Get started
nebulgym accelerates training by means of class decorators. A class decorator is a very elegant and non-intrusive method that allows nebulgym to tag your model (@accelerate_model) and your dataset (@accelerate_dataset) and add functionalities to their classes. When you run a training session, nebulgym will greatly reduce the training time of your decorated model. As simple as that!
You can find more information about nebulgym class decorators, the parameters they can take as input, and other nebulgym classes that can be used as an alternative to decorators in Advanced options.

How to use nebulgym class decorators

Put nebulgym class decorators right before defining your dataset and model classes.
  • @accelerate_dataset must be entered before the dataset definition. nebulgym will cache dataset samples in memory, so that reading these samples after the first time becomes much faster. Caching the dataset makes data loading faster and more efficient, solving what could become the main bottleneck of the whole training process.
  • @accelerate_model must be entered before the model definition. nebulgym will accelerate both forward and backward propagations by reducing the number of computationally expensive propagation steps and making computations more efficient.
To learn more about class decorators, their optional input parameters, and alternatives to decorators, see Advanced options.
Export as PDF
Copy link