{"categories":[{"label":"Artificial Intelligence","url":"https://skillfed.io/packages/category/scientific-engineering-artificial-intelligence/6"}],"enrichment":{"capability":"Provides an Adam optimizer variant using atan2 for numerical stability and scale invariance, with optional plasticity features for continual learning.","skillfed_tags":["optimizer","continual-learning","numerical-stability"],"use_cases":["Replace standard Adam in deep learning models where epsilon tuning is problematic or parameter scales vary widely.","Continual learning pipelines where maintaining plasticity across task sequences is critical.","Training scenarios where numerical stability without manual epsilon adjustment is desired.","Research comparing optimizer variants with scale-invariant update rules.","Large-scale training where removing epsilon hyperparameter reduces tuning overhead."],"what_it_does":"Adam-atan2-pytorch implements a modified Adam optimizer that replaces the epsilon term with atan2 to achieve numerical stability without requiring manual epsilon tuning. The core change is algorithmic\u2014using atan2 in the update rule removes the need for epsilon altogether while maintaining scale invariance across different parameter ranges. This is based on research from DeepMind proposing the atan2 modification.\n\nThe package also includes features for improving plasticity in continual learning scenarios, where models must learn new tasks without catastrophic forgetting of old ones. It integrates with PyTorch's standard optimizer interface, so you instantiate it with model parameters and a learning rate, then call step() and zero_grad() in your training loop like any other optimizer. The implementation depends on einops for tensor operations and torch itself.","worth_installing":"Yes, if you use PyTorch and want to experiment with a theoretically motivated Adam variant. Low install friction, permissive license, no known vulnerabilities, and active maintenance make it a low-risk addition. Best suited for researchers exploring optimizer variants or teams dealing with scale-invariance issues; standard Adam remains the safer default for most production training."},"id":"adam-atan2-pytorch","links":{"html":"https://skillfed.io/packages/adam-atan2-pytorch","md":"https://skillfed.io/packages/adam-atan2-pytorch.md","pypi":"https://pypi.org/project/adam-atan2-pytorch/"},"maintenance":{"status":"active"},"meta":{"latest_release":"2026-07-17","license_spdx":null,"license_treatment":"permissive","name":"adam-atan2-pytorch","python_support":"supports_current","summary":"Adam-atan2 for Pytorch"},"popularity":{"monthly_downloads":270109,"position":8242,"tier":"top_15000"},"security":{"n_vulnerabilities":0},"version":"0.4.0"}