c46ff4f75f
* cleaning up requirements.txt and adding versions * need to update github actions, docs, and make optional dependencies work * updated documentation * set directory * new functorch version * updated caching * fixed hash
4.2 KiB
Installation
This repository is tested with CUDA 11.3. Make sure to install Conda before preceding.
Installing Conda
This step is fairly self-explanatory, but here are the basic steps. You can also find countless tutorials online.
```bash
cd /path/to/install/miniconda
mkdir -p miniconda3
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O miniconda3/miniconda.sh
bash miniconda3/miniconda.sh -b -u -p miniconda3
rm -rf miniconda/miniconda.sh
```
Create the python environment
conda create --name nerfactory python=3.8.13
conda activate nerfactory
Clone the repo
git clone git@github.com:plenoptix/nerfactory.git
Install dependencies and nerfactory as a library
pip install -e .
Install library with CUDA development support. Change setup.py to USE_CUDA = True and then
python setup.py develop
Install tiny-cuda-nn (tcnn) to use with the graph_instant_ngp.yaml config
pip install git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch
Downloading data
Download the original NeRF dataset and unfold it in the following format. This is for the blender dataset type. We support the major datasets and allow users to create their own dataset, described in detail here.
|─ nerfactory/
├─ data/
| ├─ blender/
| ├─ fern/
| ├─ lego/
...
|- <dataset_format>/
|- <scene>
...
Training a model
To run with all the defaults, e.g. vanilla nerf method with the blender lego image
Run a vanilla nerf model.
python scripts/run_train.py
Run a faster version with instant ngp using tcnn (without the viewer).
python scripts/run_train.py --config-name=graph_instant_ngp.yaml
Run with the viewer. However, you'll have to start the viewer server first. (See viewer docs)
python scripts/run_train.py --config-name=graph_instant_ngp.yaml viewer.enable=true
With support for Hydra, you can run with other configurations by changing appropriate configs defined in configs/ or by setting flags via command-line arguments.
Visualizing training runs
If you using a fast NeRF variant (ie. Instant-NGP), we reccomend using our viewer. See our viewer docs for more details. The viewer will allow interactive visualization of training in realtime.
Additionally, if you run everything with the default configuration, by default, we use TensorBoard to log all training curves, test images, and other stats. Once the job is launched, you will be able to track training by launching the tensorboard in outputs/blender_lego/vanilla_nerf/<timestamp>/<events.tfevents>.
tensorboard --logdir outputs
Rendering a Trajectory
To evaluate the trained NeRF, we provide an evaluation script that allows you to do benchmarking (see our benchmarking workflow) or to render out the scene with a custom trajectory and save the output to a video.
python scripts/run_eval.py render-trajectory --load-config=outputs/blender_lego/instant_ngp/2022-07-07_230905/config.yml --traj=spiral --output-path=output.mp4
Please note, this quickstart allows you to preform everything in a headless manner. We also provide a web-based visualizer that allows you to easily monitor training or render out trajectories. See our viewer docs for more.
FAQ
(tiny-cuda-error)=
TinyCUDA installation errors out with cuda mismatch
While installing tiny-cuda, you run into: The detected CUDA version mismatches the version that was used to compile PyTorch (10.2). Please make sure to use the same CUDA versions.
Solution:
pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html