Improved instant-NGP model

This repo is an improved version of another Instant-NGP repo, and bases on pytorch implementation.

Dependencies

So far this has only been tested on Roger's machine. Any feedback is welcomed. Following installations are tested on a CUDA11.3 machine.

Python libraries

• Install pytorch>=1.11.0 by pip install torch==1.11.0 torchvision==0.12.0 --extra-index-url https://download.pytorch.org/whl/cu113
• Install torch-scatter following their instruction
• Install core requirements by pip install -r requirements.txt

TinyCudaNN

This repo relies on TinyCudaNN.

It is recommended to build tTinyCudaNN on a CUDA-11.3 machine.

Installation steps are as follows:

1. git clone --recursive https://github.com/NVlabs/tiny-cuda-nn.git
2. cd tiny-cuda-nn
3. Use your favorite editor to edit include/tiny-cuda-nn/common.h and set TCNN_HALF_PRECISION to 0 (see NVlabs/tiny-cuda-nn#51 for details)
4. cd bindings/torch
5. python setup.py install

Compile CUDA extension of this project

Run pip install models/csrc/ (please run this each time you pull the code)

Preparing Data

• download preprocessed data at https://drive.google.com/uc?id=10Tj-0uh_zIIXf0FZ6vT7_te90VsDnfCU
• put under datasets/TanksAndTempleBG/

Run examples!

python train.py --config configs/Playground.txt

This code will validate your model when training procedure finishes.

Resume training!

python train.py --config configs/Playground.txt --ckpt_path PATH/TO/CHECKPOINT/DIR/epoch={n}.ckpt

There is a bug of pytorch lightning regarding to progress bar(see Lightning-AI/pytorch-lightning#13124 for details).

Validate your model!

python train.py --config configs/Playground.txt --ckpt_path PATH/TO/CHECKPOINT/DIR/epoch={n}.ckpt --val_only

Renderings

python render.py --config configs/Playground.txt --weight_path PATH/TO/SLIM/CHECKPOINT/DIR/epoch={n}_slim.ckpt