Complete the 2nd NeRF Development

[naufalso]

• Fix chunk predictions and refactor NeRF utils

  • Rewrite the code to chunk the whole pipeline instead of only on MLP
  • Initialize TF Variable to store the accumulated loss and gradients
  • Refactor the NeRF compile into multiple sub-function
  • Refactor nerf.utils function into NeRFUtils class
  • Test the updated codes

• Refactor the NeRF format and add new features

  • Add white background option on image loader and ray composting
  • Output the training log as csv
  • Support for continuing the model training from last log
  • Reformat the saved NeRF model format to json and h5 for the weights
  • Refactor the NeRFUtils
  • Refactor training and inference codes

• Update Readme, reformat the NeRF outputs, and better logs

  • Remove --steps_per_epoch arguments and use all datasets for an epoch
  • Add more detail logging when --verbose is used for better debuging:
  • Output per step predictions on the callback
  • Show warning when the model gradient is zero during eager train step
  • Reformat model prediction format from set into dictionary
  • Update the test codes
  • Update the Readme to include quickstart and results examples