users\{user}\Desktop\2\Deblur-GS\output\4744193f-5
got novel_view folder with depth and rgb dirs.

but in rgb dir i do not see any deblurred images:

may be i trained wrong as initial i got bench dir like downloaded:

but there was no any error messages while training ...

[giich]

ok, i trained model with: python train.py -s C:\Users\{user}\Desktop\2\Deblur-GS\exblur_release\bench --eval got in output folder events.out.tfevents.1715678047.ws5411.864.0 - 1Gb.

executed python render_video.py -m C:\Users\{user}\Desktop\2\Deblur-GS\output\4744193f-5 got novel_view folder with depth and rgb dirs.

but in rgb dir i do not see any deblurred images:

may be i trained wrong as initial i got bench dir like downloaded:

but there was no any error messages while training ...

Have you run render.py? Not render_video.py but render.py?

[zoldaten]

Have you run render.py? Not render_video.py but render.py?

yes, i did that before render_video.py:

python render.py -m <path to trained model> # Generate renderings
python metrics.py -m <path to trained model> # Compute error metrics on renderings

[Chaphlagical]

ok, i trained model with: python train.py -s C:\Users\{user}\Desktop\2\Deblur-GS\exblur_release\bench --eval got in output folder events.out.tfevents.1715678047.ws5411.864.0 - 1Gb.

executed python render_video.py -m C:\Users\{user}\Desktop\2\Deblur-GS\output\4744193f-5 got novel_view folder with depth and rgb dirs.

but in rgb dir i do not see any deblurred images:

may be i trained wrong as initial i got bench dir like downloaded:

but there was no any error messages while training ...

You should use command line parameter --deblur to enable deblur mode

[Chaphlagical]

ok, i trained model with: python train.py -s C:\Users\{user}\Desktop\2\Deblur-GS\exblur_release\bench --eval got in output folder events.out.tfevents.1715678047.ws5411.864.0 - 1Gb.
executed python render_video.py -m C:\Users\{user}\Desktop\2\Deblur-GS\output\4744193f-5 got novel_view folder with depth and rgb dirs.
but in rgb dir i do not see any deblurred images:
may be i trained wrong as initial i got bench dir like downloaded:

but there was no any error messages while training ...

You should use command line parameter --deblur to enable deblur mode

Perhaps I should set true as default.

[zoldaten]

i started training but it seems no cuda utilization ( it will take 18 hours to train.

although in output in cfg_args it is presented:
Namespace(sh_degree=3, source_path='C:\\Users\\___\\Desktop\\2\\Deblur-GS\\exblur_release\\bench', model_path='./output/59ea1996-6', images='images', resolution=-1, white_background=False, data_device='cuda', bezier_order=7, mode='Linear', eval=True, visdom_server='127.0.0.1', visdom_port=2333)

[zoldaten]

ok. got it!
after all i join images from novel_view\rgb to video:
ffmpeg -framerate 24 -i %05d.png Project.mp4

looks good!

Project.mp4

but could you provide how to make custom dataset ? is it hard enough ?

[giich]

Use a colmap?