Hassan1278/DynaLR: DynaLR: Adaptive Learning Rate Optimizers using PID Control

DynaLR — Advanced Learning Rate Optimizers for PyTorch

Author: Hassan Al Subaidi

Version: 1.0.3

Abstract 📌

DynaLR introduces a principle of adaptive learning rate optimizers using PID control theory.

• 2.6% accuracy gain over Adam on CNN architectures
• Faster convergence (3–5% speedup)
• Architecture-aware performance (excels on CNNs)
• Four specialized variants for different use cases

Benchmark Summary (30 Epochs, 3 Seeds)

Hardware: A100 GPU (ResNet18), ve6-1 TPU (SimpleCNN)

SimpleCNN on CIFAR-10

• DynaLRMemory: 77.35% ± 0.31% (Time: 153.9s, +1.01%)
• DynaLRenhanced: 77.08% ± 0.24% (154.4s, +0.74%)
• DynaLRnoMemory: 77.11% ± 0.89% (155.5s, +0.77%)
• Adam: 76.34% ± 0.33% (159.0s, Baseline)
• DynaLRAdaptivePID: 71.95% ± 0.76% (152.0s, -4.39%)

ResNet18 on CIFAR-10

• DynaLRMemory: 87.45% ± 0.85% (271.8s, -2.19%)
• DynaLRenhanced: 87.71% ± 0.69% (271.7s, -1.93%)
• DynaLRnoMemory: 87.76% ± 0.56% (273.8s, -1.88%)
• Adam: 89.64% ± 0.30% (276.6s, Baseline)
• DynaLRAdaptivePID: 88.71% ± 0.28% (271.5s, -0.93%)

SimpleCNN on CIFAR-100

• DynaLRMemory: 45.89% ± 0.88% (154.9s, +2.64%)
• DynaLRenhanced: 45.26% ± 0.40% (156.9s, +2.01%)
• DynaLRnoMemory: 45.03% ± 0.61% (156.5s, +1.78%)
• Adam: 43.25% ± 0.76% (164.8s, Baseline)
• DynaLRAdaptivePID: 35.08% ± 0.13% (156.7s, -8.17%)

ResNet18 on CIFAR-100

• DynaLRMemory: 64.05% ± 0.72% (272.1s, -0.84%)
• DynaLRenhanced: 61.42% ± 1.30% (276.4s, -3.47%)
• DynaLRnoMemory: 63.16% ± 0.51% (277.7s, -1.73%)
• Adam: 64.89% ± 0.35% (276.7s, Baseline)
• DynaLRAdaptivePID: 65.04% ± 0.62% (276.4s, +0.15%)

Key Findings 🔍

1. CNN Dominance: Up to +2.64% accuracy over Adam
2. ResNet Specialist: DynaLRAdaptivePID beats Adam on CIFAR-100
3. Architecture Matters:
   • Memory variant best for CNNs
   • AdaptivePID best for ResNets
4. Speed Advantage: Average 2.5% speedup vs Adam

License

MIT License (c) 2025 Hassan Al Subaidi

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 the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.

GitHub: https://github.com/DynaLR