File size: 22,075 Bytes
bdd9175 | 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 | """GIFT-Eval gluonts predictor for TinyCast.
``TinyCastPredictor`` is the deployed predictor. It wraps the model in the
gluonts Predictor protocol and drives autoregressive-rollout decoding
(48-step chunks), flip-invariance symmetrization, NaN-imputation and
optional period-alignment downsampling.
Anything that changes what the predictor emits says so:
``device`` a named device must exist. ``device=None`` is the only
request that selects one for you (CUDA when present, else
CPU); the resolved device is ``predictor.device``.
``TINYCAST_INT8`` ``w8`` or ``w8a8`` post-training fake quantization, off by
default. Any other non-empty value is an error, not an
inert setting.
``TINYCAST_TILT_K`` an eval-only probe that moves the emitted median off the
quantile grid, off (``0``) by default, with
``TINYCAST_TILT_MODE`` in {``adaptive``, ``fixed``}.
The environment variables are read once, when the predictor is constructed, and
an engaged one prints a line naming itself. Reproducing the published numbers
means leaving them unset.
"""
from __future__ import annotations
import math
import os
from typing import List, Optional, Tuple
import numpy as np
import torch
import torch.nn as nn
from .checkpoint import load_checkpoint
from .normalization import WindowMinMax
from .scale import seasonal_scale_factor
try:
from torch.amp import autocast as _autocast_fp
except Exception: # pragma: no cover
_autocast_fp = None
def _resolve_device(requested: Optional[str]) -> torch.device:
"""Return the requested device, or say why it is unavailable.
``None`` is the request to choose: CUDA when it is available, CPU otherwise.
Every other value is a requirement. Falling back to CPU behind the caller's
back drops bf16 autocast for fp32 and moves every forecast by roughly 1e-3,
which is small enough to read as a regression and large enough to fail an
assertion against the published aggregates, so an absent accelerator is an
error rather than a substitution.
"""
if requested is None:
return torch.device("cuda" if torch.cuda.is_available() else "cpu")
dev = torch.device(requested)
if dev.type == "cuda":
if not torch.cuda.is_available():
raise RuntimeError(
f"device={requested!r} was requested but "
"torch.cuda.is_available() is False. Pass device='cpu' to run "
"on CPU (fp32, and about 1e-3 off the published numbers), or "
"device=None to take whichever device is present."
)
if dev.index is not None and dev.index >= torch.cuda.device_count():
raise RuntimeError(
f"device={requested!r} was requested but only "
f"{torch.cuda.device_count()} CUDA device(s) are visible."
)
if dev.type == "mps" and not torch.backends.mps.is_available():
raise RuntimeError(
f"device={requested!r} was requested but MPS is unavailable. "
"Pass device='cpu' to run on CPU."
)
return dev
_INT8_MODES = ("w8", "w8a8")
_INT8_OFF = ("", "0", "off", "no", "false", "none")
_TILT_MODES = ("adaptive", "fixed")
def _resolve_int8() -> Optional[str]:
"""Read ``TINYCAST_INT8``: the quantization mode, or ``None`` when off.
Only ``w8`` and ``w8a8`` name a scheme. A truthy value such as ``1`` names
no scheme, and picking one for the caller would quantize a run that asked
for something else, so an unrecognized value raises. The alternative is what
this used to do: leave the run in fp32 while the caller believes they are
measuring INT8.
"""
raw = os.environ.get("TINYCAST_INT8", "").strip().lower()
if raw in _INT8_OFF:
return None
if raw not in _INT8_MODES:
raise ValueError(
f"TINYCAST_INT8={raw!r} is not a quantization mode. Use 'w8' "
"(per-channel INT8 weights) or 'w8a8' (+ per-tensor dynamic INT8 "
"activations), or unset it to run in floating point."
)
return raw
def _resolve_tilt() -> Tuple[float, str]:
"""Read the tilt probe's variables, announcing an engaged one.
``TINYCAST_INT8`` prints a line when it engages; this does the same, so
neither switch can invalidate a golden array without appearing in the log.
``TINYCAST_TILT_MODE`` is checked only when the tilt is on, which keeps it
inert at ``K=0`` as documented.
"""
raw = os.environ.get("TINYCAST_TILT_K", "").strip()
try:
k = float(raw) if raw else 0.0
except ValueError:
raise ValueError(
f"TINYCAST_TILT_K={raw!r} is not a number. Unset it, or set 0, to "
"emit the model's own median."
) from None
mode = os.environ.get("TINYCAST_TILT_MODE", "adaptive").strip().lower()
if k != 0.0:
if mode not in _TILT_MODES:
raise ValueError(
f"TINYCAST_TILT_MODE={mode!r} is not a tilt rule. Use "
"'adaptive' (skew-following) or 'fixed'."
)
print(
f"[tilt] TINYCAST_TILT_K={k:g} {mode}: the emitted median is "
"re-interpolated off the quantile grid, so this is not the "
"published model's forecast (inert below 3 quantiles)",
flush=True,
)
return k, mode
def _numpy_fill(arr: np.ndarray) -> np.ndarray:
mask = np.isnan(arr)
idx = np.where(~mask, np.arange(mask.shape[1]), 0)
np.maximum.accumulate(idx, axis=1, out=idx)
return arr[np.arange(idx.shape[0])[:, None], idx]
class ARRolloutPredictor:
"""Base predictor: AR rollout + flip + NaN-imputation + period downsample.
Subclasses install ``self.model`` (a callable ``(x, x_mark, y_mark)``).
``device=None`` selects one; a named device must exist. Either way the
device that ran is ``self.device``.
Every setting that changes a forecast is a named parameter, so a keyword
that is not one raises. Absorbing unknown keywords would let a misspelling
such as ``force_flip_invarience=True`` construct a predictor with flip
symmetrization off and return a forecast that looks right and cannot be
reproduced.
"""
def __init__(
self,
prediction_length: int,
device: Optional[str] = None,
seq_len: int = 2048,
input_token_len: int = 2048,
output_token_len: int = 48,
num_samples: int = 100,
batch_size: int = 256,
use_amp: int = 1,
downsample_factor: int = 1,
force_flip_invariance: bool = False,
adaptive_ar_rollout: bool = False,
):
self.device = _resolve_device(device)
self.prediction_length = int(prediction_length)
self.num_samples = int(num_samples)
self.batch_size = int(batch_size)
self.seq_len = int(seq_len)
self.input_token_len = int(input_token_len)
self.output_token_len = int(output_token_len)
self.use_amp = int(use_amp)
self.downsample_factor = int(downsample_factor)
self.force_flip_invariance = bool(force_flip_invariance)
self.adaptive_ar_rollout = bool(adaptive_ar_rollout)
# Quantile levels emitted by the model (set by subclasses from cfg).
self.quantiles = [0.5]
self.model = None
def _downsample_if_needed(
self, series: torch.Tensor
) -> Tuple[torch.Tensor, int]:
cur = series
if self.downsample_factor > 1:
cur = cur[::self.downsample_factor]
return cur, self.downsample_factor
def _left_pad_to_len(
self, arr: np.ndarray, target_len: int
) -> Tuple[np.ndarray, int]:
if arr.shape[0] >= target_len:
return arr[-target_len:], 0
pad_len = target_len - arr.shape[0]
fill_value = arr[0] if arr.shape[0] > 0 else 0.0
padding = np.full((pad_len,), fill_value, dtype=arr.dtype)
return np.concatenate([padding, arr], axis=0), pad_len
def _prepare_context_matrix(
self, context: List[torch.Tensor]
) -> Tuple[torch.Tensor, List[int]]:
xs = []
downsample_factors = []
for c in context:
cur, df = self._downsample_if_needed(c)
downsample_factors.append(df)
cur_np = cur.detach().cpu().float().numpy()
cur_np, _ = self._left_pad_to_len(cur_np, self.seq_len)
x2d = cur_np[None, :]
x_interp = np.copy(x2d)
series = x2d[0]
if np.any(np.isnan(series)):
valid_mask = ~np.isnan(series)
if np.sum(valid_mask) >= 2:
valid_idx = np.where(valid_mask)[0]
valid_val = series[valid_mask]
x_interp[0] = np.interp(
np.arange(len(series)), valid_idx, valid_val
)
else:
x_interp = _numpy_fill(x2d)
ff = _numpy_fill(x_interp)
bf = np.flip(_numpy_fill(np.flip(x_interp, axis=1)), axis=1)
x_imp = np.where(np.isnan(ff), bf, ff)
x_imp = np.where(np.isnan(x_imp), 0.0, x_imp)
xs.append(x_imp[0])
x = torch.tensor(
np.stack(xs), device=self.device, dtype=torch.float32
).unsqueeze(-1)
return x, downsample_factors
def _decode_autoregressive(
self,
init_ctx: torch.Tensor,
use_bf16: bool,
downsample_factors: List[int],
) -> torch.Tensor:
B, _, C = init_ctx.shape
roll_len = int(self.output_token_len)
if self.adaptive_ar_rollout:
try:
from .periodogram import significant_periods
periods, _s, _nv = significant_periods(
init_ctx[:, -self.seq_len:, 0].float(),
min_period=2, max_period=self.seq_len // 2, top_k=1,
)
pos = periods[periods > 0]
if pos.numel() > 0:
p0 = int(pos.float().median().item())
if 2 <= p0 < roll_len:
roll_len = max(4, p0)
except Exception:
pass
target_pred_lens = [
int(self.prediction_length) // int(max(1, df))
for df in downsample_factors
]
max_target_pred_len = max(target_pred_lens)
steps = math.ceil(max_target_pred_len / roll_len)
preds: List[torch.Tensor] = []
batch_ctx = init_ctx
y_mark = torch.zeros(
B, self.output_token_len, C,
device=self.device, dtype=init_ctx.dtype,
)
for _ in range(steps):
x_in = batch_ctx[:, -self.seq_len:, :]
x_mark = torch.zeros_like(x_in)
if _autocast_fp is not None and self.use_amp and use_bf16:
try:
with _autocast_fp("cuda", dtype=torch.bfloat16):
out = self.model(x_in, x_mark, y_mark)
except Exception:
out = self.model(x_in, x_mark, y_mark)
else:
out = self.model(x_in, x_mark, y_mark)
chunk = out[:, -self.output_token_len:, :][:, :roll_len, :] # (B, roll, Q)
preds.append(chunk)
# Feed only the MEDIAN quantile back into the context (the AR state
# is a univariate series); keep all Q in preds for the forecast.
q_mid = chunk.shape[-1] // 2
batch_ctx = torch.cat([batch_ctx, chunk[:, :, q_mid:q_mid + 1]], dim=1)
return torch.cat(preds, dim=1) # (B, pl, Q)
@torch.no_grad()
def predict(self, test_data_input, use_bf16_if_available: bool = True):
from gluonts.itertools import batcher
from gluonts.model.forecast import SampleForecast, QuantileForecast
forecasts: List = []
use_bf16 = bool(
use_bf16_if_available
and self.device.type == "cuda"
and torch.cuda.is_available()
and torch.cuda.is_bf16_supported()
)
for batch in batcher(test_data_input, batch_size=self.batch_size):
targets = [
torch.tensor(entry["target"], dtype=torch.float32)
for entry in batch
]
batch_ctx, dfs = self._prepare_context_matrix(targets)
pred_pos = self._decode_autoregressive(batch_ctx, use_bf16, dfs) # (B,pl,Q)
if self.force_flip_invariance:
pred_neg = self._decode_autoregressive(-batch_ctx, use_bf16, dfs)
# Flip-symmetrize. For quantiles the tau-quantile of -y is
# -(the (1-tau)-quantile of y), so reverse the quantile axis on
# the negated branch. Q=1 (median) reverse is a no-op.
pred = 0.5 * (pred_pos - pred_neg.flip(dims=[-1]))
else:
pred = pred_pos
Q = pred.shape[-1]
pred_np = pred.float().detach().cpu().numpy() # (B, pl, Q)
if not np.isfinite(pred_np).all():
for qi in range(Q):
pred_np[:, :, qi] = _numpy_fill(pred_np[:, :, qi])
for i, ts in enumerate(batch):
df = int(max(1, dfs[i]))
target_pl = int(self.prediction_length) // df
arr = pred_np[i, :target_pl, :] # (target_pl, Q)
if df > 1:
new_len = int(self.prediction_length)
src = np.linspace(0, 1, arr.shape[0])
dst = np.linspace(0, 1, new_len)
arr = np.stack([np.interp(dst, src, arr[:, qi])
for qi in range(Q)], axis=1) # (new_len, Q)
start_date = ts["start"] + len(ts["target"])
if Q > 1:
# Sort across quantiles to guarantee non-crossing, then a
# QuantileForecast so gluonts scores WQL/CRPS over the deciles.
arr = np.sort(arr, axis=1)
forecasts.append(QuantileForecast(
forecast_arrays=arr.T, # (Q, pl)
start_date=start_date,
forecast_keys=[str(q) for q in self.quantiles],
))
else:
samples = np.repeat(arr[:, 0][None, :], self.num_samples, axis=0)
forecasts.append(
SampleForecast(samples=samples, start_date=start_date)
)
return forecasts
class _BackboneAdapter(nn.Module):
"""Adapts the model wrapper to the ``(x, x_mark, y_mark) -> (B, p, Q)`` contract."""
def __init__(self, backbone: nn.Module, scale_factor: float = 1.0):
super().__init__()
self.backbone = backbone
self.scale_factor = float(scale_factor)
# single-shot counterfactual: emit this many steps in ONE forward.
self.pred_len_override = None
# Resolved once, and announced when on: a forecast must not change
# because a variable was exported after the predictor was built.
self.tilt_k, self.tilt_mode = _resolve_tilt()
def forward(self, x, x_mark=None, y_mark=None, **kwargs):
y_norm, x_min, x_range = self.backbone.encode(
x, batch_first=True, scale_factor=self.scale_factor,
horizon=self.pred_len_override,
)
# Match training: the chunk loss clamps y_norm to [-5,5] before the
# pinball loss, so the model is never optimized outside that band.
# Clamp at inference too: otherwise an un-penalized overshoot feeds
# back into the AR rollout context and compounds over chunks.
y_norm = y_norm.clamp(-5.0, 5.0)
# num_quantiles > 1 => y_norm is (B, p, Q). Manual inverse since
# WindowMinMax.inverse_transform squeezes x_min/x_range for the (B, p)
# point case; x_min/x_range are (B, 1, 1) so they broadcast directly.
if y_norm.dim() == 3:
tk = self.tilt_k
Q = y_norm.shape[-1]
if tk != 0.0 and Q >= 3:
# EVAL-ONLY de-hedge probe (no training); default off (tk=0).
qm = Q // 2
lo = y_norm[..., 0]; hi = y_norm[..., -1]; med = y_norm[..., qm]
if self.tilt_mode == "fixed":
tau = torch.full_like(med, 0.5 + tk)
else: # skew-adaptive
asym = (hi + lo - 2 * med) / (hi - lo).abs().clamp(min=1e-6)
tau = 0.5 + tk * torch.tanh(asym) # >0 right-skew
tau = tau.clamp(qm / (Q + 1.0), (qm + 2) / (Q + 1.0))
idxf = (tau * (Q + 1) - 1).clamp(0, Q - 1 - 1e-4)
ilo = idxf.floor().long().clamp(0, Q - 2)
frac = (idxf - ilo.to(idxf.dtype)).clamp(0, 1)
qlo = torch.gather(y_norm, -1, ilo.unsqueeze(-1)).squeeze(-1)
qhi = torch.gather(y_norm, -1, (ilo + 1).unsqueeze(-1)).squeeze(-1)
y_norm = y_norm.clone()
y_norm[..., qm] = qlo * (1 - frac) + qhi * frac
return y_norm * x_range + x_min # (B, p, Q)
y_pred = WindowMinMax.inverse_transform(y_norm, x_min, x_range)
return y_pred.unsqueeze(-1) # (B, p, 1)
class TinyCastPredictor(ARRolloutPredictor):
"""The deployed predictor: AR-rollout decoding around the TinyCast model."""
def __init__(
self,
prediction_length: int,
checkpoint_path: str,
device: Optional[str] = None,
num_samples: int = 100,
batch_size: int = 256,
use_amp: int = 1,
downsample_factor: int = 1,
force_flip_invariance: bool = False,
freq: Optional[str] = None,
domain: Optional[str] = None,
no_daily: bool = False,
single_shot: bool = False,
adaptive_ar_rollout: bool = False,
config_path: Optional[str] = None,
):
self.single_shot = bool(single_shot)
# 1. Load the weights (safetensors + config.json).
model, cfg = load_checkpoint(checkpoint_path, config_path)
# 2. Eval-time scale_factor from (freq, domain). bizitobs_l2c has no
# daily cycle -> /7.
if freq is not None and domain is not None:
sf_eval = seasonal_scale_factor(freq, domain)
if no_daily:
sf_eval /= 7
else:
sf_eval = 1.0
self._eval_scale_factor = float(sf_eval)
# 3. Base predictor bookkeeping (drives AR-rollout / batching / flip).
super().__init__(
prediction_length=prediction_length,
device=device,
seq_len=int(cfg.seq_len),
input_token_len=int(cfg.seq_len),
output_token_len=int(cfg.output_token_len),
num_samples=num_samples,
batch_size=batch_size,
use_amp=use_amp,
downsample_factor=downsample_factor,
force_flip_invariance=force_flip_invariance,
adaptive_ar_rollout=adaptive_ar_rollout,
)
self.quantiles = list(getattr(cfg, "quantiles", [0.5])) or [0.5]
self._missing_aware = bool(getattr(cfg, "missing_channel", False))
# 4. Install the model wrapped in the adapter interface.
backbone = model.model # TinyCastForPrediction -> TinyCastBackbone
backbone.to(self.device).eval()
# Optional INT8 post-training fake-quant. Off by default;
# TINYCAST_INT8=w8 (weights) or w8a8 (+ dynamic activations).
# quantize_int8_ prints the mode it applied.
_int8 = _resolve_int8()
if _int8 is not None:
from .quant import quantize_int8_
quantize_int8_(backbone, _int8)
self.model = _BackboneAdapter(
backbone, scale_factor=self._eval_scale_factor,
).to(self.device).eval()
if self.single_shot:
self.model.pred_len_override = int(prediction_length)
def _decode_autoregressive(self, init_ctx, use_bf16, downsample_factors):
"""Single-shot override: one forward for the full horizon (no AR loop).
Default deployed behavior (single_shot=False) falls back to the base
AR rollout."""
if not self.single_shot:
return super()._decode_autoregressive(init_ctx, use_bf16, downsample_factors)
x_in = init_ctx[:, -self.seq_len:, :]
x_mark = torch.zeros_like(x_in)
use_cuda = bool(self.use_amp and use_bf16 and str(self.device).startswith("cuda"))
if use_cuda:
with torch.autocast("cuda", dtype=torch.bfloat16):
out = self.model(x_in, x_mark, None)
else:
out = self.model(x_in, x_mark, None)
return out # (B, prediction_length, 1)
def _prepare_context_matrix(self, context):
"""Missing-aware override: keep genuine gaps as NaN so the observed-mask
+ observed-only normalization see true missingness. Falls back to the
base (interpolating) behavior when the model is not missing-aware
(the deployed model is not)."""
if not getattr(self, "_missing_aware", False):
return super()._prepare_context_matrix(context)
xs, dfs = [], []
for c in context:
cur, df = self._downsample_if_needed(c)
dfs.append(df)
a = cur.detach().cpu().float().numpy()
if a.shape[0] >= self.seq_len:
a = a[-self.seq_len:]
else:
pad = np.full((self.seq_len - a.shape[0],), np.nan, dtype=a.dtype)
a = np.concatenate([pad, a], axis=0)
xs.append(a)
x = torch.tensor(
np.stack(xs), device=self.device, dtype=torch.float32
).unsqueeze(-1)
return x, dfs
|