ESWEEK 2026 — DSE Results & Plotting Scripts

This repository contains the DSE (Design Space Exploration) experiment results and plotting scripts for the ESWEEK 2026 paper submission.

---

Repository structure

.
├── bar_plots.py                # Normalized bar charts of final DSE results
├── plot_best_temperature.py    # Temperature of best feasible solution over evaluations
├── plot_convergence_grid.py    # Convergence grid across thresholds and objectives
├── plot_dse_combined.py        # Combined convergence plot (EDP and Delay side by side)
├── plot_dse_comparison.py      # Scatter plots of all evaluations per algorithm (2×2 grid)
├── dse_results.csv             # Final DSE results table (all algorithms, DNNs, objectives)
└── plot_ready/
    ├── turbo/                  # Ours — TuRBO logCEI explorer, m=4 (full runs)
    ├── turbo_m1/               # Ours — TuRBO logCEI explorer, m=1
    ├── nsga2/                  # Baseline — NSGA-II
    ├── ucb/                    # Baseline — VAESA (UCB)
    └── msa/                    # Baseline — TREAD-M3D

Each subdirectory under plot_ready/<algo>/ follows the naming convention:

<dnn>_<objective>_<temperature>/
    progress_YYYYMMDD_HHMMSS.jsonl

DNNs: mobilebert, mobilenet_v3, mobilevit_s, resnet18, resnet50 Objectives: edp, delay Temperature thresholds: 70c, 80c, 90c

---

Requirements

pip install matplotlib numpy

A LaTeX installation (e.g. texlive) is required for PGF/PDF output. PNG output works without it if you change the matplotlib backend.

---

Scripts

1. bar_plots.py

Plots normalized bar charts of final DSE results (EDP and Delay) across all DNNs and temperature thresholds. Each algorithm is normalized to the MSA (TREAD-M3D) baseline. Reads from dse_results.csv (bundled in the repo).

python bar_plots.py

Output: plots/normalized_{EDP,DELAY}_{70,80,90}C.{png,pgf}

---

2. plot_best_temperature.py

Plots the temperature of the running-best feasible solution over evaluations, with one panel per objective (EDP and Delay).

# Default: mobilevit_s, 70°C threshold
python plot_best_temperature.py

# Specify DNN and threshold
python plot_best_temperature.py --dnn mobilevit_s --temp-label 80c --T-threshold 353.15
python plot_best_temperature.py --dnn resnet18 --temp-label 70c --T-threshold 343.15

# Include only specific algorithms
python plot_best_temperature.py --dnn mobilevit_s --temp-label 80c --algos turbo turbo_m1 nsga2 ucb msa

Note: The turbo_m1 data is only available for mobilevit_s at 80c.

Output: plots/<dnn>_<temp>_best_temperature.{png,pgf}

---

3. plot_convergence_grid.py

Plots a grid of convergence curves: one row per temperature threshold, two columns for EDP and Delay objectives. All algorithms are overlaid as lines.

# Default: mobilevit_s, thresholds 80°C and 90°C
python plot_convergence_grid.py

# Specify DNN and thresholds
python plot_convergence_grid.py --dnn resnet18 --thresholds 70 80 90
python plot_convergence_grid.py --dnn mobilebert --thresholds 80 90

Output: plots/<dnn>_convergence_<thresholds>c.{png,pgf}

---

4. plot_dse_comparison.py

Scatter plots of all DSE evaluations (green=feasible, red=infeasible) with a running-best step line for each algorithm. Produces a 2×2 grid per objective (one subplot per algorithm).

python plot_dse_comparison.py --dnn mobilevit_s --T 80
python plot_dse_comparison.py --dnn resnet18 --T 70
python plot_dse_comparison.py --dnn mobilebert --T 90

Output: plots/<dnn>_dse_comparison_<T>c_{combined.png,delay.pdf,edp.pdf}

---

5. plot_dse_combined.py

Plots convergence curves for EDP and Delay side by side, for a single DNN and temperature threshold.

# Default: resnet18, 70°C threshold
python plot_dse_combined.py

# Specify DNN and threshold
python plot_dse_combined.py --dnn mobilebert --temp-label 80c --T-threshold 353.15
python plot_dse_combined.py --dnn mobilevit_s --temp-label 90c --T-threshold 363.15 --algos turbo nsga2 ucb msa

Output: plots/<dnn>_<temp>_combined.{png,pgf}

---

Temperature threshold reference

Label	Kelvin	Celsius
70c	343.15 K	70°C
80c	353.15 K	80°C
90c	363.15 K	90°C

---

Algorithm reference

Key	Label	Description
turbo	Ours (m=4)	TuRBO logCEI explorer, m=4 (full)
turbo_m1	Ours (m=1)	TuRBO logCEI explorer, m=1
nsga2	NSGA-II	Baseline
ucb	VAESA	Baseline
msa	TREAD-M3D	Baseline

---

Commands Used For The Paper Figures

The following four commands regenerate exactly the figures currently included in the paper.

python bar_plots.py

python plot_dse_comparison.py --dnn mobilevit_s --T 70

python plot_best_temperature.py --dnn mobilevit_s --objective delay --temp-labels 70c 80c 90c --T-thresholds 343.15 353.15 363.15 --algos turbo nsga2 ucb msa --out-name mobilevit_s_delay_best_temperature

python plot_convergence_grid.py --dnn mobilevit_s --thresholds 80 90