README.md

NerfLightning - Simple NERF Archive

Project Website: https://khoidoo.github.io/nerf-lightning/

Setup

python3 -m venv .env
source .env/bin/activate
python -m pip install -U pip

pip install wheel
pip3 install torch torchvision torchaudio
python -m pip install lightning
pip install omegaconf
pip install opencv-python

Training

All training can be written in a YAML file. An example is located at ./configs/tinynerf.yaml, shown below

name: nerf
save_dir: runs
trial_dir: ${get_trial_dir:${save_dir}}
save_cfg_path: ${path_append:${trial_dir},config.yaml}
seed: 0

data:
  data_source: data/src
  train_path: ${path_append:${data.data_source},training_data.pkl}
  valid_path: ${path_append:${data.data_source},testing_data.pkl}
  batch_size: 1024
  shuffle: ${get_shuffle:${train.trainer.devices}}
  num_workers: 24

system_type: TinyNerf
system:
  model_type: NerfModel
  model:
    hidden_dim: 256
    embedding_dim_pos: 10
    embedding_dim_direction: 4
  optimizer:
    name: Adam
    args:
      lr: 5.e-4
  scheduler:
    name: MultiStepLR
    args:
      milestones: [2, 4, 8]
      gamma: 0.5
      last_epoch: -1
  loss: 
    name: mse
    args:
      reduction: mean
  args:
    near: 2
    far: 6
    bins_count: 192
    eval_height: 400
    eval_width: 400

train:
  trainer:
    devices: [0]
    max_epochs: 20
    check_val_every_n_epoch: 1
    enable_progress_bar: True
    accumulate_grad_batches: 1
    log_every_n_steps: 50
    default_root_dir: ${trial_dir}
  logger:
    names: [WandbLogger, CSVLogger]
    args:
      wandb:
        name: ${get_run_id:${trial_dir}}
        project: nerf
        save_dir: ${path_append:${trial_dir},wandb}
        id: ${get_run_id:${trial_dir}}
        anonymous: True
        log_model: all
      csv:
        name: csv
        save_dir: ${trial_dir}
  callback:
    names: [ModelCheckpoint]
    args:
      modelcp:
        monitor: train/psnr
        mode: max

Note that the repo supports all Pytorch's optimizers and Pytorch's Schedulers, specifying the accurate name from Pytorch and place all available args into args dictionary. The repo also supports all Lightning's Logger and Lightning's Callbacks. To start training using CLI:

python main.py --config ./configs/tinynerf.yaml