vimarsh commited on
Commit
ef0e019
·
verified ·
1 Parent(s): 1c2530d

Update logbook: Reproduction: Adaptive Personalized Federated Learning via Multi-task Averaging of Kernel Mean Embeddings

Browse files
README.md CHANGED
@@ -1,10 +1,19 @@
1
  ---
2
- title: Repro Adaptive Personalized Fl Multi Task Averaging Kernel Mean Embedding
3
- emoji: ⚡
4
- colorFrom: green
5
- colorTo: green
6
  sdk: static
7
  pinned: false
 
 
 
 
 
 
 
8
  ---
9
 
10
- Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
 
 
 
1
  ---
2
+ title: "Reproduction: Adaptive Personalized Federated Learning via Multi-task Averaging of Kernel Mean Embeddings"
3
+ emoji: 🎯
4
+ colorFrom: yellow
5
+ colorTo: red
6
  sdk: static
7
  pinned: false
8
+ tags:
9
+ - trackio
10
+ - trackio-logbook
11
+ - open-experiment
12
+ - icml2026-repro
13
+ - paper-ADHGbxsREr
14
+ - arxiv:2603.02233
15
  ---
16
 
17
+ # Reproduction: Adaptive Personalized Federated Learning via Multi-task Averaging of Kernel Mean Embeddings
18
+
19
+ An open experiment logbook, published with [Trackio](https://github.com/gradio-app/trackio).
bucket-icon.svg ADDED
index.html CHANGED
@@ -1,19 +1,84 @@
1
  <!doctype html>
2
- <html>
3
- <head>
4
- <meta charset="utf-8" />
5
- <meta name="viewport" content="width=device-width" />
6
- <title>My static Space</title>
7
- <link rel="stylesheet" href="style.css" />
8
- </head>
9
- <body>
10
- <div class="card">
11
- <h1>Welcome to your static Space!</h1>
12
- <p>You can modify this app directly by editing <i>index.html</i> in the Files and versions tab.</p>
13
- <p>
14
- Also don't forget to check the
15
- <a href="https://huggingface.co/docs/hub/spaces" target="_blank">Spaces documentation</a>.
16
- </p>
17
- </div>
18
- </body>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
19
  </html>
 
1
  <!doctype html>
2
+ <html lang="en">
3
+ <head>
4
+ <meta charset="utf-8" />
5
+ <meta name="viewport" content="width=device-width, initial-scale=1" />
6
+ <title>Reproduction: Adaptive Personalized Federated Learning via Multi-task Averaging of Kernel Mean Embeddings</title>
7
+ <link rel="stylesheet" href="./logbook.css" />
8
+ </head>
9
+ <body>
10
+ <div id="app">
11
+ <aside id="sidebar">
12
+ <div id="book-head">
13
+ <img id="book-wordmark" src="./trackio-wordmark-dark.png" alt="" />
14
+ <div id="book-title" class="sr-only">Logbook</div>
15
+ </div>
16
+ <nav id="tree"></nav>
17
+ <div id="sidebar-foot" hidden>
18
+ <button id="connect-btn" type="button">
19
+ <span class="ico">ⓘ</span> Collaborate with your agent
20
+ </button>
21
+ </div>
22
+ </aside>
23
+ <main id="content">
24
+ <nav id="view-tabs" aria-label="Logbook views">
25
+ <a data-view="code" href="#/view/code/index">
26
+ <svg viewBox="0 0 24 24" aria-hidden="true">
27
+ <path d="m18 16 4-4-4-4" />
28
+ <path d="m6 8-4 4 4 4" />
29
+ <path d="m14.5 4-5 16" />
30
+ </svg>
31
+ <span>Logbook</span>
32
+ </a>
33
+ <a data-view="trace" href="#/view/trace">
34
+ <svg viewBox="0 0 24 24" aria-hidden="true">
35
+ <path d="M8 5h13" />
36
+ <path d="M13 12h8" />
37
+ <path d="M13 19h8" />
38
+ <path d="M3 10a2 2 0 0 0 2 2h3" />
39
+ <path d="M3 5v12a2 2 0 0 0 2 2h3" />
40
+ </svg>
41
+ <span>Traces</span>
42
+ </a>
43
+ <a data-view="workspace" href="#/view/workspace">
44
+ <svg viewBox="0 0 24 24" aria-hidden="true">
45
+ <path d="M20 20a2 2 0 0 0 2-2V8a2 2 0 0 0-2-2h-7.9a2 2 0 0 1-1.69-.9L9.6 3.9A2 2 0 0 0 7.93 3H4a2 2 0 0 0-2 2v13a2 2 0 0 0 2 2Z" />
46
+ </svg>
47
+ <span>Workspace</span>
48
+ </a>
49
+ </nav>
50
+ <header id="logbook-header">
51
+ <h1 id="logbook-title"></h1>
52
+ <div id="logbook-cli"></div>
53
+ </header>
54
+ <div id="page"></div>
55
+ </main>
56
+ </div>
57
+
58
+ <div id="modal" hidden>
59
+ <div class="modal-backdrop"></div>
60
+ <div class="modal-card" role="dialog" aria-modal="true">
61
+ <div class="modal-head">
62
+ <div class="modal-title">
63
+ <img class="modal-logo" src="./trackio-logo.png" alt="" />
64
+ Collaborate with your agent
65
+ </div>
66
+ <div class="modal-actions">
67
+ <button id="copy-agent" class="btn">Copy for agent</button>
68
+ <button id="modal-close" class="btn icon" aria-label="Close">×</button>
69
+ </div>
70
+ </div>
71
+ <div class="modal-body">
72
+ <p class="modal-intro">
73
+ Point your coding agent at this logbook. It reads a compact,
74
+ token-efficient version — and if you've given it write access to this
75
+ Space, it can add findings that sync back automatically.
76
+ </p>
77
+ <ol id="connect-steps"></ol>
78
+ </div>
79
+ </div>
80
+ </div>
81
+
82
+ <script src="./logbook.js"></script>
83
+ </body>
84
  </html>
logbook.css ADDED
@@ -0,0 +1,2142 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ :root {
2
+ --bg: #ffffff;
3
+ --paper: #ffffff;
4
+ --panel: #ffffff;
5
+ --ink: #1f2937;
6
+ --muted: #6b7280;
7
+ --line: #e5e7eb;
8
+ --accent: #f97316;
9
+ --accent-strong: #ea580c;
10
+ --accent-soft: #fff7ed;
11
+ --accent-line: rgba(249, 115, 22, 0.16);
12
+ --grid-line: rgba(31, 41, 55, 0.02);
13
+ --code-bg: #f3f4f6;
14
+ --radius: 12px;
15
+ --sidebar-width: 280px;
16
+ --content-gutter: 40px;
17
+ --serif: ui-serif, "Iowan Old Style", "Palatino Linotype", Georgia, serif;
18
+ --sans: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial,
19
+ sans-serif;
20
+ --mono: "SFMono-Regular", "Cascadia Mono", "JetBrains Mono", Menlo, Consolas,
21
+ ui-monospace, monospace;
22
+ }
23
+
24
+ * {
25
+ box-sizing: border-box;
26
+ }
27
+
28
+ html,
29
+ body {
30
+ margin: 0;
31
+ padding: 0;
32
+ }
33
+
34
+ html {
35
+ scroll-behavior: smooth;
36
+ scrollbar-gutter: stable;
37
+ }
38
+
39
+ body {
40
+ background: var(--bg);
41
+ color: var(--ink);
42
+ font-family: var(--sans);
43
+ font-size: 13px;
44
+ line-height: 1.65;
45
+ -webkit-font-smoothing: antialiased;
46
+ }
47
+
48
+ #app {
49
+ display: flex;
50
+ min-height: 100vh;
51
+ }
52
+
53
+ body[data-view="trace"] #sidebar-foot,
54
+ body[data-view="workspace"] #sidebar-foot {
55
+ display: none;
56
+ }
57
+
58
+ /* ---- sidebar (composition-book cover) ---- */
59
+ #sidebar {
60
+ width: var(--sidebar-width);
61
+ flex: 0 0 var(--sidebar-width);
62
+ background: #17181c;
63
+ color: #e7e7ea;
64
+ position: sticky;
65
+ top: 0;
66
+ height: 100vh;
67
+ overflow-y: auto;
68
+ padding: 22px 16px;
69
+ display: flex;
70
+ flex-direction: column;
71
+ }
72
+
73
+ #book-head {
74
+ display: flex;
75
+ align-items: center;
76
+ gap: 10px;
77
+ padding: 8px;
78
+ margin-bottom: 12px;
79
+ border-radius: 10px;
80
+ cursor: pointer;
81
+ transition: background 0.12s;
82
+ }
83
+ #book-head:hover {
84
+ background: rgba(255, 255, 255, 0.05);
85
+ }
86
+ #book-wordmark {
87
+ width: 154px;
88
+ height: auto;
89
+ object-fit: contain;
90
+ }
91
+ .sr-only {
92
+ position: absolute;
93
+ width: 1px;
94
+ height: 1px;
95
+ padding: 0;
96
+ margin: -1px;
97
+ overflow: hidden;
98
+ clip: rect(0, 0, 0, 0);
99
+ white-space: nowrap;
100
+ border: 0;
101
+ }
102
+
103
+ #tree {
104
+ flex: 1;
105
+ padding-top: 8px;
106
+ }
107
+
108
+ #tree .tree-label {
109
+ padding: 6px 10px 8px;
110
+ color: #777a83;
111
+ font-size: 10px;
112
+ font-weight: 700;
113
+ letter-spacing: 0.12em;
114
+ text-transform: uppercase;
115
+ }
116
+
117
+ #tree a {
118
+ display: block;
119
+ padding: 6px 10px;
120
+ border-radius: 8px;
121
+ color: #c3c4cb;
122
+ text-decoration: none;
123
+ font-size: 14px;
124
+ transition: background 0.12s, color 0.12s;
125
+ overflow: hidden;
126
+ text-overflow: ellipsis;
127
+ white-space: nowrap;
128
+ }
129
+
130
+ #tree a:hover {
131
+ background: rgba(255, 255, 255, 0.06);
132
+ color: #ffffff;
133
+ }
134
+
135
+ #tree a.active {
136
+ background: rgba(249, 115, 22, 0.16);
137
+ color: #fdba74;
138
+ font-weight: 600;
139
+ }
140
+
141
+ #tree a .tree-mark {
142
+ color: #6b6d76;
143
+ }
144
+
145
+ #tree a:hover .tree-mark,
146
+ #tree a.active .tree-mark {
147
+ color: inherit;
148
+ opacity: 0.6;
149
+ }
150
+
151
+ #tree .depth-1 {
152
+ padding-left: 22px;
153
+ }
154
+ #tree .depth-2 {
155
+ padding-left: 34px;
156
+ }
157
+ #tree .depth-3 {
158
+ padding-left: 46px;
159
+ }
160
+
161
+
162
+ /* ---- content ---- */
163
+ #content {
164
+ flex: 1;
165
+ min-width: 0;
166
+ padding: 24px
167
+ clamp(
168
+ var(--content-gutter),
169
+ calc(100vw - 960px),
170
+ calc(var(--sidebar-width) + var(--content-gutter))
171
+ )
172
+ 120px var(--content-gutter);
173
+ background-color: var(--paper);
174
+ background-image:
175
+ linear-gradient(var(--grid-line) 1px, transparent 1px),
176
+ linear-gradient(90deg, var(--grid-line) 1px, transparent 1px);
177
+ background-size: 26px 26px;
178
+ background-position: center top;
179
+ }
180
+
181
+ #logbook-header {
182
+ width: 100%;
183
+ max-width: 1080px;
184
+ margin: 0 auto 20px;
185
+ }
186
+ #logbook-title {
187
+ font-family: var(--serif);
188
+ font-size: 34px;
189
+ line-height: 1.15;
190
+ letter-spacing: -0.02em;
191
+ margin: 0 0 10px;
192
+ overflow-wrap: anywhere;
193
+ }
194
+ #logbook-cli {
195
+ display: grid;
196
+ gap: 7px;
197
+ }
198
+
199
+ #page {
200
+ width: 100%;
201
+ min-width: 0;
202
+ max-width: 1080px;
203
+ margin: 0 auto;
204
+ }
205
+
206
+ .page-section {
207
+ scroll-margin-top: 40px;
208
+ padding: 0 0 35px;
209
+ margin: 0 0 32px;
210
+ }
211
+
212
+ .page-section:last-child {
213
+ margin-bottom: 0;
214
+ }
215
+
216
+ .page-layout {
217
+ display: block;
218
+ }
219
+
220
+ .page-body {
221
+ min-width: 0;
222
+ }
223
+
224
+ .resource-anchor {
225
+ display: block;
226
+ height: 0;
227
+ overflow: hidden;
228
+ }
229
+
230
+ /* ---- pinned notes ---- */
231
+ .pinned-notes {
232
+ margin: 30px 0 32px;
233
+ }
234
+ .pinned-notes-list .cell {
235
+ margin: 0;
236
+ }
237
+ .pinned-notes-list .cell-title {
238
+ display: flex;
239
+ align-items: center;
240
+ gap: 7px;
241
+ }
242
+ .pin-ico {
243
+ flex: 0 0 auto;
244
+ width: 14px;
245
+ height: 14px;
246
+ fill: var(--accent);
247
+ stroke: none;
248
+ }
249
+ .pinned-notes-list .cell + .cell {
250
+ margin-top: 12px;
251
+ }
252
+ .book-intro.has-pinned-notes {
253
+ border-bottom: none;
254
+ padding-bottom: 22px;
255
+ margin-bottom: 30px;
256
+ }
257
+ .book-intro.book-intro-tight {
258
+ border-bottom: none;
259
+ padding-bottom: 4px;
260
+ margin-bottom: 20px;
261
+ }
262
+
263
+ #page h1 {
264
+ font-family: var(--serif);
265
+ font-size: 34px;
266
+ line-height: 1.15;
267
+ letter-spacing: -0.02em;
268
+ margin: 0 0 8px;
269
+ overflow-wrap: anywhere;
270
+ }
271
+
272
+ #page .page-section:not(.book-intro) h1 {
273
+ font-size: 26px;
274
+ }
275
+
276
+ #page h2 {
277
+ font-family: var(--serif);
278
+ font-size: 24px;
279
+ margin: 36px 0 10px;
280
+ }
281
+
282
+ #page h3 {
283
+ font-size: 17px;
284
+ font-weight: 700;
285
+ margin: 26px 0 2px;
286
+ letter-spacing: -0.01em;
287
+ }
288
+
289
+ #page h3::before {
290
+ content: "";
291
+ display: inline-block;
292
+ width: 7px;
293
+ height: 7px;
294
+ border-radius: 2px;
295
+ background: var(--accent);
296
+ margin-right: 10px;
297
+ vertical-align: middle;
298
+ transform: translateY(-1px);
299
+ }
300
+
301
+ #page p {
302
+ margin: 10px 0;
303
+ }
304
+
305
+ #page blockquote {
306
+ margin: 14px 0;
307
+ padding: 2px 16px;
308
+ border-left: 3px solid #fdba74;
309
+ color: var(--muted);
310
+ }
311
+
312
+ #page hr {
313
+ display: none;
314
+ }
315
+
316
+ #page code {
317
+ font-family: var(--mono);
318
+ font-size: 0.86em;
319
+ background: var(--code-bg);
320
+ padding: 2px 6px;
321
+ border-radius: 6px;
322
+ }
323
+
324
+ #page pre {
325
+ max-width: 100%;
326
+ background: var(--code-bg);
327
+ border: 1px solid var(--line);
328
+ border-radius: var(--radius);
329
+ padding: 14px 16px;
330
+ overflow-x: auto;
331
+ }
332
+ #page pre code {
333
+ background: none;
334
+ padding: 0;
335
+ font-size: 11.5px;
336
+ }
337
+
338
+ /* ---- code blocks + collapsible accordion ---- */
339
+ #page pre.hl {
340
+ background: #17181c;
341
+ border: none;
342
+ color: #e7e7ea;
343
+ font-size: 13px;
344
+ line-height: 1.58;
345
+ }
346
+ #page pre.hl code {
347
+ color: inherit;
348
+ font-family: var(--mono);
349
+ }
350
+ .code-accordion {
351
+ border: 1px solid rgba(249, 115, 22, 0.2);
352
+ border-radius: 8px;
353
+ overflow: hidden;
354
+ margin: 12px 0;
355
+ background: #17181c;
356
+ }
357
+ .code-accordion summary {
358
+ list-style: none;
359
+ cursor: pointer;
360
+ display: flex;
361
+ align-items: center;
362
+ gap: 9px;
363
+ padding: 9px 12px;
364
+ font-family: var(--mono);
365
+ font-size: 11.5px;
366
+ font-weight: 700;
367
+ color: #e7e7ea;
368
+ background: #1e2027;
369
+ user-select: none;
370
+ overflow-wrap: anywhere;
371
+ }
372
+ .code-accordion summary::-webkit-details-marker {
373
+ display: none;
374
+ }
375
+ .code-accordion summary::after {
376
+ content: "▸";
377
+ margin-left: auto;
378
+ color: var(--accent);
379
+ transition: transform 0.12s;
380
+ transform: rotate(180deg);
381
+ }
382
+ .code-accordion[open] summary::after {
383
+ transform: rotate(90deg);
384
+ }
385
+ .code-accordion .code-ico {
386
+ color: var(--accent);
387
+ font-weight: 700;
388
+ }
389
+ .code-accordion pre.hl {
390
+ margin: 0;
391
+ border-radius: 0;
392
+ border: none;
393
+ border-top: 1px solid rgba(249, 115, 22, 0.16);
394
+ }
395
+ .tok-comment {
396
+ color: #7a7d87;
397
+ font-style: italic;
398
+ }
399
+ .tok-string {
400
+ color: #a5d6a7;
401
+ }
402
+ .tok-keyword {
403
+ color: #fdba74;
404
+ }
405
+ .tok-number {
406
+ color: #7fd0e0;
407
+ }
408
+
409
+ #page a {
410
+ color: var(--accent);
411
+ }
412
+
413
+ #page ul {
414
+ padding-left: 20px;
415
+ }
416
+
417
+ .ts {
418
+ font-family: var(--mono);
419
+ font-size: 12px;
420
+ color: var(--muted);
421
+ background: none;
422
+ padding: 0;
423
+ }
424
+
425
+ /* ---- notebook-style cells ---- */
426
+ .cell {
427
+ max-width: 100%;
428
+ margin: 0 0 32px;
429
+ background: none;
430
+ border: none;
431
+ border-radius: 0;
432
+ box-shadow: none;
433
+ overflow: visible;
434
+ }
435
+ .cell-head {
436
+ display: flex;
437
+ justify-content: space-between;
438
+ gap: 16px;
439
+ align-items: baseline;
440
+ padding: 0 0 5px;
441
+ background: none;
442
+ border-bottom: none;
443
+ }
444
+ .cell-head.no-title {
445
+ justify-content: flex-end;
446
+ padding: 0 0 3px;
447
+ }
448
+ .cell-title {
449
+ flex: 1;
450
+ min-width: 0;
451
+ font-size: 13px;
452
+ font-weight: 650;
453
+ color: var(--ink);
454
+ line-height: 1.35;
455
+ overflow-wrap: anywhere;
456
+ }
457
+ .cell-meta {
458
+ flex: 0 0 auto;
459
+ display: flex;
460
+ align-items: center;
461
+ gap: 10px;
462
+ font-family: var(--sans);
463
+ font-size: 13px;
464
+ color: var(--muted);
465
+ }
466
+ .cell-open {
467
+ flex: 0 0 auto;
468
+ font-family: var(--mono);
469
+ font-size: 12px;
470
+ color: var(--accent);
471
+ text-decoration: none;
472
+ }
473
+ .cell-open:hover {
474
+ color: var(--accent-strong);
475
+ }
476
+ .cell-body {
477
+ min-width: 0;
478
+ padding: 0;
479
+ }
480
+ .cell.dashboard .cell-body {
481
+ padding: 0;
482
+ }
483
+ #page .cell-body h1,
484
+ #page .cell-body h2 {
485
+ font-family: var(--sans);
486
+ font-size: 17px;
487
+ font-weight: 700;
488
+ letter-spacing: -0.01em;
489
+ line-height: 1.35;
490
+ margin: 22px 0 6px;
491
+ }
492
+ #page .cell-body > :first-child {
493
+ margin-top: 0;
494
+ }
495
+ #page .cell-body > :last-child {
496
+ margin-bottom: 0;
497
+ }
498
+ .figure-fit {
499
+ position: relative;
500
+ overflow: hidden;
501
+ min-height: 160px;
502
+ border: 1px solid var(--line);
503
+ border-radius: 8px;
504
+ background: #fff;
505
+ }
506
+ .figure-fit[hidden] {
507
+ display: none;
508
+ }
509
+ .figure-fit:fullscreen,
510
+ .figure-fit:-webkit-full-screen {
511
+ width: 100%;
512
+ height: 100%;
513
+ border: none;
514
+ border-radius: 0;
515
+ }
516
+ .figure-frame {
517
+ display: block;
518
+ width: 100%;
519
+ min-height: 160px;
520
+ border: none;
521
+ background: #fff;
522
+ }
523
+ .figure-frame[hidden],
524
+ .figure-raw[hidden] {
525
+ display: none;
526
+ }
527
+ .fig-switch {
528
+ position: relative;
529
+ display: inline-flex;
530
+ flex: 0 0 auto;
531
+ border: 1px solid var(--line);
532
+ border-radius: 999px;
533
+ background: var(--code-bg);
534
+ padding: 2px;
535
+ }
536
+ .fig-switch button {
537
+ position: relative;
538
+ z-index: 1;
539
+ flex: 1;
540
+ min-width: 62px;
541
+ border: none;
542
+ background: none;
543
+ font-family: var(--sans);
544
+ font-size: 12px;
545
+ font-weight: 600;
546
+ color: var(--muted);
547
+ padding: 3px 12px;
548
+ border-radius: 999px;
549
+ cursor: pointer;
550
+ transition: color 0.15s;
551
+ }
552
+ .fig-switch button.active {
553
+ color: var(--accent-strong);
554
+ }
555
+ .fig-switch-thumb {
556
+ position: absolute;
557
+ top: 2px;
558
+ bottom: 2px;
559
+ left: 2px;
560
+ width: calc(50% - 2px);
561
+ border-radius: 999px;
562
+ background: var(--panel);
563
+ border: 1px solid rgba(249, 115, 22, 0.35);
564
+ box-shadow: 0 1px 4px rgba(31, 41, 55, 0.08);
565
+ transition: transform 0.18s ease;
566
+ }
567
+ .fig-switch.raw .fig-switch-thumb {
568
+ transform: translateX(100%);
569
+ }
570
+ #page .figure-raw pre {
571
+ margin: 0;
572
+ max-height: 420px;
573
+ overflow: auto;
574
+ font-family: var(--mono);
575
+ font-size: 13px;
576
+ line-height: 1.55;
577
+ background: var(--code-bg);
578
+ border: 1px solid var(--line);
579
+ border-radius: 8px;
580
+ padding: 12px 14px;
581
+ }
582
+ /* ---- figure fullscreen ---- */
583
+ .cell-fullscreen {
584
+ position: relative;
585
+ display: inline-flex;
586
+ flex: 0 0 auto;
587
+ }
588
+ .cell-fullscreen-btn {
589
+ display: inline-flex;
590
+ align-items: center;
591
+ justify-content: center;
592
+ width: 26px;
593
+ height: 26px;
594
+ padding: 0;
595
+ border: 1px solid var(--line);
596
+ border-radius: 999px;
597
+ background: var(--code-bg);
598
+ color: var(--muted);
599
+ cursor: pointer;
600
+ transition: color 0.15s, border-color 0.15s, background 0.15s;
601
+ }
602
+ .cell-fullscreen-btn:hover {
603
+ color: var(--accent-strong);
604
+ border-color: rgba(249, 115, 22, 0.35);
605
+ background: var(--accent-soft);
606
+ }
607
+ .cell-fullscreen-btn svg {
608
+ width: 14px;
609
+ height: 14px;
610
+ }
611
+ /* ---- copyable snippets ---- */
612
+ .snippet {
613
+ position: relative;
614
+ }
615
+ .copy-snippet {
616
+ position: absolute;
617
+ top: 7px;
618
+ right: 8px;
619
+ width: 24px;
620
+ height: 24px;
621
+ border: none;
622
+ border-radius: 6px;
623
+ background: rgba(255, 255, 255, 0.08);
624
+ color: #9a9da8;
625
+ font-size: 12px;
626
+ line-height: 1;
627
+ cursor: pointer;
628
+ opacity: 0;
629
+ transition: opacity 0.12s, color 0.12s, background 0.12s;
630
+ }
631
+ .snippet:hover .copy-snippet,
632
+ .jp-out:hover .copy-snippet,
633
+ .figure-raw:hover .copy-snippet,
634
+ .code-accordion summary:hover .copy-snippet {
635
+ opacity: 1;
636
+ }
637
+ .copy-snippet:hover {
638
+ color: #ffffff;
639
+ background: rgba(255, 255, 255, 0.16);
640
+ }
641
+ .copy-snippet.copied {
642
+ color: #52d08a;
643
+ opacity: 1;
644
+ }
645
+ .code-accordion .code-name {
646
+ user-select: text;
647
+ cursor: text;
648
+ }
649
+ .jp-out,
650
+ .figure-raw {
651
+ position: relative;
652
+ }
653
+ .jp-out .copy-snippet,
654
+ .figure-raw .copy-snippet {
655
+ background: var(--code-bg);
656
+ color: var(--muted);
657
+ border: 1px solid var(--line);
658
+ }
659
+ .jp-out .copy-snippet:hover,
660
+ .figure-raw .copy-snippet:hover {
661
+ color: var(--accent-strong);
662
+ background: var(--panel);
663
+ }
664
+
665
+ /* ---- jupyter-style code cells ---- */
666
+ .jp {
667
+ border: 1px solid var(--line);
668
+ border-radius: 10px;
669
+ overflow: hidden;
670
+ margin: 0;
671
+ background: var(--panel);
672
+ }
673
+ .jp-cmd {
674
+ display: flex;
675
+ align-items: baseline;
676
+ gap: 9px;
677
+ position: relative;
678
+ padding: 10px 16px 10px 0;
679
+ font-family: var(--mono);
680
+ font-size: 12px;
681
+ color: #8b8e98;
682
+ }
683
+ .jp-cmd-prompt {
684
+ color: var(--accent);
685
+ font-weight: 700;
686
+ }
687
+ #page .jp-cmd code {
688
+ min-width: 0;
689
+ color: #b6b9c2;
690
+ font-family: var(--mono);
691
+ font-size: 12px;
692
+ background: none;
693
+ padding: 0;
694
+ border-radius: 0;
695
+ overflow-wrap: anywhere;
696
+ }
697
+ .jp-cmd:hover .copy-snippet {
698
+ opacity: 1;
699
+ }
700
+ .jp-in-body .jp-cmd + .code-accordion,
701
+ .jp-in-body .jp-cmd + .snippet {
702
+ border-top: 1px solid rgba(255, 255, 255, 0.09);
703
+ }
704
+ .jp-gutter {
705
+ flex: 0 0 46px;
706
+ padding: 13px 0 0 13px;
707
+ font-family: var(--mono);
708
+ font-size: 10.5px;
709
+ letter-spacing: 0.07em;
710
+ text-transform: uppercase;
711
+ font-weight: 600;
712
+ user-select: none;
713
+ }
714
+ .jp-in {
715
+ display: flex;
716
+ background: #17181c;
717
+ }
718
+ .jp-in .jp-gutter {
719
+ color: #6f727d;
720
+ }
721
+ .jp-in-body {
722
+ flex: 1;
723
+ min-width: 0;
724
+ }
725
+ #page .jp-in-body pre.hl {
726
+ margin: 0;
727
+ border: none;
728
+ border-radius: 0;
729
+ background: none;
730
+ padding: 12px 16px 12px 0;
731
+ overflow-y: auto;
732
+ max-height: 26em;
733
+ }
734
+ .jp-in-body .code-accordion {
735
+ margin: 0;
736
+ border: none;
737
+ border-top: 1px solid rgba(255, 255, 255, 0.09);
738
+ border-radius: 0;
739
+ background: none;
740
+ }
741
+ .jp-in-body .code-accordion summary {
742
+ background: none;
743
+ padding: 9px 16px 9px 0;
744
+ }
745
+ .jp-in-body .code-accordion pre.hl {
746
+ border-top: 1px solid rgba(255, 255, 255, 0.09);
747
+ }
748
+ .jp-meta {
749
+ padding: 5px 14px;
750
+ font-family: var(--mono);
751
+ font-size: 11.5px;
752
+ color: var(--muted);
753
+ background: #fbfbfc;
754
+ border-top: 1px solid var(--line);
755
+ }
756
+ .jp-out {
757
+ display: flex;
758
+ border-top: 1px solid var(--line);
759
+ background: var(--panel);
760
+ }
761
+ .jp-out .jp-gutter {
762
+ color: var(--accent-strong);
763
+ }
764
+ .jp-out-body {
765
+ flex: 1;
766
+ min-width: 0;
767
+ }
768
+ #page .jp-out-pre {
769
+ min-width: 0;
770
+ margin: 0;
771
+ border: none;
772
+ border-radius: 0;
773
+ background: none;
774
+ color: var(--ink);
775
+ font-family: var(--mono);
776
+ font-size: 13px;
777
+ line-height: 1.55;
778
+ padding: 12px 16px 12px 0;
779
+ white-space: pre;
780
+ overflow-x: auto;
781
+ overflow-y: auto;
782
+ max-height: 26em;
783
+ }
784
+ .jp-artifacts {
785
+ display: flex;
786
+ flex-direction: column;
787
+ }
788
+ .jp-out-body .jp-out-pre + .jp-artifacts {
789
+ border-top: 1px solid var(--line);
790
+ }
791
+ .out-artifact {
792
+ display: flex;
793
+ align-items: baseline;
794
+ gap: 8px;
795
+ padding: 9px 16px 9px 0;
796
+ text-decoration: none;
797
+ color: inherit;
798
+ }
799
+ .out-artifact + .out-artifact {
800
+ border-top: 1px solid var(--line);
801
+ }
802
+ a.out-artifact:hover .out-artifact-name {
803
+ color: var(--accent-strong);
804
+ }
805
+ .out-artifact-ico {
806
+ flex: 0 0 auto;
807
+ font-size: 13px;
808
+ }
809
+ .out-artifact-name {
810
+ font-family: var(--mono);
811
+ font-size: 12.5px;
812
+ font-weight: 600;
813
+ color: var(--ink);
814
+ overflow: hidden;
815
+ text-overflow: ellipsis;
816
+ white-space: nowrap;
817
+ }
818
+ .out-artifact-meta {
819
+ flex: 0 0 auto;
820
+ margin-left: auto;
821
+ padding-left: 12px;
822
+ font-size: 12px;
823
+ color: var(--muted);
824
+ white-space: nowrap;
825
+ }
826
+ .out-artifact-state.open {
827
+ color: var(--accent);
828
+ font-weight: 600;
829
+ }
830
+ .trackio-embed {
831
+ border: 1px solid var(--line);
832
+ border-radius: var(--radius);
833
+ overflow: hidden;
834
+ background: var(--panel);
835
+ }
836
+ .trackio-cell-meta {
837
+ display: flex;
838
+ gap: 6px;
839
+ flex-wrap: wrap;
840
+ justify-content: flex-end;
841
+ }
842
+
843
+ /* ---- unfurl cards ---- */
844
+ .unfurl {
845
+ display: block;
846
+ border: 1px solid var(--line);
847
+ border-radius: var(--radius);
848
+ background: var(--panel);
849
+ margin: 12px 0;
850
+ overflow: hidden;
851
+ text-decoration: none;
852
+ color: inherit;
853
+ transition: border-color 0.14s, box-shadow 0.14s;
854
+ }
855
+ .unfurl:hover {
856
+ border-color: #cfcbe6;
857
+ box-shadow: 0 4px 18px rgba(30, 20, 80, 0.06);
858
+ }
859
+
860
+ .unfurl-body {
861
+ padding: 13px 16px;
862
+ display: flex;
863
+ gap: 12px;
864
+ align-items: flex-start;
865
+ }
866
+
867
+ .unfurl-ico {
868
+ font-size: 20px;
869
+ line-height: 1.3;
870
+ flex: 0 0 auto;
871
+ }
872
+
873
+ .unfurl-main {
874
+ min-width: 0;
875
+ flex: 1;
876
+ }
877
+
878
+ .unfurl-kind {
879
+ font-family: var(--mono);
880
+ font-size: 10.5px;
881
+ text-transform: uppercase;
882
+ letter-spacing: 0.08em;
883
+ color: var(--accent);
884
+ font-weight: 600;
885
+ }
886
+
887
+ .unfurl-title {
888
+ font-weight: 650;
889
+ font-size: 15px;
890
+ margin: 1px 0 2px;
891
+ white-space: nowrap;
892
+ overflow: hidden;
893
+ text-overflow: ellipsis;
894
+ }
895
+
896
+ .unfurl-desc {
897
+ color: var(--muted);
898
+ font-size: 13.5px;
899
+ line-height: 1.45;
900
+ }
901
+
902
+ .unfurl-meta {
903
+ margin-top: 6px;
904
+ display: flex;
905
+ flex-wrap: wrap;
906
+ gap: 6px;
907
+ }
908
+
909
+ .chip {
910
+ font-size: 11.5px;
911
+ background: var(--code-bg);
912
+ border-radius: 999px;
913
+ padding: 2px 9px;
914
+ color: var(--muted);
915
+ font-family: var(--mono);
916
+ }
917
+
918
+ .unfurl-raw {
919
+ font-family: var(--mono);
920
+ font-size: 11px;
921
+ color: var(--muted);
922
+ border-top: 1px solid var(--line);
923
+ padding: 7px 16px;
924
+ white-space: nowrap;
925
+ overflow: hidden;
926
+ text-overflow: ellipsis;
927
+ }
928
+
929
+ .unfurl.embed {
930
+ padding: 0;
931
+ overflow: hidden;
932
+ }
933
+ .embed-head {
934
+ display: flex;
935
+ align-items: center;
936
+ gap: 10px;
937
+ padding: 10px 14px;
938
+ border-bottom: 1px solid var(--line);
939
+ }
940
+ .embed-head .unfurl-kind {
941
+ flex: 0 0 auto;
942
+ }
943
+ .embed-title {
944
+ flex: 1;
945
+ min-width: 0;
946
+ font-weight: 650;
947
+ font-size: 14px;
948
+ color: var(--ink);
949
+ text-decoration: none;
950
+ white-space: nowrap;
951
+ overflow: hidden;
952
+ text-overflow: ellipsis;
953
+ }
954
+ .embed-title:hover {
955
+ color: var(--accent);
956
+ }
957
+ .embed-open {
958
+ flex: 0 0 auto;
959
+ font-family: var(--mono);
960
+ font-size: 12px;
961
+ color: var(--accent);
962
+ text-decoration: none;
963
+ }
964
+ .embed-frame {
965
+ display: block;
966
+ width: 100%;
967
+ height: 560px;
968
+ border: 0;
969
+ background: var(--code-bg);
970
+ }
971
+
972
+ .dashboard-shell {
973
+ display: block;
974
+ }
975
+ .dashboard-shell .dashboard-frame {
976
+ display: block;
977
+ width: 100%;
978
+ height: 900px;
979
+ border: 0;
980
+ background: var(--code-bg);
981
+ }
982
+
983
+ .unfurl.image {
984
+ padding: 0;
985
+ }
986
+ .unfurl.image img {
987
+ display: block;
988
+ width: 100%;
989
+ height: auto;
990
+ max-height: 460px;
991
+ object-fit: contain;
992
+ background: var(--code-bg);
993
+ }
994
+
995
+ .artifact-chip {
996
+ border: 1px solid var(--line);
997
+ background: var(--panel);
998
+ border-radius: var(--radius);
999
+ padding: 10px 14px;
1000
+ margin: 8px 0;
1001
+ font-size: 14px;
1002
+ }
1003
+ .cell.dashboard .artifact-chip {
1004
+ margin: 14px 18px 18px;
1005
+ }
1006
+ .artifact-chip code {
1007
+ color: var(--accent);
1008
+ }
1009
+
1010
+ /* ---- task board ---- */
1011
+ .board-wrap {
1012
+ overflow-x: auto;
1013
+ border: 1px solid var(--line);
1014
+ border-radius: var(--radius);
1015
+ margin: 12px 0 20px;
1016
+ background: var(--panel);
1017
+ }
1018
+ table.board {
1019
+ border-collapse: collapse;
1020
+ width: 100%;
1021
+ font-size: 14px;
1022
+ }
1023
+ table.board th,
1024
+ table.board td {
1025
+ text-align: left;
1026
+ padding: 9px 14px;
1027
+ border-bottom: 1px solid var(--line);
1028
+ vertical-align: top;
1029
+ }
1030
+ table.board thead th {
1031
+ background: var(--accent-soft);
1032
+ font-size: 12px;
1033
+ text-transform: uppercase;
1034
+ letter-spacing: 0.05em;
1035
+ color: #9a4a12;
1036
+ font-weight: 600;
1037
+ border-bottom: 1px solid var(--line);
1038
+ }
1039
+ table.board tbody tr:last-child td {
1040
+ border-bottom: none;
1041
+ }
1042
+ table.board .col-check {
1043
+ text-align: center;
1044
+ width: 92px;
1045
+ white-space: nowrap;
1046
+ }
1047
+ table.board tr.section-row td {
1048
+ background: var(--accent-soft);
1049
+ text-align: center;
1050
+ font-weight: 700;
1051
+ font-size: 13px;
1052
+ color: var(--accent-strong);
1053
+ padding: 7px 14px;
1054
+ letter-spacing: 0.02em;
1055
+ }
1056
+ .box {
1057
+ display: inline-flex;
1058
+ align-items: center;
1059
+ justify-content: center;
1060
+ width: 18px;
1061
+ height: 18px;
1062
+ border: 1.5px solid #cfcbe0;
1063
+ border-radius: 5px;
1064
+ font-size: 12px;
1065
+ color: #fff;
1066
+ line-height: 1;
1067
+ }
1068
+ .box.on {
1069
+ background: var(--accent);
1070
+ border-color: var(--accent);
1071
+ }
1072
+ .who-chip {
1073
+ display: inline-block;
1074
+ padding: 3px 12px;
1075
+ border-radius: 999px;
1076
+ font-size: 12.5px;
1077
+ font-weight: 600;
1078
+ white-space: nowrap;
1079
+ }
1080
+ .who-chip.muted {
1081
+ background: var(--code-bg);
1082
+ color: var(--muted);
1083
+ font-weight: 500;
1084
+ }
1085
+
1086
+ /* ---- status badges + clickable rows ---- */
1087
+ table.board .col-status {
1088
+ width: 130px;
1089
+ white-space: nowrap;
1090
+ }
1091
+ .badge {
1092
+ display: inline-block;
1093
+ padding: 3px 11px;
1094
+ border-radius: 999px;
1095
+ font-size: 12px;
1096
+ font-weight: 600;
1097
+ letter-spacing: 0.01em;
1098
+ }
1099
+ .badge.gray {
1100
+ background: var(--code-bg);
1101
+ color: var(--muted);
1102
+ }
1103
+ .badge.amber {
1104
+ background: var(--accent-soft);
1105
+ color: #b45309;
1106
+ }
1107
+ .badge.green {
1108
+ background: #e6f7ee;
1109
+ color: #1a8a55;
1110
+ }
1111
+ .badge.red {
1112
+ background: #fde8ec;
1113
+ color: #c62a4b;
1114
+ }
1115
+ table.board tr.linked-row {
1116
+ cursor: pointer;
1117
+ }
1118
+ table.board tr.linked-row:hover td {
1119
+ background: var(--accent-soft);
1120
+ }
1121
+ table.board tr.linked-row a {
1122
+ color: var(--ink);
1123
+ font-weight: 600;
1124
+ text-decoration: none;
1125
+ }
1126
+ table.board tr.linked-row:hover a {
1127
+ color: var(--accent-strong);
1128
+ }
1129
+
1130
+ /* ---- agent read hint ---- */
1131
+ .agent-hint {
1132
+ display: flex;
1133
+ align-items: center;
1134
+ flex-wrap: wrap;
1135
+ gap: 8px;
1136
+ margin: 0;
1137
+ font-size: 12.5px;
1138
+ color: var(--muted);
1139
+ }
1140
+ .agent-hint code {
1141
+ flex: 1 1 18rem;
1142
+ min-width: 0;
1143
+ background: var(--code-bg);
1144
+ padding: 2px 9px;
1145
+ border-radius: 6px;
1146
+ font-family: var(--mono);
1147
+ font-size: 12px;
1148
+ font-weight: 500;
1149
+ color: var(--ink);
1150
+ overflow: hidden;
1151
+ text-overflow: ellipsis;
1152
+ white-space: nowrap;
1153
+ }
1154
+ .agent-hint .copy {
1155
+ flex: 0 0 auto;
1156
+ background: none;
1157
+ color: var(--muted);
1158
+ border: 1px solid var(--line);
1159
+ border-radius: 6px;
1160
+ width: 22px;
1161
+ height: 22px;
1162
+ font-size: 11px;
1163
+ line-height: 1;
1164
+ cursor: pointer;
1165
+ transition: color 0.12s, border-color 0.12s;
1166
+ }
1167
+ .agent-hint .copy:hover {
1168
+ color: var(--accent-strong);
1169
+ border-color: var(--accent);
1170
+ }
1171
+ .agent-hint .copy.copied {
1172
+ color: #1a8a55;
1173
+ border-color: #1a8a55;
1174
+ }
1175
+ .agent-hint-note {
1176
+ margin-left: auto;
1177
+ font-size: 12px;
1178
+ color: var(--muted);
1179
+ }
1180
+ .hub-destination {
1181
+ display: flex;
1182
+ align-items: center;
1183
+ flex-wrap: wrap;
1184
+ gap: 8px;
1185
+ color: var(--muted);
1186
+ font-size: 12.5px;
1187
+ }
1188
+ .hub-destination a {
1189
+ display: inline-flex;
1190
+ align-items: center;
1191
+ gap: 6px;
1192
+ max-width: 100%;
1193
+ padding: 3px 9px;
1194
+ border: 1px solid var(--accent-line);
1195
+ border-radius: 999px;
1196
+ background: var(--accent-soft);
1197
+ color: var(--accent-strong);
1198
+ font-family: var(--mono);
1199
+ font-size: 12px;
1200
+ font-weight: 650;
1201
+ line-height: 1.5;
1202
+ text-decoration: none;
1203
+ overflow-wrap: anywhere;
1204
+ transition: border-color 0.12s, background 0.12s, color 0.12s;
1205
+ }
1206
+ .hub-destination a:hover {
1207
+ border-color: var(--accent);
1208
+ background: #ffedd5;
1209
+ color: #c2410c;
1210
+ }
1211
+ .hub-destination svg {
1212
+ width: 13px;
1213
+ height: 13px;
1214
+ flex: 0 0 auto;
1215
+ fill: none;
1216
+ stroke: currentColor;
1217
+ stroke-width: 1.8;
1218
+ stroke-linecap: round;
1219
+ stroke-linejoin: round;
1220
+ }
1221
+
1222
+ .index-paper-link {
1223
+ margin: 14px 0 30px;
1224
+ font-size: 19px;
1225
+ line-height: 1.35;
1226
+ font-weight: 700;
1227
+ }
1228
+ .index-paper-link a {
1229
+ text-underline-offset: 4px;
1230
+ text-decoration-thickness: 2px;
1231
+ }
1232
+ .art-ico {
1233
+ width: 1em;
1234
+ height: 1em;
1235
+ object-fit: contain;
1236
+ vertical-align: -0.15em;
1237
+ }
1238
+ .art-file-ico {
1239
+ width: 15px;
1240
+ height: 15px;
1241
+ flex: 0 0 auto;
1242
+ fill: none;
1243
+ stroke: currentColor;
1244
+ stroke-width: 1.7;
1245
+ stroke-linecap: round;
1246
+ stroke-linejoin: round;
1247
+ vertical-align: -0.2em;
1248
+ }
1249
+ .out-artifact-ico .art-file-ico {
1250
+ color: var(--muted);
1251
+ }
1252
+
1253
+ /* ---- scroll-to-resource highlight ---- */
1254
+ .res-flash {
1255
+ animation: res-flash 1.5s ease;
1256
+ border-radius: 8px;
1257
+ }
1258
+ @keyframes res-flash {
1259
+ 0%,
1260
+ 25% {
1261
+ box-shadow: 0 0 0 3px var(--accent);
1262
+ }
1263
+ 100% {
1264
+ box-shadow: 0 0 0 3px rgba(249, 115, 22, 0);
1265
+ }
1266
+ }
1267
+
1268
+ /* ---- inline resource chips ---- */
1269
+ #page .res-chip {
1270
+ display: inline-flex;
1271
+ align-items: center;
1272
+ gap: 5px;
1273
+ max-width: 100%;
1274
+ padding: 0 9px 0 6px;
1275
+ margin: 0 1px;
1276
+ border: 1px solid var(--line);
1277
+ border-radius: 999px;
1278
+ background: var(--panel);
1279
+ font-family: var(--mono);
1280
+ font-size: 0.78em;
1281
+ font-weight: 600;
1282
+ color: var(--ink);
1283
+ text-decoration: none;
1284
+ white-space: nowrap;
1285
+ overflow: hidden;
1286
+ text-overflow: ellipsis;
1287
+ vertical-align: middle;
1288
+ line-height: 1.65;
1289
+ transform: translateY(-0.08em);
1290
+ transition: border-color 0.12s, background 0.12s, color 0.12s;
1291
+ }
1292
+ .res-chip-ico {
1293
+ font-size: 1.05em;
1294
+ line-height: 1;
1295
+ }
1296
+ #page .res-chip:hover {
1297
+ border-color: var(--accent);
1298
+ background: var(--accent-soft);
1299
+ color: var(--accent-strong);
1300
+ }
1301
+
1302
+ /* ---- connect footer + modal ---- */
1303
+ #sidebar-foot {
1304
+ margin-top: auto;
1305
+ padding-top: 14px;
1306
+ border-top: 1px solid rgba(255, 255, 255, 0.1);
1307
+ }
1308
+
1309
+ #connect-btn {
1310
+ width: 100%;
1311
+ display: flex;
1312
+ align-items: center;
1313
+ gap: 8px;
1314
+ background: rgba(255, 255, 255, 0.05);
1315
+ color: #c3c4cb;
1316
+ border: 1px solid rgba(255, 255, 255, 0.12);
1317
+ border-radius: 9px;
1318
+ padding: 9px 12px;
1319
+ font-size: 13.5px;
1320
+ font-family: var(--sans);
1321
+ cursor: pointer;
1322
+ transition: background 0.12s, color 0.12s, border-color 0.12s;
1323
+ }
1324
+ #connect-btn:hover {
1325
+ background: rgba(249, 115, 22, 0.14);
1326
+ border-color: rgba(249, 115, 22, 0.4);
1327
+ color: #fdba74;
1328
+ }
1329
+ #connect-btn .ico {
1330
+ font-size: 15px;
1331
+ }
1332
+
1333
+ #modal[hidden] {
1334
+ display: none;
1335
+ }
1336
+ #modal {
1337
+ position: fixed;
1338
+ inset: 0;
1339
+ z-index: 100;
1340
+ display: flex;
1341
+ align-items: center;
1342
+ justify-content: center;
1343
+ padding: 24px;
1344
+ }
1345
+ .modal-backdrop {
1346
+ position: absolute;
1347
+ inset: 0;
1348
+ background: rgba(20, 18, 30, 0.5);
1349
+ backdrop-filter: blur(2px);
1350
+ }
1351
+ .modal-card {
1352
+ position: relative;
1353
+ background: var(--panel);
1354
+ border-radius: 16px;
1355
+ width: 100%;
1356
+ max-width: 620px;
1357
+ max-height: 85vh;
1358
+ overflow-y: auto;
1359
+ box-shadow: 0 24px 70px rgba(20, 15, 50, 0.28);
1360
+ }
1361
+ .modal-head {
1362
+ display: flex;
1363
+ align-items: center;
1364
+ justify-content: space-between;
1365
+ gap: 12px;
1366
+ padding: 18px 22px;
1367
+ border-bottom: 1px solid var(--line);
1368
+ position: sticky;
1369
+ top: 0;
1370
+ background: var(--panel);
1371
+ }
1372
+ .modal-title {
1373
+ display: flex;
1374
+ align-items: center;
1375
+ gap: 10px;
1376
+ font-family: var(--serif);
1377
+ font-size: 21px;
1378
+ letter-spacing: -0.01em;
1379
+ }
1380
+ .modal-logo {
1381
+ width: 26px;
1382
+ height: 26px;
1383
+ object-fit: contain;
1384
+ }
1385
+ .modal-actions {
1386
+ display: flex;
1387
+ align-items: center;
1388
+ gap: 8px;
1389
+ }
1390
+ .btn {
1391
+ font-family: var(--sans);
1392
+ font-size: 13.5px;
1393
+ font-weight: 600;
1394
+ border: 1px solid var(--line);
1395
+ background: var(--panel);
1396
+ color: var(--ink);
1397
+ border-radius: 9px;
1398
+ padding: 8px 13px;
1399
+ cursor: pointer;
1400
+ transition: background 0.12s, border-color 0.12s, color 0.12s;
1401
+ }
1402
+ .btn:hover {
1403
+ border-color: var(--accent);
1404
+ color: var(--accent-strong);
1405
+ }
1406
+ .btn.copied {
1407
+ border-color: #1a8a55;
1408
+ color: #1a8a55;
1409
+ }
1410
+ .btn.icon {
1411
+ font-size: 18px;
1412
+ line-height: 1;
1413
+ padding: 6px 11px;
1414
+ font-weight: 400;
1415
+ }
1416
+ .modal-body {
1417
+ padding: 20px 22px 26px;
1418
+ }
1419
+ .modal-intro {
1420
+ margin: 0 0 20px;
1421
+ color: var(--muted);
1422
+ line-height: 1.55;
1423
+ }
1424
+ #connect-steps {
1425
+ list-style: none;
1426
+ margin: 0;
1427
+ padding: 0;
1428
+ }
1429
+ #connect-steps li {
1430
+ margin-bottom: 18px;
1431
+ }
1432
+ .step-title {
1433
+ font-weight: 600;
1434
+ font-size: 14.5px;
1435
+ margin-bottom: 8px;
1436
+ }
1437
+ .codeblock {
1438
+ display: flex;
1439
+ align-items: center;
1440
+ gap: 8px;
1441
+ background: #17181c;
1442
+ border-radius: 10px;
1443
+ padding: 11px 12px 11px 15px;
1444
+ }
1445
+ .codeblock code {
1446
+ flex: 1;
1447
+ min-width: 0;
1448
+ overflow-x: auto;
1449
+ white-space: nowrap;
1450
+ font-family: var(--mono);
1451
+ font-size: 13px;
1452
+ color: #f0efff;
1453
+ background: none;
1454
+ padding: 0;
1455
+ }
1456
+ .codeblock .copy {
1457
+ flex: 0 0 auto;
1458
+ background: rgba(255, 255, 255, 0.08);
1459
+ color: #c3c4cb;
1460
+ border: 1px solid rgba(255, 255, 255, 0.14);
1461
+ border-radius: 7px;
1462
+ width: 30px;
1463
+ height: 30px;
1464
+ font-size: 14px;
1465
+ cursor: pointer;
1466
+ transition: background 0.12s, color 0.12s;
1467
+ }
1468
+ .codeblock .copy:hover {
1469
+ background: rgba(249, 115, 22, 0.2);
1470
+ color: #fdba74;
1471
+ }
1472
+ .codeblock .copy.copied {
1473
+ color: #52d08a;
1474
+ }
1475
+
1476
+ /* ---- top-level logbook views ---- */
1477
+ #view-tabs {
1478
+ position: sticky;
1479
+ top: 0;
1480
+ z-index: 30;
1481
+ width: 100%;
1482
+ max-width: 1080px;
1483
+ margin: 0 auto 24px;
1484
+ padding-top: 10px;
1485
+ display: flex;
1486
+ align-items: center;
1487
+ justify-content: flex-start;
1488
+ gap: 26px;
1489
+ border-bottom: 1px solid var(--line);
1490
+ background: var(--paper);
1491
+ }
1492
+ #view-tabs a {
1493
+ display: inline-flex;
1494
+ align-items: center;
1495
+ gap: 8px;
1496
+ min-height: 44px;
1497
+ margin-bottom: -1px;
1498
+ color: var(--muted);
1499
+ border-bottom: 2px solid transparent;
1500
+ text-decoration: none;
1501
+ font-size: 13.5px;
1502
+ font-weight: 600;
1503
+ transition: color 0.12s, border-color 0.12s;
1504
+ }
1505
+ #view-tabs a:hover {
1506
+ color: var(--ink);
1507
+ }
1508
+ #view-tabs a.active {
1509
+ color: var(--accent-strong);
1510
+ border-bottom-color: var(--accent);
1511
+ }
1512
+ #view-tabs svg {
1513
+ width: 18px;
1514
+ height: 18px;
1515
+ flex: 0 0 auto;
1516
+ fill: none;
1517
+ stroke: currentColor;
1518
+ stroke-width: 2;
1519
+ stroke-linecap: round;
1520
+ stroke-linejoin: round;
1521
+ }
1522
+ .workspace-file svg,
1523
+ .workspace-folder summary svg,
1524
+ .workspace-download svg {
1525
+ width: 17px;
1526
+ height: 17px;
1527
+ flex: 0 0 auto;
1528
+ fill: none;
1529
+ stroke: currentColor;
1530
+ stroke-width: 1.7;
1531
+ stroke-linecap: round;
1532
+ stroke-linejoin: round;
1533
+ }
1534
+
1535
+ #page.trace-page,
1536
+ #page.workspace-page {
1537
+ max-width: 1080px;
1538
+ }
1539
+ .view-loading {
1540
+ padding: 72px 0;
1541
+ color: var(--muted);
1542
+ text-align: center;
1543
+ }
1544
+ .view-empty {
1545
+ margin: 48px 0;
1546
+ padding: 44px 28px;
1547
+ border: 1px dashed #d8dbe1;
1548
+ border-radius: var(--radius);
1549
+ background: rgba(255, 255, 255, 0.72);
1550
+ text-align: center;
1551
+ }
1552
+ .view-empty h2 {
1553
+ margin: 0 0 7px;
1554
+ font-size: 18px;
1555
+ }
1556
+ .view-empty p {
1557
+ max-width: 560px;
1558
+ margin: 0 auto;
1559
+ color: var(--muted);
1560
+ }
1561
+ .view-empty code {
1562
+ display: inline-block;
1563
+ margin-top: 18px;
1564
+ padding: 7px 10px;
1565
+ border-radius: 7px;
1566
+ background: var(--code-bg);
1567
+ font-family: var(--mono);
1568
+ font-size: 12px;
1569
+ }
1570
+ #page .repo-ref-link {
1571
+ display: inline-block;
1572
+ margin-top: 18px;
1573
+ padding: 8px 14px;
1574
+ border-radius: 8px;
1575
+ background: var(--accent-strong, #2158d0);
1576
+ color: #fff;
1577
+ font-weight: 600;
1578
+ text-decoration: none;
1579
+ }
1580
+ #page .repo-ref-link:hover,
1581
+ #page .repo-ref-link:focus-visible {
1582
+ color: #fff;
1583
+ filter: brightness(0.95);
1584
+ }
1585
+ .view-eyebrow {
1586
+ margin-bottom: 4px;
1587
+ color: var(--accent-strong);
1588
+ font-family: var(--mono);
1589
+ font-size: 11px;
1590
+ font-weight: 700;
1591
+ letter-spacing: 0.12em;
1592
+ text-transform: uppercase;
1593
+ }
1594
+
1595
+ /* ---- trace ---- */
1596
+ .trace-session {
1597
+ scroll-margin-top: 24px;
1598
+ }
1599
+ .trace-session + .trace-session {
1600
+ margin-top: 44px;
1601
+ padding-top: 40px;
1602
+ border-top: 1px solid var(--line);
1603
+ }
1604
+ .trace-session-title {
1605
+ margin: 0 0 14px;
1606
+ color: var(--ink);
1607
+ font-family: var(--serif);
1608
+ font-size: 22px;
1609
+ line-height: 1.2;
1610
+ letter-spacing: -0.02em;
1611
+ overflow-wrap: anywhere;
1612
+ }
1613
+ .workspace-header h1 {
1614
+ margin: 0;
1615
+ color: var(--ink);
1616
+ font-size: 30px;
1617
+ line-height: 1.2;
1618
+ letter-spacing: -0.025em;
1619
+ }
1620
+ .trace-meta {
1621
+ display: flex;
1622
+ flex-wrap: wrap;
1623
+ gap: 9px 20px;
1624
+ margin-bottom: 34px;
1625
+ padding: 14px 16px;
1626
+ border: 1px solid var(--line);
1627
+ border-radius: 10px;
1628
+ background: rgba(255, 255, 255, 0.78);
1629
+ color: var(--muted);
1630
+ font-family: var(--mono);
1631
+ font-size: 11px;
1632
+ }
1633
+ .trace-meta strong {
1634
+ color: var(--ink);
1635
+ font-weight: 650;
1636
+ }
1637
+ .trace-source-missing {
1638
+ color: #b45309;
1639
+ }
1640
+ .trace-timeline {
1641
+ position: relative;
1642
+ }
1643
+ .trace-timeline::before {
1644
+ content: "";
1645
+ position: absolute;
1646
+ top: 0;
1647
+ bottom: 0;
1648
+ left: 82px;
1649
+ width: 1px;
1650
+ background: #dedfe3;
1651
+ }
1652
+ .trace-load-controls {
1653
+ display: flex;
1654
+ align-items: center;
1655
+ justify-content: space-between;
1656
+ gap: 16px;
1657
+ margin: 22px 0 0 100px;
1658
+ padding-top: 16px;
1659
+ border-top: 1px solid var(--line);
1660
+ }
1661
+ .trace-load-progress {
1662
+ color: var(--muted);
1663
+ font-family: var(--mono);
1664
+ font-size: 11px;
1665
+ }
1666
+ .trace-load-more {
1667
+ padding: 7px 12px;
1668
+ border: 1px solid var(--line-strong);
1669
+ border-radius: 7px;
1670
+ background: var(--paper);
1671
+ color: var(--ink);
1672
+ cursor: pointer;
1673
+ font: 650 12px/1.2 var(--sans);
1674
+ }
1675
+ .trace-load-more:hover:not(:disabled) {
1676
+ border-color: var(--accent);
1677
+ color: var(--accent-strong);
1678
+ }
1679
+ .trace-load-more:disabled {
1680
+ cursor: default;
1681
+ opacity: 0.65;
1682
+ }
1683
+ .trace-entry {
1684
+ --trace-depth: 0;
1685
+ position: relative;
1686
+ display: grid;
1687
+ grid-template-columns: 100px minmax(0, 1fr);
1688
+ margin: 0 0 18px calc(var(--trace-depth) * 24px);
1689
+ }
1690
+ .trace-rail {
1691
+ position: relative;
1692
+ min-height: 36px;
1693
+ padding: 4px 28px 0 0;
1694
+ color: #8a8d95;
1695
+ text-align: right;
1696
+ font-family: var(--mono);
1697
+ }
1698
+ .trace-number,
1699
+ .trace-elapsed {
1700
+ display: block;
1701
+ white-space: nowrap;
1702
+ }
1703
+ .trace-number {
1704
+ font-size: 12px;
1705
+ font-weight: 650;
1706
+ }
1707
+ .trace-elapsed {
1708
+ margin-top: 3px;
1709
+ font-size: 10px;
1710
+ }
1711
+ .trace-dot {
1712
+ position: absolute;
1713
+ top: 10px;
1714
+ right: 11px;
1715
+ width: 11px;
1716
+ height: 11px;
1717
+ border: 2px solid var(--paper);
1718
+ border-radius: 50%;
1719
+ background: var(--accent);
1720
+ box-shadow: 0 0 0 1px #d7d9de;
1721
+ }
1722
+ .trace-card {
1723
+ min-width: 0;
1724
+ overflow: hidden;
1725
+ border: 1px solid #dddfe4;
1726
+ border-radius: 11px;
1727
+ background: rgba(255, 255, 255, 0.92);
1728
+ }
1729
+ .trace-card > header {
1730
+ display: flex;
1731
+ align-items: center;
1732
+ gap: 10px;
1733
+ min-height: 37px;
1734
+ padding: 8px 13px;
1735
+ border-bottom: 1px solid #eceef1;
1736
+ }
1737
+ .trace-status .trace-card > header {
1738
+ border-bottom: 0;
1739
+ padding-bottom: 5px;
1740
+ }
1741
+ .trace-kind {
1742
+ font-family: var(--mono);
1743
+ font-size: 10.5px;
1744
+ font-weight: 750;
1745
+ letter-spacing: 0.08em;
1746
+ text-transform: uppercase;
1747
+ }
1748
+ .trace-turn {
1749
+ color: var(--muted);
1750
+ font: 10px var(--mono);
1751
+ }
1752
+ .trace-status-badge {
1753
+ margin-left: auto;
1754
+ padding: 1px 6px;
1755
+ border-radius: 999px;
1756
+ background: #eef0f3;
1757
+ color: var(--muted);
1758
+ font: 9.5px var(--mono);
1759
+ text-transform: uppercase;
1760
+ }
1761
+ .trace-status-badge-error,
1762
+ .trace-status-badge-failed {
1763
+ background: #fef2f2;
1764
+ color: #b91c1c;
1765
+ }
1766
+ .trace-body {
1767
+ margin: 0;
1768
+ padding: 15px 17px 17px;
1769
+ overflow-wrap: anywhere;
1770
+ white-space: pre-wrap;
1771
+ font-family: var(--sans);
1772
+ font-size: 13px;
1773
+ line-height: 1.65;
1774
+ }
1775
+ .trace-reasoning .trace-card {
1776
+ border-style: dashed;
1777
+ border-color: #d7b98a;
1778
+ background: #fffdf8;
1779
+ }
1780
+ .trace-reasoning .trace-kind {
1781
+ color: #9a6b22;
1782
+ }
1783
+ .trace-reasoning .trace-body {
1784
+ font-style: italic;
1785
+ }
1786
+ .trace-user .trace-card {
1787
+ border-left: 3px solid #f3a66d;
1788
+ }
1789
+ .trace-tool_call .trace-card,
1790
+ .trace-tool_result .trace-card {
1791
+ border-color: #2d3036;
1792
+ background: #191a1e;
1793
+ color: #ececf0;
1794
+ }
1795
+ .trace-tool_call .trace-card > header,
1796
+ .trace-tool_result .trace-card > header {
1797
+ border-bottom-color: rgba(255, 255, 255, 0.1);
1798
+ }
1799
+ .trace-tool_call .trace-kind,
1800
+ .trace-tool_result .trace-kind {
1801
+ color: #f5a66d;
1802
+ }
1803
+ .trace-tool_call .trace-turn,
1804
+ .trace-tool_result .trace-turn {
1805
+ color: #979aa3;
1806
+ }
1807
+ .trace-tool_call .trace-body,
1808
+ .trace-tool_result .trace-body,
1809
+ .trace-output pre {
1810
+ font-family: var(--mono);
1811
+ font-size: 11.5px;
1812
+ line-height: 1.6;
1813
+ }
1814
+ #page .trace-tool_call pre.trace-body,
1815
+ #page .trace-tool_result pre.trace-body {
1816
+ margin: 0;
1817
+ padding: 15px 17px 17px;
1818
+ border: 0;
1819
+ border-radius: 0;
1820
+ background: transparent;
1821
+ color: #ececf0;
1822
+ }
1823
+ .trace-output {
1824
+ border-top: 1px dashed rgba(255, 255, 255, 0.14);
1825
+ }
1826
+ .trace-output summary {
1827
+ padding: 9px 14px;
1828
+ color: #aaaeb7;
1829
+ cursor: pointer;
1830
+ font: 700 10px var(--mono);
1831
+ letter-spacing: 0.06em;
1832
+ text-transform: uppercase;
1833
+ }
1834
+ #page .trace-output pre {
1835
+ max-height: 480px;
1836
+ margin: 0;
1837
+ padding: 0 16px 16px;
1838
+ border: 0;
1839
+ border-radius: 0;
1840
+ background: transparent;
1841
+ overflow: auto;
1842
+ color: #d7d8dd;
1843
+ white-space: pre-wrap;
1844
+ }
1845
+
1846
+ /* ---- workspace ---- */
1847
+ .workspace-header {
1848
+ padding-bottom: 24px;
1849
+ }
1850
+ .workspace-header p {
1851
+ margin: 0;
1852
+ color: var(--muted);
1853
+ font-family: var(--mono);
1854
+ font-size: 11px;
1855
+ }
1856
+ .workspace-inventory {
1857
+ overflow: hidden;
1858
+ border: 1px solid var(--line);
1859
+ border-radius: 11px;
1860
+ background: rgba(255, 255, 255, 0.92);
1861
+ }
1862
+ .workspace-folder > summary {
1863
+ display: flex;
1864
+ align-items: center;
1865
+ gap: 8px;
1866
+ min-height: 39px;
1867
+ padding: 8px 13px;
1868
+ background: #fafafa;
1869
+ cursor: pointer;
1870
+ font-weight: 650;
1871
+ list-style: none;
1872
+ }
1873
+ .workspace-folder > summary::-webkit-details-marker {
1874
+ display: none;
1875
+ }
1876
+ .workspace-folder > summary::after {
1877
+ content: "›";
1878
+ margin-left: auto;
1879
+ color: #989ba2;
1880
+ transform: rotate(90deg);
1881
+ }
1882
+ .workspace-folder:not([open]) > summary::after {
1883
+ transform: rotate(0);
1884
+ }
1885
+ .workspace-folder-children {
1886
+ padding-left: 20px;
1887
+ }
1888
+ .workspace-file {
1889
+ display: grid;
1890
+ grid-template-columns: minmax(180px, 1fr) 72px 78px 180px 36px;
1891
+ align-items: center;
1892
+ min-height: 44px;
1893
+ padding: 7px 10px 7px 13px;
1894
+ color: var(--muted);
1895
+ font-family: var(--mono);
1896
+ font-size: 10.5px;
1897
+ }
1898
+ .workspace-file-name {
1899
+ display: flex;
1900
+ align-items: center;
1901
+ min-width: 0;
1902
+ gap: 8px;
1903
+ color: var(--ink);
1904
+ font-family: var(--sans);
1905
+ font-size: 12.5px;
1906
+ font-weight: 550;
1907
+ }
1908
+ .workspace-file-name span {
1909
+ overflow: hidden;
1910
+ text-overflow: ellipsis;
1911
+ white-space: nowrap;
1912
+ }
1913
+ .workspace-file-type {
1914
+ width: fit-content;
1915
+ padding: 1px 6px;
1916
+ border-radius: 999px;
1917
+ background: var(--accent-soft);
1918
+ color: var(--accent-strong);
1919
+ text-transform: uppercase;
1920
+ }
1921
+ .workspace-download {
1922
+ display: inline-flex;
1923
+ align-items: center;
1924
+ justify-content: center;
1925
+ width: 30px;
1926
+ height: 30px;
1927
+ border-radius: 7px;
1928
+ color: var(--muted);
1929
+ }
1930
+ .workspace-download:hover {
1931
+ background: var(--accent-soft);
1932
+ color: var(--accent-strong);
1933
+ }
1934
+ .workspace-unpublished {
1935
+ color: #9ca3af;
1936
+ text-align: center;
1937
+ }
1938
+
1939
+ .workspace-header {
1940
+ display: flex;
1941
+ align-items: center;
1942
+ justify-content: space-between;
1943
+ gap: 16px;
1944
+ flex-wrap: wrap;
1945
+ }
1946
+ .workspace-toggle {
1947
+ display: inline-flex;
1948
+ align-items: center;
1949
+ padding: 2px;
1950
+ border: 1px solid var(--line);
1951
+ border-radius: 999px;
1952
+ background: #fafafa;
1953
+ }
1954
+ .workspace-toggle-btn {
1955
+ padding: 4px 13px;
1956
+ border: 0;
1957
+ border-radius: 999px;
1958
+ background: transparent;
1959
+ color: var(--muted);
1960
+ font-family: var(--sans);
1961
+ font-size: 12px;
1962
+ font-weight: 600;
1963
+ cursor: pointer;
1964
+ }
1965
+ .workspace-toggle-btn:hover {
1966
+ color: var(--accent-strong);
1967
+ }
1968
+ .workspace-toggle-btn.is-active {
1969
+ background: var(--accent);
1970
+ color: #ffffff;
1971
+ }
1972
+ .workspace-group + .workspace-group {
1973
+ margin-top: 18px;
1974
+ }
1975
+ .workspace-group-head,
1976
+ .workspace-hub-group-head {
1977
+ display: flex;
1978
+ align-items: center;
1979
+ gap: 8px;
1980
+ margin: 0;
1981
+ padding: 8px 13px;
1982
+ background: #fafafa;
1983
+ border-bottom: 1px solid var(--line);
1984
+ color: var(--ink);
1985
+ font-family: var(--sans);
1986
+ font-size: 12px;
1987
+ font-weight: 650;
1988
+ text-transform: capitalize;
1989
+ }
1990
+ .workspace-group-count,
1991
+ .workspace-hub-count {
1992
+ padding: 0 7px;
1993
+ border-radius: 999px;
1994
+ background: var(--accent-soft);
1995
+ color: var(--accent-strong);
1996
+ font-family: var(--mono);
1997
+ font-size: 10.5px;
1998
+ }
1999
+ .workspace-group {
2000
+ overflow: hidden;
2001
+ border: 1px solid var(--line);
2002
+ border-radius: 11px;
2003
+ background: rgba(255, 255, 255, 0.92);
2004
+ }
2005
+
2006
+ .workspace-hub {
2007
+ margin-top: 28px;
2008
+ }
2009
+ .workspace-hub-title {
2010
+ margin: 0 0 14px;
2011
+ font-family: var(--sans);
2012
+ font-size: 16px;
2013
+ font-weight: 700;
2014
+ color: var(--ink);
2015
+ }
2016
+ .workspace-hub-group {
2017
+ overflow: hidden;
2018
+ border: 1px solid var(--line);
2019
+ border-radius: 11px;
2020
+ background: rgba(255, 255, 255, 0.92);
2021
+ }
2022
+ .workspace-hub-group + .workspace-hub-group {
2023
+ margin-top: 14px;
2024
+ }
2025
+ .workspace-hub-list {
2026
+ display: flex;
2027
+ flex-direction: column;
2028
+ }
2029
+ .workspace-hub-link {
2030
+ padding: 9px 13px;
2031
+ color: var(--accent-strong);
2032
+ font-family: var(--mono);
2033
+ font-size: 12px;
2034
+ text-decoration: none;
2035
+ overflow: hidden;
2036
+ text-overflow: ellipsis;
2037
+ white-space: nowrap;
2038
+ }
2039
+ .workspace-hub-link + .workspace-hub-link {
2040
+ border-top: 1px solid var(--line);
2041
+ }
2042
+ .workspace-hub-link:hover {
2043
+ background: var(--accent-soft);
2044
+ text-decoration: underline;
2045
+ }
2046
+
2047
+ /* --- UI nits --- */
2048
+ /* Flush group headers: #page h3/h2 (ID selectors) otherwise inject a top margin
2049
+ that, with overflow:hidden on the card, shows as whitespace above "Jobs" etc. */
2050
+ #page .workspace-hub-title {
2051
+ margin: 0 0 14px;
2052
+ }
2053
+ #page .workspace-hub-group-head,
2054
+ #page .workspace-group-head {
2055
+ margin: 0;
2056
+ }
2057
+ /* HF brand logo before the "Hugging Face artifacts" heading */
2058
+ .workspace-hub-title {
2059
+ display: flex;
2060
+ align-items: center;
2061
+ gap: 9px;
2062
+ }
2063
+ .workspace-hub-logo {
2064
+ width: 22px;
2065
+ height: 22px;
2066
+ flex: none;
2067
+ }
2068
+ /* Center empty-state placeholders (heading, body, command) */
2069
+ .view-empty {
2070
+ display: flex;
2071
+ flex-direction: column;
2072
+ align-items: center;
2073
+ }
2074
+ #page .view-empty h2,
2075
+ #page .view-empty p {
2076
+ text-align: center;
2077
+ }
2078
+
2079
+ @media (max-width: 720px) {
2080
+ #app {
2081
+ flex-direction: column;
2082
+ }
2083
+ #sidebar {
2084
+ width: 100%;
2085
+ flex: none;
2086
+ height: auto;
2087
+ position: static;
2088
+ }
2089
+ #content {
2090
+ display: block;
2091
+ width: 100%;
2092
+ padding: 28px 20px 80px;
2093
+ overflow-x: hidden;
2094
+ }
2095
+ #view-tabs {
2096
+ margin: 0 0 20px;
2097
+ gap: 18px;
2098
+ justify-content: flex-start;
2099
+ overflow-x: auto;
2100
+ }
2101
+ #view-tabs a {
2102
+ flex: 0 0 auto;
2103
+ }
2104
+ .trace-timeline::before {
2105
+ left: 16px;
2106
+ }
2107
+ .trace-entry {
2108
+ grid-template-columns: 32px minmax(0, 1fr);
2109
+ margin-left: calc(var(--trace-depth) * 10px);
2110
+ }
2111
+ .trace-rail {
2112
+ padding: 0;
2113
+ }
2114
+ .trace-number,
2115
+ .trace-elapsed {
2116
+ display: none;
2117
+ }
2118
+ .trace-dot {
2119
+ top: 10px;
2120
+ right: 10px;
2121
+ }
2122
+ .workspace-file {
2123
+ grid-template-columns: minmax(150px, 1fr) 66px 34px;
2124
+ }
2125
+ .workspace-file-size,
2126
+ .workspace-file-time {
2127
+ display: none;
2128
+ }
2129
+ #page {
2130
+ width: 100%;
2131
+ max-width: 100%;
2132
+ }
2133
+ #page h1,
2134
+ #logbook-title {
2135
+ font-size: 30px;
2136
+ }
2137
+ .cell-head {
2138
+ align-items: flex-start;
2139
+ flex-direction: column;
2140
+ gap: 4px;
2141
+ }
2142
+ }
logbook.js ADDED
The diff for this file is too large to render. See raw diff
 
logbook.json ADDED
@@ -0,0 +1,87 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 2,
3
+ "title": "Reproduction: Adaptive Personalized Federated Learning via Multi-task Averaging of Kernel Mean Embeddings",
4
+ "emoji": "🎯",
5
+ "space_id": "vimarsh/repro-adaptive-personalized-fl-multi-task-averaging-kernel-mean-embedding",
6
+ "paper": {
7
+ "arxiv_id": "2603.02233"
8
+ },
9
+ "tags": [
10
+ "icml2026-repro",
11
+ "paper-ADHGbxsREr"
12
+ ],
13
+ "updated_at": "2026-07-28T06:21:29+00:00",
14
+ "root": {
15
+ "slug": "index",
16
+ "title": "Reproduction: Adaptive Personalized Federated Learning via Multi-task Averaging of Kernel Mean Embeddings",
17
+ "file": "pages/index.md",
18
+ "children": [
19
+ {
20
+ "slug": "executive-summary",
21
+ "title": "Executive summary",
22
+ "file": "pages/executive-summary/page.md",
23
+ "children": []
24
+ },
25
+ {
26
+ "slug": "claim-1-theorem-4-4-finite-sample-kme-guarantee",
27
+ "title": "Claim 1: Theorem 4.4 finite-sample KME guarantee",
28
+ "file": "pages/claim-1-theorem-4-4-finite-sample-kme-guarantee/page.md",
29
+ "children": []
30
+ },
31
+ {
32
+ "slug": "claim-2-corollary-4-6-downstream-excess-risk-transfer",
33
+ "title": "Claim 2: Corollary 4.6 downstream excess-risk transfer",
34
+ "file": "pages/claim-2-corollary-4-6-downstream-excess-risk-transfer/page.md",
35
+ "children": []
36
+ },
37
+ {
38
+ "slug": "claim-3-example-4-5-automatic-adaptive-collaboration",
39
+ "title": "Claim 3: Example 4.5 automatic adaptive collaboration",
40
+ "file": "pages/claim-3-example-4-5-automatic-adaptive-collaboration/page.md",
41
+ "children": []
42
+ },
43
+ {
44
+ "slug": "claim-4-theorem-5-2-random-fourier-feature-tradeoff",
45
+ "title": "Claim 4: Theorem 5.2 random Fourier-feature tradeoff",
46
+ "file": "pages/claim-4-theorem-5-2-random-fourier-feature-tradeoff/page.md",
47
+ "children": []
48
+ },
49
+ {
50
+ "slug": "claim-5-section-6-1-100-agent-synthetic-concept-shift",
51
+ "title": "Claim 5: Section 6.1 100-agent synthetic concept shift",
52
+ "file": "pages/claim-5-section-6-1-100-agent-synthetic-concept-shift/page.md",
53
+ "children": []
54
+ },
55
+ {
56
+ "slug": "claim-6-section-6-3-femnist-192-agent-benchmark",
57
+ "title": "Claim 6: Section 6.3 FEMNIST 192-agent benchmark",
58
+ "file": "pages/claim-6-section-6-3-femnist-192-agent-benchmark/page.md",
59
+ "children": []
60
+ },
61
+ {
62
+ "slug": "conclusion",
63
+ "title": "Conclusion",
64
+ "file": "pages/conclusion/page.md",
65
+ "children": []
66
+ }
67
+ ]
68
+ },
69
+ "traces": [],
70
+ "workspace": {
71
+ "file": "workspace.json",
72
+ "file_count": 0,
73
+ "total_size": 0,
74
+ "bucket_id": null
75
+ },
76
+ "agent_view_tokens": 4609,
77
+ "trace_view_tokens": 10,
78
+ "workspace_view_tokens": 48,
79
+ "revision": "da220efeb6d76cd226b2",
80
+ "workspace_ref": {
81
+ "repo_id": "vimarsh/repro-adaptive-personalized-fl-multi-task-averaging-kernel-mean-embedding-artifacts",
82
+ "repo_type": "bucket",
83
+ "repo_url": "https://huggingface.co/buckets/vimarsh/repro-adaptive-personalized-fl-multi-task-averaging-kernel-mean-embedding-artifacts",
84
+ "private": true
85
+ },
86
+ "workspace_bucket": "https://huggingface.co/buckets/vimarsh/repro-adaptive-personalized-fl-multi-task-averaging-kernel-mean-embedding-artifacts"
87
+ }
pages/claim-1-theorem-4-4-finite-sample-kme-guarantee/page.md ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Claim 1: Theorem 4.4 finite-sample KME guarantee
2
+
3
+
4
+ ---
5
+ <!-- trackio-cell
6
+ {"type": "markdown", "id": "cell_069091946489", "created_at": "2026-07-19T09:32:03+00:00", "title": "Claim 1: Theorem 4.4 finite-sample KME guarantee"}
7
+ -->
8
+ Document setup, runs, and results for **Claim 1: Theorem 4.4 finite-sample KME guarantee**.
9
+
10
+
11
+ ---
12
+ <!-- trackio-cell
13
+ {"type": "markdown", "id": "cell_40b9d46f1306", "created_at": "2026-07-19T09:54:23+00:00", "title": "Evidence and status"}
14
+ -->
15
+ The source theorem was checked against the [arXiv PDF](https://arxiv.org/pdf/2603.02233), Theorem 4.4 (Eq. 14–15). For every subset containing target agent 1, it bounds expected squared MMD by a bias term , variance , and lower-order terms involving , , and effective dimension. This directly establishes the stated bias–variance tradeoff. **Status: source-verified theorem statement; not independently re-proven.**
16
+
17
+
18
+ ---
19
+ <!-- trackio-cell
20
+ {"type": "markdown", "id": "cell_7efaad82fdef", "created_at": "2026-07-19T09:56:03+00:00", "title": "Clarified evidence"}
21
+ -->
22
+ Clarification: in Eq. 14, for every V containing agent 1, the leading terms are Delta_V squared plus (Tr Sigma_1 + 2 Delta_V)/n_V. Delta_V is the largest distributional MMD within V and n_V is the combined client sample size. The remaining terms depend on log(B n_1), |V|/n_V, and effective dimension. This is a direct source audit of https://arxiv.org/pdf/2603.02233, not an independent proof.
pages/claim-2-corollary-4-6-downstream-excess-risk-transfer/page.md ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Claim 2: Corollary 4.6 downstream excess-risk transfer
2
+
3
+
4
+ ---
5
+ <!-- trackio-cell
6
+ {"type": "markdown", "id": "cell_6ed1cb9e8ba8", "created_at": "2026-07-19T09:32:03+00:00", "title": "Claim 2: Corollary 4.6 downstream excess-risk transfer"}
7
+ -->
8
+ Document setup, runs, and results for **Claim 2: Corollary 4.6 downstream excess-risk transfer**.
9
+
10
+
11
+ ---
12
+ <!-- trackio-cell
13
+ {"type": "markdown", "id": "cell_130e10d00f66", "created_at": "2026-07-19T09:54:23+00:00", "title": "Evidence and status"}
14
+ -->
15
+ [Corollary 4.6 in the arXiv PDF](https://arxiv.org/pdf/2603.02233) states . Its Appendix F.4 proof expands population and weighted empirical risks through KMEs, applies optimality of , then Cauchy–Schwarz twice. Thus the claimed direct transfer from KME estimation error to task excess risk is supported by the paper. **Status: source-verified theorem/proof audit; optimization error remains outside the corollary’s scope.**
16
+
17
+
18
+ ---
19
+ <!-- trackio-cell
20
+ {"type": "markdown", "id": "cell_efe13b52d7af", "created_at": "2026-07-19T09:56:17+00:00", "title": "Clarified evidence"}
21
+ -->
22
+ Clarification: Corollary 4.6 gives expected target excess risk at most 2 R_Theta times expected MMD between the learned empirical mixture and P_1. Appendix F.4 derives it by KME risk representation, weighted empirical-risk optimality, and Cauchy-Schwarz. Source: https://arxiv.org/pdf/2603.02233 .
pages/claim-3-example-4-5-automatic-adaptive-collaboration/page.md ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Claim 3: Example 4.5 automatic adaptive collaboration
2
+
3
+
4
+ ---
5
+ <!-- trackio-cell
6
+ {"type": "markdown", "id": "cell_3bda5e9c37b3", "created_at": "2026-07-19T09:32:03+00:00", "title": "Claim 3: Example 4.5 automatic adaptive collaboration"}
7
+ -->
8
+ Document setup, runs, and results for **Claim 3: Example 4.5 automatic adaptive collaboration**.
9
+
10
+
11
+ ---
12
+ <!-- trackio-cell
13
+ {"type": "markdown", "id": "cell_632280fed65b", "created_at": "2026-07-19T09:54:23+00:00", "title": "Evidence and status"}
14
+ -->
15
+ [Example 4.5](https://arxiv.org/pdf/2603.02233) specializes Theorem 4.4 to identical agents: the leading error becomes (up to lower-order terms), while returns the local rate. This verifies the stated oracle-style pooling/local regimes at the level of the theorem. In the independent implementation below (Claim 5), the literal RFF Q-objective did not recover the paper’s empirical transition; see that page for the non-replication result. **Status: source-verified asymptotic example; empirical adaptivity not reproduced by this implementation.**
16
+
17
+
18
+ ---
19
+ <!-- trackio-cell
20
+ {"type": "markdown", "id": "cell_d6652be82cbd", "created_at": "2026-07-19T09:56:34+00:00", "title": "Clarified evidence"}
21
+ -->
22
+ Clarification: Example 4.5 gives leading KME error Tr Sigma_1/n_V for an identical-agent set V, whereas V={1} recovers Tr Sigma_1/n_1. It establishes the pooling versus local statistical regimes in theory. Source: https://arxiv.org/pdf/2603.02233 .
pages/claim-4-theorem-5-2-random-fourier-feature-tradeoff/page.md ADDED
@@ -0,0 +1,302 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Claim 4: Theorem 5.2 random Fourier-feature tradeoff
2
+
3
+
4
+ ---
5
+ <!-- trackio-cell
6
+ {"type": "markdown", "id": "cell_3385698671bb", "created_at": "2026-07-19T09:32:03+00:00", "title": "Claim 4: Theorem 5.2 random Fourier-feature tradeoff"}
7
+ -->
8
+ Document setup, runs, and results for **Claim 4: Theorem 5.2 random Fourier-feature tradeoff**.
9
+
10
+
11
+ ---
12
+ <!-- trackio-cell
13
+ {"type": "code", "id": "cell_e08f5b2ace5f", "created_at": "2026-07-19T09:53:12+00:00", "title": "Run: python reproduce.py (exit 0)", "command": ["/home/vim/miniconda3/envs/pytorch-gpu-hf/bin/python", "src/reproduce.py", "rff", "--repeats", "100", "--rff-grid", "16", "32", "64", "128", "256", "512", "1024", "--output", "outputs/rff_tradeoff_100seeds.csv", "--device", "cuda"], "exit_code": 0, "duration_s": 1.465}
14
+ -->
15
+ ````bash
16
+ $ /home/vim/miniconda3/envs/pytorch-gpu-hf/bin/python src/reproduce.py rff --repeats 100 --rff-grid 16 32 64 128 256 512 1024 --output outputs/rff_tradeoff_100seeds.csv --device cuda
17
+ ````
18
+
19
+ exit 0 · 1.5s
20
+
21
+
22
+ ````python title=reproduce.py
23
+ #!/usr/bin/env python3
24
+ """Independent, GPU-capable checks for Fermanian et al. (2026), arXiv:2603.02233.
25
+
26
+ The implementation follows Algorithms 1--2 at a scaled experimental budget. It
27
+ uses RFF KMEs to choose a target client's collaboration weights and fits the
28
+ weighted empirical-risk minimizer exactly for linear regression. The FEMNIST
29
+ mode uses the public natural writer partition and a linear softmax model.
30
+ """
31
+ from __future__ import annotations
32
+
33
+ import argparse
34
+ import csv
35
+ import json
36
+ import math
37
+ from pathlib import Path
38
+ from time import perf_counter
39
+
40
+ import numpy as np
41
+ import torch
42
+
43
+
44
+ ROOT = Path(__file__).resolve().parents[1]
45
+
46
+
47
+ def device_from_arg(name: str) -> torch.device:
48
+ if name == "auto":
49
+ return torch.device("cuda" if torch.cuda.is_available() else "cpu")
50
+ return torch.device(name)
51
+
52
+
53
+ def rff(z: torch.Tensor, d_rff: int, generator: torch.Generator, scale: torch.Tensor) -> torch.Tensor:
54
+ """Gaussian-kernel RFF, with a supplied coordinate rescaling."""
55
+ w = torch.randn((z.shape[1], d_rff), generator=generator, device=z.device, dtype=z.dtype)
56
+ w = w * scale[:, None]
57
+ b = 2 * math.pi * torch.rand((d_rff,), generator=generator, device=z.device, dtype=z.dtype)
58
+ return math.sqrt(2.0 / d_rff) * torch.cos(z @ w + b)
59
+
60
+
61
+ def simplex_weights(means: torch.Tensor, target_phi: torch.Tensor, steps: int = 900) -> torch.Tensor:
62
+ """Solve Algorithm 1's empirical Q-aggregation objective by projected GD."""
63
+ b_agents = means.shape[0]
64
+ n1 = target_phi.shape[0]
65
+ centered = target_phi - means[0]
66
+ cov = centered.T @ centered / max(n1 - 1, 1)
67
+ trace = torch.trace(cov)
68
+ diffs = means[1:] - means[0]
69
+ q_terms = torch.sqrt(torch.clamp(torch.einsum("bd,df,bf->b", diffs, cov, diffs), min=0.0) + 1e-12)
70
+ p_terms = torch.linalg.vector_norm(diffs, dim=1)
71
+ # Table 1 uses CQ^2=CP=log B: CQ=sqrt(log B), CP=log B.
72
+ cq, cp = math.sqrt(math.log(b_agents)), math.log(b_agents)
73
+ w = torch.full((b_agents,), 1.0 / b_agents, device=means.device, dtype=means.dtype)
74
+ for _ in range(steps):
75
+ w.requires_grad_(True)
76
+ mixture = w @ means
77
+ loss = torch.sum((mixture - means[0]) ** 2) + 2.0 * w[0] * trace / n1
78
+ loss = loss + cq / math.sqrt(n1) * torch.dot(w[1:], q_terms)
79
+ loss = loss + cp / n1 * torch.dot(w[1:], p_terms)
80
+ (grad,) = torch.autograd.grad(loss, w)
81
+ with torch.no_grad():
82
+ # Euclidean projection on the simplex, Duchi et al. (2008).
83
+ v = w - 0.04 * grad
84
+ u, _ = torch.sort(v, descending=True)
85
+ cssv = torch.cumsum(u, 0) - 1
86
+ ind = torch.arange(1, b_agents + 1, device=v.device, dtype=v.dtype)
87
+ rho = torch.nonzero(u - cssv / ind > 0, as_tuple=False)[-1, 0]
88
+ theta = cssv[rho] / (rho + 1)
89
+ w = torch.clamp(v - theta, min=0)
90
+ return w.detach()
91
+
92
+
93
+ def weighted_linear_fit(xs: torch.Tensor, ys: torch.Tensor, weights: torch.Tensor, ridge: float = 1e-3) -> torch.Tensor:
94
+ """Exact weighted empirical-risk minimizer for linear regression."""
95
+ b_agents, n, dim = xs.shape
96
+ x = torch.cat([xs, torch.ones((b_agents, n, 1), device=xs.device, dtype=xs.dtype)], dim=-1)
97
+ sample_w = (weights[:, None] / n).expand(b_agents, n).reshape(-1)
98
+ flat_x, flat_y = x.reshape(-1, dim + 1), ys.reshape(-1)
99
+ lhs = flat_x.T @ (flat_x * sample_w[:, None]) + ridge * torch.eye(dim + 1, device=xs.device)
100
+ rhs = flat_x.T @ (flat_y * sample_w)
101
+ return torch.linalg.solve(lhs, rhs)
102
+
103
+
104
+ def concept_trial(seed: int, sigma2: float, d_rff: int, device: torch.device) -> dict:
105
+ g = torch.Generator(device=device).manual_seed(seed)
106
+ b_agents, n, dim, n_test = 100, 10, 20, 1000
107
+ signs = torch.where(torch.rand((b_agents,), generator=g, device=device) > .5, 1., -1.)
108
+ beta0 = torch.randn((dim,), generator=g, device=device)
109
+ eps = torch.randn((b_agents, dim), generator=g, device=device)
110
+ betas = signs[:, None] * math.sqrt(1 - sigma2) * beta0 + math.sqrt(sigma2) * eps
111
+ x = 1.0 + torch.randn((b_agents, n, dim), generator=g, device=device)
112
+ y = torch.einsum("bnd,bd->bn", x, betas) + math.sqrt(2.0) * torch.randn((b_agents, n), generator=g, device=device)
113
+ x_test = 1.0 + torch.randn((n_test, dim), generator=g, device=device)
114
+ y_test = x_test @ betas[0] + math.sqrt(2.0) * torch.randn((n_test,), generator=g, device=device)
115
+ # The paper's weighted Gaussian KME kernel gives X a 1/sqrt(d) scaling.
116
+ z = torch.cat([x, y[..., None]], dim=-1).reshape(-1, dim + 1)
117
+ scale = torch.cat([torch.full((dim,), 1 / math.sqrt(dim), device=device), torch.ones((1,), device=device)])
118
+ phi = rff(z, d_rff, g, scale).reshape(b_agents, n, d_rff)
119
+ means = phi.mean(1)
120
+ q_w = simplex_weights(means, phi[0])
121
+ local = torch.zeros(b_agents, device=device); local[0] = 1
122
+ grand = torch.full((b_agents,), 1 / b_agents, device=device)
123
+ oracle = (signs == signs[0]).float(); oracle /= oracle.sum()
124
+ result = {"sigma2": sigma2, "seed": seed, "q_local_weight": float(q_w[0]),
125
+ "q_same_group_weight": float(q_w[signs == signs[0]].sum())}
126
+ for name, weights in [("local", local), ("grandmean", grand), ("oracle", oracle), ("qaggregation", q_w)]:
127
+ beta = weighted_linear_fit(x, y, weights)
128
+ pred = x_test @ beta[:-1] + beta[-1]
129
+ result[f"{name}_mse"] = float(torch.mean((pred - y_test) ** 2))
130
+ return result
131
+
132
+
133
+ def run_concept(args: argparse.Namespace) -> list[dict]:
134
+ device = device_from_arg(args.device)
135
+ rows = []
136
+ for sigma2 in args.sigmas:
137
+ for rep in range(args.repeats):
138
+ rows.append(concept_trial(args.seed + rep * 101 + int(sigma2 * 1000), sigma2, args.rff_dim, device))
139
+ return rows
140
+
141
+
142
+ def exact_rbf_mmd(x: torch.Tensor, y: torch.Tensor, bandwidth: float = 1.0) -> torch.Tensor:
143
+ def k(a: torch.Tensor, b: torch.Tensor) -> torch.Tensor:
144
+ return torch.exp(-torch.cdist(a, b).square() / (2 * bandwidth * bandwidth))
145
+ return torch.sqrt(torch.clamp(k(x, x).mean() + k(y, y).mean() - 2 * k(x, y).mean(), min=0.0))
146
+
147
+
148
+ def run_rff(args: argparse.Namespace) -> list[dict]:
149
+ device = device_from_arg(args.device)
150
+ g = torch.Generator(device=device).manual_seed(args.seed)
151
+ x = torch.randn((300, 8), generator=g, device=device)
152
+ y = torch.randn((300, 8), generator=g, device=device) + torch.tensor([0.35] + [0.] * 7, device=device)
153
+ exact = float(exact_rbf_mmd(x, y))
154
+ rows = []
155
+ for d_rff in args.rff_grid:
156
+ errors = []
157
+ for trial in range(args.repeats):
158
+ tg = torch.Generator(device=device).manual_seed(args.seed + 1000 * d_rff + trial)
159
+ phi = rff(torch.cat([x, y]), d_rff, tg, torch.ones(8, device=device))
160
+ approx = torch.linalg.vector_norm(phi[:len(x)].mean(0) - phi[len(x):].mean(0)).item()
161
+ errors.append(abs(approx - exact))
162
+ rows.append({"rff_dim": d_rff, "exact_mmd": exact, "mean_abs_error": float(np.mean(errors)),
163
+ "std_abs_error": float(np.std(errors)), "sqrt_logB_over_D": math.sqrt(math.log(100) / d_rff),
164
+ "one_shot_floats_per_agent": d_rff})
165
+ return rows
166
+
167
+
168
+ def run_femnist(args: argparse.Namespace) -> list[dict]:
169
+ """Actual FEMNIST writer data, deliberately scaled in samples/features/epochs."""
170
+ from datasets import load_dataset
171
+ device = device_from_arg(args.device)
172
+ local_parquet = ROOT / "data" / "femnist" / "data" / "train-00000-of-00001.parquet"
173
+ dataset = load_dataset("parquet", data_files=str(local_parquet), split="train", streaming=True) if local_parquet.exists() else load_dataset("flwrlabs/femnist", split="train", streaming=True)
174
+ by_writer: dict[str, list[tuple[np.ndarray, int]]] = {}
175
+ want = args.agents
176
+ for ex in dataset:
177
+ writer = ex["writer_id"]
178
+ if writer not in by_writer and len(by_writer) >= want:
179
+ continue
180
+ if writer in by_writer or len(by_writer) < want:
181
+ by_writer.setdefault(writer, []).append((np.asarray(ex["image"], dtype=np.float32) / 255., int(ex["character"])))
182
+ if len(by_writer) >= want and all(len(v) >= args.train_per_agent + args.test_per_agent for v in by_writer.values()):
183
+ break
184
+ eligible = [v for v in by_writer.values() if len(v) >= args.train_per_agent + args.test_per_agent][:want]
185
+ if len(eligible) < want:
186
+ raise RuntimeError(f"Only found {len(eligible)} eligible writers; need {want}")
187
+ x_train = torch.tensor(np.stack([np.stack([a for a, _ in v[:args.train_per_agent]]) for v in eligible]), device=device)
188
+ y_train = torch.tensor(np.stack([[b for _, b in v[:args.train_per_agent]] for v in eligible]), device=device, dtype=torch.long)
189
+ x_test = torch.tensor(np.stack([np.stack([a for a, _ in v[args.train_per_agent:args.train_per_agent + args.test_per_agent]]) for v in eligible]), device=device)
190
+ y_test = torch.tensor(np.stack([[b for _, b in v[args.train_per_agent:args.train_per_agent + args.test_per_agent]] for v in eligible]), device=device, dtype=torch.long)
191
+ b_agents, n, raw_dim = x_train.shape[0], x_train.shape[1], 784
192
+ flat = x_train.reshape(b_agents, n, raw_dim)
193
+ # Fixed random projection reduces the RFF input cost while retaining a reproducible vision proxy.
194
+ g = torch.Generator(device=device).manual_seed(args.seed)
195
+ proj = torch.randn((raw_dim, args.feature_dim), generator=g, device=device) / math.sqrt(raw_dim)
196
+ features = flat @ proj
197
+ phi_all = rff(features.reshape(-1, args.feature_dim), args.rff_dim, g, torch.ones(args.feature_dim, device=device)).reshape(b_agents, n, args.rff_dim)
198
+ means = phi_all.mean(1)
199
+ # Fit one regularized softmax classifier for each target using its personalized weights.
200
+ rows = []
201
+ all_x = features.reshape(-1, args.feature_dim)
202
+ all_y = y_train.reshape(-1)
203
+ for target in range(b_agents):
204
+ weights = simplex_weights(means, phi_all[target], steps=300)
205
+ schemes = {"local": torch.nn.functional.one_hot(torch.tensor(target, device=device), b_agents).float(),
206
+ "grandmean": torch.full((b_agents,), 1 / b_agents, device=device), "qaggregation": weights}
207
+ test_feat = x_test[target].reshape(args.test_per_agent, raw_dim) @ proj
208
+ for name, w in schemes.items():
209
+ model = torch.zeros((args.feature_dim, 62), device=device, requires_grad=True)
210
+ opt = torch.optim.Adam([model], lr=args.lr, weight_decay=1e-4)
211
+ sample_w = (w[:, None] / n).expand(b_agents, n).reshape(-1)
212
+ for _ in range(args.epochs):
213
+ opt.zero_grad()
214
+ losses = torch.nn.functional.cross_entropy(all_x @ model, all_y, reduction="none")
215
+ (losses * sample_w).sum().backward(); opt.step()
216
+ acc = ((test_feat @ model).argmax(1) == y_test[target]).float().mean().item()
217
+ rows.append({"agent": target, "method": name, "accuracy": acc, "q_local_weight": float(weights[target])})
218
+ return rows
219
+
220
+
221
+ def write_rows(rows: list[dict], out: Path) -> None:
222
+ out.parent.mkdir(parents=True, exist_ok=True)
223
+ with out.open("w", newline="") as f:
224
+ writer = csv.DictWriter(f, fieldnames=list(rows[0]))
225
+ writer.writeheader(); writer.writerows(rows)
226
+ out.with_suffix(".json").write_text(json.dumps(rows, indent=2))
227
+
228
+
229
+ def main() -> None:
230
+ p = argparse.ArgumentParser()
231
+ p.add_argument("mode", choices=["concept", "rff", "femnist"])
232
+ p.add_argument("--output", type=Path, required=True)
233
+ p.add_argument("--device", default="auto")
234
+ p.add_argument("--seed", type=int, default=20260719)
235
+ p.add_argument("--repeats", type=int, default=30)
236
+ p.add_argument("--rff-dim", type=int, default=500)
237
+ p.add_argument("--sigmas", type=float, nargs="+", default=[0., .1, .25, .5, .75, 1.])
238
+ p.add_argument("--rff-grid", type=int, nargs="+", default=[16, 32, 64, 128, 256, 512, 1024])
239
+ p.add_argument("--agents", type=int, default=192)
240
+ p.add_argument("--train-per-agent", type=int, default=32)
241
+ p.add_argument("--test-per-agent", type=int, default=16)
242
+ p.add_argument("--feature-dim", type=int, default=64)
243
+ p.add_argument("--epochs", type=int, default=20)
244
+ p.add_argument("--lr", type=float, default=.08)
245
+ args = p.parse_args()
246
+ started = perf_counter()
247
+ rows = {"concept": run_concept, "rff": run_rff, "femnist": run_femnist}[args.mode](args)
248
+ write_rows(rows, args.output)
249
+ if args.mode == "concept":
250
+ summary = {str(s): {k: float(np.mean([r[k] for r in rows if r["sigma2"] == s])) for k in ["local_mse", "grandmean_mse", "oracle_mse", "qaggregation_mse", "q_local_weight", "q_same_group_weight"]} for s in args.sigmas}
251
+ print(json.dumps({"concept_summary": summary}, indent=2))
252
+ elif args.mode == "rff":
253
+ print(json.dumps({"rff_summary": {str(r["rff_dim"]): r["mean_abs_error"] for r in rows}}, indent=2))
254
+ elif args.mode == "femnist":
255
+ summary = {m: float(np.mean([r["accuracy"] for r in rows if r["method"] == m])) for m in ["local", "grandmean", "qaggregation"]}
256
+ print(json.dumps({"femnist_accuracy": summary}, indent=2))
257
+ print(json.dumps({"mode": args.mode, "rows": len(rows), "output": str(args.output), "seconds": round(perf_counter() - started, 2), "device": str(device_from_arg(args.device))}, indent=2))
258
+
259
+
260
+ if __name__ == "__main__":
261
+ main()
262
+
263
+ ````
264
+
265
+
266
+ ````output
267
+ {
268
+ "rff_summary": {
269
+ "16": 0.011847271621227264,
270
+ "32": 0.008519105277955533,
271
+ "64": 0.005433569923043251,
272
+ "128": 0.004183858707547188,
273
+ "256": 0.003424282595515251,
274
+ "512": 0.0020784523338079454,
275
+ "1024": 0.0013805751502513886
276
+ }
277
+ }
278
+ {
279
+ "mode": "rff",
280
+ "rows": 7,
281
+ "output": "outputs/rff_tradeoff_100seeds.csv",
282
+ "seconds": 0.47,
283
+ "device": "cuda"
284
+ }
285
+
286
+ ````
287
+
288
+
289
+ ---
290
+ <!-- trackio-cell
291
+ {"type": "artifact", "id": "cell_ffe4264c28a2", "created_at": "2026-07-19T09:53:12+00:00", "title": "Artifact: rff_tradeoff_100seeds.csv", "path": "outputs/rff_tradeoff_100seeds.csv", "size": 721, "artifact_type": "dataset", "auto": true}
292
+ -->
293
+ **📦 Artifact** `outputs/rff_tradeoff_100seeds.csv` · dataset · 721 B
294
+
295
+ https://huggingface.co/buckets/vimarsh/repro-adaptive-personalized-fl-multi-task-averaging-kernel-mean-embedding-artifacts#logbook-files/outputs/rff_tradeoff_100seeds.csv
296
+
297
+
298
+ ---
299
+ <!-- trackio-cell
300
+ {"type": "markdown", "id": "cell_e4032fbdf86b", "created_at": "2026-07-19T09:56:47+00:00", "title": "Clarified evidence"}
301
+ -->
302
+ Theorem 5.2 source audit: https://arxiv.org/pdf/2603.02233 adds an RFF approximation contribution of order sqrt(log B / D), so increasing D trades one-shot communication for statistical precision. CUDA 100-seed check: mean absolute MMD error falls from 0.01185 at D=16 to 0.00138 at D=1024, while transmitted floats/client grow from 16 to 1024. Status: mechanism reproduced.
pages/claim-5-section-6-1-100-agent-synthetic-concept-shift/page.md ADDED
@@ -0,0 +1,1310 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Claim 5: Section 6.1 100-agent synthetic concept shift
2
+
3
+
4
+ ---
5
+ <!-- trackio-cell
6
+ {"type": "markdown", "id": "cell_c62136c62a13", "created_at": "2026-07-19T09:32:03+00:00", "title": "Claim 5: Section 6.1 100-agent synthetic concept shift"}
7
+ -->
8
+ Document setup, runs, and results for **Claim 5: Section 6.1 100-agent synthetic concept shift**.
9
+
10
+
11
+ ---
12
+ <!-- trackio-cell
13
+ {"type": "code", "id": "cell_afaa63e40a29", "created_at": "2026-07-19T09:45:31+00:00", "title": "Run: echo (exit 0)", "command": ["echo", "TRACKIO_CAPTURE_CHECK"], "exit_code": 0, "duration_s": 0.001}
14
+ -->
15
+ ````bash
16
+ $ echo TRACKIO_CAPTURE_CHECK
17
+ ````
18
+
19
+ exit 0 · 0.0s
20
+
21
+
22
+ ````output
23
+ TRACKIO_CAPTURE_CHECK
24
+
25
+ ````
26
+
27
+
28
+ ---
29
+ <!-- trackio-cell
30
+ {"type": "code", "id": "cell_ab96ac0a6224", "created_at": "2026-07-19T09:47:33+00:00", "title": "Run: python reproduce.py (exit 0)", "command": ["/home/vim/miniconda3/envs/pytorch-gpu-hf/bin/python", "src/reproduce.py", "concept", "--repeats", "100", "--rff-dim", "500", "--output", "outputs/concept_100agents_100reps.csv", "--device", "cuda"], "exit_code": 0, "duration_s": 287.498}
31
+ -->
32
+ ````bash
33
+ $ /home/vim/miniconda3/envs/pytorch-gpu-hf/bin/python src/reproduce.py concept --repeats 100 --rff-dim 500 --output outputs/concept_100agents_100reps.csv --device cuda
34
+ ````
35
+
36
+ exit 0 · 287.5s
37
+
38
+
39
+ ````python title=reproduce.py
40
+ #!/usr/bin/env python3
41
+ """Independent, GPU-capable checks for Fermanian et al. (2026), arXiv:2603.02233.
42
+
43
+ The implementation follows Algorithms 1--2 at a scaled experimental budget. It
44
+ uses RFF KMEs to choose a target client's collaboration weights and fits the
45
+ weighted empirical-risk minimizer exactly for linear regression. The FEMNIST
46
+ mode uses the public natural writer partition and a linear softmax model.
47
+ """
48
+ from __future__ import annotations
49
+
50
+ import argparse
51
+ import csv
52
+ import json
53
+ import math
54
+ from pathlib import Path
55
+ from time import perf_counter
56
+
57
+ import numpy as np
58
+ import torch
59
+
60
+
61
+ ROOT = Path(__file__).resolve().parents[1]
62
+
63
+
64
+ def device_from_arg(name: str) -> torch.device:
65
+ if name == "auto":
66
+ return torch.device("cuda" if torch.cuda.is_available() else "cpu")
67
+ return torch.device(name)
68
+
69
+
70
+ def rff(z: torch.Tensor, d_rff: int, generator: torch.Generator, scale: torch.Tensor) -> torch.Tensor:
71
+ """Gaussian-kernel RFF, with a supplied coordinate rescaling."""
72
+ w = torch.randn((z.shape[1], d_rff), generator=generator, device=z.device, dtype=z.dtype)
73
+ w = w * scale[:, None]
74
+ b = 2 * math.pi * torch.rand((d_rff,), generator=generator, device=z.device, dtype=z.dtype)
75
+ return math.sqrt(2.0 / d_rff) * torch.cos(z @ w + b)
76
+
77
+
78
+ def simplex_weights(means: torch.Tensor, target_phi: torch.Tensor, steps: int = 900) -> torch.Tensor:
79
+ """Solve Algorithm 1's empirical Q-aggregation objective by projected GD."""
80
+ b_agents = means.shape[0]
81
+ n1 = target_phi.shape[0]
82
+ centered = target_phi - means[0]
83
+ cov = centered.T @ centered / max(n1 - 1, 1)
84
+ trace = torch.trace(cov)
85
+ diffs = means[1:] - means[0]
86
+ q_terms = torch.sqrt(torch.clamp(torch.einsum("bd,df,bf->b", diffs, cov, diffs), min=0.0) + 1e-12)
87
+ p_terms = torch.linalg.vector_norm(diffs, dim=1)
88
+ # Table 1 uses CQ^2=CP=log B: CQ=sqrt(log B), CP=log B.
89
+ cq, cp = math.sqrt(math.log(b_agents)), math.log(b_agents)
90
+ w = torch.full((b_agents,), 1.0 / b_agents, device=means.device, dtype=means.dtype)
91
+ for _ in range(steps):
92
+ w.requires_grad_(True)
93
+ mixture = w @ means
94
+ loss = torch.sum((mixture - means[0]) ** 2) + 2.0 * w[0] * trace / n1
95
+ loss = loss + cq / math.sqrt(n1) * torch.dot(w[1:], q_terms)
96
+ loss = loss + cp / n1 * torch.dot(w[1:], p_terms)
97
+ (grad,) = torch.autograd.grad(loss, w)
98
+ with torch.no_grad():
99
+ # Euclidean projection on the simplex, Duchi et al. (2008).
100
+ v = w - 0.04 * grad
101
+ u, _ = torch.sort(v, descending=True)
102
+ cssv = torch.cumsum(u, 0) - 1
103
+ ind = torch.arange(1, b_agents + 1, device=v.device, dtype=v.dtype)
104
+ rho = torch.nonzero(u - cssv / ind > 0, as_tuple=False)[-1, 0]
105
+ theta = cssv[rho] / (rho + 1)
106
+ w = torch.clamp(v - theta, min=0)
107
+ return w.detach()
108
+
109
+
110
+ def weighted_linear_fit(xs: torch.Tensor, ys: torch.Tensor, weights: torch.Tensor, ridge: float = 1e-3) -> torch.Tensor:
111
+ """Exact weighted empirical-risk minimizer for linear regression."""
112
+ b_agents, n, dim = xs.shape
113
+ x = torch.cat([xs, torch.ones((b_agents, n, 1), device=xs.device, dtype=xs.dtype)], dim=-1)
114
+ sample_w = (weights[:, None] / n).expand(b_agents, n).reshape(-1)
115
+ flat_x, flat_y = x.reshape(-1, dim + 1), ys.reshape(-1)
116
+ lhs = flat_x.T @ (flat_x * sample_w[:, None]) + ridge * torch.eye(dim + 1, device=xs.device)
117
+ rhs = flat_x.T @ (flat_y * sample_w)
118
+ return torch.linalg.solve(lhs, rhs)
119
+
120
+
121
+ def concept_trial(seed: int, sigma2: float, d_rff: int, device: torch.device) -> dict:
122
+ g = torch.Generator(device=device).manual_seed(seed)
123
+ b_agents, n, dim, n_test = 100, 10, 20, 1000
124
+ signs = torch.where(torch.rand((b_agents,), generator=g, device=device) > .5, 1., -1.)
125
+ beta0 = torch.randn((dim,), generator=g, device=device)
126
+ eps = torch.randn((b_agents, dim), generator=g, device=device)
127
+ betas = signs[:, None] * math.sqrt(1 - sigma2) * beta0 + math.sqrt(sigma2) * eps
128
+ x = 1.0 + torch.randn((b_agents, n, dim), generator=g, device=device)
129
+ y = torch.einsum("bnd,bd->bn", x, betas) + math.sqrt(2.0) * torch.randn((b_agents, n), generator=g, device=device)
130
+ x_test = 1.0 + torch.randn((n_test, dim), generator=g, device=device)
131
+ y_test = x_test @ betas[0] + math.sqrt(2.0) * torch.randn((n_test,), generator=g, device=device)
132
+ # The paper's weighted Gaussian KME kernel gives X a 1/sqrt(d) scaling.
133
+ z = torch.cat([x, y[..., None]], dim=-1).reshape(-1, dim + 1)
134
+ scale = torch.cat([torch.full((dim,), 1 / math.sqrt(dim), device=device), torch.ones((1,), device=device)])
135
+ phi = rff(z, d_rff, g, scale).reshape(b_agents, n, d_rff)
136
+ means = phi.mean(1)
137
+ q_w = simplex_weights(means, phi[0])
138
+ local = torch.zeros(b_agents, device=device); local[0] = 1
139
+ grand = torch.full((b_agents,), 1 / b_agents, device=device)
140
+ oracle = (signs == signs[0]).float(); oracle /= oracle.sum()
141
+ result = {"sigma2": sigma2, "seed": seed, "q_local_weight": float(q_w[0]),
142
+ "q_same_group_weight": float(q_w[signs == signs[0]].sum())}
143
+ for name, weights in [("local", local), ("grandmean", grand), ("oracle", oracle), ("qaggregation", q_w)]:
144
+ beta = weighted_linear_fit(x, y, weights)
145
+ pred = x_test @ beta[:-1] + beta[-1]
146
+ result[f"{name}_mse"] = float(torch.mean((pred - y_test) ** 2))
147
+ return result
148
+
149
+
150
+ def run_concept(args: argparse.Namespace) -> list[dict]:
151
+ device = device_from_arg(args.device)
152
+ rows = []
153
+ for sigma2 in args.sigmas:
154
+ for rep in range(args.repeats):
155
+ rows.append(concept_trial(args.seed + rep * 101 + int(sigma2 * 1000), sigma2, args.rff_dim, device))
156
+ return rows
157
+
158
+
159
+ def exact_rbf_mmd(x: torch.Tensor, y: torch.Tensor, bandwidth: float = 1.0) -> torch.Tensor:
160
+ def k(a: torch.Tensor, b: torch.Tensor) -> torch.Tensor:
161
+ return torch.exp(-torch.cdist(a, b).square() / (2 * bandwidth * bandwidth))
162
+ return torch.sqrt(torch.clamp(k(x, x).mean() + k(y, y).mean() - 2 * k(x, y).mean(), min=0.0))
163
+
164
+
165
+ def run_rff(args: argparse.Namespace) -> list[dict]:
166
+ device = device_from_arg(args.device)
167
+ g = torch.Generator(device=device).manual_seed(args.seed)
168
+ x = torch.randn((300, 8), generator=g, device=device)
169
+ y = torch.randn((300, 8), generator=g, device=device) + torch.tensor([0.35] + [0.] * 7, device=device)
170
+ exact = float(exact_rbf_mmd(x, y))
171
+ rows = []
172
+ for d_rff in args.rff_grid:
173
+ errors = []
174
+ for trial in range(args.repeats):
175
+ tg = torch.Generator(device=device).manual_seed(args.seed + 1000 * d_rff + trial)
176
+ phi = rff(torch.cat([x, y]), d_rff, tg, torch.ones(8, device=device))
177
+ approx = torch.linalg.vector_norm(phi[:len(x)].mean(0) - phi[len(x):].mean(0)).item()
178
+ errors.append(abs(approx - exact))
179
+ rows.append({"rff_dim": d_rff, "exact_mmd": exact, "mean_abs_error": float(np.mean(errors)),
180
+ "std_abs_error": float(np.std(errors)), "sqrt_logB_over_D": math.sqrt(math.log(100) / d_rff),
181
+ "one_shot_floats_per_agent": d_rff})
182
+ return rows
183
+
184
+
185
+ def run_femnist(args: argparse.Namespace) -> list[dict]:
186
+ """Actual FEMNIST writer data, deliberately scaled in samples/features/epochs."""
187
+ from datasets import load_dataset
188
+ device = device_from_arg(args.device)
189
+ local_parquet = ROOT / "data" / "femnist" / "data" / "train-00000-of-00001.parquet"
190
+ dataset = load_dataset("parquet", data_files=str(local_parquet), split="train", streaming=True) if local_parquet.exists() else load_dataset("flwrlabs/femnist", split="train", streaming=True)
191
+ by_writer: dict[str, list[tuple[np.ndarray, int]]] = {}
192
+ want = args.agents
193
+ for ex in dataset:
194
+ writer = ex["writer_id"]
195
+ if writer not in by_writer and len(by_writer) >= want:
196
+ continue
197
+ if writer in by_writer or len(by_writer) < want:
198
+ by_writer.setdefault(writer, []).append((np.asarray(ex["image"], dtype=np.float32) / 255., int(ex["character"])))
199
+ if len(by_writer) >= want and all(len(v) >= args.train_per_agent + args.test_per_agent for v in by_writer.values()):
200
+ break
201
+ eligible = [v for v in by_writer.values() if len(v) >= args.train_per_agent + args.test_per_agent][:want]
202
+ if len(eligible) < want:
203
+ raise RuntimeError(f"Only found {len(eligible)} eligible writers; need {want}")
204
+ x_train = torch.tensor(np.stack([np.stack([a for a, _ in v[:args.train_per_agent]]) for v in eligible]), device=device)
205
+ y_train = torch.tensor(np.stack([[b for _, b in v[:args.train_per_agent]] for v in eligible]), device=device, dtype=torch.long)
206
+ x_test = torch.tensor(np.stack([np.stack([a for a, _ in v[args.train_per_agent:args.train_per_agent + args.test_per_agent]]) for v in eligible]), device=device)
207
+ y_test = torch.tensor(np.stack([[b for _, b in v[args.train_per_agent:args.train_per_agent + args.test_per_agent]] for v in eligible]), device=device, dtype=torch.long)
208
+ b_agents, n, raw_dim = x_train.shape[0], x_train.shape[1], 784
209
+ flat = x_train.reshape(b_agents, n, raw_dim)
210
+ # Fixed random projection reduces the RFF input cost while retaining a reproducible vision proxy.
211
+ g = torch.Generator(device=device).manual_seed(args.seed)
212
+ proj = torch.randn((raw_dim, args.feature_dim), generator=g, device=device) / math.sqrt(raw_dim)
213
+ features = flat @ proj
214
+ phi_all = rff(features.reshape(-1, args.feature_dim), args.rff_dim, g, torch.ones(args.feature_dim, device=device)).reshape(b_agents, n, args.rff_dim)
215
+ means = phi_all.mean(1)
216
+ # Fit one regularized softmax classifier for each target using its personalized weights.
217
+ rows = []
218
+ all_x = features.reshape(-1, args.feature_dim)
219
+ all_y = y_train.reshape(-1)
220
+ for target in range(b_agents):
221
+ weights = simplex_weights(means, phi_all[target], steps=300)
222
+ schemes = {"local": torch.nn.functional.one_hot(torch.tensor(target, device=device), b_agents).float(),
223
+ "grandmean": torch.full((b_agents,), 1 / b_agents, device=device), "qaggregation": weights}
224
+ test_feat = x_test[target].reshape(args.test_per_agent, raw_dim) @ proj
225
+ for name, w in schemes.items():
226
+ model = torch.zeros((args.feature_dim, 62), device=device, requires_grad=True)
227
+ opt = torch.optim.Adam([model], lr=args.lr, weight_decay=1e-4)
228
+ sample_w = (w[:, None] / n).expand(b_agents, n).reshape(-1)
229
+ for _ in range(args.epochs):
230
+ opt.zero_grad()
231
+ losses = torch.nn.functional.cross_entropy(all_x @ model, all_y, reduction="none")
232
+ (losses * sample_w).sum().backward(); opt.step()
233
+ acc = ((test_feat @ model).argmax(1) == y_test[target]).float().mean().item()
234
+ rows.append({"agent": target, "method": name, "accuracy": acc, "q_local_weight": float(weights[target])})
235
+ return rows
236
+
237
+
238
+ def write_rows(rows: list[dict], out: Path) -> None:
239
+ out.parent.mkdir(parents=True, exist_ok=True)
240
+ with out.open("w", newline="") as f:
241
+ writer = csv.DictWriter(f, fieldnames=list(rows[0]))
242
+ writer.writeheader(); writer.writerows(rows)
243
+ out.with_suffix(".json").write_text(json.dumps(rows, indent=2))
244
+
245
+
246
+ def main() -> None:
247
+ p = argparse.ArgumentParser()
248
+ p.add_argument("mode", choices=["concept", "rff", "femnist"])
249
+ p.add_argument("--output", type=Path, required=True)
250
+ p.add_argument("--device", default="auto")
251
+ p.add_argument("--seed", type=int, default=20260719)
252
+ p.add_argument("--repeats", type=int, default=30)
253
+ p.add_argument("--rff-dim", type=int, default=500)
254
+ p.add_argument("--sigmas", type=float, nargs="+", default=[0., .1, .25, .5, .75, 1.])
255
+ p.add_argument("--rff-grid", type=int, nargs="+", default=[16, 32, 64, 128, 256, 512, 1024])
256
+ p.add_argument("--agents", type=int, default=192)
257
+ p.add_argument("--train-per-agent", type=int, default=32)
258
+ p.add_argument("--test-per-agent", type=int, default=16)
259
+ p.add_argument("--feature-dim", type=int, default=64)
260
+ p.add_argument("--epochs", type=int, default=20)
261
+ p.add_argument("--lr", type=float, default=.08)
262
+ args = p.parse_args()
263
+ started = perf_counter()
264
+ rows = {"concept": run_concept, "rff": run_rff, "femnist": run_femnist}[args.mode](args)
265
+ write_rows(rows, args.output)
266
+ if args.mode == "concept":
267
+ summary = {str(s): {k: float(np.mean([r[k] for r in rows if r["sigma2"] == s])) for k in ["local_mse", "grandmean_mse", "oracle_mse", "qaggregation_mse", "q_local_weight", "q_same_group_weight"]} for s in args.sigmas}
268
+ print(json.dumps({"concept_summary": summary}, indent=2))
269
+ elif args.mode == "rff":
270
+ print(json.dumps({"rff_summary": {str(r["rff_dim"]): r["mean_abs_error"] for r in rows}}, indent=2))
271
+ elif args.mode == "femnist":
272
+ summary = {m: float(np.mean([r["accuracy"] for r in rows if r["method"] == m])) for m in ["local", "grandmean", "qaggregation"]}
273
+ print(json.dumps({"femnist_accuracy": summary}, indent=2))
274
+ print(json.dumps({"mode": args.mode, "rows": len(rows), "output": str(args.output), "seconds": round(perf_counter() - started, 2), "device": str(device_from_arg(args.device))}, indent=2))
275
+
276
+
277
+ if __name__ == "__main__":
278
+ main()
279
+
280
+ ````
281
+
282
+
283
+ ````output
284
+ {
285
+ "concept_summary": {
286
+ "0.0": {
287
+ "local_mse": 15.748318042755127,
288
+ "grandmean_mse": 40.606961507797244,
289
+ "oracle_mse": 2.0862513756752015,
290
+ "qaggregation_mse": 17.83958231449127,
291
+ "q_local_weight": 0.9771137142181396,
292
+ "q_same_group_weight": 0.9975556522607804
293
+ },
294
+ "0.1": {
295
+ "local_mse": 15.361917610168456,
296
+ "grandmean_mse": 45.73134832382202,
297
+ "oracle_mse": 6.119768922328949,
298
+ "qaggregation_mse": 18.07122127532959,
299
+ "q_local_weight": 0.9787336409091949,
300
+ "q_same_group_weight": 0.9981670784950256
301
+ },
302
+ "0.25": {
303
+ "local_mse": 15.827756361961365,
304
+ "grandmean_mse": 39.008377923965455,
305
+ "oracle_mse": 11.351614267826081,
306
+ "qaggregation_mse": 25.339054584503174,
307
+ "q_local_weight": 0.9825072956085205,
308
+ "q_same_group_weight": 0.99599778175354
309
+ },
310
+ "0.5": {
311
+ "local_mse": 14.459182372093201,
312
+ "grandmean_mse": 45.512830362319946,
313
+ "oracle_mse": 24.322485909461975,
314
+ "qaggregation_mse": 25.430337595939637,
315
+ "q_local_weight": 0.9876203048229217,
316
+ "q_same_group_weight": 0.9948384702205658
317
+ },
318
+ "0.75": {
319
+ "local_mse": 15.87236515045166,
320
+ "grandmean_mse": 39.63199781417847,
321
+ "oracle_mse": 30.853181953430177,
322
+ "qaggregation_mse": 33.74232290267944,
323
+ "q_local_weight": 0.9877495408058167,
324
+ "q_same_group_weight": 0.9943676108121872
325
+ },
326
+ "1.0": {
327
+ "local_mse": 15.390748682022094,
328
+ "grandmean_mse": 39.16119129180908,
329
+ "oracle_mse": 40.00151491165161,
330
+ "qaggregation_mse": 30.293080587387085,
331
+ "q_local_weight": 0.9896031093597412,
332
+ "q_same_group_weight": 0.9947122848033905
333
+ }
334
+ }
335
+ }
336
+ {
337
+ "mode": "concept",
338
+ "rows": 600,
339
+ "output": "outputs/concept_100agents_100reps.csv",
340
+ "seconds": 286.58,
341
+ "device": "cuda"
342
+ }
343
+
344
+ ````
345
+
346
+
347
+ ---
348
+ <!-- trackio-cell
349
+ {"type": "artifact", "id": "cell_edfa78971b70", "created_at": "2026-07-19T09:47:34+00:00", "title": "Artifact: concept_100agents_100reps.csv", "path": "outputs/concept_100agents_100reps.csv", "size": 63718, "artifact_type": "dataset", "auto": true}
350
+ -->
351
+ **📦 Artifact** `outputs/concept_100agents_100reps.csv` · dataset · 63.7 kB
352
+
353
+ https://huggingface.co/buckets/vimarsh/repro-adaptive-personalized-fl-multi-task-averaging-kernel-mean-embedding-artifacts#logbook-files/outputs/concept_100agents_100reps.csv
354
+
355
+
356
+ ---
357
+ <!-- trackio-cell
358
+ {"type": "code", "id": "cell_1d5aa5c25cad", "created_at": "2026-07-19T09:49:48+00:00", "title": "Run: python reproduce.py (exit 0)", "command": ["/home/vim/miniconda3/envs/pytorch-gpu-hf/bin/python", "src/reproduce.py", "concept", "--repeats", "100", "--rff-dim", "500", "--output", "outputs/concept_100agents_100reps.csv", "--device", "cuda"], "exit_code": 0, "duration_s": 333.017}
359
+ -->
360
+ ````bash
361
+ $ /home/vim/miniconda3/envs/pytorch-gpu-hf/bin/python src/reproduce.py concept --repeats 100 --rff-dim 500 --output outputs/concept_100agents_100reps.csv --device cuda
362
+ ````
363
+
364
+ exit 0 · 333.0s
365
+
366
+
367
+ ````python title=reproduce.py
368
+ #!/usr/bin/env python3
369
+ """Independent, GPU-capable checks for Fermanian et al. (2026), arXiv:2603.02233.
370
+
371
+ The implementation follows Algorithms 1--2 at a scaled experimental budget. It
372
+ uses RFF KMEs to choose a target client's collaboration weights and fits the
373
+ weighted empirical-risk minimizer exactly for linear regression. The FEMNIST
374
+ mode uses the public natural writer partition and a linear softmax model.
375
+ """
376
+ from __future__ import annotations
377
+
378
+ import argparse
379
+ import csv
380
+ import json
381
+ import math
382
+ from pathlib import Path
383
+ from time import perf_counter
384
+
385
+ import numpy as np
386
+ import torch
387
+
388
+
389
+ ROOT = Path(__file__).resolve().parents[1]
390
+
391
+
392
+ def device_from_arg(name: str) -> torch.device:
393
+ if name == "auto":
394
+ return torch.device("cuda" if torch.cuda.is_available() else "cpu")
395
+ return torch.device(name)
396
+
397
+
398
+ def rff(z: torch.Tensor, d_rff: int, generator: torch.Generator, scale: torch.Tensor) -> torch.Tensor:
399
+ """Gaussian-kernel RFF, with a supplied coordinate rescaling."""
400
+ w = torch.randn((z.shape[1], d_rff), generator=generator, device=z.device, dtype=z.dtype)
401
+ w = w * scale[:, None]
402
+ b = 2 * math.pi * torch.rand((d_rff,), generator=generator, device=z.device, dtype=z.dtype)
403
+ return math.sqrt(2.0 / d_rff) * torch.cos(z @ w + b)
404
+
405
+
406
+ def simplex_weights(means: torch.Tensor, target_phi: torch.Tensor, steps: int = 900) -> torch.Tensor:
407
+ """Solve Algorithm 1's empirical Q-aggregation objective by projected GD."""
408
+ b_agents = means.shape[0]
409
+ n1 = target_phi.shape[0]
410
+ centered = target_phi - means[0]
411
+ cov = centered.T @ centered / max(n1 - 1, 1)
412
+ trace = torch.trace(cov)
413
+ diffs = means[1:] - means[0]
414
+ q_terms = torch.sqrt(torch.clamp(torch.einsum("bd,df,bf->b", diffs, cov, diffs), min=0.0) + 1e-12)
415
+ p_terms = torch.linalg.vector_norm(diffs, dim=1)
416
+ # Table 1 uses CQ^2=CP=log B: CQ=sqrt(log B), CP=log B.
417
+ cq, cp = math.sqrt(math.log(b_agents)), math.log(b_agents)
418
+ w = torch.full((b_agents,), 1.0 / b_agents, device=means.device, dtype=means.dtype)
419
+ for _ in range(steps):
420
+ w.requires_grad_(True)
421
+ mixture = w @ means
422
+ loss = torch.sum((mixture - means[0]) ** 2) + 2.0 * w[0] * trace / n1
423
+ loss = loss + cq / math.sqrt(n1) * torch.dot(w[1:], q_terms)
424
+ loss = loss + cp / n1 * torch.dot(w[1:], p_terms)
425
+ (grad,) = torch.autograd.grad(loss, w)
426
+ with torch.no_grad():
427
+ # Euclidean projection on the simplex, Duchi et al. (2008).
428
+ v = w - 0.04 * grad
429
+ u, _ = torch.sort(v, descending=True)
430
+ cssv = torch.cumsum(u, 0) - 1
431
+ ind = torch.arange(1, b_agents + 1, device=v.device, dtype=v.dtype)
432
+ rho = torch.nonzero(u - cssv / ind > 0, as_tuple=False)[-1, 0]
433
+ theta = cssv[rho] / (rho + 1)
434
+ w = torch.clamp(v - theta, min=0)
435
+ return w.detach()
436
+
437
+
438
+ def weighted_linear_fit(xs: torch.Tensor, ys: torch.Tensor, weights: torch.Tensor, ridge: float = 1e-3) -> torch.Tensor:
439
+ """Exact weighted empirical-risk minimizer for linear regression."""
440
+ b_agents, n, dim = xs.shape
441
+ x = torch.cat([xs, torch.ones((b_agents, n, 1), device=xs.device, dtype=xs.dtype)], dim=-1)
442
+ sample_w = (weights[:, None] / n).expand(b_agents, n).reshape(-1)
443
+ flat_x, flat_y = x.reshape(-1, dim + 1), ys.reshape(-1)
444
+ lhs = flat_x.T @ (flat_x * sample_w[:, None]) + ridge * torch.eye(dim + 1, device=xs.device)
445
+ rhs = flat_x.T @ (flat_y * sample_w)
446
+ return torch.linalg.solve(lhs, rhs)
447
+
448
+
449
+ def concept_trial(seed: int, sigma2: float, d_rff: int, device: torch.device) -> dict:
450
+ g = torch.Generator(device=device).manual_seed(seed)
451
+ b_agents, n, dim, n_test = 100, 10, 20, 1000
452
+ signs = torch.where(torch.rand((b_agents,), generator=g, device=device) > .5, 1., -1.)
453
+ beta0 = torch.randn((dim,), generator=g, device=device)
454
+ eps = torch.randn((b_agents, dim), generator=g, device=device)
455
+ betas = signs[:, None] * math.sqrt(1 - sigma2) * beta0 + math.sqrt(sigma2) * eps
456
+ x = 1.0 + torch.randn((b_agents, n, dim), generator=g, device=device)
457
+ y = torch.einsum("bnd,bd->bn", x, betas) + math.sqrt(2.0) * torch.randn((b_agents, n), generator=g, device=device)
458
+ x_test = 1.0 + torch.randn((n_test, dim), generator=g, device=device)
459
+ y_test = x_test @ betas[0] + math.sqrt(2.0) * torch.randn((n_test,), generator=g, device=device)
460
+ # The paper's weighted Gaussian KME kernel gives X a 1/sqrt(d) scaling.
461
+ z = torch.cat([x, y[..., None]], dim=-1).reshape(-1, dim + 1)
462
+ scale = torch.cat([torch.full((dim,), 1 / math.sqrt(dim), device=device), torch.ones((1,), device=device)])
463
+ phi = rff(z, d_rff, g, scale).reshape(b_agents, n, d_rff)
464
+ means = phi.mean(1)
465
+ q_w = simplex_weights(means, phi[0])
466
+ local = torch.zeros(b_agents, device=device); local[0] = 1
467
+ grand = torch.full((b_agents,), 1 / b_agents, device=device)
468
+ oracle = (signs == signs[0]).float(); oracle /= oracle.sum()
469
+ result = {"sigma2": sigma2, "seed": seed, "q_local_weight": float(q_w[0]),
470
+ "q_same_group_weight": float(q_w[signs == signs[0]].sum())}
471
+ for name, weights in [("local", local), ("grandmean", grand), ("oracle", oracle), ("qaggregation", q_w)]:
472
+ beta = weighted_linear_fit(x, y, weights)
473
+ pred = x_test @ beta[:-1] + beta[-1]
474
+ result[f"{name}_mse"] = float(torch.mean((pred - y_test) ** 2))
475
+ return result
476
+
477
+
478
+ def run_concept(args: argparse.Namespace) -> list[dict]:
479
+ device = device_from_arg(args.device)
480
+ rows = []
481
+ for sigma2 in args.sigmas:
482
+ for rep in range(args.repeats):
483
+ rows.append(concept_trial(args.seed + rep * 101 + int(sigma2 * 1000), sigma2, args.rff_dim, device))
484
+ return rows
485
+
486
+
487
+ def exact_rbf_mmd(x: torch.Tensor, y: torch.Tensor, bandwidth: float = 1.0) -> torch.Tensor:
488
+ def k(a: torch.Tensor, b: torch.Tensor) -> torch.Tensor:
489
+ return torch.exp(-torch.cdist(a, b).square() / (2 * bandwidth * bandwidth))
490
+ return torch.sqrt(torch.clamp(k(x, x).mean() + k(y, y).mean() - 2 * k(x, y).mean(), min=0.0))
491
+
492
+
493
+ def run_rff(args: argparse.Namespace) -> list[dict]:
494
+ device = device_from_arg(args.device)
495
+ g = torch.Generator(device=device).manual_seed(args.seed)
496
+ x = torch.randn((300, 8), generator=g, device=device)
497
+ y = torch.randn((300, 8), generator=g, device=device) + torch.tensor([0.35] + [0.] * 7, device=device)
498
+ exact = float(exact_rbf_mmd(x, y))
499
+ rows = []
500
+ for d_rff in args.rff_grid:
501
+ errors = []
502
+ for trial in range(args.repeats):
503
+ tg = torch.Generator(device=device).manual_seed(args.seed + 1000 * d_rff + trial)
504
+ phi = rff(torch.cat([x, y]), d_rff, tg, torch.ones(8, device=device))
505
+ approx = torch.linalg.vector_norm(phi[:len(x)].mean(0) - phi[len(x):].mean(0)).item()
506
+ errors.append(abs(approx - exact))
507
+ rows.append({"rff_dim": d_rff, "exact_mmd": exact, "mean_abs_error": float(np.mean(errors)),
508
+ "std_abs_error": float(np.std(errors)), "sqrt_logB_over_D": math.sqrt(math.log(100) / d_rff),
509
+ "one_shot_floats_per_agent": d_rff})
510
+ return rows
511
+
512
+
513
+ def run_femnist(args: argparse.Namespace) -> list[dict]:
514
+ """Actual FEMNIST writer data, deliberately scaled in samples/features/epochs."""
515
+ from datasets import load_dataset
516
+ device = device_from_arg(args.device)
517
+ local_parquet = ROOT / "data" / "femnist" / "data" / "train-00000-of-00001.parquet"
518
+ dataset = load_dataset("parquet", data_files=str(local_parquet), split="train", streaming=True) if local_parquet.exists() else load_dataset("flwrlabs/femnist", split="train", streaming=True)
519
+ by_writer: dict[str, list[tuple[np.ndarray, int]]] = {}
520
+ want = args.agents
521
+ for ex in dataset:
522
+ writer = ex["writer_id"]
523
+ if writer not in by_writer and len(by_writer) >= want:
524
+ continue
525
+ if writer in by_writer or len(by_writer) < want:
526
+ by_writer.setdefault(writer, []).append((np.asarray(ex["image"], dtype=np.float32) / 255., int(ex["character"])))
527
+ if len(by_writer) >= want and all(len(v) >= args.train_per_agent + args.test_per_agent for v in by_writer.values()):
528
+ break
529
+ eligible = [v for v in by_writer.values() if len(v) >= args.train_per_agent + args.test_per_agent][:want]
530
+ if len(eligible) < want:
531
+ raise RuntimeError(f"Only found {len(eligible)} eligible writers; need {want}")
532
+ x_train = torch.tensor(np.stack([np.stack([a for a, _ in v[:args.train_per_agent]]) for v in eligible]), device=device)
533
+ y_train = torch.tensor(np.stack([[b for _, b in v[:args.train_per_agent]] for v in eligible]), device=device, dtype=torch.long)
534
+ x_test = torch.tensor(np.stack([np.stack([a for a, _ in v[args.train_per_agent:args.train_per_agent + args.test_per_agent]]) for v in eligible]), device=device)
535
+ y_test = torch.tensor(np.stack([[b for _, b in v[args.train_per_agent:args.train_per_agent + args.test_per_agent]] for v in eligible]), device=device, dtype=torch.long)
536
+ b_agents, n, raw_dim = x_train.shape[0], x_train.shape[1], 784
537
+ flat = x_train.reshape(b_agents, n, raw_dim)
538
+ # Fixed random projection reduces the RFF input cost while retaining a reproducible vision proxy.
539
+ g = torch.Generator(device=device).manual_seed(args.seed)
540
+ proj = torch.randn((raw_dim, args.feature_dim), generator=g, device=device) / math.sqrt(raw_dim)
541
+ features = flat @ proj
542
+ phi_all = rff(features.reshape(-1, args.feature_dim), args.rff_dim, g, torch.ones(args.feature_dim, device=device)).reshape(b_agents, n, args.rff_dim)
543
+ means = phi_all.mean(1)
544
+ # Fit one regularized softmax classifier for each target using its personalized weights.
545
+ rows = []
546
+ all_x = features.reshape(-1, args.feature_dim)
547
+ all_y = y_train.reshape(-1)
548
+ for target in range(b_agents):
549
+ weights = simplex_weights(means, phi_all[target], steps=300)
550
+ schemes = {"local": torch.nn.functional.one_hot(torch.tensor(target, device=device), b_agents).float(),
551
+ "grandmean": torch.full((b_agents,), 1 / b_agents, device=device), "qaggregation": weights}
552
+ test_feat = x_test[target].reshape(args.test_per_agent, raw_dim) @ proj
553
+ for name, w in schemes.items():
554
+ model = torch.zeros((args.feature_dim, 62), device=device, requires_grad=True)
555
+ opt = torch.optim.Adam([model], lr=args.lr, weight_decay=1e-4)
556
+ sample_w = (w[:, None] / n).expand(b_agents, n).reshape(-1)
557
+ for _ in range(args.epochs):
558
+ opt.zero_grad()
559
+ losses = torch.nn.functional.cross_entropy(all_x @ model, all_y, reduction="none")
560
+ (losses * sample_w).sum().backward(); opt.step()
561
+ acc = ((test_feat @ model).argmax(1) == y_test[target]).float().mean().item()
562
+ rows.append({"agent": target, "method": name, "accuracy": acc, "q_local_weight": float(weights[target])})
563
+ return rows
564
+
565
+
566
+ def write_rows(rows: list[dict], out: Path) -> None:
567
+ out.parent.mkdir(parents=True, exist_ok=True)
568
+ with out.open("w", newline="") as f:
569
+ writer = csv.DictWriter(f, fieldnames=list(rows[0]))
570
+ writer.writeheader(); writer.writerows(rows)
571
+ out.with_suffix(".json").write_text(json.dumps(rows, indent=2))
572
+
573
+
574
+ def main() -> None:
575
+ p = argparse.ArgumentParser()
576
+ p.add_argument("mode", choices=["concept", "rff", "femnist"])
577
+ p.add_argument("--output", type=Path, required=True)
578
+ p.add_argument("--device", default="auto")
579
+ p.add_argument("--seed", type=int, default=20260719)
580
+ p.add_argument("--repeats", type=int, default=30)
581
+ p.add_argument("--rff-dim", type=int, default=500)
582
+ p.add_argument("--sigmas", type=float, nargs="+", default=[0., .1, .25, .5, .75, 1.])
583
+ p.add_argument("--rff-grid", type=int, nargs="+", default=[16, 32, 64, 128, 256, 512, 1024])
584
+ p.add_argument("--agents", type=int, default=192)
585
+ p.add_argument("--train-per-agent", type=int, default=32)
586
+ p.add_argument("--test-per-agent", type=int, default=16)
587
+ p.add_argument("--feature-dim", type=int, default=64)
588
+ p.add_argument("--epochs", type=int, default=20)
589
+ p.add_argument("--lr", type=float, default=.08)
590
+ args = p.parse_args()
591
+ started = perf_counter()
592
+ rows = {"concept": run_concept, "rff": run_rff, "femnist": run_femnist}[args.mode](args)
593
+ write_rows(rows, args.output)
594
+ if args.mode == "concept":
595
+ summary = {str(s): {k: float(np.mean([r[k] for r in rows if r["sigma2"] == s])) for k in ["local_mse", "grandmean_mse", "oracle_mse", "qaggregation_mse", "q_local_weight", "q_same_group_weight"]} for s in args.sigmas}
596
+ print(json.dumps({"concept_summary": summary}, indent=2))
597
+ elif args.mode == "rff":
598
+ print(json.dumps({"rff_summary": {str(r["rff_dim"]): r["mean_abs_error"] for r in rows}}, indent=2))
599
+ elif args.mode == "femnist":
600
+ summary = {m: float(np.mean([r["accuracy"] for r in rows if r["method"] == m])) for m in ["local", "grandmean", "qaggregation"]}
601
+ print(json.dumps({"femnist_accuracy": summary}, indent=2))
602
+ print(json.dumps({"mode": args.mode, "rows": len(rows), "output": str(args.output), "seconds": round(perf_counter() - started, 2), "device": str(device_from_arg(args.device))}, indent=2))
603
+
604
+
605
+ if __name__ == "__main__":
606
+ main()
607
+
608
+ ````
609
+
610
+
611
+ ````output
612
+ {
613
+ "concept_summary": {
614
+ "0.0": {
615
+ "local_mse": 15.748318042755127,
616
+ "grandmean_mse": 40.606961507797244,
617
+ "oracle_mse": 2.0862513756752015,
618
+ "qaggregation_mse": 17.83958231449127,
619
+ "q_local_weight": 0.9771137142181396,
620
+ "q_same_group_weight": 0.9975556522607804
621
+ },
622
+ "0.1": {
623
+ "local_mse": 15.361917610168456,
624
+ "grandmean_mse": 45.73134832382202,
625
+ "oracle_mse": 6.119768922328949,
626
+ "qaggregation_mse": 18.07122127532959,
627
+ "q_local_weight": 0.9787336409091949,
628
+ "q_same_group_weight": 0.9981670784950256
629
+ },
630
+ "0.25": {
631
+ "local_mse": 15.827756361961365,
632
+ "grandmean_mse": 39.008377923965455,
633
+ "oracle_mse": 11.351614267826081,
634
+ "qaggregation_mse": 25.339054584503174,
635
+ "q_local_weight": 0.9825072956085205,
636
+ "q_same_group_weight": 0.99599778175354
637
+ },
638
+ "0.5": {
639
+ "local_mse": 14.459182372093201,
640
+ "grandmean_mse": 45.512830362319946,
641
+ "oracle_mse": 24.322485909461975,
642
+ "qaggregation_mse": 25.430337595939637,
643
+ "q_local_weight": 0.9876203048229217,
644
+ "q_same_group_weight": 0.9948384702205658
645
+ },
646
+ "0.75": {
647
+ "local_mse": 15.87236515045166,
648
+ "grandmean_mse": 39.63199781417847,
649
+ "oracle_mse": 30.853181953430177,
650
+ "qaggregation_mse": 33.74232290267944,
651
+ "q_local_weight": 0.9877495408058167,
652
+ "q_same_group_weight": 0.9943676108121872
653
+ },
654
+ "1.0": {
655
+ "local_mse": 15.390748682022094,
656
+ "grandmean_mse": 39.16119129180908,
657
+ "oracle_mse": 40.00151491165161,
658
+ "qaggregation_mse": 30.293080587387085,
659
+ "q_local_weight": 0.9896031093597412,
660
+ "q_same_group_weight": 0.9947122848033905
661
+ }
662
+ }
663
+ }
664
+ {
665
+ "mode": "concept",
666
+ "rows": 600,
667
+ "output": "outputs/concept_100agents_100reps.csv",
668
+ "seconds": 332.1,
669
+ "device": "cuda"
670
+ }
671
+
672
+ ````
673
+
674
+
675
+ ---
676
+ <!-- trackio-cell
677
+ {"type": "artifact", "id": "cell_5838f11deee2", "created_at": "2026-07-19T09:49:49+00:00", "title": "Artifact: concept_100agents_100reps.csv", "path": "outputs/concept_100agents_100reps.csv", "size": 63718, "artifact_type": "dataset", "auto": true}
678
+ -->
679
+ **📦 Artifact** `outputs/concept_100agents_100reps.csv` · dataset · 63.7 kB
680
+
681
+ https://huggingface.co/buckets/vimarsh/repro-adaptive-personalized-fl-multi-task-averaging-kernel-mean-embedding-artifacts#logbook-files/outputs/concept_100agents_100reps.csv
682
+
683
+
684
+ ---
685
+ <!-- trackio-cell
686
+ {"type": "code", "id": "cell_0e566181e79c", "created_at": "2026-07-19T09:50:42+00:00", "title": "Run: python reproduce.py (exit 0)", "command": ["/home/vim/miniconda3/envs/pytorch-gpu-hf/bin/python", "src/reproduce.py", "concept", "--repeats", "100", "--rff-dim", "500", "--output", "outputs/concept_100agents_100reps.csv", "--device", "cuda"], "exit_code": 0, "duration_s": 300.077}
687
+ -->
688
+ ````bash
689
+ $ /home/vim/miniconda3/envs/pytorch-gpu-hf/bin/python src/reproduce.py concept --repeats 100 --rff-dim 500 --output outputs/concept_100agents_100reps.csv --device cuda
690
+ ````
691
+
692
+ exit 0 · 300.1s
693
+
694
+
695
+ ````python title=reproduce.py
696
+ #!/usr/bin/env python3
697
+ """Independent, GPU-capable checks for Fermanian et al. (2026), arXiv:2603.02233.
698
+
699
+ The implementation follows Algorithms 1--2 at a scaled experimental budget. It
700
+ uses RFF KMEs to choose a target client's collaboration weights and fits the
701
+ weighted empirical-risk minimizer exactly for linear regression. The FEMNIST
702
+ mode uses the public natural writer partition and a linear softmax model.
703
+ """
704
+ from __future__ import annotations
705
+
706
+ import argparse
707
+ import csv
708
+ import json
709
+ import math
710
+ from pathlib import Path
711
+ from time import perf_counter
712
+
713
+ import numpy as np
714
+ import torch
715
+
716
+
717
+ ROOT = Path(__file__).resolve().parents[1]
718
+
719
+
720
+ def device_from_arg(name: str) -> torch.device:
721
+ if name == "auto":
722
+ return torch.device("cuda" if torch.cuda.is_available() else "cpu")
723
+ return torch.device(name)
724
+
725
+
726
+ def rff(z: torch.Tensor, d_rff: int, generator: torch.Generator, scale: torch.Tensor) -> torch.Tensor:
727
+ """Gaussian-kernel RFF, with a supplied coordinate rescaling."""
728
+ w = torch.randn((z.shape[1], d_rff), generator=generator, device=z.device, dtype=z.dtype)
729
+ w = w * scale[:, None]
730
+ b = 2 * math.pi * torch.rand((d_rff,), generator=generator, device=z.device, dtype=z.dtype)
731
+ return math.sqrt(2.0 / d_rff) * torch.cos(z @ w + b)
732
+
733
+
734
+ def simplex_weights(means: torch.Tensor, target_phi: torch.Tensor, steps: int = 900) -> torch.Tensor:
735
+ """Solve Algorithm 1's empirical Q-aggregation objective by projected GD."""
736
+ b_agents = means.shape[0]
737
+ n1 = target_phi.shape[0]
738
+ centered = target_phi - means[0]
739
+ cov = centered.T @ centered / max(n1 - 1, 1)
740
+ trace = torch.trace(cov)
741
+ diffs = means[1:] - means[0]
742
+ q_terms = torch.sqrt(torch.clamp(torch.einsum("bd,df,bf->b", diffs, cov, diffs), min=0.0) + 1e-12)
743
+ p_terms = torch.linalg.vector_norm(diffs, dim=1)
744
+ # Table 1 uses CQ^2=CP=log B: CQ=sqrt(log B), CP=log B.
745
+ cq, cp = math.sqrt(math.log(b_agents)), math.log(b_agents)
746
+ w = torch.full((b_agents,), 1.0 / b_agents, device=means.device, dtype=means.dtype)
747
+ for _ in range(steps):
748
+ w.requires_grad_(True)
749
+ mixture = w @ means
750
+ loss = torch.sum((mixture - means[0]) ** 2) + 2.0 * w[0] * trace / n1
751
+ loss = loss + cq / math.sqrt(n1) * torch.dot(w[1:], q_terms)
752
+ loss = loss + cp / n1 * torch.dot(w[1:], p_terms)
753
+ (grad,) = torch.autograd.grad(loss, w)
754
+ with torch.no_grad():
755
+ # Euclidean projection on the simplex, Duchi et al. (2008).
756
+ v = w - 0.04 * grad
757
+ u, _ = torch.sort(v, descending=True)
758
+ cssv = torch.cumsum(u, 0) - 1
759
+ ind = torch.arange(1, b_agents + 1, device=v.device, dtype=v.dtype)
760
+ rho = torch.nonzero(u - cssv / ind > 0, as_tuple=False)[-1, 0]
761
+ theta = cssv[rho] / (rho + 1)
762
+ w = torch.clamp(v - theta, min=0)
763
+ return w.detach()
764
+
765
+
766
+ def weighted_linear_fit(xs: torch.Tensor, ys: torch.Tensor, weights: torch.Tensor, ridge: float = 1e-3) -> torch.Tensor:
767
+ """Exact weighted empirical-risk minimizer for linear regression."""
768
+ b_agents, n, dim = xs.shape
769
+ x = torch.cat([xs, torch.ones((b_agents, n, 1), device=xs.device, dtype=xs.dtype)], dim=-1)
770
+ sample_w = (weights[:, None] / n).expand(b_agents, n).reshape(-1)
771
+ flat_x, flat_y = x.reshape(-1, dim + 1), ys.reshape(-1)
772
+ lhs = flat_x.T @ (flat_x * sample_w[:, None]) + ridge * torch.eye(dim + 1, device=xs.device)
773
+ rhs = flat_x.T @ (flat_y * sample_w)
774
+ return torch.linalg.solve(lhs, rhs)
775
+
776
+
777
+ def concept_trial(seed: int, sigma2: float, d_rff: int, device: torch.device) -> dict:
778
+ g = torch.Generator(device=device).manual_seed(seed)
779
+ b_agents, n, dim, n_test = 100, 10, 20, 1000
780
+ signs = torch.where(torch.rand((b_agents,), generator=g, device=device) > .5, 1., -1.)
781
+ beta0 = torch.randn((dim,), generator=g, device=device)
782
+ eps = torch.randn((b_agents, dim), generator=g, device=device)
783
+ betas = signs[:, None] * math.sqrt(1 - sigma2) * beta0 + math.sqrt(sigma2) * eps
784
+ x = 1.0 + torch.randn((b_agents, n, dim), generator=g, device=device)
785
+ y = torch.einsum("bnd,bd->bn", x, betas) + math.sqrt(2.0) * torch.randn((b_agents, n), generator=g, device=device)
786
+ x_test = 1.0 + torch.randn((n_test, dim), generator=g, device=device)
787
+ y_test = x_test @ betas[0] + math.sqrt(2.0) * torch.randn((n_test,), generator=g, device=device)
788
+ # The paper's weighted Gaussian KME kernel gives X a 1/sqrt(d) scaling.
789
+ z = torch.cat([x, y[..., None]], dim=-1).reshape(-1, dim + 1)
790
+ scale = torch.cat([torch.full((dim,), 1 / math.sqrt(dim), device=device), torch.ones((1,), device=device)])
791
+ phi = rff(z, d_rff, g, scale).reshape(b_agents, n, d_rff)
792
+ means = phi.mean(1)
793
+ q_w = simplex_weights(means, phi[0])
794
+ local = torch.zeros(b_agents, device=device); local[0] = 1
795
+ grand = torch.full((b_agents,), 1 / b_agents, device=device)
796
+ oracle = (signs == signs[0]).float(); oracle /= oracle.sum()
797
+ result = {"sigma2": sigma2, "seed": seed, "q_local_weight": float(q_w[0]),
798
+ "q_same_group_weight": float(q_w[signs == signs[0]].sum())}
799
+ for name, weights in [("local", local), ("grandmean", grand), ("oracle", oracle), ("qaggregation", q_w)]:
800
+ beta = weighted_linear_fit(x, y, weights)
801
+ pred = x_test @ beta[:-1] + beta[-1]
802
+ result[f"{name}_mse"] = float(torch.mean((pred - y_test) ** 2))
803
+ return result
804
+
805
+
806
+ def run_concept(args: argparse.Namespace) -> list[dict]:
807
+ device = device_from_arg(args.device)
808
+ rows = []
809
+ for sigma2 in args.sigmas:
810
+ for rep in range(args.repeats):
811
+ rows.append(concept_trial(args.seed + rep * 101 + int(sigma2 * 1000), sigma2, args.rff_dim, device))
812
+ return rows
813
+
814
+
815
+ def exact_rbf_mmd(x: torch.Tensor, y: torch.Tensor, bandwidth: float = 1.0) -> torch.Tensor:
816
+ def k(a: torch.Tensor, b: torch.Tensor) -> torch.Tensor:
817
+ return torch.exp(-torch.cdist(a, b).square() / (2 * bandwidth * bandwidth))
818
+ return torch.sqrt(torch.clamp(k(x, x).mean() + k(y, y).mean() - 2 * k(x, y).mean(), min=0.0))
819
+
820
+
821
+ def run_rff(args: argparse.Namespace) -> list[dict]:
822
+ device = device_from_arg(args.device)
823
+ g = torch.Generator(device=device).manual_seed(args.seed)
824
+ x = torch.randn((300, 8), generator=g, device=device)
825
+ y = torch.randn((300, 8), generator=g, device=device) + torch.tensor([0.35] + [0.] * 7, device=device)
826
+ exact = float(exact_rbf_mmd(x, y))
827
+ rows = []
828
+ for d_rff in args.rff_grid:
829
+ errors = []
830
+ for trial in range(args.repeats):
831
+ tg = torch.Generator(device=device).manual_seed(args.seed + 1000 * d_rff + trial)
832
+ phi = rff(torch.cat([x, y]), d_rff, tg, torch.ones(8, device=device))
833
+ approx = torch.linalg.vector_norm(phi[:len(x)].mean(0) - phi[len(x):].mean(0)).item()
834
+ errors.append(abs(approx - exact))
835
+ rows.append({"rff_dim": d_rff, "exact_mmd": exact, "mean_abs_error": float(np.mean(errors)),
836
+ "std_abs_error": float(np.std(errors)), "sqrt_logB_over_D": math.sqrt(math.log(100) / d_rff),
837
+ "one_shot_floats_per_agent": d_rff})
838
+ return rows
839
+
840
+
841
+ def run_femnist(args: argparse.Namespace) -> list[dict]:
842
+ """Actual FEMNIST writer data, deliberately scaled in samples/features/epochs."""
843
+ from datasets import load_dataset
844
+ device = device_from_arg(args.device)
845
+ local_parquet = ROOT / "data" / "femnist" / "data" / "train-00000-of-00001.parquet"
846
+ dataset = load_dataset("parquet", data_files=str(local_parquet), split="train", streaming=True) if local_parquet.exists() else load_dataset("flwrlabs/femnist", split="train", streaming=True)
847
+ by_writer: dict[str, list[tuple[np.ndarray, int]]] = {}
848
+ want = args.agents
849
+ for ex in dataset:
850
+ writer = ex["writer_id"]
851
+ if writer not in by_writer and len(by_writer) >= want:
852
+ continue
853
+ if writer in by_writer or len(by_writer) < want:
854
+ by_writer.setdefault(writer, []).append((np.asarray(ex["image"], dtype=np.float32) / 255., int(ex["character"])))
855
+ if len(by_writer) >= want and all(len(v) >= args.train_per_agent + args.test_per_agent for v in by_writer.values()):
856
+ break
857
+ eligible = [v for v in by_writer.values() if len(v) >= args.train_per_agent + args.test_per_agent][:want]
858
+ if len(eligible) < want:
859
+ raise RuntimeError(f"Only found {len(eligible)} eligible writers; need {want}")
860
+ x_train = torch.tensor(np.stack([np.stack([a for a, _ in v[:args.train_per_agent]]) for v in eligible]), device=device)
861
+ y_train = torch.tensor(np.stack([[b for _, b in v[:args.train_per_agent]] for v in eligible]), device=device, dtype=torch.long)
862
+ x_test = torch.tensor(np.stack([np.stack([a for a, _ in v[args.train_per_agent:args.train_per_agent + args.test_per_agent]]) for v in eligible]), device=device)
863
+ y_test = torch.tensor(np.stack([[b for _, b in v[args.train_per_agent:args.train_per_agent + args.test_per_agent]] for v in eligible]), device=device, dtype=torch.long)
864
+ b_agents, n, raw_dim = x_train.shape[0], x_train.shape[1], 784
865
+ flat = x_train.reshape(b_agents, n, raw_dim)
866
+ # Fixed random projection reduces the RFF input cost while retaining a reproducible vision proxy.
867
+ g = torch.Generator(device=device).manual_seed(args.seed)
868
+ proj = torch.randn((raw_dim, args.feature_dim), generator=g, device=device) / math.sqrt(raw_dim)
869
+ features = flat @ proj
870
+ phi_all = rff(features.reshape(-1, args.feature_dim), args.rff_dim, g, torch.ones(args.feature_dim, device=device)).reshape(b_agents, n, args.rff_dim)
871
+ means = phi_all.mean(1)
872
+ # Fit one regularized softmax classifier for each target using its personalized weights.
873
+ rows = []
874
+ all_x = features.reshape(-1, args.feature_dim)
875
+ all_y = y_train.reshape(-1)
876
+ for target in range(b_agents):
877
+ weights = simplex_weights(means, phi_all[target], steps=300)
878
+ schemes = {"local": torch.nn.functional.one_hot(torch.tensor(target, device=device), b_agents).float(),
879
+ "grandmean": torch.full((b_agents,), 1 / b_agents, device=device), "qaggregation": weights}
880
+ test_feat = x_test[target].reshape(args.test_per_agent, raw_dim) @ proj
881
+ for name, w in schemes.items():
882
+ model = torch.zeros((args.feature_dim, 62), device=device, requires_grad=True)
883
+ opt = torch.optim.Adam([model], lr=args.lr, weight_decay=1e-4)
884
+ sample_w = (w[:, None] / n).expand(b_agents, n).reshape(-1)
885
+ for _ in range(args.epochs):
886
+ opt.zero_grad()
887
+ losses = torch.nn.functional.cross_entropy(all_x @ model, all_y, reduction="none")
888
+ (losses * sample_w).sum().backward(); opt.step()
889
+ acc = ((test_feat @ model).argmax(1) == y_test[target]).float().mean().item()
890
+ rows.append({"agent": target, "method": name, "accuracy": acc, "q_local_weight": float(weights[target])})
891
+ return rows
892
+
893
+
894
+ def write_rows(rows: list[dict], out: Path) -> None:
895
+ out.parent.mkdir(parents=True, exist_ok=True)
896
+ with out.open("w", newline="") as f:
897
+ writer = csv.DictWriter(f, fieldnames=list(rows[0]))
898
+ writer.writeheader(); writer.writerows(rows)
899
+ out.with_suffix(".json").write_text(json.dumps(rows, indent=2))
900
+
901
+
902
+ def main() -> None:
903
+ p = argparse.ArgumentParser()
904
+ p.add_argument("mode", choices=["concept", "rff", "femnist"])
905
+ p.add_argument("--output", type=Path, required=True)
906
+ p.add_argument("--device", default="auto")
907
+ p.add_argument("--seed", type=int, default=20260719)
908
+ p.add_argument("--repeats", type=int, default=30)
909
+ p.add_argument("--rff-dim", type=int, default=500)
910
+ p.add_argument("--sigmas", type=float, nargs="+", default=[0., .1, .25, .5, .75, 1.])
911
+ p.add_argument("--rff-grid", type=int, nargs="+", default=[16, 32, 64, 128, 256, 512, 1024])
912
+ p.add_argument("--agents", type=int, default=192)
913
+ p.add_argument("--train-per-agent", type=int, default=32)
914
+ p.add_argument("--test-per-agent", type=int, default=16)
915
+ p.add_argument("--feature-dim", type=int, default=64)
916
+ p.add_argument("--epochs", type=int, default=20)
917
+ p.add_argument("--lr", type=float, default=.08)
918
+ args = p.parse_args()
919
+ started = perf_counter()
920
+ rows = {"concept": run_concept, "rff": run_rff, "femnist": run_femnist}[args.mode](args)
921
+ write_rows(rows, args.output)
922
+ if args.mode == "concept":
923
+ summary = {str(s): {k: float(np.mean([r[k] for r in rows if r["sigma2"] == s])) for k in ["local_mse", "grandmean_mse", "oracle_mse", "qaggregation_mse", "q_local_weight", "q_same_group_weight"]} for s in args.sigmas}
924
+ print(json.dumps({"concept_summary": summary}, indent=2))
925
+ elif args.mode == "rff":
926
+ print(json.dumps({"rff_summary": {str(r["rff_dim"]): r["mean_abs_error"] for r in rows}}, indent=2))
927
+ elif args.mode == "femnist":
928
+ summary = {m: float(np.mean([r["accuracy"] for r in rows if r["method"] == m])) for m in ["local", "grandmean", "qaggregation"]}
929
+ print(json.dumps({"femnist_accuracy": summary}, indent=2))
930
+ print(json.dumps({"mode": args.mode, "rows": len(rows), "output": str(args.output), "seconds": round(perf_counter() - started, 2), "device": str(device_from_arg(args.device))}, indent=2))
931
+
932
+
933
+ if __name__ == "__main__":
934
+ main()
935
+
936
+ ````
937
+
938
+
939
+ ````output
940
+ {
941
+ "concept_summary": {
942
+ "0.0": {
943
+ "local_mse": 15.748318042755127,
944
+ "grandmean_mse": 40.606961507797244,
945
+ "oracle_mse": 2.0862513756752015,
946
+ "qaggregation_mse": 17.83958231449127,
947
+ "q_local_weight": 0.9771137142181396,
948
+ "q_same_group_weight": 0.9975556522607804
949
+ },
950
+ "0.1": {
951
+ "local_mse": 15.361917610168456,
952
+ "grandmean_mse": 45.73134832382202,
953
+ "oracle_mse": 6.119768922328949,
954
+ "qaggregation_mse": 18.07122127532959,
955
+ "q_local_weight": 0.9787336409091949,
956
+ "q_same_group_weight": 0.9981670784950256
957
+ },
958
+ "0.25": {
959
+ "local_mse": 15.827756361961365,
960
+ "grandmean_mse": 39.008377923965455,
961
+ "oracle_mse": 11.351614267826081,
962
+ "qaggregation_mse": 25.339054584503174,
963
+ "q_local_weight": 0.9825072956085205,
964
+ "q_same_group_weight": 0.99599778175354
965
+ },
966
+ "0.5": {
967
+ "local_mse": 14.459182372093201,
968
+ "grandmean_mse": 45.512830362319946,
969
+ "oracle_mse": 24.322485909461975,
970
+ "qaggregation_mse": 25.430337595939637,
971
+ "q_local_weight": 0.9876203048229217,
972
+ "q_same_group_weight": 0.9948384702205658
973
+ },
974
+ "0.75": {
975
+ "local_mse": 15.87236515045166,
976
+ "grandmean_mse": 39.63199781417847,
977
+ "oracle_mse": 30.853181953430177,
978
+ "qaggregation_mse": 33.74232290267944,
979
+ "q_local_weight": 0.9877495408058167,
980
+ "q_same_group_weight": 0.9943676108121872
981
+ },
982
+ "1.0": {
983
+ "local_mse": 15.390748682022094,
984
+ "grandmean_mse": 39.16119129180908,
985
+ "oracle_mse": 40.00151491165161,
986
+ "qaggregation_mse": 30.293080587387085,
987
+ "q_local_weight": 0.9896031093597412,
988
+ "q_same_group_weight": 0.9947122848033905
989
+ }
990
+ }
991
+ }
992
+ {
993
+ "mode": "concept",
994
+ "rows": 600,
995
+ "output": "outputs/concept_100agents_100reps.csv",
996
+ "seconds": 299.14,
997
+ "device": "cuda"
998
+ }
999
+
1000
+ ````
1001
+
1002
+
1003
+ ---
1004
+ <!-- trackio-cell
1005
+ {"type": "artifact", "id": "cell_40a11b9e84e9", "created_at": "2026-07-19T09:50:42+00:00", "title": "Artifact: concept_100agents_100reps.csv", "path": "outputs/concept_100agents_100reps.csv", "size": 63718, "artifact_type": "dataset", "auto": true}
1006
+ -->
1007
+ **📦 Artifact** `outputs/concept_100agents_100reps.csv` · dataset · 63.7 kB
1008
+
1009
+ https://huggingface.co/buckets/vimarsh/repro-adaptive-personalized-fl-multi-task-averaging-kernel-mean-embedding-artifacts#logbook-files/outputs/concept_100agents_100reps.csv
1010
+
1011
+
1012
+ ---
1013
+ <!-- trackio-cell
1014
+ {"type": "code", "id": "cell_09cae6dd6213", "created_at": "2026-07-19T09:51:35+00:00", "title": "Run: python reproduce.py (exit 0)", "command": ["/home/vim/miniconda3/envs/pytorch-gpu-hf/bin/python", "src/reproduce.py", "--help"], "exit_code": 0, "duration_s": 0.941}
1015
+ -->
1016
+ ````bash
1017
+ $ /home/vim/miniconda3/envs/pytorch-gpu-hf/bin/python src/reproduce.py --help
1018
+ ````
1019
+
1020
+ exit 0 · 0.9s
1021
+
1022
+
1023
+ ````python title=reproduce.py
1024
+ #!/usr/bin/env python3
1025
+ """Independent, GPU-capable checks for Fermanian et al. (2026), arXiv:2603.02233.
1026
+
1027
+ The implementation follows Algorithms 1--2 at a scaled experimental budget. It
1028
+ uses RFF KMEs to choose a target client's collaboration weights and fits the
1029
+ weighted empirical-risk minimizer exactly for linear regression. The FEMNIST
1030
+ mode uses the public natural writer partition and a linear softmax model.
1031
+ """
1032
+ from __future__ import annotations
1033
+
1034
+ import argparse
1035
+ import csv
1036
+ import json
1037
+ import math
1038
+ from pathlib import Path
1039
+ from time import perf_counter
1040
+
1041
+ import numpy as np
1042
+ import torch
1043
+
1044
+
1045
+ ROOT = Path(__file__).resolve().parents[1]
1046
+
1047
+
1048
+ def device_from_arg(name: str) -> torch.device:
1049
+ if name == "auto":
1050
+ return torch.device("cuda" if torch.cuda.is_available() else "cpu")
1051
+ return torch.device(name)
1052
+
1053
+
1054
+ def rff(z: torch.Tensor, d_rff: int, generator: torch.Generator, scale: torch.Tensor) -> torch.Tensor:
1055
+ """Gaussian-kernel RFF, with a supplied coordinate rescaling."""
1056
+ w = torch.randn((z.shape[1], d_rff), generator=generator, device=z.device, dtype=z.dtype)
1057
+ w = w * scale[:, None]
1058
+ b = 2 * math.pi * torch.rand((d_rff,), generator=generator, device=z.device, dtype=z.dtype)
1059
+ return math.sqrt(2.0 / d_rff) * torch.cos(z @ w + b)
1060
+
1061
+
1062
+ def simplex_weights(means: torch.Tensor, target_phi: torch.Tensor, steps: int = 900) -> torch.Tensor:
1063
+ """Solve Algorithm 1's empirical Q-aggregation objective by projected GD."""
1064
+ b_agents = means.shape[0]
1065
+ n1 = target_phi.shape[0]
1066
+ centered = target_phi - means[0]
1067
+ cov = centered.T @ centered / max(n1 - 1, 1)
1068
+ trace = torch.trace(cov)
1069
+ diffs = means[1:] - means[0]
1070
+ q_terms = torch.sqrt(torch.clamp(torch.einsum("bd,df,bf->b", diffs, cov, diffs), min=0.0) + 1e-12)
1071
+ p_terms = torch.linalg.vector_norm(diffs, dim=1)
1072
+ # Table 1 uses CQ^2=CP=log B: CQ=sqrt(log B), CP=log B.
1073
+ cq, cp = math.sqrt(math.log(b_agents)), math.log(b_agents)
1074
+ w = torch.full((b_agents,), 1.0 / b_agents, device=means.device, dtype=means.dtype)
1075
+ for _ in range(steps):
1076
+ w.requires_grad_(True)
1077
+ mixture = w @ means
1078
+ loss = torch.sum((mixture - means[0]) ** 2) + 2.0 * w[0] * trace / n1
1079
+ loss = loss + cq / math.sqrt(n1) * torch.dot(w[1:], q_terms)
1080
+ loss = loss + cp / n1 * torch.dot(w[1:], p_terms)
1081
+ (grad,) = torch.autograd.grad(loss, w)
1082
+ with torch.no_grad():
1083
+ # Euclidean projection on the simplex, Duchi et al. (2008).
1084
+ v = w - 0.04 * grad
1085
+ u, _ = torch.sort(v, descending=True)
1086
+ cssv = torch.cumsum(u, 0) - 1
1087
+ ind = torch.arange(1, b_agents + 1, device=v.device, dtype=v.dtype)
1088
+ rho = torch.nonzero(u - cssv / ind > 0, as_tuple=False)[-1, 0]
1089
+ theta = cssv[rho] / (rho + 1)
1090
+ w = torch.clamp(v - theta, min=0)
1091
+ return w.detach()
1092
+
1093
+
1094
+ def weighted_linear_fit(xs: torch.Tensor, ys: torch.Tensor, weights: torch.Tensor, ridge: float = 1e-3) -> torch.Tensor:
1095
+ """Exact weighted empirical-risk minimizer for linear regression."""
1096
+ b_agents, n, dim = xs.shape
1097
+ x = torch.cat([xs, torch.ones((b_agents, n, 1), device=xs.device, dtype=xs.dtype)], dim=-1)
1098
+ sample_w = (weights[:, None] / n).expand(b_agents, n).reshape(-1)
1099
+ flat_x, flat_y = x.reshape(-1, dim + 1), ys.reshape(-1)
1100
+ lhs = flat_x.T @ (flat_x * sample_w[:, None]) + ridge * torch.eye(dim + 1, device=xs.device)
1101
+ rhs = flat_x.T @ (flat_y * sample_w)
1102
+ return torch.linalg.solve(lhs, rhs)
1103
+
1104
+
1105
+ def concept_trial(seed: int, sigma2: float, d_rff: int, device: torch.device) -> dict:
1106
+ g = torch.Generator(device=device).manual_seed(seed)
1107
+ b_agents, n, dim, n_test = 100, 10, 20, 1000
1108
+ signs = torch.where(torch.rand((b_agents,), generator=g, device=device) > .5, 1., -1.)
1109
+ beta0 = torch.randn((dim,), generator=g, device=device)
1110
+ eps = torch.randn((b_agents, dim), generator=g, device=device)
1111
+ betas = signs[:, None] * math.sqrt(1 - sigma2) * beta0 + math.sqrt(sigma2) * eps
1112
+ x = 1.0 + torch.randn((b_agents, n, dim), generator=g, device=device)
1113
+ y = torch.einsum("bnd,bd->bn", x, betas) + math.sqrt(2.0) * torch.randn((b_agents, n), generator=g, device=device)
1114
+ x_test = 1.0 + torch.randn((n_test, dim), generator=g, device=device)
1115
+ y_test = x_test @ betas[0] + math.sqrt(2.0) * torch.randn((n_test,), generator=g, device=device)
1116
+ # The paper's weighted Gaussian KME kernel gives X a 1/sqrt(d) scaling.
1117
+ z = torch.cat([x, y[..., None]], dim=-1).reshape(-1, dim + 1)
1118
+ scale = torch.cat([torch.full((dim,), 1 / math.sqrt(dim), device=device), torch.ones((1,), device=device)])
1119
+ phi = rff(z, d_rff, g, scale).reshape(b_agents, n, d_rff)
1120
+ means = phi.mean(1)
1121
+ q_w = simplex_weights(means, phi[0])
1122
+ local = torch.zeros(b_agents, device=device); local[0] = 1
1123
+ grand = torch.full((b_agents,), 1 / b_agents, device=device)
1124
+ oracle = (signs == signs[0]).float(); oracle /= oracle.sum()
1125
+ result = {"sigma2": sigma2, "seed": seed, "q_local_weight": float(q_w[0]),
1126
+ "q_same_group_weight": float(q_w[signs == signs[0]].sum())}
1127
+ for name, weights in [("local", local), ("grandmean", grand), ("oracle", oracle), ("qaggregation", q_w)]:
1128
+ beta = weighted_linear_fit(x, y, weights)
1129
+ pred = x_test @ beta[:-1] + beta[-1]
1130
+ result[f"{name}_mse"] = float(torch.mean((pred - y_test) ** 2))
1131
+ return result
1132
+
1133
+
1134
+ def run_concept(args: argparse.Namespace) -> list[dict]:
1135
+ device = device_from_arg(args.device)
1136
+ rows = []
1137
+ for sigma2 in args.sigmas:
1138
+ for rep in range(args.repeats):
1139
+ rows.append(concept_trial(args.seed + rep * 101 + int(sigma2 * 1000), sigma2, args.rff_dim, device))
1140
+ return rows
1141
+
1142
+
1143
+ def exact_rbf_mmd(x: torch.Tensor, y: torch.Tensor, bandwidth: float = 1.0) -> torch.Tensor:
1144
+ def k(a: torch.Tensor, b: torch.Tensor) -> torch.Tensor:
1145
+ return torch.exp(-torch.cdist(a, b).square() / (2 * bandwidth * bandwidth))
1146
+ return torch.sqrt(torch.clamp(k(x, x).mean() + k(y, y).mean() - 2 * k(x, y).mean(), min=0.0))
1147
+
1148
+
1149
+ def run_rff(args: argparse.Namespace) -> list[dict]:
1150
+ device = device_from_arg(args.device)
1151
+ g = torch.Generator(device=device).manual_seed(args.seed)
1152
+ x = torch.randn((300, 8), generator=g, device=device)
1153
+ y = torch.randn((300, 8), generator=g, device=device) + torch.tensor([0.35] + [0.] * 7, device=device)
1154
+ exact = float(exact_rbf_mmd(x, y))
1155
+ rows = []
1156
+ for d_rff in args.rff_grid:
1157
+ errors = []
1158
+ for trial in range(args.repeats):
1159
+ tg = torch.Generator(device=device).manual_seed(args.seed + 1000 * d_rff + trial)
1160
+ phi = rff(torch.cat([x, y]), d_rff, tg, torch.ones(8, device=device))
1161
+ approx = torch.linalg.vector_norm(phi[:len(x)].mean(0) - phi[len(x):].mean(0)).item()
1162
+ errors.append(abs(approx - exact))
1163
+ rows.append({"rff_dim": d_rff, "exact_mmd": exact, "mean_abs_error": float(np.mean(errors)),
1164
+ "std_abs_error": float(np.std(errors)), "sqrt_logB_over_D": math.sqrt(math.log(100) / d_rff),
1165
+ "one_shot_floats_per_agent": d_rff})
1166
+ return rows
1167
+
1168
+
1169
+ def run_femnist(args: argparse.Namespace) -> list[dict]:
1170
+ """Actual FEMNIST writer data, deliberately scaled in samples/features/epochs."""
1171
+ from datasets import load_dataset
1172
+ device = device_from_arg(args.device)
1173
+ local_parquet = ROOT / "data" / "femnist" / "data" / "train-00000-of-00001.parquet"
1174
+ dataset = load_dataset("parquet", data_files=str(local_parquet), split="train", streaming=True) if local_parquet.exists() else load_dataset("flwrlabs/femnist", split="train", streaming=True)
1175
+ by_writer: dict[str, list[tuple[np.ndarray, int]]] = {}
1176
+ want = args.agents
1177
+ for ex in dataset:
1178
+ writer = ex["writer_id"]
1179
+ if writer not in by_writer and len(by_writer) >= want:
1180
+ continue
1181
+ if writer in by_writer or len(by_writer) < want:
1182
+ by_writer.setdefault(writer, []).append((np.asarray(ex["image"], dtype=np.float32) / 255., int(ex["character"])))
1183
+ if len(by_writer) >= want and all(len(v) >= args.train_per_agent + args.test_per_agent for v in by_writer.values()):
1184
+ break
1185
+ eligible = [v for v in by_writer.values() if len(v) >= args.train_per_agent + args.test_per_agent][:want]
1186
+ if len(eligible) < want:
1187
+ raise RuntimeError(f"Only found {len(eligible)} eligible writers; need {want}")
1188
+ x_train = torch.tensor(np.stack([np.stack([a for a, _ in v[:args.train_per_agent]]) for v in eligible]), device=device)
1189
+ y_train = torch.tensor(np.stack([[b for _, b in v[:args.train_per_agent]] for v in eligible]), device=device, dtype=torch.long)
1190
+ x_test = torch.tensor(np.stack([np.stack([a for a, _ in v[args.train_per_agent:args.train_per_agent + args.test_per_agent]]) for v in eligible]), device=device)
1191
+ y_test = torch.tensor(np.stack([[b for _, b in v[args.train_per_agent:args.train_per_agent + args.test_per_agent]] for v in eligible]), device=device, dtype=torch.long)
1192
+ b_agents, n, raw_dim = x_train.shape[0], x_train.shape[1], 784
1193
+ flat = x_train.reshape(b_agents, n, raw_dim)
1194
+ # Fixed random projection reduces the RFF input cost while retaining a reproducible vision proxy.
1195
+ g = torch.Generator(device=device).manual_seed(args.seed)
1196
+ proj = torch.randn((raw_dim, args.feature_dim), generator=g, device=device) / math.sqrt(raw_dim)
1197
+ features = flat @ proj
1198
+ phi_all = rff(features.reshape(-1, args.feature_dim), args.rff_dim, g, torch.ones(args.feature_dim, device=device)).reshape(b_agents, n, args.rff_dim)
1199
+ means = phi_all.mean(1)
1200
+ # Fit one regularized softmax classifier for each target using its personalized weights.
1201
+ rows = []
1202
+ all_x = features.reshape(-1, args.feature_dim)
1203
+ all_y = y_train.reshape(-1)
1204
+ for target in range(b_agents):
1205
+ weights = simplex_weights(means, phi_all[target], steps=300)
1206
+ schemes = {"local": torch.nn.functional.one_hot(torch.tensor(target, device=device), b_agents).float(),
1207
+ "grandmean": torch.full((b_agents,), 1 / b_agents, device=device), "qaggregation": weights}
1208
+ test_feat = x_test[target].reshape(args.test_per_agent, raw_dim) @ proj
1209
+ for name, w in schemes.items():
1210
+ model = torch.zeros((args.feature_dim, 62), device=device, requires_grad=True)
1211
+ opt = torch.optim.Adam([model], lr=args.lr, weight_decay=1e-4)
1212
+ sample_w = (w[:, None] / n).expand(b_agents, n).reshape(-1)
1213
+ for _ in range(args.epochs):
1214
+ opt.zero_grad()
1215
+ losses = torch.nn.functional.cross_entropy(all_x @ model, all_y, reduction="none")
1216
+ (losses * sample_w).sum().backward(); opt.step()
1217
+ acc = ((test_feat @ model).argmax(1) == y_test[target]).float().mean().item()
1218
+ rows.append({"agent": target, "method": name, "accuracy": acc, "q_local_weight": float(weights[target])})
1219
+ return rows
1220
+
1221
+
1222
+ def write_rows(rows: list[dict], out: Path) -> None:
1223
+ out.parent.mkdir(parents=True, exist_ok=True)
1224
+ with out.open("w", newline="") as f:
1225
+ writer = csv.DictWriter(f, fieldnames=list(rows[0]))
1226
+ writer.writeheader(); writer.writerows(rows)
1227
+ out.with_suffix(".json").write_text(json.dumps(rows, indent=2))
1228
+
1229
+
1230
+ def main() -> None:
1231
+ p = argparse.ArgumentParser()
1232
+ p.add_argument("mode", choices=["concept", "rff", "femnist"])
1233
+ p.add_argument("--output", type=Path, required=True)
1234
+ p.add_argument("--device", default="auto")
1235
+ p.add_argument("--seed", type=int, default=20260719)
1236
+ p.add_argument("--repeats", type=int, default=30)
1237
+ p.add_argument("--rff-dim", type=int, default=500)
1238
+ p.add_argument("--sigmas", type=float, nargs="+", default=[0., .1, .25, .5, .75, 1.])
1239
+ p.add_argument("--rff-grid", type=int, nargs="+", default=[16, 32, 64, 128, 256, 512, 1024])
1240
+ p.add_argument("--agents", type=int, default=192)
1241
+ p.add_argument("--train-per-agent", type=int, default=32)
1242
+ p.add_argument("--test-per-agent", type=int, default=16)
1243
+ p.add_argument("--feature-dim", type=int, default=64)
1244
+ p.add_argument("--epochs", type=int, default=20)
1245
+ p.add_argument("--lr", type=float, default=.08)
1246
+ args = p.parse_args()
1247
+ started = perf_counter()
1248
+ rows = {"concept": run_concept, "rff": run_rff, "femnist": run_femnist}[args.mode](args)
1249
+ write_rows(rows, args.output)
1250
+ if args.mode == "concept":
1251
+ summary = {str(s): {k: float(np.mean([r[k] for r in rows if r["sigma2"] == s])) for k in ["local_mse", "grandmean_mse", "oracle_mse", "qaggregation_mse", "q_local_weight", "q_same_group_weight"]} for s in args.sigmas}
1252
+ print(json.dumps({"concept_summary": summary}, indent=2))
1253
+ elif args.mode == "rff":
1254
+ print(json.dumps({"rff_summary": {str(r["rff_dim"]): r["mean_abs_error"] for r in rows}}, indent=2))
1255
+ elif args.mode == "femnist":
1256
+ summary = {m: float(np.mean([r["accuracy"] for r in rows if r["method"] == m])) for m in ["local", "grandmean", "qaggregation"]}
1257
+ print(json.dumps({"femnist_accuracy": summary}, indent=2))
1258
+ print(json.dumps({"mode": args.mode, "rows": len(rows), "output": str(args.output), "seconds": round(perf_counter() - started, 2), "device": str(device_from_arg(args.device))}, indent=2))
1259
+
1260
+
1261
+ if __name__ == "__main__":
1262
+ main()
1263
+
1264
+ ````
1265
+
1266
+
1267
+ ````output
1268
+ usage: reproduce.py [-h] --output OUTPUT [--device DEVICE] [--seed SEED]
1269
+ [--repeats REPEATS] [--rff-dim RFF_DIM]
1270
+ [--sigmas SIGMAS [SIGMAS ...]]
1271
+ [--rff-grid RFF_GRID [RFF_GRID ...]] [--agents AGENTS]
1272
+ [--train-per-agent TRAIN_PER_AGENT]
1273
+ [--test-per-agent TEST_PER_AGENT]
1274
+ [--feature-dim FEATURE_DIM] [--epochs EPOCHS] [--lr LR]
1275
+ {concept,rff,femnist}
1276
+
1277
+ positional arguments:
1278
+ {concept,rff,femnist}
1279
+
1280
+ options:
1281
+ -h, --help show this help message and exit
1282
+ --output OUTPUT
1283
+ --device DEVICE
1284
+ --seed SEED
1285
+ --repeats REPEATS
1286
+ --rff-dim RFF_DIM
1287
+ --sigmas SIGMAS [SIGMAS ...]
1288
+ --rff-grid RFF_GRID [RFF_GRID ...]
1289
+ --agents AGENTS
1290
+ --train-per-agent TRAIN_PER_AGENT
1291
+ --test-per-agent TEST_PER_AGENT
1292
+ --feature-dim FEATURE_DIM
1293
+ --epochs EPOCHS
1294
+ --lr LR
1295
+
1296
+ ````
1297
+
1298
+
1299
+ ---
1300
+ <!-- trackio-cell
1301
+ {"type": "markdown", "id": "cell_8356b325c920", "created_at": "2026-07-19T09:54:23+00:00", "title": "Evidence and status"}
1302
+ -->
1303
+ Independent CUDA reproduction used the paper’s , , , noise variance 2, , and 100 repetitions ([paper](https://arxiv.org/pdf/2603.02233), Sec. 6.1/Table 1). It uses an independent projected-gradient implementation of Algorithm 1 and exact weighted linear-risk minimization. Contrary to the paper’s transition claim, its Q weights remained almost local (mean target weight 0.977–0.990) and Q MSE was worse than Local at all tested (e.g. 17.84 vs 15.75 at 0; 25.43 vs 14.46 at .5). **Status: not reproduced by this independent implementation.** The full configuration, CUDA run, and CSV are captured here; no official code was found.
1304
+
1305
+
1306
+ ---
1307
+ <!-- trackio-cell
1308
+ {"type": "markdown", "id": "cell_4a35924ff734", "created_at": "2026-07-19T09:57:06+00:00", "title": "Clarified evidence"}
1309
+ -->
1310
+ Clarification of measured scope: B=100, n=10/client, d=20, noise variance 2, D=500, 100 repetitions, CUDA. Q target weight was 0.977 to 0.990 across sigma_c squared values and Q MSE was worse than local at every tested value. This independent literal implementation does not reproduce the reported transition; it is not presented as a refutation because official code was unavailable.
pages/claim-6-section-6-3-femnist-192-agent-benchmark/page.md ADDED
@@ -0,0 +1,298 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Claim 6: Section 6.3 FEMNIST 192-agent benchmark
2
+
3
+
4
+ ---
5
+ <!-- trackio-cell
6
+ {"type": "markdown", "id": "cell_80ab313b85c0", "created_at": "2026-07-19T09:32:03+00:00", "title": "Claim 6: Section 6.3 FEMNIST 192-agent benchmark"}
7
+ -->
8
+ Document setup, runs, and results for **Claim 6: Section 6.3 FEMNIST 192-agent benchmark**.
9
+
10
+
11
+ ---
12
+ <!-- trackio-cell
13
+ {"type": "code", "id": "cell_5675f10a2abb", "created_at": "2026-07-19T09:52:42+00:00", "title": "Run: env reproduce.py (exit 0)", "command": ["env", "CUDA_VISIBLE_DEVICES=1", "/home/vim/miniconda3/envs/pytorch-gpu-hf/bin/python", "src/reproduce.py", "femnist", "--agents", "192", "--train-per-agent", "32", "--test-per-agent", "16", "--feature-dim", "64", "--rff-dim", "1000", "--epochs", "20", "--lr", "0.08", "--output", "outputs/femnist_192agents_scaled.csv", "--device", "cuda"], "exit_code": 0, "duration_s": 33.426}
14
+ -->
15
+ ````bash
16
+ $ env CUDA_VISIBLE_DEVICES=1 /home/vim/miniconda3/envs/pytorch-gpu-hf/bin/python src/reproduce.py femnist --agents 192 --train-per-agent 32 --test-per-agent 16 --feature-dim 64 --rff-dim 1000 --epochs 20 --lr 0.08 --output outputs/femnist_192agents_scaled.csv --device cuda
17
+ ````
18
+
19
+ exit 0 · 33.4s
20
+
21
+
22
+ ````python title=reproduce.py
23
+ #!/usr/bin/env python3
24
+ """Independent, GPU-capable checks for Fermanian et al. (2026), arXiv:2603.02233.
25
+
26
+ The implementation follows Algorithms 1--2 at a scaled experimental budget. It
27
+ uses RFF KMEs to choose a target client's collaboration weights and fits the
28
+ weighted empirical-risk minimizer exactly for linear regression. The FEMNIST
29
+ mode uses the public natural writer partition and a linear softmax model.
30
+ """
31
+ from __future__ import annotations
32
+
33
+ import argparse
34
+ import csv
35
+ import json
36
+ import math
37
+ from pathlib import Path
38
+ from time import perf_counter
39
+
40
+ import numpy as np
41
+ import torch
42
+
43
+
44
+ ROOT = Path(__file__).resolve().parents[1]
45
+
46
+
47
+ def device_from_arg(name: str) -> torch.device:
48
+ if name == "auto":
49
+ return torch.device("cuda" if torch.cuda.is_available() else "cpu")
50
+ return torch.device(name)
51
+
52
+
53
+ def rff(z: torch.Tensor, d_rff: int, generator: torch.Generator, scale: torch.Tensor) -> torch.Tensor:
54
+ """Gaussian-kernel RFF, with a supplied coordinate rescaling."""
55
+ w = torch.randn((z.shape[1], d_rff), generator=generator, device=z.device, dtype=z.dtype)
56
+ w = w * scale[:, None]
57
+ b = 2 * math.pi * torch.rand((d_rff,), generator=generator, device=z.device, dtype=z.dtype)
58
+ return math.sqrt(2.0 / d_rff) * torch.cos(z @ w + b)
59
+
60
+
61
+ def simplex_weights(means: torch.Tensor, target_phi: torch.Tensor, steps: int = 900) -> torch.Tensor:
62
+ """Solve Algorithm 1's empirical Q-aggregation objective by projected GD."""
63
+ b_agents = means.shape[0]
64
+ n1 = target_phi.shape[0]
65
+ centered = target_phi - means[0]
66
+ cov = centered.T @ centered / max(n1 - 1, 1)
67
+ trace = torch.trace(cov)
68
+ diffs = means[1:] - means[0]
69
+ q_terms = torch.sqrt(torch.clamp(torch.einsum("bd,df,bf->b", diffs, cov, diffs), min=0.0) + 1e-12)
70
+ p_terms = torch.linalg.vector_norm(diffs, dim=1)
71
+ # Table 1 uses CQ^2=CP=log B: CQ=sqrt(log B), CP=log B.
72
+ cq, cp = math.sqrt(math.log(b_agents)), math.log(b_agents)
73
+ w = torch.full((b_agents,), 1.0 / b_agents, device=means.device, dtype=means.dtype)
74
+ for _ in range(steps):
75
+ w.requires_grad_(True)
76
+ mixture = w @ means
77
+ loss = torch.sum((mixture - means[0]) ** 2) + 2.0 * w[0] * trace / n1
78
+ loss = loss + cq / math.sqrt(n1) * torch.dot(w[1:], q_terms)
79
+ loss = loss + cp / n1 * torch.dot(w[1:], p_terms)
80
+ (grad,) = torch.autograd.grad(loss, w)
81
+ with torch.no_grad():
82
+ # Euclidean projection on the simplex, Duchi et al. (2008).
83
+ v = w - 0.04 * grad
84
+ u, _ = torch.sort(v, descending=True)
85
+ cssv = torch.cumsum(u, 0) - 1
86
+ ind = torch.arange(1, b_agents + 1, device=v.device, dtype=v.dtype)
87
+ rho = torch.nonzero(u - cssv / ind > 0, as_tuple=False)[-1, 0]
88
+ theta = cssv[rho] / (rho + 1)
89
+ w = torch.clamp(v - theta, min=0)
90
+ return w.detach()
91
+
92
+
93
+ def weighted_linear_fit(xs: torch.Tensor, ys: torch.Tensor, weights: torch.Tensor, ridge: float = 1e-3) -> torch.Tensor:
94
+ """Exact weighted empirical-risk minimizer for linear regression."""
95
+ b_agents, n, dim = xs.shape
96
+ x = torch.cat([xs, torch.ones((b_agents, n, 1), device=xs.device, dtype=xs.dtype)], dim=-1)
97
+ sample_w = (weights[:, None] / n).expand(b_agents, n).reshape(-1)
98
+ flat_x, flat_y = x.reshape(-1, dim + 1), ys.reshape(-1)
99
+ lhs = flat_x.T @ (flat_x * sample_w[:, None]) + ridge * torch.eye(dim + 1, device=xs.device)
100
+ rhs = flat_x.T @ (flat_y * sample_w)
101
+ return torch.linalg.solve(lhs, rhs)
102
+
103
+
104
+ def concept_trial(seed: int, sigma2: float, d_rff: int, device: torch.device) -> dict:
105
+ g = torch.Generator(device=device).manual_seed(seed)
106
+ b_agents, n, dim, n_test = 100, 10, 20, 1000
107
+ signs = torch.where(torch.rand((b_agents,), generator=g, device=device) > .5, 1., -1.)
108
+ beta0 = torch.randn((dim,), generator=g, device=device)
109
+ eps = torch.randn((b_agents, dim), generator=g, device=device)
110
+ betas = signs[:, None] * math.sqrt(1 - sigma2) * beta0 + math.sqrt(sigma2) * eps
111
+ x = 1.0 + torch.randn((b_agents, n, dim), generator=g, device=device)
112
+ y = torch.einsum("bnd,bd->bn", x, betas) + math.sqrt(2.0) * torch.randn((b_agents, n), generator=g, device=device)
113
+ x_test = 1.0 + torch.randn((n_test, dim), generator=g, device=device)
114
+ y_test = x_test @ betas[0] + math.sqrt(2.0) * torch.randn((n_test,), generator=g, device=device)
115
+ # The paper's weighted Gaussian KME kernel gives X a 1/sqrt(d) scaling.
116
+ z = torch.cat([x, y[..., None]], dim=-1).reshape(-1, dim + 1)
117
+ scale = torch.cat([torch.full((dim,), 1 / math.sqrt(dim), device=device), torch.ones((1,), device=device)])
118
+ phi = rff(z, d_rff, g, scale).reshape(b_agents, n, d_rff)
119
+ means = phi.mean(1)
120
+ q_w = simplex_weights(means, phi[0])
121
+ local = torch.zeros(b_agents, device=device); local[0] = 1
122
+ grand = torch.full((b_agents,), 1 / b_agents, device=device)
123
+ oracle = (signs == signs[0]).float(); oracle /= oracle.sum()
124
+ result = {"sigma2": sigma2, "seed": seed, "q_local_weight": float(q_w[0]),
125
+ "q_same_group_weight": float(q_w[signs == signs[0]].sum())}
126
+ for name, weights in [("local", local), ("grandmean", grand), ("oracle", oracle), ("qaggregation", q_w)]:
127
+ beta = weighted_linear_fit(x, y, weights)
128
+ pred = x_test @ beta[:-1] + beta[-1]
129
+ result[f"{name}_mse"] = float(torch.mean((pred - y_test) ** 2))
130
+ return result
131
+
132
+
133
+ def run_concept(args: argparse.Namespace) -> list[dict]:
134
+ device = device_from_arg(args.device)
135
+ rows = []
136
+ for sigma2 in args.sigmas:
137
+ for rep in range(args.repeats):
138
+ rows.append(concept_trial(args.seed + rep * 101 + int(sigma2 * 1000), sigma2, args.rff_dim, device))
139
+ return rows
140
+
141
+
142
+ def exact_rbf_mmd(x: torch.Tensor, y: torch.Tensor, bandwidth: float = 1.0) -> torch.Tensor:
143
+ def k(a: torch.Tensor, b: torch.Tensor) -> torch.Tensor:
144
+ return torch.exp(-torch.cdist(a, b).square() / (2 * bandwidth * bandwidth))
145
+ return torch.sqrt(torch.clamp(k(x, x).mean() + k(y, y).mean() - 2 * k(x, y).mean(), min=0.0))
146
+
147
+
148
+ def run_rff(args: argparse.Namespace) -> list[dict]:
149
+ device = device_from_arg(args.device)
150
+ g = torch.Generator(device=device).manual_seed(args.seed)
151
+ x = torch.randn((300, 8), generator=g, device=device)
152
+ y = torch.randn((300, 8), generator=g, device=device) + torch.tensor([0.35] + [0.] * 7, device=device)
153
+ exact = float(exact_rbf_mmd(x, y))
154
+ rows = []
155
+ for d_rff in args.rff_grid:
156
+ errors = []
157
+ for trial in range(args.repeats):
158
+ tg = torch.Generator(device=device).manual_seed(args.seed + 1000 * d_rff + trial)
159
+ phi = rff(torch.cat([x, y]), d_rff, tg, torch.ones(8, device=device))
160
+ approx = torch.linalg.vector_norm(phi[:len(x)].mean(0) - phi[len(x):].mean(0)).item()
161
+ errors.append(abs(approx - exact))
162
+ rows.append({"rff_dim": d_rff, "exact_mmd": exact, "mean_abs_error": float(np.mean(errors)),
163
+ "std_abs_error": float(np.std(errors)), "sqrt_logB_over_D": math.sqrt(math.log(100) / d_rff),
164
+ "one_shot_floats_per_agent": d_rff})
165
+ return rows
166
+
167
+
168
+ def run_femnist(args: argparse.Namespace) -> list[dict]:
169
+ """Actual FEMNIST writer data, deliberately scaled in samples/features/epochs."""
170
+ from datasets import load_dataset
171
+ device = device_from_arg(args.device)
172
+ local_parquet = ROOT / "data" / "femnist" / "data" / "train-00000-of-00001.parquet"
173
+ dataset = load_dataset("parquet", data_files=str(local_parquet), split="train", streaming=True) if local_parquet.exists() else load_dataset("flwrlabs/femnist", split="train", streaming=True)
174
+ by_writer: dict[str, list[tuple[np.ndarray, int]]] = {}
175
+ want = args.agents
176
+ for ex in dataset:
177
+ writer = ex["writer_id"]
178
+ if writer not in by_writer and len(by_writer) >= want:
179
+ continue
180
+ if writer in by_writer or len(by_writer) < want:
181
+ by_writer.setdefault(writer, []).append((np.asarray(ex["image"], dtype=np.float32) / 255., int(ex["character"])))
182
+ if len(by_writer) >= want and all(len(v) >= args.train_per_agent + args.test_per_agent for v in by_writer.values()):
183
+ break
184
+ eligible = [v for v in by_writer.values() if len(v) >= args.train_per_agent + args.test_per_agent][:want]
185
+ if len(eligible) < want:
186
+ raise RuntimeError(f"Only found {len(eligible)} eligible writers; need {want}")
187
+ x_train = torch.tensor(np.stack([np.stack([a for a, _ in v[:args.train_per_agent]]) for v in eligible]), device=device)
188
+ y_train = torch.tensor(np.stack([[b for _, b in v[:args.train_per_agent]] for v in eligible]), device=device, dtype=torch.long)
189
+ x_test = torch.tensor(np.stack([np.stack([a for a, _ in v[args.train_per_agent:args.train_per_agent + args.test_per_agent]]) for v in eligible]), device=device)
190
+ y_test = torch.tensor(np.stack([[b for _, b in v[args.train_per_agent:args.train_per_agent + args.test_per_agent]] for v in eligible]), device=device, dtype=torch.long)
191
+ b_agents, n, raw_dim = x_train.shape[0], x_train.shape[1], 784
192
+ flat = x_train.reshape(b_agents, n, raw_dim)
193
+ # Fixed random projection reduces the RFF input cost while retaining a reproducible vision proxy.
194
+ g = torch.Generator(device=device).manual_seed(args.seed)
195
+ proj = torch.randn((raw_dim, args.feature_dim), generator=g, device=device) / math.sqrt(raw_dim)
196
+ features = flat @ proj
197
+ phi_all = rff(features.reshape(-1, args.feature_dim), args.rff_dim, g, torch.ones(args.feature_dim, device=device)).reshape(b_agents, n, args.rff_dim)
198
+ means = phi_all.mean(1)
199
+ # Fit one regularized softmax classifier for each target using its personalized weights.
200
+ rows = []
201
+ all_x = features.reshape(-1, args.feature_dim)
202
+ all_y = y_train.reshape(-1)
203
+ for target in range(b_agents):
204
+ weights = simplex_weights(means, phi_all[target], steps=300)
205
+ schemes = {"local": torch.nn.functional.one_hot(torch.tensor(target, device=device), b_agents).float(),
206
+ "grandmean": torch.full((b_agents,), 1 / b_agents, device=device), "qaggregation": weights}
207
+ test_feat = x_test[target].reshape(args.test_per_agent, raw_dim) @ proj
208
+ for name, w in schemes.items():
209
+ model = torch.zeros((args.feature_dim, 62), device=device, requires_grad=True)
210
+ opt = torch.optim.Adam([model], lr=args.lr, weight_decay=1e-4)
211
+ sample_w = (w[:, None] / n).expand(b_agents, n).reshape(-1)
212
+ for _ in range(args.epochs):
213
+ opt.zero_grad()
214
+ losses = torch.nn.functional.cross_entropy(all_x @ model, all_y, reduction="none")
215
+ (losses * sample_w).sum().backward(); opt.step()
216
+ acc = ((test_feat @ model).argmax(1) == y_test[target]).float().mean().item()
217
+ rows.append({"agent": target, "method": name, "accuracy": acc, "q_local_weight": float(weights[target])})
218
+ return rows
219
+
220
+
221
+ def write_rows(rows: list[dict], out: Path) -> None:
222
+ out.parent.mkdir(parents=True, exist_ok=True)
223
+ with out.open("w", newline="") as f:
224
+ writer = csv.DictWriter(f, fieldnames=list(rows[0]))
225
+ writer.writeheader(); writer.writerows(rows)
226
+ out.with_suffix(".json").write_text(json.dumps(rows, indent=2))
227
+
228
+
229
+ def main() -> None:
230
+ p = argparse.ArgumentParser()
231
+ p.add_argument("mode", choices=["concept", "rff", "femnist"])
232
+ p.add_argument("--output", type=Path, required=True)
233
+ p.add_argument("--device", default="auto")
234
+ p.add_argument("--seed", type=int, default=20260719)
235
+ p.add_argument("--repeats", type=int, default=30)
236
+ p.add_argument("--rff-dim", type=int, default=500)
237
+ p.add_argument("--sigmas", type=float, nargs="+", default=[0., .1, .25, .5, .75, 1.])
238
+ p.add_argument("--rff-grid", type=int, nargs="+", default=[16, 32, 64, 128, 256, 512, 1024])
239
+ p.add_argument("--agents", type=int, default=192)
240
+ p.add_argument("--train-per-agent", type=int, default=32)
241
+ p.add_argument("--test-per-agent", type=int, default=16)
242
+ p.add_argument("--feature-dim", type=int, default=64)
243
+ p.add_argument("--epochs", type=int, default=20)
244
+ p.add_argument("--lr", type=float, default=.08)
245
+ args = p.parse_args()
246
+ started = perf_counter()
247
+ rows = {"concept": run_concept, "rff": run_rff, "femnist": run_femnist}[args.mode](args)
248
+ write_rows(rows, args.output)
249
+ if args.mode == "concept":
250
+ summary = {str(s): {k: float(np.mean([r[k] for r in rows if r["sigma2"] == s])) for k in ["local_mse", "grandmean_mse", "oracle_mse", "qaggregation_mse", "q_local_weight", "q_same_group_weight"]} for s in args.sigmas}
251
+ print(json.dumps({"concept_summary": summary}, indent=2))
252
+ elif args.mode == "rff":
253
+ print(json.dumps({"rff_summary": {str(r["rff_dim"]): r["mean_abs_error"] for r in rows}}, indent=2))
254
+ elif args.mode == "femnist":
255
+ summary = {m: float(np.mean([r["accuracy"] for r in rows if r["method"] == m])) for m in ["local", "grandmean", "qaggregation"]}
256
+ print(json.dumps({"femnist_accuracy": summary}, indent=2))
257
+ print(json.dumps({"mode": args.mode, "rows": len(rows), "output": str(args.output), "seconds": round(perf_counter() - started, 2), "device": str(device_from_arg(args.device))}, indent=2))
258
+
259
+
260
+ if __name__ == "__main__":
261
+ main()
262
+
263
+ ````
264
+
265
+
266
+ ````output
267
+ {
268
+ "femnist_accuracy": {
269
+ "local": 0.2503255208333333,
270
+ "grandmean": 0.24283854166666666,
271
+ "qaggregation": 0.17578125
272
+ }
273
+ }
274
+ {
275
+ "mode": "femnist",
276
+ "rows": 576,
277
+ "output": "outputs/femnist_192agents_scaled.csv",
278
+ "seconds": 32.19,
279
+ "device": "cuda"
280
+ }
281
+
282
+ ````
283
+
284
+
285
+ ---
286
+ <!-- trackio-cell
287
+ {"type": "artifact", "id": "cell_a404932005b6", "created_at": "2026-07-19T09:52:42+00:00", "title": "Artifact: femnist_192agents_scaled.csv", "path": "outputs/femnist_192agents_scaled.csv", "size": 14226, "artifact_type": "dataset", "auto": true}
288
+ -->
289
+ **�� Artifact** `outputs/femnist_192agents_scaled.csv` · dataset · 14.2 kB
290
+
291
+ https://huggingface.co/buckets/vimarsh/repro-adaptive-personalized-fl-multi-task-averaging-kernel-mean-embedding-artifacts#logbook-files/outputs/femnist_192agents_scaled.csv
292
+
293
+
294
+ ---
295
+ <!-- trackio-cell
296
+ {"type": "markdown", "id": "cell_87c9b71c7614", "created_at": "2026-07-19T09:57:23+00:00", "title": "Clarified evidence"}
297
+ -->
298
+ Actual public dataset: https://huggingface.co/datasets/flwrlabs/femnist . Scope: 192 natural writer partitions, D=1000, CUDA; scaled to 32 train and 16 test examples/writer, 64 projected features, and 20 linear-softmax epochs. Mean accuracy: local 0.2503, grand mean 0.2428, Q aggregation 0.1758. The full paper MLP uses 2000 epochs, so this is a non-reproducing scaled check, not a full replication.
pages/conclusion/page.md ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Conclusion
2
+
3
+
4
+ ---
5
+ <!-- trackio-cell
6
+ {"type": "markdown", "id": "cell_5ef64984e3ab", "created_at": "2026-07-19T09:32:03+00:00", "title": "Reproduction bundle"}
7
+ -->
8
+ The reproduction bundle contains the independent GPU-capable implementation, CSV results, rendered Posterly artifact, gate report, and rerun instructions. It excludes local caches and the replaceable FEMNIST Parquet download.
9
+
10
+ After publication, download the bundle from the artifact cell, open `README.md`, and run its CUDA commands. Public data source: https://huggingface.co/datasets/flwrlabs/femnist .
11
+
12
+
13
+ ---
14
+ <!-- trackio-cell
15
+ {"type": "artifact", "id": "cell_175c3d4d70bd", "created_at": "2026-07-19T10:07:56+00:00", "title": "Reproduction bundle", "artifact": "repro-adaptive-personalized-federated-learning-via-multi-task-averaging-of-kernel-mean-embedding/repro-bundle:v0", "artifact_type": "dataset"}
16
+ -->
17
+ **📦 Artifact** `repro-adaptive-personalized-federated-learning-via-multi-task-averaging-of-kernel-mean-embedding/repro-bundle:v0` · dataset
18
+
19
+ https://huggingface.co/buckets/vimarsh/repro-adaptive-personalized-fl-multi-task-averaging-kernel-mean-embedding-artifacts#repro-adaptive-personalized-federated-learning-via-multi-task-averaging-of-kernel-mean-embedding/repro-bundle:v0
pages/executive-summary/page.md ADDED
The diff for this file is too large to render. See raw diff
 
pages/index.md ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Reproduction: Adaptive Personalized Federated Learning via Multi-task Averaging of Kernel Mean Embeddings
2
+
3
+ [OpenReview paper](https://openreview.net/forum?id=ADHGbxsREr)
4
+
5
+ ## Pages
6
+
7
+ | Page |
8
+ | --- |
9
+ | [Executive summary](#/executive-summary) |
10
+ | [Claim 1: Theorem 4.4 finite-sample KME guarantee](#/claim-1-theorem-4-4-finite-sample-kme-guarantee) |
11
+ | [Claim 2: Corollary 4.6 downstream excess-risk transfer](#/claim-2-corollary-4-6-downstream-excess-risk-transfer) |
12
+ | [Claim 3: Example 4.5 automatic adaptive collaboration](#/claim-3-example-4-5-automatic-adaptive-collaboration) |
13
+ | [Claim 4: Theorem 5.2 random Fourier-feature tradeoff](#/claim-4-theorem-5-2-random-fourier-feature-tradeoff) |
14
+ | [Claim 5: Section 6.1 100-agent synthetic concept shift](#/claim-5-section-6-1-100-agent-synthetic-concept-shift) |
15
+ | [Claim 6: Section 6.3 FEMNIST 192-agent benchmark](#/claim-6-section-6-3-femnist-192-agent-benchmark) |
16
+ | [Conclusion](#/conclusion) |
trackio-logo-light.png ADDED
trackio-logo.png ADDED
trackio-wordmark-dark.png ADDED
workspace.json ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "file_count": 0,
4
+ "total_size": 0,
5
+ "files": [],
6
+ "hub_refs": [
7
+ {
8
+ "url": "https://huggingface.co/jobs/vimarsh/6a5c9e7cbee6ee1cf4ed0df8",
9
+ "type": "Jobs",
10
+ "label": "vimarsh/6a5c9e7cbee6ee1cf4ed0df8"
11
+ },
12
+ {
13
+ "url": "https://huggingface.co/buckets/vimarsh/jobs-artifacts",
14
+ "type": "Buckets",
15
+ "label": "vimarsh/jobs-artifacts"
16
+ },
17
+ {
18
+ "url": "https://huggingface.co/buckets/vimarsh/repro-adaptive-personalized-fl-multi-task-averaging-kernel-mean-embedding-artifacts#logbook-files/outputs/rff_tradeoff_100seeds.csv",
19
+ "type": "Buckets",
20
+ "label": "vimarsh/repro-adaptive-personalized-fl-multi-task-averaging-kernel-mean-embedding-artifacts"
21
+ },
22
+ {
23
+ "url": "https://huggingface.co/datasets/flwrlabs/femnist",
24
+ "type": "Datasets",
25
+ "label": "flwrlabs/femnist"
26
+ }
27
+ ],
28
+ "reference_only": true
29
+ }