Spaces:
Sleeping
Sleeping
File size: 8,166 Bytes
e317359 | 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 | from __future__ import annotations
from dataclasses import dataclass
from typing import Any, Iterable
import numpy as np
import pandas as pd
from gluonts.dataset.common import ListDataset
from gluonts.dataset.split import split
from gluonts.model import Predictor, evaluate_model
from gluonts.time_feature import get_seasonality
from tsfm_bench.data.interface import TimeSeriesDataSource
from tsfm_bench.data.open_meteo import OpenMeteoDataSource
from tsfm_bench.data.ts_bench import TsBenchDataSource
from tsfm_bench.eval.metrics import RESULT_COLUMNS, get_eval_metrics
@dataclass
class OnlineEvalResult:
dataset: str
model: str
metrics: dict[str, float]
domain: str
num_variates: int
data_fetched_at: str
context_length: int
prediction_length: int
forecast_snapshot: dict | None = None
def _records_to_entries(
record,
to_univariate: bool,
) -> list[dict[str, Any]]:
target = np.asarray(record.target, dtype=np.float32)
start = pd.Period(record.start, freq=record.freq)
entries: list[dict[str, Any]] = []
if target.ndim == 1 or not to_univariate:
entries.append(
{
"item_id": record.item_id,
"start": start,
"target": target if target.ndim == 1 else target,
"freq": record.freq,
}
)
return entries
for idx in range(target.shape[0]):
entries.append(
{
"item_id": f"{record.item_id}_v{idx}",
"start": start,
"target": target[idx],
"freq": record.freq,
}
)
return entries
def resolve_seasonality(freq: str, series_length: int, prediction_length: int) -> int:
"""Cap seasonality so MASE/MSIS scaling is defined on short live windows."""
nominal = int(get_seasonality(freq))
train_length = series_length - prediction_length
# Need at least two seasonal periods in the training window.
capped = min(nominal, max(1, train_length // 2))
return max(1, capped)
def evaluate_online_window(
predictor: Predictor,
entries: list[dict[str, Any]],
prediction_length: int,
freq: str,
) -> dict[str, float]:
dataset = ListDataset(entries, freq=freq)
_, test_template = split(dataset, offset=-prediction_length)
test_data = test_template.generate_instances(
prediction_length=prediction_length,
windows=1,
distance=prediction_length,
)
series_length = int(np.asarray(entries[0]["target"]).shape[-1])
seasonality = resolve_seasonality(freq, series_length, prediction_length)
evaluated = evaluate_model(
predictor,
test_data=test_data,
metrics=get_eval_metrics(),
batch_size=64,
axis=None,
mask_invalid_label=True,
allow_nan_forecast=True,
seasonality=seasonality,
)
result = {key: float(value.iloc[0]) for key, value in evaluated.items()}
return result
def _capture_forecast_snapshot(
predictor: Predictor,
entries: list[dict],
pred_len: int,
freq: str,
evaluated_at: str,
dataset: str,
model: str,
) -> dict | None:
"""Call predictor once to capture p10/p50/p90 and actuals for visualization.
Uses Option A: one extra API call per dataset, minimal code invasiveness.
Returns None silently if anything fails (snapshot is best-effort).
"""
try:
ds = ListDataset(entries, freq=freq)
_, test_template = split(ds, offset=-pred_len)
test_data = list(test_template.generate_instances(
prediction_length=pred_len, windows=1, distance=pred_len,
))
if not test_data:
return None
input_entry, label_entry = test_data[0]
context_vals = np.asarray(input_entry["target"], dtype=np.float64).flatten().tolist()
actuals = np.asarray(label_entry["target"], dtype=np.float64).flatten().tolist()
input_ds = ListDataset([input_entry], freq=freq)
forecasts = list(predictor.predict(input_ds))
if not forecasts:
return None
fc = forecasts[0]
fallback = fc.mean.tolist()
quantile_levels = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]
quantiles = {
f"{level:g}": (
fc.quantile(level).tolist() if hasattr(fc, "quantile") else fallback
)
for level in quantile_levels
}
p10 = quantiles["0.1"]
p50 = quantiles["0.5"]
p90 = quantiles["0.9"]
n_context = len(context_vals)
n_future = len(actuals)
timestamps = []
try:
import pandas as pd
start_period = input_entry["start"]
dr = pd.period_range(start=start_period, periods=n_context + n_future, freq=freq)
timestamps = [str(p.to_timestamp()) for p in dr]
except Exception:
pass
return {
"dataset": dataset,
"model": model,
"evaluated_at": evaluated_at,
"freq": freq,
"context_length": n_context,
"prediction_length": n_future,
"context": context_vals,
"actuals": actuals,
"p50": p50,
"p10": p10,
"p90": p90,
"mean": fallback,
"quantiles": quantiles,
"timestamps": timestamps,
}
except Exception:
return None
def run_online_eval_for_dataset(
source: TimeSeriesDataSource,
ds_name: str,
predictor: Predictor,
prediction_length: int | None = None,
) -> OnlineEvalResult:
meta = source.get_metadata(ds_name)
record = next(iter(source.stream(ds_name)))
if isinstance(source, OpenMeteoDataSource):
pred_len = prediction_length or source._settings.prediction_length
fetched_at = source.fetched_at()
elif isinstance(source, TsBenchDataSource):
pred_len = prediction_length or source.get_prediction_length(ds_name)
fetched_at = source.fetched_at()
else:
pred_len = prediction_length or 24
fetched_at = pd.Timestamp.utcnow().isoformat()
to_univariate = meta.num_variates > 1
entries = _records_to_entries(record, to_univariate=to_univariate)
metrics = evaluate_online_window(predictor, entries, pred_len, meta.frequency)
ds_key = ds_name.split("/")[0].lower()
if isinstance(source, TsBenchDataSource):
config_name = source.leaderboard_dataset_name(ds_name)
else:
config_name = f"{ds_key}/{meta.frequency}/short"
model_name = getattr(predictor, "leaderboard_name", predictor.__class__.__name__)
context_length = int(np.asarray(entries[0]["target"]).shape[-1]) - pred_len
evaluated_at = pd.Timestamp.utcnow().isoformat()
forecast_snapshot = _capture_forecast_snapshot(
predictor,
entries,
pred_len,
meta.frequency,
evaluated_at,
config_name,
model_name,
)
return OnlineEvalResult(
dataset=config_name,
model=model_name,
metrics=metrics,
domain=meta.domain,
num_variates=1 if to_univariate else meta.num_variates,
data_fetched_at=fetched_at,
context_length=context_length,
prediction_length=pred_len,
forecast_snapshot=forecast_snapshot,
)
def result_to_row(result: OnlineEvalResult) -> list[Any]:
return [
result.dataset,
result.model,
result.metrics["MSE[mean]"],
result.metrics["MSE[0.5]"],
result.metrics["MAE[0.5]"],
result.metrics["MASE[0.5]"],
result.metrics["MAPE[0.5]"],
result.metrics["sMAPE[0.5]"],
result.metrics["MSIS"],
result.metrics["RMSE[mean]"],
result.metrics["NRMSE[mean]"],
result.metrics["ND[0.5]"],
result.metrics["mean_weighted_sum_quantile_loss"],
result.domain,
result.num_variates,
]
def results_to_dataframe(results: Iterable[OnlineEvalResult]) -> pd.DataFrame:
rows = [result_to_row(result) for result in results]
return pd.DataFrame(rows, columns=RESULT_COLUMNS)
|