Kompress

Overview

Kompress provides a set of compression techniques that are tailored for specific deployment constraints. Users get the flexibility to choose and combine multiple techniques to achieve the best trade-off between model performance and deployment constraints. Leveraging cutting-edge techniques like pruning, quantization, distillation, etc, Kompress achieves exceptional model compression levels on a variety of language and vision models.

Basic Workflow

Step 1 - Import a model and dataset.

Check Import Data and Import Model for the exact steps. Make sure that the model-dataset combination provided is valid and of the same task.

Step 2 - Choose an Algorithm.

Check available Algorithms and respective hyperparameters. By default, most optimal hyperparameters are chosen, however, depending upon the task, model and dataset another set of hyperparameters can work better too.

Step 3 - Monitor Logs and Export.

By default, Job logs and model checkpoints are saved in Kompress/user_data/, however, users can view the logs and download their compressed models locally using the export functionality.