dysweep 0.1.6

Use Weights and Biases Sweeps for Dynamic Configuration generation.

Dysweep

Dysweep is a Python library enhancing the functionalities of the Weights and Biases sweep library. It allows entire experiments to be executed using a configuration dictionary (YAML/JSON).

Features

• Checkpointing for the Sweep Server: Dysweep introduces checkpointing that allows resuming certain runs, useful when only a small fraction of runs fail, eliminating the need to re-run the entire sweep.

• Running Sweeps Over Hierarchies: Dysweep supports hierarchically structured parameters, thereby eliminating the need for hard-coding the selection between different classes.

Dysweep is inspired by DyPy, offering a versatile configuration set that empowers defining experiments at any layer of abstraction.

Applications

Dysweep aids in large-scale hyperparameter tuning across various models/methods and running models over different configurations and datasets. It provides a systematic way to define a sweep in WandB, allowing parallel execution of experiments.