Keepprob

Keepprob is an open‑source Python library designed to provide flexible control over keep probability parameters used in neural network regularization techniques such as dropout and stochastic depth. It was introduced in 2022 by a small team of researchers at a machine learning laboratory in order to simplify the process of scheduling and adjusting keep probabilities during training. The library supports both static and dynamic schedules, allowing users to define keep probabilities that can change over epochs, batches, or even individual forward passes.

The core functionality of keepprob revolves around two main components: the Scheduler class and the KeepProbCallback hook for integration with popular deep learning frameworks such as PyTorch and TensorFlow. The Scheduler can use linear, exponential, or custom interpolation functions to generate keep probability trajectories. It also provides mechanisms for clamping values to a specified range and for applying noise to encourage exploration during training. The KeepProbCallback automatically updates the dropout or stochastic depth modules in a model’s forward pass based on the current schedule, ensuring a seamless training experience.

Keepprob has been adopted by several research projects that require fine‑tuned regularization, particularly in the areas of computer vision and natural language processing. The library has an MIT license and is maintained on GitHub where contributions through pull requests and issue reports are regularly reviewed. Documentation includes example notebooks, API references, and guidelines for extending the library with custom scheduling strategies. Its lightweight design makes it suitable for use in both research prototypes and production training pipelines.