Spaces:
Sleeping
Sleeping
File size: 22,613 Bytes
b89c27d 4a535ea b89c27d 901a0ed b89c27d 901a0ed b89c27d 901a0ed b89c27d | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 | """FastAPI endpoints used by the React frontend.
Provides:
- /api/landscape build a template and return a Plotly contour + hints
- /api/baseline_race run 4 LR-tuned baselines and return plots + summary
- /api/arena full Phase-D evaluation of a user optimizer vs Adam
- /api/llm_run SSE-streamed LLM-driven episode
"""
from __future__ import annotations
import asyncio
import json
import re
import time
from typing import Any, Optional
import numpy as np
import requests
from fastapi import APIRouter, Query
from fastapi.responses import StreamingResponse
from pydantic import BaseModel
try:
from ..arena import auto_test_draft, run_arena, ArenaResult
from ..landscapes import BUILDERS, build_landscape, structural_hints
from ..reference_optimizers import (
run_baseline_tuned, tune_adam_lr,
)
from ..rewards import ast_novelty_score, compute_optcoder_reward
from ..sandbox import SandboxError, compile_optimizer
from ..models import LandscapeforgeAction
from ..prompts import build_prompt, parse_action
from .landscapeforge_environment import LandscapeforgeEnvironment
except ImportError: # flat layout
from arena import auto_test_draft, run_arena, ArenaResult # type: ignore
from landscapes import BUILDERS, build_landscape, structural_hints # type: ignore
from reference_optimizers import ( # type: ignore
run_baseline_tuned, tune_adam_lr,
)
from rewards import ast_novelty_score, compute_optcoder_reward # type: ignore
from sandbox import SandboxError, compile_optimizer # type: ignore
from models import LandscapeforgeAction # type: ignore
from prompts import build_prompt, parse_action # type: ignore
from server.landscapeforge_environment import LandscapeforgeEnvironment # type: ignore
router = APIRouter(prefix="/api", tags=["lf-frontend"])
# ---------- palette constants for Plotly layouts ----------
_PLOTLY_LAYOUT = dict(
font=dict(family="Inter", color="#f3f0e8", size=12),
paper_bgcolor="#2a2824", plot_bgcolor="#1f1d1a",
hoverlabel=dict(bgcolor="#f3f0e8", font_color="#1f1d1a"),
legend=dict(bgcolor="rgba(31,29,26,0.85)",
bordercolor="#403b34", borderwidth=1,
font=dict(color="#f3f0e8")),
)
_AXIS = dict(gridcolor="#403b34", zerolinecolor="#554e45",
showline=True, linecolor="#554e45",
tickfont=dict(color="#b5ada0"))
_DEFAULT_MARGIN = dict(l=60, r=30, t=60, b=55)
_TITLE = dict(x=0.02, xanchor="left", font=dict(size=14, color="#f3f0e8"))
OPT_COLORS = {
"sgd": "#c05450",
"momentum": "#d9865b",
"adam": "#5b7a6b",
"lbfgs": "#556b99",
"custom": "#e28763",
}
# ---------- shared plot helpers ----------
def _color(name: str) -> str:
return OPT_COLORS.get(name.split("(")[0].strip(), "#e28763")
def _contour_fig(ls, trajectories=None, title=None):
import numpy as np
if ls.dim != 2:
return _empty_fig(f"{ls.name} · dim={ls.dim}\nContour is 2-D only", 480)
CLIP = 8.0
xs_all, ys_all = [0.0], [0.0]
for traj in (trajectories or {}).values():
arr = np.asarray(traj)
if arr.size == 0:
continue
mask = (np.abs(arr) <= CLIP).all(axis=1) & np.isfinite(arr).all(axis=1)
good = arr[mask]
if good.size:
xs_all.extend(good[:, 0].tolist())
ys_all.extend(good[:, 1].tolist())
x_min = max(min(xs_all) - 1.5, -CLIP); x_max = min(max(xs_all) + 1.5, CLIP)
y_min = max(min(ys_all) - 1.5, -CLIP); y_max = min(max(ys_all) + 1.5, CLIP)
x_min, x_max = min(x_min, -3.5), max(x_max, 3.5)
y_min, y_max = min(y_min, -3.5), max(y_max, 3.5)
g = 70
xs = np.linspace(x_min, x_max, g)
ys = np.linspace(y_min, y_max, g)
X, Y = np.meshgrid(xs, ys)
Z = np.empty_like(X)
for i in range(g):
for j in range(g):
Z[i, j] = ls.f(np.array([X[i, j], Y[i, j]]))
finite = Z[np.isfinite(Z)]
lo, hi = map(float, np.percentile(finite, [2, 95]))
data = [dict(
type="contour", x=xs.tolist(), y=ys.tolist(), z=Z.tolist(),
zmin=lo, zmax=hi,
colorscale=[
[0.0, "#1f1d1a"], [0.15, "#2f2a22"], [0.3, "#4a2f22"],
[0.5, "#7a4229"], [0.7, "#c25a3a"], [0.85, "#e28763"],
[1.0, "#f4d6c5"],
],
contours=dict(coloring="heatmap", showlabels=False),
line=dict(width=0.5, color="rgba(243,240,232,0.12)"),
colorbar=dict(title=dict(text="f(x)",
font=dict(size=11, color="#f3f0e8")),
thickness=12, len=0.85,
tickfont=dict(size=10, color="#b5ada0"),
outlinewidth=0),
hovertemplate="x₁=%{x:.3f}<br>x₂=%{y:.3f}<br>f=%{z:.3f}<extra></extra>",
)]
if trajectories:
for name, traj in trajectories.items():
arr = np.asarray(traj)
if not arr.size:
continue
mask = (np.abs(arr) <= CLIP).all(axis=1) & np.isfinite(arr).all(axis=1)
diverged = not mask.all()
arr = arr[mask]
if arr.shape[0] == 0:
continue
color = _color(name)
label = f"{name} · diverged" if diverged else name
data.append(dict(
type="scatter", mode="lines+markers",
x=arr[:, 0].tolist(), y=arr[:, 1].tolist(),
name=label,
line=dict(color=color, width=2.5, dash="dash" if diverged else "solid"),
marker=dict(size=4, color=color,
line=dict(color="#ffffff", width=0.8)),
hovertemplate="step %{pointNumber}<br>x₁=%{x:.3f}<br>x₂=%{y:.3f}"
"<extra>" + label + "</extra>",
))
data.append(dict(type="scatter", mode="markers",
x=[arr[0, 0].item()], y=[arr[0, 1].item()],
showlegend=False,
marker=dict(size=12, color=color, symbol="circle-open",
line=dict(color=color, width=2.5)),
hoverinfo="skip"))
end_sym = "x" if diverged else "star"
data.append(dict(type="scatter", mode="markers",
x=[arr[-1, 0].item()], y=[arr[-1, 1].item()],
showlegend=False,
marker=dict(size=14 if diverged else 16,
color=color, symbol=end_sym,
line=dict(color="#ffffff", width=1.2)),
hoverinfo="skip"))
layout = {
**_PLOTLY_LAYOUT,
"title": {"text": title or f"{ls.name} (dim=2)", **_TITLE},
"height": 480, "margin": _DEFAULT_MARGIN,
"xaxis": {"title": "x₁", "range": [x_min, x_max], **_AXIS},
"yaxis": {"title": "x₂", "range": [y_min, y_max],
"scaleanchor": "x", "scaleratio": 1, **_AXIS},
}
return {"data": data, "layout": layout}
def _empty_fig(msg: str, h: int = 480):
return {"data": [], "layout": {
**_PLOTLY_LAYOUT, "height": h, "margin": _DEFAULT_MARGIN,
"xaxis": {"visible": False}, "yaxis": {"visible": False},
"annotations": [{"text": msg, "showarrow": False,
"x": 0.5, "y": 0.5, "xref": "paper", "yref": "paper",
"font": {"size": 14, "color": "#b5ada0"}}],
}}
def _curves_fig(curves, title):
data = []
for name, fs in curves.items():
if not fs:
continue
color = _color(name)
data.append(dict(
type="scatter", mode="lines+markers", name=name,
x=list(range(len(fs))),
y=[v if np.isfinite(v) else None for v in fs],
line=dict(color=color, width=2.2, shape="spline"),
marker=dict(size=4, color=color),
hovertemplate="step=%{x}<br>f=%{y:.4g}<extra>" + name + "</extra>",
connectgaps=False,
))
layout = {
**_PLOTLY_LAYOUT, "title": {"text": title, **_TITLE},
"height": 360, "margin": _DEFAULT_MARGIN,
"xaxis": {"title": "optimizer step", **_AXIS},
"yaxis": {"title": "f(x) (symlog)", "type": "log", **_AXIS},
}
return {"data": data, "layout": layout}
def _bar_fig(values, title, ylabel):
names = list(values.keys())
vs = [values[n] for n in names]
data = [dict(type="bar", x=names, y=vs,
marker=dict(color=[_color(n) for n in names],
line=dict(color="#ffffff", width=1)),
text=[f"{v:.3g}" for v in vs], textposition="outside",
textfont=dict(size=11, color="#f3f0e8"),
hovertemplate="%{x}<br>" + ylabel + "=%{y:.4g}<extra></extra>")]
layout = {
**_PLOTLY_LAYOUT, "title": {"text": title, **_TITLE},
"height": 280, "margin": _DEFAULT_MARGIN,
"xaxis": {**_AXIS},
"yaxis": {"title": ylabel, **_AXIS},
"showlegend": False,
}
return {"data": data, "layout": layout}
# ---------- request/response models ----------
class LandscapeReq(BaseModel):
template: str
dim: int = 2
seed: int = 0
class BaselineReq(BaseModel):
template: str
seed: int = 1
class ArenaReq(BaseModel):
template: str
dim: int = 5
seed: int = 42
code: str
# ---------- /api/landscape ----------
def _landscape_params(template: str) -> dict:
if template == "quadratic": return {"cond": 10.0}
if template == "gaussian_mix": return {"k": 3, "sigma": 0.5, "spread": 2.0}
return {}
@router.post("/landscape")
def api_landscape(req: LandscapeReq):
rng = np.random.default_rng(req.seed)
dim = 2 if req.template == "himmelblau" else req.dim
ls = build_landscape(template=req.template, dim=dim,
params=_landscape_params(req.template), rng=rng)
hints = structural_hints(ls, rng=rng)
hints_rows = [[k, f"{v:.4g}" if isinstance(v, float) else str(v)]
for k, v in hints.items()]
hints_rows.append(["dim", str(ls.dim)])
hints_rows.append(["f_min (known)", f"{ls.f_min:.4g}"])
hints_rows.append(["description", ls.description])
return {
"contour": _contour_fig(ls, title=f"{req.template} · dim={ls.dim}"),
"hints": hints_rows,
}
# ---------- /api/baseline_race ----------
@router.post("/baseline_race")
def api_baseline_race(req: BaselineReq):
rng = np.random.default_rng(req.seed)
ls = build_landscape(template=req.template, dim=2,
params=_landscape_params(req.template), rng=rng)
x0 = np.random.default_rng(req.seed + 999).normal(0.0, 0.5, size=2)
traj_2d, curves, finals, lrs = {}, {}, {}, {}
for name in ["sgd", "momentum", "adam", "lbfgs"]:
r = run_baseline_tuned(name, ls.f, ls.grad, x0, steps=50)
lrs[name] = r["lr"]
traj = [s for s in r["trajectory"] if s.get("x") is not None]
traj_2d[name] = [(s["x"][0], s["x"][1]) for s in traj]
curves[name] = [s["f"] for s in traj if s.get("f") is not None]
finals[name] = curves[name][-1] if curves[name] else float("inf")
lr_list = " · ".join(f"<code>{n}</code>: <code>{lr:g}</code>"
for n, lr in lrs.items())
best = min(finals, key=finals.get)
return {
"contour": _contour_fig(ls, trajectories=traj_2d,
title=f"{req.template} — baselines racing (LR-tuned)"),
"curves": _curves_fig(curves, "f(x) vs step"),
"finals": _bar_fig(finals, "Final f after 50 steps",
"f(x) at step 50"),
"summary_md": (
f"<p><strong>{ls.description}</strong></p>"
f"<p>Tuned LR per baseline (7-point sweep, 30 steps): {lr_list}</p>"
f"<p>Best baseline: <code>{best}</code> at f = "
f"<code>{finals[best]:.4f}</code></p>"
),
}
# ---------- /api/arena ----------
ADAM_TEMPLATE = """\
class Optimizer:
def __init__(self, dim):
self.lr = {lr}
self.b1, self.b2, self.eps = 0.9, 0.999, 1e-8
self.m = np.zeros(dim); self.v = np.zeros(dim); self.t = 0
def step(self, x, f_val, grad):
self.t += 1
self.m = self.b1*self.m + (1-self.b1)*grad
self.v = self.b2*self.v + (1-self.b2)*grad*grad
mh = self.m/(1-self.b1**self.t); vh = self.v/(1-self.b2**self.t)
return x - self.lr * mh / (np.sqrt(vh) + self.eps)
"""
ARENA_SEEDS = [101, 202, 303, 404, 505, 606, 707, 808, 909, 1010]
@router.post("/arena")
def api_arena(req: ArenaReq):
rng = np.random.default_rng(req.seed)
dim = 2 if req.template == "himmelblau" else req.dim
ls = build_landscape(template=req.template, dim=dim,
params=_landscape_params(req.template), rng=rng)
tune_x0 = np.random.default_rng(0).normal(0.0, 0.5, size=dim)
best_lr = tune_adam_lr(ls.f, ls.grad, tune_x0, sweep_steps=30)
adam_src = ADAM_TEMPLATE.format(lr=best_lr)
try:
opt = compile_optimizer(req.code, dim=dim)
except SandboxError as e:
return {"error": str(e)}
test = auto_test_draft(opt, ls, seed=req.seed, steps=20)
user_arena = run_arena(opt, ls, seeds=ARENA_SEEDS, steps=200)
adam_opt = compile_optimizer(adam_src, dim=dim)
adam_arena = run_arena(adam_opt, ls, seeds=ARENA_SEEDS, steps=200)
reward = compute_optcoder_reward(
arena=user_arena, adam_arena=adam_arena,
actions_used_cost=0, budget_total=12,
novelty_score=ast_novelty_score(req.code, [adam_src]),
convergence_step=None, arena_steps=200,
)
# 2-D contour if applicable
contour = None
if dim == 2:
try:
from ..reference_optimizers import run_baseline as _rb
except ImportError:
from reference_optimizers import run_baseline as _rb # type: ignore
user_traj = [(s["x"][0], s["x"][1]) for s in test["detail"]]
adam_run = _rb("adam", ls.f, ls.grad,
np.random.default_rng(req.seed).normal(0.0, 0.5, 2),
steps=50)
adam_traj = [(s["x"][0], s["x"][1]) for s in adam_run["trajectory"]
if s.get("x") is not None]
contour = _contour_fig(ls,
trajectories={"custom": user_traj, "adam": adam_traj},
title=f"{req.template} — your optimizer vs tuned Adam")
bk = reward.breakdown
speedup = bk.get("speedup_vs_adam", 0.0)
# Narrate the reward decomposition so users aren't confused when reward
# is positive despite speedup≈1× (r_convergence + r_robustness contribute
# independently of beating Adam; see §9.1 of LANDSCAPEFORGE_DESIGN.md).
parts = []
if abs(bk["r_regret"] * 1.0) > 0.01:
parts.append(f"regret {bk['r_regret']*1.0:+.3f}")
if abs(bk["r_convergence"] * 0.3) > 0.01:
parts.append(f"convergence {bk['r_convergence']*0.3:+.3f}")
if abs(bk["r_robustness"] * 0.3) > 0.01:
parts.append(f"robustness {bk['r_robustness']*0.3:+.3f}")
if abs(bk["r_novelty"] * 0.1) > 0.01:
parts.append(f"novelty {bk['r_novelty']*0.1:+.3f}")
if abs(bk["r_budget"] * 0.05) > 0.01:
parts.append(f"budget {-bk['r_budget']*0.05:+.3f}")
if abs(bk["r_eval_failures"] * 0.5) > 0.01:
parts.append(f"eval {-bk['r_eval_failures']*0.5:+.3f}")
# Speedup phrasing — avoid nonsense like "0.00×" when diverged
my_p, adam_p = user_arena.mean_progress, adam_arena.mean_progress
if my_p < 0:
speedup_line = "your optimizer <strong>diverged</strong> (f moved uphill)"
elif adam_p <= 0:
speedup_line = (f"Adam made no progress on this landscape; "
f"your progress: <code>{my_p:.3g}</code>")
else:
speedup_line = (f"Speedup vs Adam: <code>{speedup:.3g}×</code> "
f"(your descent <code>{my_p:.3g}</code>, Adam's "
f"<code>{adam_p:.3g}</code>)")
return {
"contour": contour or _empty_fig(f"{req.template} · dim={dim}\nContour is 2-D only"),
"progress": _bar_fig(
{"custom": user_arena.mean_progress,
"adam (tuned)": adam_arena.mean_progress},
"Arena mean progress",
"mean(f₀ − f_N) over 10 seeds",
),
"breakdown": bk,
"total": reward.r_total,
"summary_md": (
f"<h3>Results</h3>"
f"<ul>"
f"<li>{speedup_line}</li>"
f"<li>Tuned Adam LR: <code>{best_lr:g}</code></li>"
f"<li>Your crash fraction: <code>{user_arena.crash_fraction:.0%}</code></li>"
f"<li><strong>Total reward: <code>{reward.r_total:+.3f}</code></strong>"
+ (f"<span style='color:#b5ada0'> "
f"= {' + '.join(parts)}</span>" if parts else "")
+ "</li>"
f"</ul>"
),
}
# ---------- /api/llm_run (SSE stream) ----------
def _sse(event: str, data: dict) -> str:
return f"event: {event}\ndata: {json.dumps(data, default=str)}\n\n"
@router.get("/llm_run")
def api_llm_run(
base_url: str = Query(...),
api_key: str = "",
model: str = Query(...),
tier: str = "T0",
seed: int = 42,
temperature: float = 0.7,
max_turns: int = 10,
):
"""SSE-streamed LLM-driven episode. One event per turn."""
def gen():
url = base_url.rstrip("/") + "/chat/completions"
headers = {"Content-Type": "application/json"}
if api_key:
headers["Authorization"] = f"Bearer {api_key}"
env = LandscapeforgeEnvironment(tier=tier, seed=int(seed))
obs = env.reset()
yield _sse("message", {
"kind": "header", "model": model, "base_url": base_url,
"landscape": obs.landscape_description,
"dim": obs.dim, "budget": obs.budget_remaining,
})
for turn in range(1, int(max_turns) + 1):
messages = build_prompt(obs)
t0 = time.time()
try:
r = requests.post(url, headers=headers, json={
"model": model, "messages": messages,
"temperature": float(temperature),
"max_tokens": 1200, "stream": False,
}, timeout=180)
if r.status_code >= 400:
yield _sse("message", {
"kind": "error",
"message": f"LLM {r.status_code}: {r.text[:400]}",
})
return
raw = r.json()["choices"][0]["message"]["content"]
except Exception as e:
yield _sse("message", {
"kind": "error",
"message": f"request failed: {type(e).__name__}: {e}",
})
return
dt = time.time() - t0
try:
action = parse_action(raw)
except Exception as e:
yield _sse("message", {
"kind": "error",
"message": f"parse error: {e}. Raw: {raw[:400]}",
})
return
obs = env.step(action)
lar = obs.last_action_result or {}
output_chips = []
if lar.get("compile_error"):
output_chips.append({"kind": "bad", "text": "compile error"})
if lar.get("summary"):
s = lar["summary"]
if s.get("converged"):
output_chips.append({"kind": "good", "text": "auto-test converged"})
elif s.get("diverged"):
output_chips.append({"kind": "warn", "text": "auto-test diverged"})
if s.get("final_f") is not None:
output_chips.append({
"kind": "info",
"text": f"<code>final_f</code>=<b>{s['final_f']:.3g}</b>",
})
if action.kind == "run_baseline" and lar.get("final_f") is not None:
output_chips.append({
"kind": "info",
"text": f"<code>final_f</code>=<b>{lar['final_f']:.3g}</b>",
})
for k, v in (lar.get("feedback") or {}).items():
output_chips.append({
"kind": "good" if v >= 0 else "warn",
"text": f"<code>{k}</code> <b>{v:+.3f}</b>",
})
if action.kind == "draft":
action_str = f"draft ({len(action.code or '')} chars)"
elif action.kind == "run_baseline":
action_str = f"run_baseline({action.baseline_name})"
elif action.kind == "inspect":
action_str = (f"inspect(draft={action.draft_idx}, "
f"[{action.step_range_start},{action.step_range_end}])")
else:
action_str = "commit"
yield _sse("message", {
"kind": "turn",
"turn": turn, "kind_of": action.kind,
"action_str": action_str, "output": output_chips,
"duration_s": dt,
"budget_remaining": obs.budget_remaining,
"code": action.code if action.kind == "draft" else None,
})
if obs.done:
bk = obs.r_optcoder_breakdown or {}
yield _sse("message", {
"kind": "done",
"reason": (obs.last_action_result or {}).get("reason"),
"reward": obs.r_optcoder or 0.0,
"final_regret": obs.final_regret or 0.0,
"my_progress": bk.get("my_progress", 0.0),
"adam_progress": bk.get("adam_progress", 0.0),
"speedup_vs_adam": bk.get("speedup_vs_adam", 0.0),
"breakdown": bk,
})
yield "event: end\ndata: {}\n\n"
return
yield _sse("message", {
"kind": "error",
"message": f"reached MAX_TURNS ({max_turns}) without commit",
})
yield "event: end\ndata: {}\n\n"
return StreamingResponse(gen(), media_type="text/event-stream")
|