Hyperparameter Search

Pointrix supports searching over all parameters specified in the configuration files. Example code can be found in examples/gaussian_splatting_sweep.

Here, we’ll use the colmap_sweep.yaml configuration file as an example:

method: "random"
name: "colmap_sweep"
metric: {goal: "maximize", name: "psnr"}

parameters:
  controller:
    densify_grad_threshold: 
      min: 0.0002
      max: 0.0004
    densify_stop_iter:
      values: [10000, 15000]
  optimizer:
    optimizer_1:
      params:
        point_cloud.position:
          lr: 
            min: 0.00016
            max: 0.00032

This configuration file specifies the parameters to be searched within colmap.yaml. For example, densify_grad_threshold ranges from 0.0002 to 0.0004, and densify_stop_iter can take values of 10000 or 15000, among others.

The method can be random, grid, or bayes, corresponding to random search, grid search, and Bayesian optimization, respectively. Users can also configure the search range for any other parameters in colmap.yaml.

The metric specifies the goal to optimize. In this example, we aim to maximize the psnr parameter. Related parameters can be read in launch_sweep.py.

Once colmap_sweep.yaml and colmap.yaml are configured, run the following command:

cd examples/gaussian_splatting_sweep
python launch_sweep.py --config configs/colmap.yaml --config_sweep configs/colmap_sweep.yaml trainer.datapipeline.dataset.data_path=your_data_path trainer.output_path=your_log_path

During the run, you can visit the provided URL to view the analysis results of the hyperparameters and metrics.