Skip to content

Latest commit

 

History

History
76 lines (54 loc) · 2.66 KB

USAGE.md

File metadata and controls

76 lines (54 loc) · 2.66 KB
Raw

How to Use

Demo

Once the package is installed, you can run an SDS experiment using the following command:

python main.py --config-name sds

To enable logging with Weights & Biases (wandb), specify your wandb API key like this:

python main.py --config-name sds +wandb_key=your_wandb_key

If you prefer not to use wandb, you can view logs locally with TensorBoard. Logs are saved in the PROJECT_ROOT/output directory:

tensorboard --logdir output

Additional Reference

Example configuration files are available in the PROJECT_ROOT/conf directory. These configurations follow a hierarchical structure.

Configuration Example

When you run the command:

python main.py --config-name sds

The configuration file PROJECT_ROOT/conf/sds.yaml is loaded. Here’s an example content of the file:

name: sds  # Name for instantiation from the registry
group: ${name}
note: ""

defaults:
  - base  # Inherit from PROJECT_ROOT/conf/base.yaml
  - rasterizer: pixels  # Use pixels as the 2D representation, inherit from PROJECT_ROOT/conf/rasterizer/pixels.yaml
  - wt_schedule: dreamfusion  # Use DreamFusion schedule, inherit from PROJECT_ROOT/conf/wt_schedule/dreamfusion.yaml
  - algorithm: sds_algorithm  # Use the SDS optimization algorithm, inherit from PROJECT_ROOT/conf/algorithm/sds_algorithm.yaml
  - _self_  # Override default values with those defined below

prompt: "a photograph of an astronaut riding a horse."
neg_prompt: ""

iterations: 2000
log_interval: 200

algorithm:
  guidance_scale: 7.5

rasterizer:
  batch_size: 4
  optimizer:
    opt_args:
      lr: 0.1

Notes:

  • Hierarchical structure: You can refer more about the hierarchical config structure in the Hydra documentation.
  • Custom settings: You can override values as needed by modifying the YAML files or passing arguments directly via the command line. Also refer to the above links for more details.

This structure makes it easy to experiment with different combinations of settings, algorithms, or representations.

Warning

The framework automatically scans and loads modules from PROJECT_ROOT/sds_2d/algorithm/, PROJECT_ROOT/sds_2d/rasterizer/, and PROJECT_ROOT/sds_2d/wt_schedule/. This offers convenience in extending the framework with new algorithms, representations, or weight and timestep schedules. However, ensure that server files are synchronized with local files, and no redundant Python files exist in these directories on the server, as redundant files may be unintentionally loaded and cause errors. This is particularly critical when using IDEs like PyCharm.