---
license: mit
language:
  - en
library_name: pytorch
tags:
  - lstm
  - attention
  - time-series-forecasting
  - food-science
  - walnut
  - rancidity
  - oxidation
  - shelf-life
  - india
  - multi-task-learning
pipeline_tag: time-series-forecasting
datasets:
  - Arko007/walnut-rancidity-predictor
metrics:
  - mae
  - rmse
  - auc
model-index:
  - name: WalnutRancidityLSTMAttention
    results:
      - task:
          type: time-series-forecasting
          name: Walnut Rancidity & Shelf-Life Prediction
        dataset:
          name: Walnut Storage Timeseries (Indian Conditions)
          type: Arko007/walnut-rancidity-predictor
        metrics:
          - type: auc
            value: 0.9+
            name: Rancidity AUC
          - type: mae
            value: 0.389
            name: Best Val Loss
---

# Walnut Rancidity Predictor

Production-grade **Stacked LSTM + Attention** model for predicting walnut rancidity probability
and remaining shelf life from storage condition time-series.

Trained on 90,000 synthetic sequences simulating **Indian storage environments**
using Arrhenius-based lipid oxidation kinetics.

## Model Architecture

```
Input (30 days × 8 features)
  → LSTM layer 1  (hidden=64, dropout=0.2)
  → LSTM layer 2  (hidden=64, dropout=0.2)
  → LSTM layer 3  (hidden=64)
  → Attention (soft weighted context)
  → Dropout (0.2)
  ┌──────────────────────────────────┐
  │  Head 1: rancidity_probability   │  Linear → ReLU → Linear → Sigmoid
  │  Head 2: shelf_life_remaining    │  Linear → ReLU → Linear
  │  Head 3: decay_curve_value       │  Linear → ReLU → Linear → Sigmoid
  └──────────────────────────────────┘
```

| Property | Value |
|---|---|
| Parameters | ~85 K |
| Input features | 8 |
| Sequence length | 30 days |
| LSTM layers | 3 |
| Hidden size | 64 |
| Dropout | 0.2 |
| Epochs trained | 12 / 20 |
| Best val loss | 0.3892 |

## Training Data

- **Dataset**: [Arko007/walnut-rancidity-predictor](https://huggingface.co/datasets/Arko007/walnut-rancidity-predictor)
- 90,000 sequences · 5.4 M rows
- 4 Indian storage scenarios (cold warehouse, hill region, ambient, hot transport)

## Input Features

| Feature | Unit | Description |
|---|---|---|
| `temperature` | °C | Ambient temperature (2–40 °C) |
| `humidity` | % RH | Relative humidity (30–85 %) |
| `moisture` | % | Walnut moisture content (3–8 %) |
| `oxygen` | fraction | Oxygen exposure (0.18–0.23) |
| `peroxide_value` | meq/kg | Primary oxidation marker |
| `free_fatty_acids` | % | FFA content |
| `hexanal_level` | ppm | Volatile oxidation byproduct |
| `oxidation_index` | — | Composite oxidation score |

## Outputs

| Output | Type | Description |
|---|---|---|
| `rancidity_probability` | [0, 1] | Probability walnuts are rancid |
| `shelf_life_remaining_days` | float | Days until PV > 5 meq/kg |
| `decay_curve_value` | [0, 1] | Normalised peroxide value |

## Rancidity Threshold

Walnuts are rancid when **Peroxide Value > 5 meq/kg** (FSSAI / Codex standard).

```
rancidity_probability = sigmoid(PV − 5)
```

## Quick Start

```python
from huggingface_hub import hf_hub_download
import torch, joblib, numpy as np

# Download artifacts
model_path  = hf_hub_download("Arko007/walnut-rancidity-predictor", "models/walnut_rancidity_lstm_attention.pt")
scaler_path = hf_hub_download("Arko007/walnut-rancidity-predictor", "models/feature_scaler.pkl")

# Or clone and use the inference helper directly:
# from model.predict import predict_storage_risk

# Build a 30-day sequence: [temp, humidity, moisture, oxygen, PV, FFA, hexanal, ox_index]
seq = np.column_stack([
    np.full(30, 5.0),             # cold storage
    np.full(30, 50.0),
    np.full(30, 4.0),
    np.full(30, 0.20),
    np.linspace(0.5, 1.2, 30),
    np.linspace(0.05, 0.10, 30),
    np.linspace(0.1,  0.3,  30),
    np.linspace(0.2,  0.5,  30),
])

result = predict_storage_risk(seq)
# → {'rancidity_probability': 0.0312, 'shelf_life_remaining_days': 143.7, 'risk_level': 'LOW'}
```

## Risk Classification

| `risk_level` | `rancidity_probability` |
|---|---|
| `LOW` | < 0.30 |
| `MEDIUM` | 0.30 – 0.70 |
| `HIGH` | > 0.70 |

## Chemistry Model

```
k(T)   = A · exp(−Ea / (R·T))    A=1.5×10¹², Ea=80 kJ/mol, R=8.314 J/mol·K
PV(t)  = PV₀ · exp(k · t)
```

Humidity and moisture accelerate oxidation via empirical correction factors.

## Files

```
models/
  walnut_rancidity_lstm_attention.pt  ← PyTorch checkpoint (epoch 12)
  feature_scaler.pkl                  ← StandardScaler for inference
  metrics.json                        ← Training metrics
model/
  predict.py                          ← Inference API
train.py                              ← Training script (resume from checkpoint)
generate_data.py                      ← Synthetic data generator
requirements.txt
```

## Resume Training

```bash
pip install -r requirements.txt
python train.py          # resumes from saved checkpoint automatically
```

## License

MIT