Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
-
# app.py — Minimal
|
2 |
import math, json, random, time, threading
|
3 |
from dataclasses import dataclass, asdict
|
4 |
from typing import List, Tuple, Dict, Any, Optional
|
@@ -16,22 +16,28 @@ import torch.optim as optim
|
|
16 |
|
17 |
from data_utils import load_piqa, load_hellaswag, hash_vectorize
|
18 |
|
19 |
-
#
|
|
|
|
|
20 |
CUSTOM_CSS = """
|
21 |
-
:root { --radius: 14px; --fg:#
|
22 |
-
* { font-family: Inter, ui-sans-serif, system-ui, -apple-system, Segoe UI, Roboto, Helvetica Neue, Arial
|
23 |
-
.gradio-container { max-width:
|
24 |
-
#header { border-radius: var(--radius); padding:
|
25 |
h1, h2, h3, .gr-markdown { color: var(--fg); }
|
26 |
-
.gr-button { border-radius: 10px }
|
27 |
-
.controls .gr-group, .panel { border: 1px solid
|
28 |
.panel { padding: 10px; }
|
29 |
-
#stats {
|
30 |
#stats strong { font-weight: 500; }
|
31 |
-
.small { font-size:
|
|
|
|
|
32 |
"""
|
33 |
|
34 |
-
#
|
|
|
|
|
35 |
@dataclass
|
36 |
class Genome:
|
37 |
d_model: int
|
@@ -86,11 +92,12 @@ def crossover(a: Genome, b: Genome, rng: random.Random) -> Genome:
|
|
86 |
memory_tokens = a.memory_tokens if rng.random()<0.5 else b.memory_tokens,
|
87 |
dropout = a.dropout if rng.random()<0.5 else b.dropout,
|
88 |
species = a.species if rng.random()<0.5 else b.species,
|
89 |
-
fitness = float("inf"),
|
90 |
-
acc = None
|
91 |
)
|
92 |
|
93 |
-
#
|
|
|
|
|
94 |
def rastrigin(x: np.ndarray) -> float:
|
95 |
A, n = 10.0, x.shape[0]
|
96 |
return A * n + np.sum(x**2 - A * np.cos(2 * math.pi * x))
|
@@ -107,9 +114,9 @@ class TinyMLP(nn.Module):
|
|
107 |
)
|
108 |
def forward(self, x): return self.net(x).squeeze(-1)
|
109 |
|
|
|
110 |
@lru_cache(maxsize=4)
|
111 |
def _cached_dataset(name: str):
|
112 |
-
# Defensive: if loading fails (e.g., datasets version / no internet), return None
|
113 |
try:
|
114 |
if name.startswith("PIQA"): return load_piqa(subset=800, seed=42)
|
115 |
if name.startswith("HellaSwag"): return load_hellaswag(subset=800, seed=42)
|
@@ -117,10 +124,9 @@ def _cached_dataset(name: str):
|
|
117 |
return None
|
118 |
return None
|
119 |
|
120 |
-
def _train_eval_proxy(genome: Genome, dataset_name: str, explore: float, device: str
|
121 |
data = _cached_dataset(dataset_name)
|
122 |
if data is None:
|
123 |
-
# Fallback to surrogate to keep the UI alive
|
124 |
v = genome.vector() * 2 - 1
|
125 |
base = rastrigin(v)
|
126 |
parsimony = 0.001 * (genome.d_model + 50*genome.n_layers + 20*genome.n_heads + 100*genome.memory_tokens)
|
@@ -132,17 +138,14 @@ def _train_eval_proxy(genome: Genome, dataset_name: str, explore: float, device:
|
|
132 |
Xtr = hash_vectorize(Xtr_txt, n_features=nfeat, seed=1234)
|
133 |
Xva = hash_vectorize(Xva_txt, n_features=nfeat, seed=5678)
|
134 |
|
135 |
-
Xtr_t = torch.from_numpy(Xtr)
|
136 |
-
|
137 |
-
Xva_t = torch.from_numpy(Xva)
|
138 |
-
yva_t = torch.from_numpy(yva.astype(np.float32))
|
139 |
|
140 |
model = TinyMLP(nfeat, genome).to(device)
|
141 |
opt = optim.AdamW(model.parameters(), lr=2e-3)
|
142 |
lossf = nn.BCEWithLogitsLoss()
|
143 |
|
144 |
-
model.train()
|
145 |
-
steps, bs, N = 120, 256, Xtr_t.size(0)
|
146 |
for _ in range(steps):
|
147 |
idx = torch.randint(0, N, (bs,))
|
148 |
xb = Xtr_t[idx].to(device); yb = ytr_t[idx].to(device)
|
@@ -157,12 +160,12 @@ def _train_eval_proxy(genome: Genome, dataset_name: str, explore: float, device:
|
|
157 |
probs = torch.sigmoid(logits).cpu().numpy()
|
158 |
|
159 |
if dataset_name.startswith("PIQA"):
|
160 |
-
probs = probs.reshape(-1,
|
161 |
pred = (probs[:,0] > probs[:,1]).astype(np.int64)
|
162 |
truth = (yva2[:,0] == 1).astype(np.int64)
|
163 |
acc = float((pred == truth).mean())
|
164 |
else:
|
165 |
-
probs = probs.reshape(-1,
|
166 |
pred = probs.argmax(axis=1); truth = yva2.argmax(axis=1)
|
167 |
acc = float((pred == truth).mean())
|
168 |
|
@@ -178,39 +181,53 @@ def evaluate_genome(genome: Genome, dataset: str, explore: float):
|
|
178 |
parsimony = 0.001 * (genome.d_model + 50*genome.n_layers + 20*genome.n_heads + 100*genome.memory_tokens)
|
179 |
noise = np.random.normal(scale=0.05 * max(0.0, min(1.0, explore)))
|
180 |
return float(base + parsimony + noise), None
|
181 |
-
if dataset.startswith("PIQA"):
|
182 |
-
|
183 |
-
if dataset.startswith("HellaSwag"):
|
184 |
-
return _train_eval_proxy(genome, "HellaSwag", explore)
|
185 |
-
# Fallback
|
186 |
v = genome.vector() * 2 - 1
|
187 |
return float(rastrigin(v)), None
|
188 |
|
189 |
-
#
|
190 |
-
|
191 |
-
|
|
|
|
|
|
|
192 |
|
193 |
def sphere_project(points: np.ndarray) -> np.ndarray:
|
194 |
rng = np.random.RandomState(42)
|
195 |
W = rng.normal(size=(points.shape[1], 3)).astype(np.float32)
|
196 |
Y = points @ W
|
197 |
norms = np.linalg.norm(Y, axis=1, keepdims=True) + 1e-8
|
198 |
-
return (Y / norms) * 1.
|
199 |
|
200 |
-
def
|
201 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
202 |
|
203 |
def make_sphere_figure(points3d: np.ndarray, genomes: List[Genome], gen_idx: int) -> go.Figure:
|
204 |
-
|
205 |
-
colors = _species_colors(species)
|
206 |
custom = np.array([[g.d_model, g.n_layers, g.n_heads, g.ffn_mult, g.memory_tokens, g.dropout,
|
207 |
g.species, g.fitness, (g.acc if g.acc is not None else -1.0)]
|
208 |
for g in genomes], dtype=np.float32)
|
209 |
-
|
210 |
scatter = go.Scatter3d(
|
211 |
x=points3d[:,0], y=points3d[:,1], z=points3d[:,2],
|
212 |
mode='markers',
|
213 |
-
marker=dict(size=
|
214 |
customdata=custom,
|
215 |
hovertemplate=(
|
216 |
"<b>Genome</b><br>"
|
@@ -221,35 +238,10 @@ def make_sphere_figure(points3d: np.ndarray, genomes: List[Genome], gen_idx: int
|
|
221 |
"accuracy=%{customdata[8]:.3f}<extra></extra>"
|
222 |
)
|
223 |
)
|
224 |
-
|
225 |
-
|
226 |
-
|
227 |
-
|
228 |
-
r = 1.2
|
229 |
-
xs = r*np.outer(np.cos(u), np.sin(v))
|
230 |
-
ys = r*np.outer(np.sin(u), np.sin(v))
|
231 |
-
zs = r*np.outer(np.ones_like(u), np.cos(v))
|
232 |
-
sphere = go.Surface(
|
233 |
-
x=xs, y=ys, z=zs,
|
234 |
-
opacity=0.08,
|
235 |
-
showscale=False,
|
236 |
-
colorscale=[[0, "#cbd5e1"], [1, "#cbd5e1"]],
|
237 |
-
hoverinfo="skip"
|
238 |
-
)
|
239 |
-
|
240 |
-
layout = go.Layout(
|
241 |
-
paper_bgcolor=BG, plot_bgcolor=BG,
|
242 |
-
title=f"Evo Architecture Sphere — Gen {gen_idx}",
|
243 |
-
scene=dict(
|
244 |
-
xaxis=dict(visible=False), yaxis=dict(visible=False), zaxis=dict(visible=False),
|
245 |
-
bgcolor=BG
|
246 |
-
),
|
247 |
-
margin=dict(l=0, r=0, t=36, b=0),
|
248 |
-
showlegend=False,
|
249 |
-
height=720,
|
250 |
-
font=dict(family="Inter, Arial, sans-serif", size=14)
|
251 |
-
)
|
252 |
-
return go.Figure(data=[sphere, scatter], layout=layout)
|
253 |
|
254 |
def make_history_figure(history: List[Tuple[int,float,float]], metric: str) -> go.Figure:
|
255 |
xs = [h[0] for h in history]
|
@@ -259,37 +251,36 @@ def make_history_figure(history: List[Tuple[int,float,float]], metric: str) -> g
|
|
259 |
else:
|
260 |
ys = [h[1] for h in history]
|
261 |
title, ylab = "Best Fitness per Generation", "Fitness (↓ better)"
|
262 |
-
fig = go.Figure(data=[go.Scatter(x=xs, y=ys, mode="lines+markers", line=dict(width=2))])
|
263 |
fig.update_layout(
|
264 |
-
paper_bgcolor=BG, plot_bgcolor=BG,
|
265 |
title=title, xaxis_title="Generation", yaxis_title=ylab,
|
266 |
-
margin=dict(l=30, r=10, t=36, b=30),
|
267 |
-
height=340,
|
268 |
-
font=dict(family="Inter, Arial, sans-serif", size=14)
|
269 |
)
|
|
|
270 |
return fig
|
271 |
|
272 |
def fig_to_html(fig: go.Figure) -> str:
|
273 |
-
|
274 |
-
return pio.to_html(
|
275 |
-
fig,
|
276 |
-
include_plotlyjs=True, # IMPORTANT: inline JS so the sphere always renders
|
277 |
-
full_html=False,
|
278 |
-
config=dict(displaylogo=False)
|
279 |
-
)
|
280 |
|
281 |
def approx_params(g: Genome) -> int:
|
282 |
per_layer = (4.0 + 2.0 * float(g.ffn_mult)) * (g.d_model ** 2)
|
283 |
total = per_layer * g.n_layers + 1000 * g.memory_tokens
|
284 |
return int(total)
|
285 |
|
286 |
-
#
|
|
|
|
|
287 |
class EvoRunner:
|
288 |
def __init__(self):
|
289 |
self.lock = threading.Lock()
|
290 |
self.running = False
|
291 |
self.stop_flag = False
|
292 |
self.state: Dict[str, Any] = {}
|
|
|
|
|
|
|
|
|
293 |
|
294 |
def run(self, dataset, pop_size, generations, mutation_rate, explore, exploit, seed, pace_ms, metric_choice):
|
295 |
rng = random.Random(int(seed))
|
@@ -302,16 +293,12 @@ class EvoRunner:
|
|
302 |
g.fitness, g.acc = fit, acc
|
303 |
|
304 |
history: List[Tuple[int,float,float]] = []
|
305 |
-
best_overall: Optional[Genome] = None
|
306 |
|
307 |
for gen in range(1, generations+1):
|
308 |
if self.stop_flag: break
|
309 |
|
310 |
k = max(2, int(2 + exploit * 5))
|
311 |
-
parents = []
|
312 |
-
for _ in range(pop_size):
|
313 |
-
sample = rng.sample(pop, k=k)
|
314 |
-
parents.append(min(sample, key=lambda x: x.fitness))
|
315 |
|
316 |
children = []
|
317 |
for i in range(0, pop_size, 2):
|
@@ -331,8 +318,6 @@ class EvoRunner:
|
|
331 |
pop[-elite_n:] = elites
|
332 |
|
333 |
best = min(pop, key=lambda x: x.fitness)
|
334 |
-
if best_overall is None or best.fitness < best_overall.fitness: best_overall = best
|
335 |
-
|
336 |
history.append((gen, best.fitness, (best.acc if best.acc is not None else float("nan"))))
|
337 |
|
338 |
P = np.stack([g.vector() for g in pop], axis=0)
|
@@ -341,22 +326,14 @@ class EvoRunner:
|
|
341 |
hist_fig = make_history_figure(history, metric_choice)
|
342 |
|
343 |
top = sorted(pop, key=lambda x: x.fitness)[: min(12, len(pop))]
|
344 |
-
top_table = [
|
345 |
-
|
346 |
-
|
347 |
-
|
348 |
-
|
349 |
-
|
350 |
-
|
351 |
-
|
352 |
-
"ffn_mult": t.ffn_mult,
|
353 |
-
"mem": t.memory_tokens,
|
354 |
-
"dropout": t.dropout,
|
355 |
-
"species": t.species,
|
356 |
-
"params_approx": approx_params(t)
|
357 |
-
} for t in top
|
358 |
-
]
|
359 |
-
best_card = top_table[0] if len(top_table) else {}
|
360 |
|
361 |
with self.lock:
|
362 |
self.state = {
|
@@ -376,24 +353,42 @@ class EvoRunner:
|
|
376 |
if self.running: return
|
377 |
t = threading.Thread(target=self.run, args=args, kwargs=kwargs, daemon=True)
|
378 |
t.start()
|
|
|
379 |
def stop(self): self.stop_flag = True
|
380 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
381 |
runner = EvoRunner()
|
382 |
|
383 |
-
#
|
|
|
|
|
384 |
def start_evo(dataset, pop, gens, mut, explore, exploit, seed, pace_ms, metric_choice):
|
385 |
runner.start(dataset, int(pop), int(gens), float(mut), float(explore), float(exploit), int(seed), int(pace_ms), metric_choice)
|
386 |
-
return (gr.update(interactive=False), gr.update(interactive=True))
|
387 |
|
388 |
def stop_evo():
|
389 |
runner.stop()
|
390 |
-
return (gr.update(interactive=True), gr.update(interactive=False))
|
|
|
|
|
|
|
|
|
|
|
|
|
391 |
|
392 |
def poll_state():
|
393 |
with runner.lock:
|
394 |
s = runner.state.copy()
|
395 |
-
sphere_html = s.get("sphere_html",
|
396 |
-
history_html = s.get("history_html", "")
|
397 |
best = s.get("best", {})
|
398 |
gen = s.get("gen", 0)
|
399 |
dataset = s.get("dataset", "Demo (Surrogate)")
|
@@ -411,7 +406,7 @@ def poll_state():
|
|
411 |
f"**~Params (rough):** {best.get('params_approx'):,}"
|
412 |
)
|
413 |
else:
|
414 |
-
stats_md = "
|
415 |
df = pd.DataFrame(top)
|
416 |
return sphere_html, history_html, stats_md, df
|
417 |
|
@@ -424,10 +419,12 @@ def export_snapshot():
|
|
424 |
f.write(payload)
|
425 |
return path
|
426 |
|
427 |
-
#
|
428 |
-
|
|
|
|
|
429 |
with gr.Column(elem_id="header"):
|
430 |
-
gr.Markdown("
|
431 |
|
432 |
with gr.Row():
|
433 |
with gr.Column(scale=1, elem_classes=["controls"]):
|
@@ -436,7 +433,7 @@ with gr.Blocks(css=CUSTOM_CSS, theme=gr.themes.Soft()) as demo:
|
|
436 |
label="Dataset",
|
437 |
choices=["Demo (Surrogate)", "PIQA (Phase 2)", "HellaSwag (Phase 2)"],
|
438 |
value="Demo (Surrogate)",
|
439 |
-
info="PIQA/HellaSwag compute
|
440 |
)
|
441 |
pop = gr.Slider(8, 80, value=24, step=2, label="Population size")
|
442 |
gens = gr.Slider(5, 200, value=60, step=1, label="Max generations")
|
@@ -445,14 +442,16 @@ with gr.Blocks(css=CUSTOM_CSS, theme=gr.themes.Soft()) as demo:
|
|
445 |
explore = gr.Slider(0.0, 1.0, value=0.35, step=0.05, label="Exploration")
|
446 |
exploit = gr.Slider(0.0, 1.0, value=0.65, step=0.05, label="Exploitation")
|
447 |
seed = gr.Number(value=42, label="Seed", precision=0)
|
448 |
-
pace = gr.Slider(0, 1000, value=120, step=10, label="Pace (ms
|
449 |
metric_choice = gr.Radio(choices=["Accuracy", "Fitness"], value="Accuracy", label="History Metric")
|
|
|
450 |
with gr.Row():
|
451 |
-
start = gr.Button("▶ Start
|
452 |
-
stop = gr.Button("⏹ Stop", variant="secondary")
|
|
|
453 |
|
454 |
with gr.Group(elem_classes=["panel"]):
|
455 |
-
stats_md = gr.Markdown("
|
456 |
|
457 |
with gr.Group(elem_classes=["panel"]):
|
458 |
export_btn = gr.Button("Export Snapshot (JSON)")
|
@@ -466,12 +465,13 @@ with gr.Blocks(css=CUSTOM_CSS, theme=gr.themes.Soft()) as demo:
|
|
466 |
with gr.Group(elem_classes=["panel"]):
|
467 |
top_df = gr.Dataframe(label="Top Genomes (live)", wrap=True, interactive=False)
|
468 |
|
469 |
-
#
|
470 |
-
start.click(start_evo, [dataset, pop, gens, mut, explore, exploit, seed, pace, metric_choice], [start, stop])
|
471 |
-
stop.click(stop_evo, [], [start, stop])
|
|
|
472 |
export_btn.click(export_snapshot, [], [export_file])
|
473 |
|
474 |
-
#
|
475 |
demo.load(poll_state, None, [sphere_html, hist_html, stats_md, top_df])
|
476 |
gr.Timer(0.7).tick(poll_state, None, [sphere_html, hist_html, stats_md, top_df])
|
477 |
|
|
|
1 |
+
# app.py — Minimal dark UI, default idle sphere, Clear button, inline Plotly
|
2 |
import math, json, random, time, threading
|
3 |
from dataclasses import dataclass, asdict
|
4 |
from typing import List, Tuple, Dict, Any, Optional
|
|
|
16 |
|
17 |
from data_utils import load_piqa, load_hellaswag, hash_vectorize
|
18 |
|
19 |
+
# =========================
|
20 |
+
# STYLE — calm, dark, thin
|
21 |
+
# =========================
|
22 |
CUSTOM_CSS = """
|
23 |
+
:root { --radius: 14px; --fg:#E5E7EB; --muted:#94A3B8; --line:#111827; --bg:#0F1A24; }
|
24 |
+
* { font-family: Inter, ui-sans-serif, system-ui, -apple-system, Segoe UI, Roboto, Helvetica Neue, Arial; font-weight: 300; }
|
25 |
+
.gradio-container { max-width: 1140px !important; background: var(--bg); }
|
26 |
+
#header { border-radius: var(--radius); padding: 6px 2px; }
|
27 |
h1, h2, h3, .gr-markdown { color: var(--fg); }
|
28 |
+
.gr-button { border-radius: 10px; }
|
29 |
+
.controls .gr-group, .panel { border: 1px solid #1f2b36; border-radius: var(--radius); background: #0c161f; }
|
30 |
.panel { padding: 10px; }
|
31 |
+
#stats { color: var(--fg); }
|
32 |
#stats strong { font-weight: 500; }
|
33 |
+
.small { font-size: 12px; color: var(--muted); }
|
34 |
+
label, .gradio-container * { color: var(--fg); }
|
35 |
+
input, textarea, select { color: var(--fg) !important; }
|
36 |
"""
|
37 |
|
38 |
+
# =========================
|
39 |
+
# GENOME
|
40 |
+
# =========================
|
41 |
@dataclass
|
42 |
class Genome:
|
43 |
d_model: int
|
|
|
92 |
memory_tokens = a.memory_tokens if rng.random()<0.5 else b.memory_tokens,
|
93 |
dropout = a.dropout if rng.random()<0.5 else b.dropout,
|
94 |
species = a.species if rng.random()<0.5 else b.species,
|
95 |
+
fitness = float("inf"), acc=None
|
|
|
96 |
)
|
97 |
|
98 |
+
# =========================
|
99 |
+
# PROXY FITNESS
|
100 |
+
# =========================
|
101 |
def rastrigin(x: np.ndarray) -> float:
|
102 |
A, n = 10.0, x.shape[0]
|
103 |
return A * n + np.sum(x**2 - A * np.cos(2 * math.pi * x))
|
|
|
114 |
)
|
115 |
def forward(self, x): return self.net(x).squeeze(-1)
|
116 |
|
117 |
+
from functools import lru_cache
|
118 |
@lru_cache(maxsize=4)
|
119 |
def _cached_dataset(name: str):
|
|
|
120 |
try:
|
121 |
if name.startswith("PIQA"): return load_piqa(subset=800, seed=42)
|
122 |
if name.startswith("HellaSwag"): return load_hellaswag(subset=800, seed=42)
|
|
|
124 |
return None
|
125 |
return None
|
126 |
|
127 |
+
def _train_eval_proxy(genome: Genome, dataset_name: str, explore: float, device: str="cpu"):
|
128 |
data = _cached_dataset(dataset_name)
|
129 |
if data is None:
|
|
|
130 |
v = genome.vector() * 2 - 1
|
131 |
base = rastrigin(v)
|
132 |
parsimony = 0.001 * (genome.d_model + 50*genome.n_layers + 20*genome.n_heads + 100*genome.memory_tokens)
|
|
|
138 |
Xtr = hash_vectorize(Xtr_txt, n_features=nfeat, seed=1234)
|
139 |
Xva = hash_vectorize(Xva_txt, n_features=nfeat, seed=5678)
|
140 |
|
141 |
+
Xtr_t = torch.from_numpy(Xtr); ytr_t = torch.from_numpy(ytr.astype(np.float32))
|
142 |
+
Xva_t = torch.from_numpy(Xva); yva_t = torch.from_numpy(yva.astype(np.float32))
|
|
|
|
|
143 |
|
144 |
model = TinyMLP(nfeat, genome).to(device)
|
145 |
opt = optim.AdamW(model.parameters(), lr=2e-3)
|
146 |
lossf = nn.BCEWithLogitsLoss()
|
147 |
|
148 |
+
model.train(); steps, bs, N = 120, 256, Xtr_t.size(0)
|
|
|
149 |
for _ in range(steps):
|
150 |
idx = torch.randint(0, N, (bs,))
|
151 |
xb = Xtr_t[idx].to(device); yb = ytr_t[idx].to(device)
|
|
|
160 |
probs = torch.sigmoid(logits).cpu().numpy()
|
161 |
|
162 |
if dataset_name.startswith("PIQA"):
|
163 |
+
probs = probs.reshape(-1,2); yva2 = yva.reshape(-1,2)
|
164 |
pred = (probs[:,0] > probs[:,1]).astype(np.int64)
|
165 |
truth = (yva2[:,0] == 1).astype(np.int64)
|
166 |
acc = float((pred == truth).mean())
|
167 |
else:
|
168 |
+
probs = probs.reshape(-1,4); yva2 = yva.reshape(-1,4)
|
169 |
pred = probs.argmax(axis=1); truth = yva2.argmax(axis=1)
|
170 |
acc = float((pred == truth).mean())
|
171 |
|
|
|
181 |
parsimony = 0.001 * (genome.d_model + 50*genome.n_layers + 20*genome.n_heads + 100*genome.memory_tokens)
|
182 |
noise = np.random.normal(scale=0.05 * max(0.0, min(1.0, explore)))
|
183 |
return float(base + parsimony + noise), None
|
184 |
+
if dataset.startswith("PIQA"): return _train_eval_proxy(genome, "PIQA", explore)
|
185 |
+
if dataset.startswith("HellaSwag"): return _train_eval_proxy(genome, "HellaSwag", explore)
|
|
|
|
|
|
|
186 |
v = genome.vector() * 2 - 1
|
187 |
return float(rastrigin(v)), None
|
188 |
|
189 |
+
# =========================
|
190 |
+
# VIZ — big transparent sphere
|
191 |
+
# =========================
|
192 |
+
BG = "#0F1A24"
|
193 |
+
DOT = "#93C5FD" # soft blue dot
|
194 |
+
SPHERE = "#cbd5e1" # subtle sphere tint
|
195 |
|
196 |
def sphere_project(points: np.ndarray) -> np.ndarray:
|
197 |
rng = np.random.RandomState(42)
|
198 |
W = rng.normal(size=(points.shape[1], 3)).astype(np.float32)
|
199 |
Y = points @ W
|
200 |
norms = np.linalg.norm(Y, axis=1, keepdims=True) + 1e-8
|
201 |
+
return (Y / norms) * 1.22
|
202 |
|
203 |
+
def make_idle_sphere() -> go.Figure:
|
204 |
+
# empty scatter, only sphere
|
205 |
+
u = np.linspace(0, 2*np.pi, 72)
|
206 |
+
v = np.linspace(0, np.pi, 36)
|
207 |
+
r = 1.22
|
208 |
+
xs = r*np.outer(np.cos(u), np.sin(v))
|
209 |
+
ys = r*np.outer(np.sin(u), np.sin(v))
|
210 |
+
zs = r*np.outer(np.ones_like(u), np.cos(v))
|
211 |
+
sphere = go.Surface(x=xs, y=ys, z=zs, opacity=0.06, showscale=False,
|
212 |
+
colorscale=[[0, SPHERE],[1, SPHERE]], hoverinfo="skip")
|
213 |
+
layout = go.Layout(
|
214 |
+
paper_bgcolor=BG, plot_bgcolor=BG,
|
215 |
+
title="Architecture Sphere (idle)", titlefont=dict(color="#E5E7EB"),
|
216 |
+
scene=dict(xaxis=dict(visible=False), yaxis=dict(visible=False), zaxis=dict(visible=False), bgcolor=BG),
|
217 |
+
margin=dict(l=0, r=0, t=36, b=0), showlegend=False, height=720,
|
218 |
+
font=dict(family="Inter, Arial, sans-serif", size=14, color="#E5E7EB")
|
219 |
+
)
|
220 |
+
return go.Figure(data=[sphere], layout=layout)
|
221 |
|
222 |
def make_sphere_figure(points3d: np.ndarray, genomes: List[Genome], gen_idx: int) -> go.Figure:
|
223 |
+
# single-color dots for a sober look
|
|
|
224 |
custom = np.array([[g.d_model, g.n_layers, g.n_heads, g.ffn_mult, g.memory_tokens, g.dropout,
|
225 |
g.species, g.fitness, (g.acc if g.acc is not None else -1.0)]
|
226 |
for g in genomes], dtype=np.float32)
|
|
|
227 |
scatter = go.Scatter3d(
|
228 |
x=points3d[:,0], y=points3d[:,1], z=points3d[:,2],
|
229 |
mode='markers',
|
230 |
+
marker=dict(size=7.2, color=DOT, opacity=0.92),
|
231 |
customdata=custom,
|
232 |
hovertemplate=(
|
233 |
"<b>Genome</b><br>"
|
|
|
238 |
"accuracy=%{customdata[8]:.3f}<extra></extra>"
|
239 |
)
|
240 |
)
|
241 |
+
idle = make_idle_sphere()
|
242 |
+
layout = idle.layout.update(title=f"Evo Architecture Sphere — Gen {gen_idx}")
|
243 |
+
fig = go.Figure(data=idle.data + (scatter,), layout=layout)
|
244 |
+
return fig
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
245 |
|
246 |
def make_history_figure(history: List[Tuple[int,float,float]], metric: str) -> go.Figure:
|
247 |
xs = [h[0] for h in history]
|
|
|
251 |
else:
|
252 |
ys = [h[1] for h in history]
|
253 |
title, ylab = "Best Fitness per Generation", "Fitness (↓ better)"
|
254 |
+
fig = go.Figure(data=[go.Scatter(x=xs, y=ys, mode="lines+markers", line=dict(width=2), marker=dict(color=DOT))])
|
255 |
fig.update_layout(
|
256 |
+
paper_bgcolor=BG, plot_bgcolor=BG, font=dict(color="#E5E7EB"),
|
257 |
title=title, xaxis_title="Generation", yaxis_title=ylab,
|
258 |
+
margin=dict(l=30, r=10, t=36, b=30), height=340
|
|
|
|
|
259 |
)
|
260 |
+
fig.update_xaxes(gridcolor="#1f2b36"); fig.update_yaxes(gridcolor="#1f2b36")
|
261 |
return fig
|
262 |
|
263 |
def fig_to_html(fig: go.Figure) -> str:
|
264 |
+
return pio.to_html(fig, include_plotlyjs=True, full_html=False, config=dict(displaylogo=False))
|
|
|
|
|
|
|
|
|
|
|
|
|
265 |
|
266 |
def approx_params(g: Genome) -> int:
|
267 |
per_layer = (4.0 + 2.0 * float(g.ffn_mult)) * (g.d_model ** 2)
|
268 |
total = per_layer * g.n_layers + 1000 * g.memory_tokens
|
269 |
return int(total)
|
270 |
|
271 |
+
# =========================
|
272 |
+
# RUNNER
|
273 |
+
# =========================
|
274 |
class EvoRunner:
|
275 |
def __init__(self):
|
276 |
self.lock = threading.Lock()
|
277 |
self.running = False
|
278 |
self.stop_flag = False
|
279 |
self.state: Dict[str, Any] = {}
|
280 |
+
# seed the idle sphere immediately
|
281 |
+
idle = fig_to_html(make_idle_sphere())
|
282 |
+
self.state = {"sphere_html": idle, "history_html": fig_to_html(make_history_figure([], "Accuracy")),
|
283 |
+
"top": [], "best": {}, "gen": 0, "dataset": "Demo (Surrogate)", "metric": "Accuracy"}
|
284 |
|
285 |
def run(self, dataset, pop_size, generations, mutation_rate, explore, exploit, seed, pace_ms, metric_choice):
|
286 |
rng = random.Random(int(seed))
|
|
|
293 |
g.fitness, g.acc = fit, acc
|
294 |
|
295 |
history: List[Tuple[int,float,float]] = []
|
|
|
296 |
|
297 |
for gen in range(1, generations+1):
|
298 |
if self.stop_flag: break
|
299 |
|
300 |
k = max(2, int(2 + exploit * 5))
|
301 |
+
parents = [min(rng.sample(pop, k=k), key=lambda x: x.fitness) for _ in range(pop_size)]
|
|
|
|
|
|
|
302 |
|
303 |
children = []
|
304 |
for i in range(0, pop_size, 2):
|
|
|
318 |
pop[-elite_n:] = elites
|
319 |
|
320 |
best = min(pop, key=lambda x: x.fitness)
|
|
|
|
|
321 |
history.append((gen, best.fitness, (best.acc if best.acc is not None else float("nan"))))
|
322 |
|
323 |
P = np.stack([g.vector() for g in pop], axis=0)
|
|
|
326 |
hist_fig = make_history_figure(history, metric_choice)
|
327 |
|
328 |
top = sorted(pop, key=lambda x: x.fitness)[: min(12, len(pop))]
|
329 |
+
top_table = [{
|
330 |
+
"gen": gen, "fitness": round(t.fitness, 4),
|
331 |
+
"accuracy": (None if t.acc is None else round(float(t.acc), 4)),
|
332 |
+
"d_model": t.d_model, "layers": t.n_layers, "heads": t.n_heads,
|
333 |
+
"ffn_mult": t.ffn_mult, "mem": t.memory_tokens, "dropout": t.dropout,
|
334 |
+
"params_approx": approx_params(t)
|
335 |
+
} for t in top]
|
336 |
+
best_card = top_table[0] if top_table else {}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
337 |
|
338 |
with self.lock:
|
339 |
self.state = {
|
|
|
353 |
if self.running: return
|
354 |
t = threading.Thread(target=self.run, args=args, kwargs=kwargs, daemon=True)
|
355 |
t.start()
|
356 |
+
|
357 |
def stop(self): self.stop_flag = True
|
358 |
|
359 |
+
def clear(self):
|
360 |
+
# stop and reset to idle sphere
|
361 |
+
self.stop_flag = True
|
362 |
+
idle = fig_to_html(make_idle_sphere())
|
363 |
+
with self.lock:
|
364 |
+
self.running = False
|
365 |
+
self.state = {"sphere_html": idle, "history_html": fig_to_html(make_history_figure([], "Accuracy")),
|
366 |
+
"top": [], "best": {}, "gen": 0, "dataset": "Demo (Surrogate)", "metric": "Accuracy"}
|
367 |
+
|
368 |
runner = EvoRunner()
|
369 |
|
370 |
+
# =========================
|
371 |
+
# UI CALLBACKS
|
372 |
+
# =========================
|
373 |
def start_evo(dataset, pop, gens, mut, explore, exploit, seed, pace_ms, metric_choice):
|
374 |
runner.start(dataset, int(pop), int(gens), float(mut), float(explore), float(exploit), int(seed), int(pace_ms), metric_choice)
|
375 |
+
return (gr.update(interactive=False), gr.update(interactive=True), gr.update(interactive=False))
|
376 |
|
377 |
def stop_evo():
|
378 |
runner.stop()
|
379 |
+
return (gr.update(interactive=True), gr.update(interactive=False), gr.update(interactive=True))
|
380 |
+
|
381 |
+
def clear_evo():
|
382 |
+
runner.clear()
|
383 |
+
# return updated visuals + reset buttons
|
384 |
+
sphere_html, history_html, stats_md, df = poll_state()
|
385 |
+
return sphere_html, history_html, stats_md, df, gr.update(interactive=True), gr.update(interactive=False), gr.update(interactive=True)
|
386 |
|
387 |
def poll_state():
|
388 |
with runner.lock:
|
389 |
s = runner.state.copy()
|
390 |
+
sphere_html = s.get("sphere_html", fig_to_html(make_idle_sphere()))
|
391 |
+
history_html = s.get("history_html", fig_to_html(make_history_figure([], "Accuracy")))
|
392 |
best = s.get("best", {})
|
393 |
gen = s.get("gen", 0)
|
394 |
dataset = s.get("dataset", "Demo (Surrogate)")
|
|
|
406 |
f"**~Params (rough):** {best.get('params_approx'):,}"
|
407 |
)
|
408 |
else:
|
409 |
+
stats_md = "Ready. Press **Start** to evolve, or **Clear** anytime."
|
410 |
df = pd.DataFrame(top)
|
411 |
return sphere_html, history_html, stats_md, df
|
412 |
|
|
|
419 |
f.write(payload)
|
420 |
return path
|
421 |
|
422 |
+
# =========================
|
423 |
+
# BUILD UI
|
424 |
+
# =========================
|
425 |
+
with gr.Blocks(css=CUSTOM_CSS) as demo:
|
426 |
with gr.Column(elem_id="header"):
|
427 |
+
gr.Markdown("### Evo Playground — Live Evolution (clean dark)")
|
428 |
|
429 |
with gr.Row():
|
430 |
with gr.Column(scale=1, elem_classes=["controls"]):
|
|
|
433 |
label="Dataset",
|
434 |
choices=["Demo (Surrogate)", "PIQA (Phase 2)", "HellaSwag (Phase 2)"],
|
435 |
value="Demo (Surrogate)",
|
436 |
+
info="PIQA/HellaSwag compute proxy accuracy; Demo is a fast surrogate."
|
437 |
)
|
438 |
pop = gr.Slider(8, 80, value=24, step=2, label="Population size")
|
439 |
gens = gr.Slider(5, 200, value=60, step=1, label="Max generations")
|
|
|
442 |
explore = gr.Slider(0.0, 1.0, value=0.35, step=0.05, label="Exploration")
|
443 |
exploit = gr.Slider(0.0, 1.0, value=0.65, step=0.05, label="Exploitation")
|
444 |
seed = gr.Number(value=42, label="Seed", precision=0)
|
445 |
+
pace = gr.Slider(0, 1000, value=120, step=10, label="Pace (ms)")
|
446 |
metric_choice = gr.Radio(choices=["Accuracy", "Fitness"], value="Accuracy", label="History Metric")
|
447 |
+
|
448 |
with gr.Row():
|
449 |
+
start = gr.Button("▶ Start", variant="primary")
|
450 |
+
stop = gr.Button("⏹ Stop", variant="secondary", interactive=False)
|
451 |
+
clear = gr.Button("↺ Clear", variant="secondary")
|
452 |
|
453 |
with gr.Group(elem_classes=["panel"]):
|
454 |
+
stats_md = gr.Markdown("Ready. Press **Start** to evolve, or **Clear** anytime.", elem_id="stats")
|
455 |
|
456 |
with gr.Group(elem_classes=["panel"]):
|
457 |
export_btn = gr.Button("Export Snapshot (JSON)")
|
|
|
465 |
with gr.Group(elem_classes=["panel"]):
|
466 |
top_df = gr.Dataframe(label="Top Genomes (live)", wrap=True, interactive=False)
|
467 |
|
468 |
+
# wiring
|
469 |
+
start.click(start_evo, [dataset, pop, gens, mut, explore, exploit, seed, pace, metric_choice], [start, stop, clear])
|
470 |
+
stop.click(stop_evo, [], [start, stop, clear])
|
471 |
+
clear.click(clear_evo, [], [sphere_html, hist_html, stats_md, top_df, start, stop, clear])
|
472 |
export_btn.click(export_snapshot, [], [export_file])
|
473 |
|
474 |
+
# initial paint + polling
|
475 |
demo.load(poll_state, None, [sphere_html, hist_html, stats_md, top_df])
|
476 |
gr.Timer(0.7).tick(poll_state, None, [sphere_html, hist_html, stats_md, top_df])
|
477 |
|