Getting Started

To train a simple model with a standard training loop with Kerosene you’ll need to define two things:

• A Pytorch model
• A config file

For more complex use cases, like multiple models, or fancy architecture with complex training strategy, a Trainer should be provided.

Model Trainer

Kerosene supports standard Pytorch model (i.e. torch.nn.Module). There is no need to extend any class or add any logging dependencies to your model. We believe that your model should only contain model related code. It should be clean and lean.

Instead, the model is wrapped in a kerosene.training.trainers.ModelTrainer class which handles

Trainer

The trainer object can be seen as the training orchestrator. The trainer handles the following:

• Main training and validation loops;
• Data loading (from the data loaders) and the distribution to the good device (GPU/CPU);
• Training event handling

For a simple use case where a single model is used with standard forward and backward passes, the predefined kerosene.training.trainers.SimpleTrainer can be used. It will handle all the training and validation process for you.

Configuration

class Python(object):
    def __init__(self):
        pass