Upload compactts/agent/agent.py with huggingface_hub
Browse files- compactts/agent/agent.py +121 -0
compactts/agent/agent.py
ADDED
|
@@ -0,0 +1,121 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Autonomous Agentic Forecaster (CompactTS-Agent).
|
| 2 |
+
|
| 3 |
+
Combines real-time series profiling, local foundation backbones,
|
| 4 |
+
harmonic Fourier basis correction, and domain-calibrated arbitration.
|
| 5 |
+
"""
|
| 6 |
+
|
| 7 |
+
from typing import Dict, Any, List, Optional
|
| 8 |
+
import numpy as np
|
| 9 |
+
import pandas as pd
|
| 10 |
+
import torch
|
| 11 |
+
|
| 12 |
+
from gluonts.model.forecast import QuantileForecast
|
| 13 |
+
from gluonts.time_feature import get_seasonality
|
| 14 |
+
from compactts.agent.profiler import SeriesProfiler
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
class CompactTSAgent:
|
| 18 |
+
"""CompactTS Autonomous Agent for Time Series Forecasting.
|
| 19 |
+
|
| 20 |
+
Implements:
|
| 21 |
+
1. Autonomous Series Profiling (Seasonality, Trend, Kurtosis, Volatility)
|
| 22 |
+
2. Dynamic Context-Window Allocation
|
| 23 |
+
3. Harmonic Basis Augmentation for highly periodic regimes
|
| 24 |
+
4. Probabilistic Quantile Monotonicity Enforcement
|
| 25 |
+
5. Domain-Aware Arbitration Policy
|
| 26 |
+
"""
|
| 27 |
+
|
| 28 |
+
QUANTILES = ["0.1", "0.2", "0.3", "0.4", "0.5", "0.6", "0.7", "0.8", "0.9"]
|
| 29 |
+
|
| 30 |
+
def __init__(
|
| 31 |
+
self,
|
| 32 |
+
backbone_pipeline=None,
|
| 33 |
+
prediction_length: int = 24,
|
| 34 |
+
freq: str = "H",
|
| 35 |
+
batch_size: int = 128,
|
| 36 |
+
domain: Optional[str] = None,
|
| 37 |
+
):
|
| 38 |
+
self.pipeline = backbone_pipeline
|
| 39 |
+
self.H = prediction_length
|
| 40 |
+
self.freq = freq
|
| 41 |
+
self.batch_size = batch_size
|
| 42 |
+
self.domain = domain
|
| 43 |
+
self.profiler = SeriesProfiler(freq=freq)
|
| 44 |
+
self.S = max(self.profiler.base_seasonality, 1)
|
| 45 |
+
|
| 46 |
+
def forecast_batch(self, series_list: List[np.ndarray], start_dates: List[Any]) -> List[QuantileForecast]:
|
| 47 |
+
"""Generate calibrated probabilistic quantile forecasts for a batch of series."""
|
| 48 |
+
total = len(series_list)
|
| 49 |
+
forecasts = []
|
| 50 |
+
|
| 51 |
+
for i in range(0, total, self.batch_size):
|
| 52 |
+
batch_y = series_list[i : i + self.batch_size]
|
| 53 |
+
batch_starts = start_dates[i : i + self.batch_size]
|
| 54 |
+
batch_tensors = []
|
| 55 |
+
batch_profiles = []
|
| 56 |
+
|
| 57 |
+
for y_raw in batch_y:
|
| 58 |
+
prof = self.profiler.profile(y_raw)
|
| 59 |
+
batch_profiles.append(prof)
|
| 60 |
+
|
| 61 |
+
y = np.asarray(y_raw, dtype=np.float32)
|
| 62 |
+
if np.isnan(y).any() or np.isinf(y).any():
|
| 63 |
+
s = pd.Series(y).ffill().bfill().to_numpy()
|
| 64 |
+
y = np.nan_to_num(s, nan=0.0).astype(np.float32)
|
| 65 |
+
|
| 66 |
+
# Agentic Dynamic Context Window:
|
| 67 |
+
# Allocates 2*S up to 512 context steps
|
| 68 |
+
ctx_len = min(len(y), max(2 * self.S, 512))
|
| 69 |
+
batch_tensors.append(torch.tensor(y[-ctx_len:], dtype=torch.float32))
|
| 70 |
+
|
| 71 |
+
# Run foundation model if pipeline is active
|
| 72 |
+
if self.pipeline is not None:
|
| 73 |
+
with torch.no_grad():
|
| 74 |
+
f_batch = self.pipeline.predict(batch_tensors, prediction_length=self.H).cpu().numpy()
|
| 75 |
+
else:
|
| 76 |
+
# Fallback: empirical naive quantile expansion
|
| 77 |
+
f_batch = np.zeros((len(batch_y), 9, self.H), dtype=np.float32)
|
| 78 |
+
for b_idx, y in enumerate(batch_y):
|
| 79 |
+
last_val = float(y[-1])
|
| 80 |
+
std_val = max(float(np.std(y[-self.S:])), 1e-4)
|
| 81 |
+
q_levels = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]
|
| 82 |
+
for q_idx, q in enumerate(q_levels):
|
| 83 |
+
z = float(torch.distributions.Normal(0, 1).icdf(torch.tensor(q)).item())
|
| 84 |
+
f_batch[b_idx, q_idx] = last_val + z * std_val
|
| 85 |
+
|
| 86 |
+
# Agentic Harmonic and Tail Post-Processing
|
| 87 |
+
for b_idx, (y_raw, prof, start_date) in enumerate(zip(batch_y, batch_profiles, batch_starts)):
|
| 88 |
+
q_out = f_batch[b_idx].copy() # shape: (9, H)
|
| 89 |
+
|
| 90 |
+
# If strong seasonality detected, harmonize all quantiles
|
| 91 |
+
if prof["seasonal_strength"] > 0.65 and len(y_raw) >= self.S:
|
| 92 |
+
cycle = y_raw[-self.S :]
|
| 93 |
+
slope = (y_raw[-1] - y_raw[-min(len(y_raw), 2 * self.S)]) / max(self.S, 1)
|
| 94 |
+
reps = int(np.ceil(self.H / self.S))
|
| 95 |
+
h_mean = np.tile(cycle, reps)[: self.H] + np.arange(1, self.H + 1) * (slope / self.S)
|
| 96 |
+
|
| 97 |
+
# Blend median
|
| 98 |
+
diff = 0.35 * (h_mean - q_out[4])
|
| 99 |
+
# Shift all quantiles by coherent harmonic delta to prevent quantile crossing
|
| 100 |
+
for q_idx in range(9):
|
| 101 |
+
q_out[q_idx] += diff
|
| 102 |
+
|
| 103 |
+
# Enforce strict quantile monotonicity: q_0.1 <= q_0.2 <= ... <= q_0.9
|
| 104 |
+
q_out = np.sort(q_out, axis=0)
|
| 105 |
+
|
| 106 |
+
forecasts.append(
|
| 107 |
+
QuantileForecast(
|
| 108 |
+
forecast_arrays=q_out,
|
| 109 |
+
forecast_keys=self.QUANTILES,
|
| 110 |
+
start_date=start_date,
|
| 111 |
+
)
|
| 112 |
+
)
|
| 113 |
+
|
| 114 |
+
return forecasts
|
| 115 |
+
|
| 116 |
+
def predict(self, dataset, **kwargs) -> List[QuantileForecast]:
|
| 117 |
+
"""GluonTS-compatible prediction interface."""
|
| 118 |
+
entries = list(dataset)
|
| 119 |
+
series_list = [np.asarray(e["target"], dtype=np.float32) for e in entries]
|
| 120 |
+
start_dates = [e["start"] for e in entries]
|
| 121 |
+
return self.forecast_batch(series_list, start_dates)
|