TFT Report Generator

Automated Temporal Fusion Transformer modeling with publication-ready report generation.

Features

• Automatic Data Analysis: EDA with GPT-powered insights
• TFT Architecture Visualization: Model structure diagram with hyperparameters
• Model Training: Automated training with progress tracking
• Bootstrap Significance Testing: Parameter stability analysis
• Variable Importance Analysis: SHAP-like interpretability
• Prediction Visualization: Actual vs Predicted with uncertainty bands
• Word Document Report: Complete academic-style report

Installation

pip install git+https://github.com/sdkparkforbi/tft-report.git

Requirements

Minimum System Requirements

Component	Requirement
Python	>= 3.9
RAM	>= 8 GB (16 GB recommended)
GPU	Optional (CPU works, GPU faster)
Storage	>= 2 GB free space

Dependencies

pytorch>=2.0.0
pytorch-forecasting>=1.0.0
lightning>=2.0.0
pandas>=1.5.0
numpy>=1.21.0
matplotlib>=3.5.0
python-docx>=0.8.0
openai>=1.0.0

API Keys

• OpenAI API Key: Required for GPT-powered explanations
  • Set as environment variable: OPENAI_API_KEY
  • Or pass directly: api_key='your-key'
  • Colab: Store in Secrets as OPENAI_API_KEY

Quick Start

Google Colab (Recommended)

# Install
!pip install git+https://github.com/sdkparkforbi/tft-report.git -q

# Import
from tft_report import generate_tft_report
import pandas as pd
import numpy as np

# Sample data
np.random.seed(42)
n_patients, n_days = 5, 60

df = pd.DataFrame({
    'time_idx': np.tile(range(n_days), n_patients),
    'patient': np.repeat([f'P{i}' for i in range(n_patients)], n_days),
    'glucose': np.random.normal(140, 20, n_patients * n_days),
    'insulin': np.random.randint(8, 16, n_patients * n_days),
    'dayofweek': np.tile(np.arange(n_days) % 7, n_patients),
})

# Generate report
generate_tft_report(
    df,
    target='glucose',
    group_id='patient',
    time_idx='time_idx',
    known_reals=['insulin', 'dayofweek'],
    download=True
)

Local Python

from tft_report import generate_tft_report
import pandas as pd

df = pd.read_csv('your_data.csv')

generate_tft_report(
    df,
    target='glucose',
    group_id='patient',
    time_idx='time_idx',
    api_key='your-openai-api-key'
)

Configuration

generate_tft_report(
    df,
    target='glucose',
    group_id='patient',
    time_idx='time_idx',
    known_reals=['insulin', 'dayofweek'],  # Known future variables
    unknown_reals=['glucose'],              # Unknown future variables
    static_categoricals=['patient'],        # Static categorical variables
    config={
        'max_encoder_length': 14,           # Lookback window
        'max_prediction_length': 3,         # Forecast horizon
        'hidden_size': 32,                  # Model hidden size
        'attention_head_size': 2,           # Number of attention heads
        'lstm_layers': 2,                   # Number of LSTM layers
        'dropout': 0.1,                     # Dropout rate
        'batch_size': 32,                   # Training batch size
        'max_epochs': 50,                   # Maximum training epochs
        'n_bootstrap': 100,                 # Bootstrap iterations
    },
    verbose=True,
    download=True
)

Data Format

Your DataFrame must have:

Column	Description	Example
time_idx	Integer time index	0, 1, 2, ...
group_id	Group identifier	'P0', 'P1', ...
target	Target variable to predict	glucose values
Other columns	Covariates	insulin, dayofweek

Example Data Structure

   time_idx patient  glucose  insulin  dayofweek
0         0      P0   145.23       10          0
1         1      P0   138.45       12          1
2         2      P0   142.67       11          2
...

Output

The generated zip file contains:

1. Word Document (tft_analysis_report.docx)

   • Introduction
   • Data Description with tables and figures
   • Model Specification (architecture, hyperparameters)
   • Results (performance metrics, variable importance)
   • Bootstrap significance analysis
   • Conclusion

2. Figures (PNG, 300 DPI)

   • Target distribution histogram
   • Time series plots
   • TFT architecture diagram
   • Prediction vs Actual scatter
   • Uncertainty visualization
   • Variable importance bar chart
   • Attention weights (if available)

Report Structure

1. Introduction
   - Dataset overview
   - TFT methodology rationale

2. Data Description
   - Table 1: Dataset characteristics
   - Figure 1: Target distribution
   - Figure 2: Time series by group

3. Model Specification
   - 3.1 TFT Architecture (with diagram)
   - 3.2 Hyperparameters (Table 2)

4. Results
   - 4.1 Model Performance (MAE, RMSE)
   - 4.2 Variable Importance
   - 4.3 Bootstrap Significance

5. Conclusion

GPU Acceleration

For faster training, use GPU:

Google Colab

• Runtime → Change runtime type → GPU

Local

import torch
print(f"GPU available: {torch.cuda.is_available()}")

Troubleshooting

"OpenAI API key not found"

# Option 1: Pass directly
generate_tft_report(df, ..., api_key='sk-...')

# Option 2: Environment variable
import os
os.environ['OPENAI_API_KEY'] = 'sk-...'

# Option 3: Colab Secrets
# Add OPENAI_API_KEY in Colab's Secrets (key icon)

"CUDA out of memory"

# Reduce batch size and model size
generate_tft_report(df, ..., config={
    'batch_size': 16,
    'hidden_size': 16,
    'attention_head_size': 1
})

"Module not found"

pip install pytorch-forecasting lightning python-docx openai

License

MIT License

Citation

If you use this package in your research, please cite:

@software{tft_report,
  title = {TFT Report Generator},
  author = {Your Name},
  year = {2025},
  url = {https://github.com/sdkparkforbi/tft-report}
}