Example Notebooks

The notebooks below are rendered directly from the examples/ folder in the repository. Each one can be run interactively on Binder (no local installation required) or downloaded and run locally.

Before you run — Backend & Import Order

Every notebook (and any script using BaseAttentive) requires the backend to be set before Keras or BaseAttentive are imported. Follow this order every time:

# 1. Set the backend environment variables FIRST
import os
os.environ["BASE_ATTENTIVE_BACKEND"] = "tensorflow"   # or "torch" / "jax"
os.environ["KERAS_BACKEND"]          = "tensorflow"   # must match above

# 2. Import Keras SECOND
import keras

# 3. Import BaseAttentive THIRD
import base_attentive
from base_attentive import BaseAttentive

If you skip step 1 or import in a different order, BaseAttentive will raise a BackendConfigurationError because it cannot detect the backend after Keras has already initialised.

Binder users: open a notebook, execute the very first cell (it already contains the environment setup), and then run the remaining cells in order. Do not skip the first cell.

Local users: the os.environ calls work only when they run before any Keras import in the same Python process. If you have already imported Keras in a prior cell or session, restart the kernel and re-run from the top.

#	Notebook	Topics covered
01	BaseAttentive: Quick Start Guide	Model creation, configuration inspection, save/load
02	BaseAttentive: Hybrid vs Transformer Architectures	Hybrid vs. Transformer objective comparison
03	BaseAttentive: Attention Stack Configuration	Attention levels, cross / hierarchical / memory-augmented
04	BaseAttentive Standalone Applications	Domain application patterns, multi-output forecasting
05	BaseAttentive as a Kernel for Robust Neural Networks	Kernel-robust training, DTW alignment, regularisation
06	CRPSLoss & Probabilistic Forecasting	CRPSLoss — quantile / gaussian / mixture modes, MC sampling
07	V2 Spec & Registry: Declarative Configuration	V2 Spec & Registry — BaseAttentiveSpec, ComponentRegistry, custom encoders
08	Financial Forecasting	Financial ML — walk-forward validation, IC/ICIR/Sharpe/drawdown, regime analysis, gradient saliency
09	Attention Interpretability	Interpretability — VSN weights, cross/hierarchical attention heatmaps, integrated gradients, multi-head diversity
10	Benchmarking	Benchmarking — 7 architecture variants vs baselines, efficiency frontier, hyperparameter sensitivity, noise robustness, statistical significance
11	Landslide Susceptibility Mapping with Physics-Informed BaseAttentive	Landslide Susceptibility — physics-informed FS regularisation, depth-profile attention, ensemble uncertainty, scenario-conditioned hazard curves, method comparison (LR/RF/BA)
12	ICU Sepsis Early Warning with SOFA-Informed BaseAttentive	ICU Sepsis Early Warning — SOFA-informed regularisation, multi-horizon risk curves (+6 h/+12 h/+24 h), temporal attention heatmaps, ensemble epistemic uncertainty, calibration curves, method comparison (LR/RF/BA); PhysioNet 2019 integration guide
12b	ICU Sepsis Early Warning — Real Data: PhysioNet Challenge 2019	ICU Sepsis — PhysioNet 2019 Real Data — full 5-component SOFA score (SpO₂/FiO₂ respiratory proxy, Platelets, Bilirubin, MAP, Creatinine; GCS absent — noted), demo-mode fallback, SOFA-consistency plots, multi-horizon BA vs LR/RF on ~40 k real patients
13	Flood Early Warning System with Physics-Informed BaseAttentive	Flood Early Warning System — FSI physics prior (Manning bankfull ratio), multi-horizon alerts (+1 h/+3 h/+6 h/+12 h/+24 h), NWP future covariates, horizon-conditioned attention saliency, ensemble epistemic uncertainty, decision-curve analysis, confidence-gated alarm API (REST/MQTT)
13b	Flood Early Warning — Real Data Integration Guide	Flood EWS — Real Data Integration Guide — drop-in loaders for CAMELS-US, USGS NWIS, ERA5-Land, GloFAS, GRDC, and UK NRFA; unified build_nb13_arrays() pipeline; demo-mode fallback; comparison table vs published CAMELS LSTM baseline
14	PADR-Net Flood Forecasting	PADR-Net Flood Forecasting — validated PADRNetConfig, TensorFlow training loop, physics-aware residuals, regional WAF/EAF/SAF synthetic events, NSE/CSI/TSS/mass-bias evaluation, hydrograph and spatial-style interpretation maps
15	Advanced PADR-Net Workflow: Transfer, Uncertainty, and Stress Testing	Advanced PADR-Net Workflow — leave-one-region-out transfer, physics-loss ablation, threshold calibration, MC-dropout uncertainty, and rainfall-intensification stress testing

Run on Binder

Launch any notebook interactively (no local install needed):