---
id: adam-atan2-pytorch
version: "0.4.0"
license: MIT License Copyright (c) 2024 Phil Wang Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in… (full text in the JSON record)
license_treatment: permissive
maintenance: active
---

# adam-atan2-pytorch — Adam-atan2 for Pytorch

License: permissive · Maintenance: active · Downloads: 270.1K/mo

## What it is and 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—using 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.

The 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.

Use it for:

- 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.

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

Provides an Adam optimizer variant using atan2 for numerical stability and scale invariance, with optional plasticity features for continual learning.

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.

## Install

pip install adam-atan2-pytorch
uv add adam-atan2-pytorch
poetry add adam-atan2-pytorch

## Installing adam-atan2-pytorch

Before you install: Low install friction with only two runtime dependencies (einops and torch). Active maintenance status with a recent release 28 days ago.

License in practice: MIT License permits unrestricted use, modification, and distribution with only attribution and liability disclaimer required—no restrictions on commercial or private use.

Quickstart:

pip install adam-atan2-pytorch

import torch
from adam_atan2_pytorch import AdamAtan2

model = torch.nn.Linear(10, 1)
opt = AdamAtan2(model.parameters(), lr=1e-4)
loss = model(torch.randn(10))
loss.backward()
opt.step()
opt.zero_grad()

Requires PyPI installation of torch and einops; Python 3.9 or later.

Verify before relying:

- Whether plasticity features are enabled by default or require explicit configuration.
- Performance comparison with standard Adam on typical training tasks.
- Specific continual learning scenarios where plasticity features provide measurable benefit.

## Package facts

- License: MIT License Copyright (c) 2024 Phil Wang Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in… (full text in the JSON record) (permissive)
- Python support: supports_current
- Install friction: low
- Maintenance: active
- Downloads: 270.1K/month (top 15,000 on PyPI)
- Known vulnerabilities: none known

## Tags

adam optimizer pytorch, atan2 optimizer, numerical stability optimizer, scale invariant optimizer, continual learning optimizer, pytorch optimizer implementation, deep learning optimization, gradient descent variant, optimizer, continual-learning, numerical-stability

[View on SkillFed](https://skillfed.io/packages/adam-atan2-pytorch) · [View on PyPI](https://pypi.org/project/adam-atan2-pytorch/)