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 🔍 CNN Dominance: Up to +2.64% accuracy over Adam ResNet Specialist: DynaLRAdaptivePID beats Adam on CIFAR-100 Architecture Matters: Memory variant best for CNNs AdaptivePID best for ResNets 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