pidtransformer 1.0.3

A Transformer architecture with internal PID control for stabilizing learning dynamics.

PID-Transformer: A Control-Theoretic Approach for Stabilizing Large-Scale Neural Networks

Control-Theoretic Stabilization for Large-Scale Transformers

A novel Transformer architecture with embedded PID control logic for robust and interpretable representation dynamics.
Includes Group-wise PID, AdaptiveDim switching, and Gating-based trajectory feedback.

---

Installation

You can install the package in the following ways:

PyPI

pip install pidtransformer

From wheel or tar.gz

pip install pidtransformer-1.0.0-py3-none-any.whl or pip install pidtransformer-1.0.0.tar.gz

---

Quick Start

1. Clone and prepare the environment

• git clone https://github.com/your-org/pid-transformer.git

• cd pid-transformer

• python3 -m venv venv

• source venv/bin/activate

• pip install -r requirements.txt

---

2. Run a training experiment

python experiments/train_baseline.py
--experiment_name "PID_Optimal_Filtered"
--kp 0.002852 --ki 0.001240 --kd 0.012805
--d_filter 3

---

3. Visualize hidden state trajectories

python experiments/plot_trajectory.py
PID_Off_history.json PID_Optimal_Filtered_history.json
--output_file trajectory_comparison.png

---

Core Components

Module	Description
GeometricPIDController	A vector-based controller applying P/I/D logic to model-internal error signals.
PIDLayer	Combines FFN + projection to control space + PID feedback loop.
AdaptiveDim	Dynamically switches dimensionality (e.g. 256→128) to optimize stability over time.
Group-wise PID	PID applied across subgroups of hidden dimensions.
Trajectory Tracker	Captures hidden state evolution to visualize smoothness/curvature via PCA.

---

Interactive Visualization

Launch in Browser

Go to Live Demo

You’ll see:

PCA latent space trajectories

Loss/PID Norm curves

GSA bar chart

Diagrammatic explanation of PIDLayer and controller design

---

Reproducible Experiments

Baseline Model (no control)

python experiments/train_baseline.py --experiment_name "PID_Off" --kp 0 --ki 0 --kd 0

Full Control Model

python experiments/train_baseline.py --experiment_name "AdaptiveDim"
--kp 0.1 --ki 0.01 --kd 0.05 --use_adaptive_dim --use_group_pid --ortho_weight 0.01

Endurance-scale

python experiments/train_baseline.py --experiment_name "Endurance_AdaptiveDim"
--num_steps 5000 --log_freq 50 --use_adaptive_dim --use_group_pid

---

Citation

If you use this work in your research, please cite:

@article{pidtransformer2025, title={AdaptiveDim+Gating PID-Transformer: A Control-Theoretic Approach for Stabilizing Large-Scale Neural Networks}, author={KANG JA IL}, year={2025} }