File size: 24,734 Bytes
b00faa3 c4b7a94 b00faa3 5c4da67 b00faa3 5c4da67 b00faa3 7e7a097 b00faa3 c4b7a94 e3bb26c c4b7a94 ba5aabf c4b7a94 b00faa3 7e7a097 c4b7a94 f132d17 6519ca2 03bc9fe 4b323a5 f132d17 2f0437f 6519ca2 c4b7a94 b00faa3 bb2faab b00faa3 c4b7a94 b00faa3 c4b7a94 f132d17 03bc9fe f132d17 b00faa3 03bc9fe bb2faab c4b7a94 b00faa3 0a17d1c b00faa3 c4b7a94 b00faa3 c4b7a94 b00faa3 03bc9fe 2f0437f 03bc9fe 2f0437f c4b7a94 6519ca2 2f0437f c4b7a94 9e5b157 5c4da67 57e3799 03bc9fe 5c4da67 c4b7a94 b00faa3 2f0437f c4b7a94 03bc9fe 5c4da67 c4b7a94 03bc9fe 4cfd118 2f0437f 03bc9fe 4cfd118 c4b7a94 4cfd118 c4b7a94 03bc9fe c4b7a94 03bc9fe 614270d 03bc9fe 7c1a4d6 e3bb26c cb00551 f0a2503 c4b7a94 4b323a5 ba5aabf b620a9c 26b0dfb 5c4da67 d6d17d4 616bb9e d6d17d4 17ce467 5c4da67 d6d17d4 17ce467 d6d17d4 4b323a5 d6d17d4 17ce467 8949df4 17ce467 4b323a5 17ce467 4b323a5 17ce467 4b323a5 17ce467 5c4da67 f132d17 7e7a097 f132d17 5c4da67 f132d17 5c4da67 c4b7a94 5c4da67 b00faa3 0fbcb31 17ce467 b00faa3 7c1a4d6 7a5f49c b00faa3 7c1a4d6 c4b7a94 7c1a4d6 6fd1ae5 7c1a4d6 e3bb26c 5877a68 c4b7a94 7e7a097 c4b7a94 7c1a4d6 616bb9e 4b323a5 616bb9e 7c1a4d6 0fbcb31 683db75 7c1a4d6 4b323a5 683db75 4b323a5 7c1a4d6 683db75 4b323a5 7c1a4d6 4b323a5 683db75 7c1a4d6 683db75 3e54b83 2b7eb68 7c1a4d6 3e54b83 7c1a4d6 3e54b83 26b0dfb 03bc9fe b00faa3 f132d17 5c4da67 7a5f49c | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 | """Toto weather forecasting demo β Gradio app for HuggingFace Spaces.
Pulls live data from an Ecowitt GW3000B home weather station, runs Datadog's
Toto 2.0 (4M) for the next N hours, and shows it next to the National Weather
Service hourly forecast β plus a rolling Toto-vs-NWS scoreboard once enough
forecasts have aged into actuals.
"""
from __future__ import annotations
import os
import threading
import time
import traceback
from datetime import datetime, timedelta, timezone
import gradio as gr
import pandas as pd
from src import ecowitt, forecast_log, nws, persist, storage, sync
from src.forecast import forecast_series
from src.weather_ui import (
aligned_comparison_markdown,
combined_figure,
hero_markdown,
residual_figure,
)
AUTO_REFRESH_SECONDS = 15 * 60 # background tick + archive sync
CACHE_TTL_SECONDS = AUTO_REFRESH_SECONDS - 60 # so autorefresh always refetches
DISPLAY_TZ = os.environ.get("DISPLAY_TZ", "America/New_York")
PLACE_NAME = os.environ.get("PLACE_NAME", "Yaphank, NY")
# Single canonical view. History is read from the local SQLite archive
# (data/ecowitt.db) instead of hitting the Ecowitt API on every page load.
# The archive is kept current by the autorefresh thread + the per-Space-tick
# sync, so over time we accumulate true 5-min granularity beyond Ecowitt's
# own 24-30h 5-min retention window.
#
# context_days = how much past data we feed Toto. display_days = how much
# we show on the chart. Feeding the model a longer context than we display
# keeps the forecast informed by the full week without making the chart
# noisy.
VIEW_WEEK = {
"label": "Past 5 days Β· 48h forecast (5-min display, Toto fed hourly)",
"cycle_type": "5min",
"resample": "5min", # display cadence on the chart
"forecast_resample": "1h", # cadence Toto actually consumes
"display_days": 5,
"context_days": 7,
"horizon_hours": 48,
}
METRICS = [
{"col": "temp_f", "title": "Outdoor temperature", "y": "Β°F", "nws_col": "temp_f"},
{"col": "rain_in_hr", "title": "Rainfall rate", "y": "in/hr", "nws_col": None},
{"col": "humidity", "title": "Outdoor humidity", "y": "%", "nws_col": "humidity"},
{"col": "pressure_inhg", "title": "Barometric pressure", "y": "inHg", "nws_col": None},
]
# --- TTL cache (multi-arg keys) ------------------------------------------
_cache: dict[tuple, tuple[float, object]] = {}
def cached(ttl: int):
def deco(fn):
def wrapper(*args, **kwargs):
key = (fn.__name__, args, tuple(sorted(kwargs.items())))
now = time.time()
hit = _cache.get(key)
if hit and now - hit[0] < ttl:
return hit[1]
out = fn(*args, **kwargs)
_cache[key] = (now, out)
return out
return wrapper
return deco
# --- data fetchers --------------------------------------------------------
ECOWITT_ARCHIVE_DB_PATH = "data/ecowitt.db"
def fetch_history(cycle_type: str, resample: str, hours: float) -> pd.DataFrame:
"""Read history from the local SQLite archive. If the archive is empty
(cold start before the first sync), fall back to a one-shot API pull so
the page still renders something."""
now_unix = int(time.time())
since_unix = now_unix - int(hours * 3600)
conn = storage.connect(ECOWITT_ARCHIVE_DB_PATH)
try:
df = storage.read_history_dataframe(
conn, since_unix=since_unix, until_unix=now_unix,
cycle_type=cycle_type, resample=resample,
)
finally:
conn.close()
if not df.empty:
return df
# Cold-start fallback: pull a small slice directly from the API so the
# page isn't blank on the very first visit before sync has run.
cfg = ecowitt.EcowittConfig.from_env()
end = datetime.now(timezone.utc).replace(tzinfo=None)
start = end - timedelta(hours=hours)
raw = ecowitt.fetch_history(
cfg, start, end, cycle_type=cycle_type,
call_back="outdoor,pressure,rainfall_piezo",
)
return ecowitt.history_to_dataframe(raw, resample=resample)
@cached(60) # short TTL β real-time is the freshness path
def fetch_realtime_snapshot() -> dict:
"""Return the most recent reading from /device/real_time as a flat dict.
The hourly history bucket only fills once Ecowitt has at least one
reading inside that hour, which can lag real time by up to 30 min on
the GW3000B. /device/real_time returns the device's last reading with
its own minute-resolution timestamp, so we use it for the live hero.
"""
cfg = ecowitt.EcowittConfig.from_env()
body = ecowitt.fetch_real_time(cfg, call_back="outdoor,pressure,rainfall_piezo")
data = body.get("data") or {}
out = {}
def _val(node):
if isinstance(node, dict) and "value" in node:
try:
return float(node["value"])
except (TypeError, ValueError):
return None
return None
def _ts(node):
if isinstance(node, dict) and "time" in node:
try:
return pd.to_datetime(int(node["time"]), unit="s", utc=True)
except (TypeError, ValueError):
return None
return None
out["temp_f"] = _val(data.get("outdoor", {}).get("temperature"))
out["humidity"] = _val(data.get("outdoor", {}).get("humidity"))
out["pressure_inhg"] = _val(data.get("pressure", {}).get("relative"))
out["rain_in_hr"] = _val(data.get("rainfall_piezo", {}).get("rain_rate"))
out["last_ts"] = _ts(data.get("outdoor", {}).get("temperature"))
return out
@cached(CACHE_TTL_SECONDS)
def fetch_nws(horizon_hours: int) -> pd.DataFrame:
lat = float(os.environ["LAT"])
lon = float(os.environ["LON"])
return nws.hourly_forecast_df(lat, lon, hours=horizon_hours)
def _resample_hours(resample: str) -> float:
return pd.to_timedelta(resample).total_seconds() / 3600.0
def _resample_nws_to(nws_df: pd.DataFrame, resample: str) -> pd.DataFrame:
"""NWS gives hourly periods. For coarser cadences, average. For finer
(e.g. 30-min), forward-fill. Either way return on a regular index."""
if nws_df.empty:
return nws_df
target_h = _resample_hours(resample)
if target_h >= 1:
return nws_df.resample(resample).mean(numeric_only=True)
# Sub-hourly: upsample with forward-fill on the numeric columns.
return nws_df.select_dtypes("number").resample(resample).ffill()
# --- main refresh ---------------------------------------------------------
def _build_view(view: dict, log_conn, log_to_scoreboard: bool) -> dict:
"""Fetch + forecast for one view config. Returns intermediate pieces so
the caller can stitch the page together."""
cycle_type = view["cycle_type"]
display_resample = view["resample"]
forecast_resample = view.get("forecast_resample", display_resample)
display_step_hours = _resample_hours(display_resample)
forecast_step_hours = _resample_hours(forecast_resample)
horizon_hours = view["horizon_hours"]
horizon_steps = max(1, int(round(horizon_hours / forecast_step_hours)))
# context_hours = what we feed Toto; display_hours = what we show on
# the chart. Fall back to old keys for backward compatibility.
context_hours = view.get(
"context_hours",
view.get("context_days", view.get("history_days", 0)) * 24
or view.get("history_hours", 0),
)
display_hours = view.get(
"display_hours",
view.get("display_days", view.get("history_days", 0)) * 24
or view.get("history_hours", 0),
)
history = fetch_history(cycle_type, display_resample, context_hours)
# Coarser series for Toto inference: keeps the input length and
# forecast horizon short enough for the 4M model to predict cleanly,
# while the chart still shows the full 5-min granularity.
if forecast_resample != display_resample:
history_for_toto = history.resample(forecast_resample).mean()
else:
history_for_toto = history
nws_df_raw = fetch_nws(horizon_hours)
nws_df = _resample_nws_to(nws_df_raw, display_resample)
last_actual = history.dropna(how="all").index.max()
nws_future = nws_df[nws_df.index > last_actual] if last_actual is not None else nws_df
if log_to_scoreboard:
forecast_log.record_actuals(log_conn, history)
totos: dict[str, object] = {}
nws_aligned: dict[str, pd.Series] = {}
for m in METRICS:
series = history_for_toto[m["col"]].dropna()
if series.empty:
continue
toto = forecast_series(series, horizon=horizon_steps)
totos[m["col"]] = toto
if log_to_scoreboard:
forecast_log.record_toto(log_conn, m["col"], toto)
if m["nws_col"] and m["nws_col"] in nws_future.columns:
ns = nws_future[m["nws_col"]].dropna()
nws_aligned[m["col"]] = ns
if log_to_scoreboard:
forecast_log.record_nws(log_conn, m["col"], ns)
now = pd.Timestamp.now(tz="UTC").floor(display_resample)
visible_steps = int(round(display_hours / display_step_hours))
visible_history = history.tail(visible_steps)
fig = combined_figure(
history=visible_history,
totos=totos,
nws_df=nws_future,
metrics=METRICS,
now=now,
)
return {
"fig": fig,
"history": history,
"totos": totos,
"nws_aligned": nws_aligned,
"nws_df_raw": nws_df_raw,
}
def refresh():
realtime = fetch_realtime_snapshot()
log_conn = forecast_log.connect()
week = _build_view(VIEW_WEEK, log_conn, log_to_scoreboard=True)
# Hero uses the weekly history + the NWS period containing "now".
nws_df_raw = week["nws_df_raw"]
now_utc = pd.Timestamp.now(tz="UTC")
if not nws_df_raw.empty:
covering = nws_df_raw[nws_df_raw.index <= now_utc]
nws_first = covering.tail(1) if not covering.empty else nws_df_raw.head(1)
else:
nws_first = None
hero = hero_markdown(
PLACE_NAME, week["history"], nws_first, DISPLAY_TZ,
realtime=realtime,
toto_temp=week["totos"].get("temp_f"),
nws_temp=week["nws_aligned"].get("temp_f"),
horizon_hours=1,
)
if "temp_f" in week["totos"]:
comparison_md = aligned_comparison_markdown(
toto=week["totos"]["temp_f"],
nws_temp=week["nws_aligned"].get("temp_f"),
tz=DISPLAY_TZ,
)
else:
comparison_md = ""
scoreboard = render_scoreboard(log_conn)
# Residual chart β 1 h-ahead, last 48h, matches the first row of the
# rolling scoreboard.
resid_df = forecast_log.residuals(log_conn, metric="temp_f", window_hours=48, lag_hours=1.0)
resid_fig = residual_figure(resid_df) if not resid_df.empty else None
persist.push_db_async()
return hero, comparison_md, week["fig"], scoreboard, resid_fig
# --- scoreboard ----------------------------------------------------------
SCOREBOARD_HORIZONS_H = [1, 3, 12]
# Pressure has no NWS counterpart, so the head-to-head scoreboard skips it.
SCOREBOARD_METRICS = [
("temp_f", "Temperature", "Β°F"),
("humidity", "Humidity", "%"),
]
def _per_metric_mae_tables(conn, window_hours: int | None):
"""For each scored metric, return a Markdown table of per-lookahead MAE
over the requested window. window_hours=None means lifetime."""
parts: list[str] = []
any_data = False
for metric, label, unit in SCOREBOARD_METRICS:
rows: list[str] = []
for lag_h in SCOREBOARD_HORIZONS_H:
df = forecast_log.scoreboard_at_lag(
conn, metric=metric, lag_hours=lag_h, window_hours=window_hours,
)
if df.empty:
continue
any_data = True
by = {r["source"]: r for _, r in df.iterrows()}
toto = by.get("toto")
nws_row = by.get("nws")
t_cell = f"**{toto['mae']:.2f} {unit}** _(n={int(toto['n'])})_" if toto is not None else "β"
n_cell = f"**{nws_row['mae']:.2f} {unit}** _(n={int(nws_row['n'])})_" if nws_row is not None else "β"
if toto is not None and nws_row is not None:
diff = toto["mae"] - nws_row["mae"]
winner = "π€ Toto" if diff < 0 else "π NWS"
d_cell = f"**{winner}** by {abs(diff):.2f} {unit}"
else:
d_cell = "β"
rows.append(f"| **{lag_h}h-ahead** | {t_cell} | {n_cell} | {d_cell} |")
if rows:
parts.append(
"\n".join(
[
f"**{label}**",
"",
"| Lookahead | π€ Toto MAE | π NWS MAE | Ξ |",
"|---|---|---|---|",
*rows,
]
)
)
return parts, any_data
def render_scoreboard(conn) -> str:
"""Two MAE tables β rolling 48 h and lifetime β each with per-metric
rows for 1 h / 3 h / 12 h forecast lookahead."""
started_row = conn.execute(
"SELECT MIN(forecast_made_at) FROM forecast_snapshots"
).fetchone()
started_unix = started_row[0] if started_row and started_row[0] else None
started_str = (
datetime.fromtimestamp(started_unix, tz=timezone.utc)
.astimezone(__import__("zoneinfo").ZoneInfo(DISPLAY_TZ))
.strftime("%b %-d, %Y %-I:%M %p %Z")
if started_unix else "β"
)
rolling_parts, rolling_has = _per_metric_mae_tables(conn, window_hours=48)
lifetime_parts, lifetime_has = _per_metric_mae_tables(conn, window_hours=None)
lines = [
"### Rolling 48h MAE β lower is better",
(
"<span style='opacity:0.6'>**n** = number of past hours scored in the rolling 48h window.</span>"
),
]
if rolling_has:
lines.extend(rolling_parts)
else:
lines.append("_No scored forecasts yet in the last 48h._")
lines.append("### π° Lifetime MAE")
lines.append(
"<span style='opacity:0.6'>"
f"Every forecast we've ever logged. Scoreboard started {started_str}."
"</span>"
)
if lifetime_has:
lines.extend(lifetime_parts)
else:
lines.append("_No scored forecasts yet._")
return "\n\n".join(lines)
# --- auto-refresh background thread --------------------------------------
ECOWITT_ARCHIVE_DB = ECOWITT_ARCHIVE_DB_PATH # alias
def _sync_archive_all_cycles() -> None:
"""Refresh the SQLite archive (data/ecowitt.db) for every cycle_type
so the local mirror of Ecowitt's storage stays current."""
try:
cfg = ecowitt.EcowittConfig.from_env()
except RuntimeError:
return
conn = storage.connect(ECOWITT_ARCHIVE_DB)
try:
for cycle in sync.CYCLES:
try:
sync.sync_cycle(cfg, conn, cycle, verbose=False)
except ecowitt.EcowittRateLimitError as err:
print(f"[autorefresh] rate-limited on {cycle.name}: {err} β skipping rest")
break
except Exception: # noqa: BLE001
print(f"[autorefresh] sync error on {cycle.name}:")
traceback.print_exc()
finally:
conn.close()
def _sync_5min_only() -> None:
"""Quick sync of just the 5-min cycle (the one the display reads from).
Called on startup so the first visitor sees fresh data."""
try:
cfg = ecowitt.EcowittConfig.from_env()
except RuntimeError:
return
conn = storage.connect(ECOWITT_ARCHIVE_DB)
try:
cycle = next(c for c in sync.CYCLES if c.name == "5min")
sync.sync_cycle(cfg, conn, cycle, verbose=False)
except Exception: # noqa: BLE001
print("[startup] 5-min sync error:")
traceback.print_exc()
finally:
conn.close()
def _autorefresh_loop():
while True:
try:
_sync_archive_all_cycles() # bring the local archive up to date first
refresh() # read fresh data from DB, write forecast log
persist.push_all_async() # ship both DBs to the HF Dataset
except Exception: # noqa: BLE001
print("[autorefresh] error during refresh:")
traceback.print_exc()
time.sleep(AUTO_REFRESH_SECONDS)
def _start_autorefresh():
threading.Thread(target=_autorefresh_loop, daemon=True, name="autorefresh").start()
print(f"[autorefresh] started, interval={AUTO_REFRESH_SECONDS}s")
# --- UI -------------------------------------------------------------------
HOOK = (
"**Language models predict the next token. "
"What if you could predict the future with the same technology?**"
)
SUBTITLE = (
"Live readings from my Ecowitt GW3000 gateway + WS90 7-in-1 sensor, plus a probabilistic forecast "
"from [Datadog's Toto 2.0 (22M)](https://huggingface.co/Datadog/Toto-2.0-22m), compared against "
"the [NWS hourly forecast](https://www.weather.gov/documentation/services-web-api), "
"with a scoreboard tracking who's been more accurate."
)
SYSTEM_FONT = [
gr.themes.GoogleFont("Inter"),
"system-ui", "-apple-system", "BlinkMacSystemFont",
"Segoe UI", "Roboto", "Helvetica", "Arial", "sans-serif",
]
HEADER_CSS = """
.gradio-container h1 { font-size: 2.2rem !important; line-height: 1.1; margin-bottom: 0.3em; }
.gradio-container h2 {
font-size: 1.65rem !important;
margin-top: 1.6em; margin-bottom: 0.4em;
padding-top: 0.6em; padding-bottom: 0.2em;
border-top: 1px solid #e3e3e3;
}
.gradio-container h3 {
font-size: 1.15rem !important;
margin-top: 0.8em; margin-bottom: 0.3em;
opacity: 0.9;
}
/* Make the two collapsible explainer titles read like real section
headings instead of small UI chrome. */
.gradio-container .label-wrap > button,
.gradio-container .accordion > .label-wrap,
.gradio-container details > summary {
font-size: 1.1rem !important;
font-weight: 600 !important;
}
"""
with gr.Blocks(title="Toto Weather Forecast") as demo:
gr.Markdown("# Toto on my home weather station")
gr.Markdown(HOOK)
gr.Markdown(SUBTITLE)
gr.Markdown("## βοΈ Live now")
hero_md = gr.Markdown()
gr.Markdown("## π
Weekly forecast")
gr.Markdown(f"### {VIEW_WEEK['label']}")
week_plot = gr.Plot(label="Weekly")
gr.Markdown("## π Toto vs NWS β near-term forecast")
comparison_md = gr.Markdown()
gr.HTML(
# KOKX is the NWS radar site at Upton, NY β covers Long Island incl.
# Westhampton Beach. The looped GIF refreshes itself on the browser
# so the map stays live without us doing anything.
'<div style="text-align:center;margin:0.5em 0;">'
'<a href="https://radar.weather.gov/station/kokx/standard" target="_blank" rel="noopener">'
'<img src="https://radar.weather.gov/ridge/standard/KOKX_loop.gif" '
'alt="NWS radar loop, Long Island (KOKX)" '
'style="max-width:560px;width:100%;border-radius:6px;border:1px solid #e0e0e0;" />'
'</a>'
'<div style="opacity:0.55;font-size:0.85em;margin-top:0.3em;">'
'NWS radar loop Β· station KOKX (Upton, NY) Β· refreshes ~6 min'
'</div>'
'</div>'
)
gr.Markdown(
"<span style='opacity:0.55'>π Live data + forecast auto-refresh every 15 minutes.</span>"
)
gr.Markdown("## π Results")
residual_plot = gr.Plot(label="Forecast residual")
scoreboard_md = gr.Markdown()
gr.Markdown("## π§ How it's made")
gr.Image(
"assets/ws90.jpeg",
show_label=False,
container=False,
interactive=False,
height=380,
)
gr.Markdown(
"<div style='text-align:center;opacity:0.6;font-size:0.9em;"
"margin-top:-0.3em;margin-bottom:0.8em;'>"
"my ecowitt weather station β Ecowitt GW3000 gateway + WS90 7-in-1 sensor"
"</div>"
)
with gr.Accordion("How the scoreboard is calculated", open=False):
gr.Markdown(
"We score each model on **how close its prediction was to the actual Ecowitt reading** "
"for the same hour, at three fixed forecast lookaheads β **1 h, 3 h, and 12 h ahead** "
"β averaged over the rolling last 48h, plus a lifetime view over every snapshot since "
"the scoreboard started.\n\n"
"**Picking which forecast counts.** Every refresh logs both models' forecasts for the "
"next 48h along with `forecast_made_at` and `target_ts`. For each past target hour "
"and each lookahead N, we pick the forecast whose `forecast_made_at` is closest to "
"`target_ts β N hours`. Same lookahead for both models = fair comparison.\n\n"
"**The math.** For each metric (temperature, humidity), per source, per lookahead:\n\n"
" `abs_err = |p50 β actual|`\n\n"
" `MAE = mean(abs_err)` over target hours in the window (48h or lifetime)\n\n"
" `n` = number of past target hours with both a forecast and an Ecowitt actual\n\n"
"The lower MAE wins. Barometric pressure is omitted from the scoreboard because NWS "
"doesn't expose a pressure forecast β there's nothing to compare against.\n\n"
"**Residual chart** (above the rolling table). Same picking rule, pinned to the "
"**1h-ahead** lookahead. Each point is `prediction β actual`; zero = perfect.\n\n"
"**What this is NOT.** We score the point prediction (p50) β which throws away Toto's "
"uncertainty. A scoring rule like CRPS or pinball loss would credit a well-calibrated "
"10β90% band; MAE doesn't.\n\n"
"Full spec: [`docs/toto-inference.md`](https://huggingface.co/spaces/bitsofchris/time-series-ai-weather-forecast/blob/main/docs/toto-inference.md#scoreboard--how-the-accuracy-is-calculated)."
)
with gr.Accordion("How the forecast is made", open=False):
gr.Markdown(
"**Model.** [Datadog/Toto-2.0-22m](https://huggingface.co/Datadog/Toto-2.0-22m) "
"(~22 M params, CPU). Second-smallest variant of Toto 2.0; the larger 313 M / 1 B / "
"2.5 B models would tighten the band further.\n\n"
"**Input.** Univariate per metric β temperature, humidity, pressure, rain rate run "
"independently. The Space pulls Ecowitt's `cycle_type=5min` history into a local "
"SQLite archive (`data/ecowitt.db`) every 15 min and accumulates true 5-min cadence "
"over time. The **chart** displays the 5-min series; **Toto** is fed the same series "
"downsampled to hourly so the input length stays in the model's sweet spot.\n\n"
"**Context length.** 7 days Γ hourly = up to 168 points. Toto requires the context "
"to be a multiple of its `patch_size` (read from `model.config`), so we truncate the "
"oldest points to the largest multiple that fits β or, if we have fewer points than "
"one patch, left-pad and set `target_mask=False` on the padding so the model ignores it.\n\n"
"**Output.** `model.forecast(...)` returns 9 analytical quantiles "
"(`[0.1, 0.2, β¦, 0.9]`) for each future step β **no Monte-Carlo sampling**. "
"We plot the p10βp90 band and the p50 median.\n\n"
"**Horizon.** 48 hourly steps = 48 hours into the future. Per-metric inference takes "
"well under a second on the free-tier CPU.\n\n"
"**Cadence.** A daemon thread inside the Space runs sync β inference β push every "
"15 minutes; a 10-minute GitHub Actions cron pings the public URL to keep the Space "
"warm. Both DBs (forecasts + raw archive) are backed up to a private HF Dataset "
"(`bitsofchris/toto-weather-forecast-log`), so the scoreboard survives Space "
"rebuilds.\n\n"
"Full spec: [`docs/toto-inference.md`](https://huggingface.co/spaces/bitsofchris/time-series-ai-weather-forecast/blob/main/docs/toto-inference.md)."
)
outputs = [hero_md, comparison_md, week_plot, scoreboard_md, residual_plot]
demo.load(refresh, outputs=outputs)
if __name__ == "__main__":
persist.pull_all() # bootstrap forecast log + archive from the HF Dataset
_sync_5min_only() # ensure the archive has fresh 5-min data before first paint
_start_autorefresh()
demo.launch(theme=gr.themes.Default(font=SYSTEM_FONT), css=HEADER_CSS)
|