PlatformX: End-to-End Transferable Platform for Energy-Efficient Neural Architecture Search

PlatformX is an end-to-end, fully automated platform for hardware-aware Neural Architecture Search (HW-NAS) targeting energy-efficient deep neural networks on mobile and edge devices.

This repository accompanies the paper:

PlatformX: An End-to-End Transferable Platform for Energy-Efficient Neural Architecture Search
(To appear at ACM/IEEE SEC 2025)

---

Key Features

• Model generation & search space design
• Transferable energy prediction (LitePred-style, RF/MLP-based, etc.)
• Multi-objective NAS algorithms
  • GD-Pareto
  • Bayesian Optimization
  • Evolutionary (Greedy+Random)
  • UCB
  • Random
• On-device inference & Monsoon Power Monitor measurement
• Latency + enercy live profiling

---

Repository Structure

PlatformX/
│
├── benchmark_tools/                 # Android/TFLite benchmark tools & wrappers
│
├── cifar10/
│   └── cifar-10-batches-py/         # CIFAR-10 dataset
│
├── demo/                            # Quickstart demos
│
├── latency_acc_live/                # Live latency and accuracy profiling
│   ├── scripts/
│   └── utils/
│
├── model_generation/                # Architecture and search space generation
│
├── model_search/                    # NAS algorithms (GD/BO/EVO/UCB/Random)
│   ├── Model_search_compare_mn.py   # Main NAS script
│   ├── search_utils.py
│   ├── pareto_utils.py
│   └── configs/
│
├── monsoon/                         # Monsoon HVPM measurement pipeline
│   ├── server_orchestrator_phone.py
│   ├── monsoon_measurement.py
│   └── HV Main Example.csv
│
├── pred_ckpts/                      # Pretrained predictor checkpoints
│   └── pixel7/
│
├── prediction/                      # Energy predictor & accuracy proxy training
│
├── prediction_results/              # Predictor outputs
│
├── results/                         # Final NAS results, plots, logs
│
├── energy_genotype.txt
└── README.md

---

Requirements

• PyTorch
• TensorFlow + TFLite
• Monsoon Power API
• adb (Android Platform Tools)

---

Running NAS Search

Example for GD-Pareto:

python model_search/Model_search_compare_mn.py \
    --energy_json prediction_results/energy.json \
    --config_json prediction_results/config.json \
    --backend cpu \
    --data_root cifar10/cifar-10-batches-py \
    --base_out results/ \
    --pool_size 100000 \
    --max_evals none \
    --acc_proxy zero_cost \
    --probe_epochs 1 \
    --probe_batches 80 \
    --method gd_pareto \
    --use_energy_predictor true \
    --acc_threshold 0.80

Supported Methods

--method gd_pareto
--method bo
--method evo
--method ucb
--method random

Enable/Disable Energy Predictor

--use_energy_predictor true
--use_energy_predictor false

---

Real-Device Power Measurement (Android + Monsoon)

python monsoon/energy_prifiling.py

This script:

• Pushes TFLite models to phone
• Runs TFLite benchmark
• Collects high-frequency power traces with Monsoon HVPM
• Aligns timestamps
• Outputs latency, power, energy CSVs

---

🎬 Demo Video

Citation

@inproceedings{tu2025platformx,
  title     = {PlatformX: An End-to-End Transferable Platform for Energy-Efficient Neural Architecture Search},
  author    = {Xiaolong Tu and others},
  booktitle = {Proceedings of the ACM/IEEE Symposium on Edge Computing (SEC)},
  year      = {2025}
}

---