Techiiot's picture
download
raw
7.71 kB
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8" />
<title>Reproduction poster — DFMPQ</title>
<style>
html, body { margin: 0; padding: 0; background: #f7f7f6;
font-family: "Helvetica Neue", "Arial", sans-serif; color: #1f2937; }
.poster { width: 100%; max-width: 1300px; margin: 0 auto; background: #fff;
border: 1px solid #e5e7eb; border-radius: 14px; overflow: hidden; }
.masthead { background: linear-gradient(120deg, #0F6070 0%, #022733 100%);
color: #fff; padding: 24px 32px; }
.masthead .eyebrow { font-size: 12px; letter-spacing: .18em;
text-transform: uppercase; color: #cffafe;
font-weight: 700; }
.masthead h1 { font-size: 26px; line-height: 1.15; margin: 8px 0 4px;
font-weight: 700; }
.masthead .venue { font-size: 13px; opacity: .9; }
.grid { display: grid; grid-template-columns: 1fr 1fr 1fr 1fr;
gap: 16px; padding: 20px 24px; }
.card { background: #f8fafc; border: 1px solid #e5e7eb; border-radius: 10px;
padding: 14px 16px; min-height: 140px; }
.card.full { grid-column: 1 / span 4; }
.card .tag { font-size: 10px; letter-spacing: .15em;
text-transform: uppercase; color: #64748b;
font-weight: 700; }
.card h3 { margin: 6px 0 8px; font-size: 16px; color: #0F6070; }
.card p, .card li { font-size: 12.5px; line-height: 1.5; margin: 0 0 6px;
color: #1f2937; }
.card ul { padding-left: 16px; margin: 0 0 6px; }
.card .big { font-size: 22px; font-weight: 800; color: #0F6070;
line-height: 1; }
.card table { font-size: 11.5px; width: 100%; border-collapse: collapse; }
.card th, .card td { padding: 4px 6px; text-align: right;
border-bottom: 1px dotted #e2e8f0; }
.card th { text-align: left; color: #475569; font-weight: 600; }
.strip { background: #0F6070; color: #fff; padding: 16px 24px;
font-size: 13px; line-height: 1.5; }
.strip b { color: #cffafe; }
.target { cursor: pointer; outline: none; border: 0; background: transparent;
color: inherit; font: inherit; padding: 0; margin: 0;
text-decoration: none; }
.target:hover { box-shadow: inset 0 -2px 0 0 #cffafe; }
.footnote { font-size: 10.5px; color: #64748b; padding: 0 24px 18px; }
</style>
</head>
<body>
<section class="poster" data-measure-role="poster">
<div class="masthead">
<div class="eyebrow">ICML 2026 reproduction · OpenReview nHbvQF35ch</div>
<h1>No Retraining at Edge: Efficient Resource-Aware Mixed-Precision Quantization
via Federated Supernet Learning</h1>
<div class="venue">Independent toy-scale reproduction · paper authors:
Lianbo Ma, Yonghui Su, Nan Li, Xingwei Wang</div>
</div>
<section class="grid">
<div class="card" data-logbook-target="executive-summary">
<div class="tag">Executive summary ↗</div>
<h3>Headline outcome</h3>
<p><span class="big">Partial</span> reproduction at toy scale.</p>
<p>Claim&nbsp;1 reproduces via <b>SAGS</b> (paper §4.3): fed-Avg-trained
supernet serves <b>20 deployment budgets</b> derivation-only
(zero gradient steps). Claim&nbsp;2 partially reproduces:
DFMPQ mixed-bit beats same-size uniform subnet in several budgets.</p>
</div>
<div class="card" data-logbook-target="claim-1-dfmpq-enables-retraining-free-deployment-for-dynamic-edge-computing-by-deriving-resource-aware-quantized-subnets-on-demand">
<div class="tag">Claim 1 ↗</div>
<h3>Retraining-free at the edge</h3>
<p>4 supernets (ResNet-8 + tiny MobileNetV2 × CIFAR-10/100). Load
checkpoints, derive subnets for 5 memory budgets per supernet via
<b>SAGS</b> (paper §4.3 — sensitivity-ranked greedy downgrade,
top-K=3, no gradient steps) — <b>zero training-time gradient
steps</b>, ~30 s wall-clock total.</p>
<p>Implements <b>CCSA</b> (§4.1 — linear-MMD class-conditional
prototype alignment, λ=0.1) and <b>SAHA</b> (§4.2 —
semantic-confidence aggregation on crucial stage-output layers,
τ=1.0) for supernet training, matching the paper's algorithm.</p>
</div>
<div class="card" data-logbook-target="claim-2-achieves-competitive-accuracy-with-significantly-reduced-computational-cost-across-multiple-datasets-and-network-architectures">
<div class="tag">Claim 2 ↗</div>
<h3>Competitive vs reduced compute</h3>
<p>Per (dataset × arch), mean Δ(DFMPQ − best uniform) — toy scale:</p>
<table>
<tr><th>pair</th><th>Δ (pp)</th></tr>
<tr><td>c10/ResNet-8</td><td>+6.06</td></tr>
<tr><td>c10/MobileV2</td><td>+2.38</td></tr>
<tr><td>c100/ResNet-8</td><td>+0.61</td></tr>
<tr><td>c100/MobileV2</td><td>+1.23</td></tr>
</table>
<p>Absolute acc far below paper — paper-scale (10×50×5) expected to close gap.</p>
</div>
<div class="card">
<div class="tag">Scope &amp; cost</div>
<h3>What we spent</h3>
<p><b>HW</b>: 1× RTX 3050 laptop GPU (4 GB) + 16 CPU cores.</p>
<p><b>Wall</b>: ≈ 22 min (~20 min training 4 supernets with CCSA+SAHA +
~1 min SAGS derivation).</p>
<p><b>Cost</b>: ≈ $0 (local). HF Jobs returned HTTP&nbsp;402
(account out of credits).</p>
<p><b>Scope gap</b>: paper ResNet-18/MNetV2/EffNet-Lite0,
10 clients × 50 rounds × 5 epochs, batch 256; we ResNet-8/tiny-MNetV2,
5 × 15 × 2, batch 128, 2 datasets.</p>
</div>
<div class="card full" data-logbook-target="conclusion">
<div class="tag">Reproduction bundle ↗</div>
<h3>What we publish</h3>
<p>Self-contained bundle (3.6 MB) containing
<code>repro_dfmpq/dfmpq.py</code> (STE quantizer + QResNet8 / QMobileTiny
supernets + Dirichlet non-IID partition + <b>CCSA</b> regularizer +
<b>SAHA</b> semantic aggregation + <b>SAGS</b> search + FedAvg),
<code>repro_dfmpq/run.py</code> (with flags --no-ccsa, --no-saha,
--lambda-ccsa, --saha-tau), <code>repro_dfmpq/derive_only.py</code>
(Claim 1 demo, SAGS derivation only), <code>repro_dfmpq/make_plots.py</code>,
<b>4 trained supernet checkpoints</b> (one per dataset × arch),
<b>results.csv</b> (every (dataset × arch × budget × method) row),
<b>Plotly Pareto-frontier HTML figures</b>, and a verbatim training
log. Bundled as a Trackio <b>dataset artifact</b> on the Conclusion
page; rerunnable in ≈ 25 min on any single GPU.</p>
</div>
</section>
<div class="strip">
Bottom line: <b>Claim 1 mechanism reproduces</b> — one federated supernet
(CCSA + SAHA during training) serves many edge budgets by <b>SAGS</b>
derivation only, no retraining. <b>Claim 2 partially reproduces</b> at
toy scale (DFMPQ &gt; same-size uniform baseline on all 4 (dataset × arch)
pairs); absolute accuracy far below paper because paper is 10 clients ×
50 rounds × 5 epochs on A100 with ResNet-18; we are 5 × 15 × 2 on a 4 GB
laptop GPU with ResNet-8. Gap to per-budget-retrain reference (~10–18 pp)
is the expected toy-scale signature.
</div>
<div class="footnote">
Built with the OpenResearch reproduction agent using the
<a href="https://github.com/Chenruishuo/posterely">Chenruishuo/posterly</a>
HTML poster skill. Hover over an underlined <code></code> tag and click to
jump to the matching logbook page; this poster embeds no external assets
— it is fully self-contained.
</div>
</section>
</body>
</html>

Xet Storage Details

Size:
7.71 kB
·
Xet hash:
dee3546e4cf6081a0f57295d91a13da15ba5c9e5fff36c5c96dfddca2579493a

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.