Skip to content

Latest commit

 

History

History
104 lines (102 loc) · 6.33 KB

changelog.md

File metadata and controls

104 lines (102 loc) · 6.33 KB
Raw

Changelog

2024.04.07 - 2024.08.06

  • Fix .vscode/launch.json template for newer versions
  • More examplified dataset configurations
  • More specific configurations for networks
  • More documentation on installation
  • Change the recommended invokation method from evc -t xxx to evc-xxx, for example: evc-gui
  • Better dataset management
    • Better ratio / scaling managemnet
    • Supports a COLMAP undistorter, which somehow works better for COLMAP camera parameters. Myths.
    • Add support for disk-based dataset
      • Yes, originally we only supported RAM-based dataset, which took a long while to load during every training start
  • Better importing system
    • Note that missing imports are not considered fatal since we'll include all module files, some of which might have optional dependencies
    • Instead, only a warning for those specific modules are put our
    • However, if you later try to use these modules, a fatal error will be raised
  • Faster real-time viewer
    • Removing unnecessary data passing between CPU and GPU
    • Better time management for the GUI
  • Better entry points
    • Warn user of duplicate EasyVolcap installation
    • Better commandline interfacese
    • Fixing some entry point definitions in the wrap.py file
  • Better config system
    • Will now support fully installed easyvolcap + local configurations
    • Since we'll not only search easyvolcap/configs, but also ./configs
  • Updated to MIT License for all our own code
  • Implemented a faster fused Adam: MyFusedAdam
    • I don't know why the PyTorch version of the fused adam is even slower than a regular Adam
    • This operation also introduces the logic for compiling CUDA extensions silently in the codebase
    • Very handy since we're doing massively parallelized work
  • Implemented a mechanism for clearing the host memory for PyTorch: easyvolcap/utils/host_utils.py
  • Fix some typos for frustum
  • Better supervisors
    • Now supervisors are modulized and can be chained together using the SequentialSupervisor
    • You get the output and batch for computing loss in any supervisors
    • Add a windowed SSIM implementation inspired by 3DGS
  • More performance model function
    • The handling of camera parameters is more optimized for real-time operations
  • Adds a CustomViewer class that can be easily inherited to use the viewer of EasyVolcap
    • Implemented a few dummy objects to make this happen since we originally only supported in house algorithms
    • Now you can totally do something like in tests/custom_gui_tests.py
  • Better visualization
    • Fixed some bugs related to the visualizer's thread pool
    • Add option for generating videos using a different backend, libx265 produces the best and smallest file but CUDA might be faster
  • Better training control
    • Added a mechanism in VolumetricVideoRunner to pause the training and open up a debugger
    • Just run evc-sig -h to check out the options
    • We use the signal mechanism to achieve this
    • Now, during training of very large models, if you're feeling insecure or something wrong with the model saving code happens (disk failure?)
    • You can just manually same the model to another location without losing hours of training progress
  • Added new mechanism for async data preprocessing on other streams
    • This is exposed as the prepare_params method for every network modules
    • This method will be called at the same time as the backward pass
    • Thus if there're some data copy operations, it can overlap with the GPU compute happening in the backward pass
    • Note that asynchronously copying to CPU might corrupt the PyTorch state, leading to weird undefined behavior
    • Explicit synchronization doesn't help, I'm guessing PyTorch bugs
  • Fix bugs in the viewer
    • Fix bugs related to t sampling
    • Fix bugs related to frame index computation
  • Add a WebSocket viewer that allows you to render on the server and view the content locally
    • The server is run like a regular gui, only appending the config configs/specs/server.yaml
    • The client is invoked as evc-ws or additionally with other parameters like shown in evc-ws -h
  • Adding a custom cache type for numpy arrays
  • Adding some custom data type conversion logic between PyTorch tensors and numpy arrays
  • Better loggin system
    • Now with more customizations, check out console_utils.py
    • Now, by default, there're much less clutter for modules in the logs
    • Making it easier to use for smaller operations
  • Better scripting system
    • A new scripting argument system introduced by the function build_parser
    • This can be used like args = dotdict(vars(build_parser(args, description=__doc__).parse_args()))
    • Define the parameters in a dict called args, and we'll handle the colorization and help messages
  • Better utility functions for saving and loading point clouds
  • Better utility functions for saving and loading images
  • Better utility scripts for point cloud, COLMAP and IBR preprocessing
  • Support for a faster gaussian rasterizer: https://github.com/dendenxu/fast-gaussian-rasterization
  • More optimized 3DGS model
  • Better timing utilities
    • Faster timer
    • Add the ability to store recorded time onto a file
    • And maybe perform further processing of it (like exporting a plot to see trends)
  • Add the ability to directly load a 3DGS model from a regular PLY file
  • Add better warning messages for splat files
  • More streamlined metric computation system
    • Now the metrics used are the same as the loss functions for training networks
    • You can select the type of LPIPS, alex (better for training) or vgg (better for validation, slow, memory hungry)
  • Fix some strange slowness caused by torch.jit.script
  • Better network loading and saving procedure
  • Add support for the RealityCapture software that recently became free to use and can be much faster than COLMAP
  • Reform all utility functions for easier control and imports
  • Better documentation
  • Fix the colmap_to_easyvolcap.py, easyvolcap_to_colmap and easymocap_to_colmap scripts
  • Add script for computing camera alignment
  • Add support for glomap in the run_colmap.py script
  • Add a standalone script for camera undistortion: undistort.py
  • Fix bugs for Google Drive batch upload
  • Add a script for compressing images
  • Better script for extracting images from videos
  • Add some more tests