BaseAttentive Documentation

A modular encoder-decoder architecture for sequence-to-sequence time series forecasting with layered attention mechanisms — version 2.0.1.

BaseAttentive is a modular encoder-decoder architecture designed to process three distinct types of inputs:

Static features — constant across time (e.g., geographical coordinates, site properties)
Dynamic past features — historical time series (e.g., sensor readings, observations)
Known future features — forecast-period exogenous variables (e.g., weather forecasts)

It combines these inputs through a configurable attention stack for forecasting experiments and applied workflows. The v2 architecture introduces a registry / resolver / assembly system making every component pluggable and backend-neutral.

Main Elements

Architecture options

Hybrid mode: Multi-scale LSTM with attention (objective="hybrid")
Transformer mode: self-attention encoder (objective="transformer")
Operational mode shortcuts: TFT-like (mode="tft"), PIHALNet-like (mode="pihal")
Declarative attention stack via attention_levels

Core components

Variable selection networks for feature weighting
Multi-scale LSTM for temporal aggregation (scales, multi_scale_agg)
Cross, hierarchical, and memory-augmented attention
Transformer encoder/decoder blocks
Quantile and probabilistic forecast heads

V2 system

BaseAttentiveSpec / BaseAttentiveComponentSpec for backend-neutral config
ComponentRegistry and ModelRegistry for pluggable components
BaseAttentiveV2Assembly resolver/assembler pattern
Multi-backend: TensorFlow, Torch, and JAX (all supported in v2)

Runtime support

Keras 3 based implementation
make_fast_predict_fn for traced TF inference
Input validation utilities
TorchDeviceManager for CUDA/MPS device management

Quick Example

import numpy as np
from base_attentive import BaseAttentive

# Create a model
model = BaseAttentive(
    static_input_dim=4,        # Site properties
    dynamic_input_dim=8,       # Historical observations
    future_input_dim=6,        # Known future features
    output_dim=2,              # Forecast targets
    forecast_horizon=24,       # 24-step forecast
    quantiles=[0.1, 0.5, 0.9], # Uncertainty quantiles
    embed_dim=32,
    num_heads=8,
    dropout_rate=0.15,
)

# Prepare inputs
batch_size = 32
x_static  = np.random.randn(batch_size, 4).astype('float32')
x_dynamic = np.random.randn(batch_size, 100, 8).astype('float32')
x_future  = np.random.randn(batch_size, 24, 6).astype('float32')

# Get predictions
predictions = model([x_static, x_dynamic, x_future])
# shape: (32, 24, 3, 2) — batch, horizon, quantiles, output_dim

User Guide

Examples

Example Notebooks

Resources

Release Notes

BaseAttentive Documentation

Main Elements

Quick Example

Indices and tables