Spaces:
Running
Running
File size: 16,958 Bytes
f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 6075521 1261d17 ef9f423 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 f565769 1261d17 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 | import marimo
__generated_with = "0.18.4"
app = marimo.App(width="medium", app_title="Mhc Exploration")
with app.setup(hide_code=True):
import sys
import subprocess
import tempfile
from pathlib import Path
import numpy as np
import altair as alt
import pandas as pd
import marimo as mo
from mhc import run_comparison, sinkhorn_knopp, compute_all_metrics
@app.cell(hide_code=True)
def _():
mo.md(r"""
# Exploring Manifold-Constrained Hyper-Connections (mHC)
[](https://molab.marimo.io/github/bassrehab/mhc-visualizer/blob/main/notebook/mhc_exploration.py)
This interactive notebook demonstrates the key insight from DeepSeek's mHC paper:
**unconstrained residual mixing matrices cause signal explosion in deep networks,
while doubly stochastic constraints keep signals bounded.**
Use the controls below to explore how Sinkhorn iterations affect stability!
**Paper:** https://arxiv.org/abs/2512.24880
**Viz Author (under MIT):** Subhadip Mitra
**Implementation/ Viz Repository:** https://github.com/bassrehab/mhc-visualizer
""")
return
@app.cell(hide_code=True)
def _():
mo.md(r"""
## The Sinkhorn-Knopp Algorithm
The Sinkhorn-Knopp algorithm projects any positive matrix onto the set of **doubly stochastic matrices** - matrices where all rows and columns sum to 1.
### Why Doubly Stochastic?
Doubly stochastic matrices have a crucial property: **they are closed under multiplication**. This means:
- If A and B are doubly stochastic, then A @ B is also doubly stochastic
- The spectral norm is bounded: ||A|| <= 1
- Signal propagation stays bounded even through many layers
### The Algorithm
Starting from any positive matrix, we alternate between:
1. Normalizing rows to sum to 1
2. Normalizing columns to sum to 1
This converges to a doubly stochastic matrix!
""")
return
@app.cell(hide_code=True)
def _():
mo.md(r"""
## Interactive Controls
Adjust the parameters below to see how they affect signal propagation stability.
""")
return
@app.cell(hide_code=True)
def _(active_config):
# Sliders initialized from active config (re-created when preset clicked)
sinkhorn_slider = mo.ui.slider(
start=0,
stop=30,
value=active_config["k"],
step=1,
label="Sinkhorn Iterations (k)",
show_value=True,
)
depth_slider = mo.ui.slider(
start=10,
stop=200,
value=active_config["depth"],
step=10,
label="Network Depth",
show_value=True,
)
n_dropdown = mo.ui.dropdown(
options={"2": 2, "4": 4, "8": 8},
value=active_config["n"],
label="Number of Streams (n)",
)
seed_input = mo.ui.number(
value=active_config["seed"], start=0, stop=10000, label="Random Seed"
)
controls = mo.hstack(
[sinkhorn_slider, depth_slider, n_dropdown, seed_input], justify="start", gap=2
)
controls
return depth_slider, n_dropdown, seed_input, sinkhorn_slider
@app.cell(hide_code=True)
def _():
# Preset buttons - clicking triggers re-run with new defaults
preset_default = mo.ui.run_button(label="Default")
preset_explosion = mo.ui.run_button(label="HC Explosion (k=0)")
preset_minimal = mo.ui.run_button(label="Minimal Projection (k=5)")
preset_deep = mo.ui.run_button(label="Deep Network (200)")
randomize_btn = mo.ui.run_button(label="Randomize Seed")
presets = mo.hstack(
[preset_default, preset_explosion, preset_minimal, preset_deep, randomize_btn],
justify="start",
gap=1,
)
presets
return preset_deep, preset_explosion, preset_minimal, randomize_btn
@app.cell
def _(depth_slider, n_dropdown, seed_input, sinkhorn_slider):
# Run the simulation with current parameters
results = run_comparison(
depth=depth_slider.value,
n=int(n_dropdown.value),
sinkhorn_iters=sinkhorn_slider.value,
seed=seed_input.value,
)
return (results,)
@app.cell(hide_code=True)
def _():
mo.md(r"""
## Signal Propagation: The Explosion Problem
The real issue isn't single-layer behavior - it's what happens when we **compose many layers**.
In a deep network, the effective transformation is:
$$H_{composite} = H_L \cdot H_{L-1} \cdot ... \cdot H_1$$
Watch the chart below: **HC (red) explodes exponentially, while mHC (blue) stays bounded!**
""")
return
@app.cell
def _(base_chart):
mo.ui.altair_chart(base_chart)
return
@app.cell(hide_code=True)
def _():
mo.md(r"""
## Stability Metrics
The table below shows metrics at the selected layer:
- **Forward Gain**: Maximum row sum (worst-case signal amplification)
- **Backward Gain**: Maximum column sum (gradient flow)
- **Spectral Norm**: Largest singular value (operator norm)
For doubly stochastic matrices (mHC), all these should be close to 1!
""")
return
@app.cell(hide_code=True)
def _(layer_selector, results):
layer_idx = layer_selector.value - 1
baseline_m = results["baseline"]["composite"][layer_idx]
hc_m = results["hc"]["composite"][layer_idx]
mhc_m = results["mhc"]["composite"][layer_idx]
def _format_gain(g):
if g > 1000:
return f"{g:.2e}"
return f"{g:.4f}"
def _status_badge(g):
if g < 2:
return mo.md(f"**{_format_gain(g)}** :green_circle:")
elif g < 10:
return mo.md(f"**{_format_gain(g)}** :yellow_circle:")
else:
return mo.md(f"**{_format_gain(g)}** :red_circle:")
metrics_md = mo.md(f"""
### Metrics at Layer {layer_selector.value}
| Metric | Baseline | HC (Unconstrained) | mHC (Sinkhorn) |
|--------|----------|-------------------|----------------|
| Forward Gain | {_format_gain(baseline_m["forward_gain"])} | {_format_gain(hc_m["forward_gain"])} | {_format_gain(mhc_m["forward_gain"])} |
| Backward Gain | {_format_gain(baseline_m["backward_gain"])} | {_format_gain(hc_m["backward_gain"])} | {_format_gain(mhc_m["backward_gain"])} |
| Spectral Norm | {_format_gain(baseline_m["spectral_norm"])} | {_format_gain(hc_m["spectral_norm"])} | {_format_gain(mhc_m["spectral_norm"])} |
| Row Sum Dev | {baseline_m["row_sum_max_dev"]:.2e} | {hc_m["row_sum_max_dev"]:.2e} | {mhc_m["row_sum_max_dev"]:.2e} |
| Col Sum Dev | {baseline_m["col_sum_max_dev"]:.2e} | {hc_m["col_sum_max_dev"]:.2e} | {mhc_m["col_sum_max_dev"]:.2e} |
""")
metrics_md
return
@app.cell(hide_code=True)
def _():
mo.md(r"""
## Matrix Visualization
Compare a sample residual mixing matrix before and after Sinkhorn projection.
- **HC (left)**: Random matrix with arbitrary row/column sums
- **mHC (right)**: Sinkhorn-projected doubly stochastic matrix (all rows and columns sum to 1)
""")
return
@app.cell
def _(heatmaps):
heatmaps
return
@app.cell(hide_code=True)
def _(hc_sample, mhc_sample):
# Show row and column sums
hc_row_sums = hc_sample.sum(axis=1)
hc_col_sums = hc_sample.sum(axis=0)
mhc_row_sums = mhc_sample.sum(axis=1)
mhc_col_sums = mhc_sample.sum(axis=0)
sums_md = mo.md(f"""
**Row/Column Sums:**
| | HC | mHC |
|---|---|---|
| Row Sums | {np.array2string(hc_row_sums, precision=2)} | {np.array2string(mhc_row_sums, precision=3)} |
| Col Sums | {np.array2string(hc_col_sums, precision=2)} | {np.array2string(mhc_col_sums, precision=3)} |
| Max Row Dev from 1 | {np.abs(hc_row_sums - 1).max():.4f} | {np.abs(mhc_row_sums - 1).max():.2e} |
| Max Col Dev from 1 | {np.abs(hc_col_sums - 1).max():.4f} | {np.abs(mhc_col_sums - 1).max():.2e} |
Notice how mHC row/column sums are all ~1.0 (doubly stochastic)!
""")
sums_md
return
@app.cell(hide_code=True)
def _():
mo.md(r"""
## The Manifold Dial: Varying Sinkhorn Iterations
Watch how the matrix transforms as we increase iterations:
- **k=0**: Raw random matrix (same as HC) - unconstrained
- **k=1-5**: Partial projection, rapid stabilization
- **k=10-20**: Fully doubly stochastic
""")
return
@app.cell(hide_code=True)
def _(seed_input):
# Show progression of Sinkhorn iterations
iters_to_show = [0, 1, 2, 5, 10, 20]
rng_dial = np.random.default_rng(seed_input.value)
dial_base = rng_dial.standard_normal((4, 4))
dial_data = []
for k in iters_to_show:
if k == 0:
mat = dial_base # Raw random matrix (same as HC)
else:
mat = sinkhorn_knopp(dial_base, iterations=k)
for _i in range(4):
for _j in range(4):
dial_data.append(
{
"row": str(_i),
"col": str(_j),
"value": float(mat[_i, _j]),
"k": f"k={k}",
}
)
dial_df = pd.DataFrame(dial_data)
dial_chart = (
alt.Chart(dial_df)
.mark_rect()
.encode(
x=alt.X("col:O", title=None, axis=alt.Axis(labels=False)),
y=alt.Y("row:O", title=None, axis=alt.Axis(labels=False)),
color=alt.Color(
"value:Q", scale=alt.Scale(scheme="blues", domain=[0, 0.6]), legend=None
),
tooltip=["row", "col", alt.Tooltip("value:Q", format=".3f")],
)
.properties(width=100, height=100)
.facet(
column=alt.Column("k:N", title="Sinkhorn Iterations", sort=iters_to_show)
)
.properties(
title="The Manifold Dial: Sinkhorn Iterations Transform Random to Doubly Stochastic"
)
)
dial_chart
return
@app.cell(hide_code=True)
def _():
mo.md(r"""
## Why Does This Work?
### The Mathematics of Stability
Doubly stochastic matrices have three key properties:
1. **Spectral norm <= 1**: The matrix doesn't amplify signals
2. **Closed under multiplication**: Products remain doubly stochastic
3. **Convex combinations of permutations**: Acts like a weighted average (Birkhoff-von Neumann theorem)
When you multiply many doubly stochastic matrices together, the result stays bounded because each multiplication is **non-expansive**.
In contrast, unconstrained matrices compound their gains exponentially:
- If each matrix has gain 1.1, after 64 layers: $1.1^{64} \approx 300$
- If each matrix has gain 1.5, after 64 layers: $1.5^{64} \approx 10^{11}$!
""")
return
@app.cell(hide_code=True)
def _():
mo.md(r"""
## Key Takeaways
1. **HC (Hyper-Connections)** use unconstrained residual mixing matrices
- Each layer's matrix can have arbitrary row/column sums
- Over many layers, these compound into **exponential explosion**
2. **mHC (Manifold-Constrained HC)** projects matrices onto the Birkhoff polytope
- Uses Sinkhorn-Knopp to ensure doubly stochastic matrices
- Spectral norm <= 1, so products stay bounded - **stable signals**
3. **The "Manifold Dial"** is the Sinkhorn iteration count (k)
- k=0: Unconstrained (like HC) - unstable
- k>=10: Well-projected - stable
- Sweet spot around k=20 for most applications
---
**Try it yourself!** Modify the sliders above and re-run to explore:
- Different depths (try 100, 200)
- Different matrix sizes (n=2, 8)
- Different random seeds
**Interactive Demo:** https://github.com/bassrehab/mhc-visualizer
""")
return
@app.cell(hide_code=True)
def _():
mo.md(r"""
---
## References
- **mHC Paper**: [DeepSeek-AI, arXiv:2512.24880](https://arxiv.org/abs/2512.24880)
- **Sinkhorn-Knopp Algorithm**: Sinkhorn & Knopp (1967)
- **This Notebook**: [github.com/bassrehab/mhc-visualizer](https://github.com/bassrehab/mhc-visualizer)
**Author**: Subhadip Mitra (contact@subhadipmitra.com)
""")
return
@app.cell
def _(depth_slider, results):
# Build DataFrame from results
data = []
for method in ["baseline", "hc", "mhc"]:
for _i, _m in enumerate(results[method]["composite"]):
data.append(
{
"layer": _i + 1,
"gain": _m["forward_gain"],
"method": method.upper() if method != "mhc" else "mHC",
}
)
df = pd.DataFrame(data)
# Layer selector
layer_selector = mo.ui.slider(
start=1,
stop=depth_slider.value,
value=depth_slider.value,
label="Inspect Layer",
show_value=True,
)
# Altair chart with log scale
base_chart = (
alt.Chart(df)
.mark_line(strokeWidth=2.5)
.encode(
x=alt.X(
"layer:Q",
title="Layer Depth",
scale=alt.Scale(domain=[1, depth_slider.value]),
),
y=alt.Y(
"gain:Q",
scale=alt.Scale(type="log"),
title="Composite Forward Gain (log scale)",
),
color=alt.Color(
"method:N",
scale=alt.Scale(
domain=["BASELINE", "HC", "mHC"],
range=["#10b981", "#ef4444", "#3b82f6"],
),
legend=alt.Legend(title="Method"),
),
strokeDash=alt.StrokeDash(
"method:N",
scale=alt.Scale(
domain=["BASELINE", "HC", "mHC"], range=[[1, 0], [1, 0], [1, 0]]
),
legend=None,
),
)
.properties(
width=700,
height=400,
title="The Manifold Dial: HC Explosion vs mHC Stability",
)
)
return base_chart, layer_selector
@app.cell
def _(n_dropdown, seed_input, sinkhorn_slider):
n_val = int(n_dropdown.value)
rng = np.random.default_rng(seed_input.value)
# Generate matrices
base_matrix = rng.standard_normal((n_val, n_val))
hc_sample = base_matrix # Unconstrained
mhc_sample = sinkhorn_knopp(base_matrix, iterations=sinkhorn_slider.value)
# Create heatmap data
def matrix_to_df(mat, name):
rows = []
for i in range(mat.shape[0]):
for j in range(mat.shape[1]):
rows.append(
{
"row": str(i),
"col": str(j),
"value": float(mat[i, j]),
"type": name,
}
)
return pd.DataFrame(rows)
hc_df = matrix_to_df(hc_sample, "HC")
mhc_df = matrix_to_df(mhc_sample, "mHC")
# HC heatmap - use diverging colorscale
hc_heatmap = (
alt.Chart(hc_df)
.mark_rect()
.encode(
x=alt.X("col:O", title="Column", axis=alt.Axis(labelAngle=0)),
y=alt.Y("row:O", title="Row"),
color=alt.Color(
"value:Q",
scale=alt.Scale(scheme="redblue", domain=[-2, 2]),
legend=alt.Legend(title="Value"),
),
tooltip=["row", "col", alt.Tooltip("value:Q", format=".3f")],
)
.properties(width=180, height=180, title=f"HC (Random)")
)
# mHC heatmap - use sequential colorscale
mhc_heatmap = (
alt.Chart(mhc_df)
.mark_rect()
.encode(
x=alt.X("col:O", title="Column", axis=alt.Axis(labelAngle=0)),
y=alt.Y("row:O", title="Row"),
color=alt.Color(
"value:Q",
scale=alt.Scale(scheme="blues", domain=[0, 0.6]),
legend=alt.Legend(title="Value"),
),
tooltip=["row", "col", alt.Tooltip("value:Q", format=".3f")],
)
.properties(width=180, height=180, title=f"mHC (k={sinkhorn_slider.value})")
)
heatmaps = hc_heatmap | mhc_heatmap
return hc_sample, heatmaps, mhc_sample
@app.cell(hide_code=True)
def _(preset_deep, preset_explosion, preset_minimal, randomize_btn):
# Active config based on which preset button was clicked
if preset_explosion.value:
active_config = {"k": 0, "depth": 64, "n": "4", "seed": 42}
elif preset_minimal.value:
active_config = {"k": 5, "depth": 64, "n": "4", "seed": 42}
elif preset_deep.value:
active_config = {"k": 20, "depth": 200, "n": "4", "seed": 42}
elif randomize_btn.value:
active_config = {
"k": 20,
"depth": 64,
"n": "4",
"seed": int(np.random.randint(0, 10000)),
}
else: # default (including preset_default.value)
active_config = {"k": 20, "depth": 64, "n": "4", "seed": 42}
return (active_config,)
if __name__ == "__main__":
app.run() |