This is an AI-based signal prediction system that combines a spatiotemporal graph network and a Transformer architecture. It can effectively capture spatiotemporal dependencies between nodes and perform accurate signal prediction。
project/
├── main.py # Main program entry
├── config.py # Configuration parameter management
├── README.md # Project documentation
├── data/
│ └── dataset.py # Data processing
├── models/
│ ├── generator.py # Generator model
│ ├── discriminator.py # Discriminator model
│ └── components.py # Model components
├── training/
│ ├── trainer.py # Training code
│ └── evaluator.py # Evaluation code
└── checkpoints/ # Model save location
1.Environment Requirements
（1）Reference Hardware Configuration
GPU: NVIDIA GeForce RTX 3090
OS: Ubuntu 18.04
2. Software Requirements：
Pytorch 2.0.0
Puthon 3.9
Opencv 4.12.0.88
Pillow 11.3.0
scikit-learn 1.6.1
scikit-image 0.24.0
numpy 1.23.5
3.Data Preparation
Data files should be tab-separated text files containing the following columns：
Frame ID (frame_id); Node ID (node_id); Node Feature (node_feature)
To ensure code execution, we provide a little batch dataset https://pan.baidu.com/s/1YtAYj5VZafe6jMCXNIh4_A code:xg6f. dataset.py provides data processing related workflows.
4.Quick Start
(1) Configure Parameters
Modify relevant parameters in config.py:
DATA_CONFIG = {
'obs_len': 90, # Observation sequence length
'pred_len': 450, # Prediction sequence length
'num_nodes': 22, # Number of nodes
'batch_size': 32, # Batch size
}
MODEL_CONFIG = {
'hidden_size': 32, # Hidden layer dimension
'num_heads': 4, # Number of attention heads
# ... Other parameters
}
(2) Train Model
python main.py
(3) Test Model
After training completes, automatic evaluation on test set will be performed
Training Features:
Adversarial Training + Content Loss + sum Loss
Early stopping mechanism and model checkpoints
Learning rate scheduling and gradient clippin
5.Configuration Details
Data Configuration (config.py)
DATA_CONFIG = {
'obs_len': 90, # Number of observed frames
'pred_len': 450, # Number of predicted frames
'num_nodes': 22, # Number of nodes
'batch_size': 32, # Batch size
'train_dir': "data/train", # Training set path
'val_dir': "data/val", # Validation set path
'test_dir': "data/test" # Test set path
}
Model Configuration
MODEL_CONFIG = {
'hidden_size': 32, # Hidden layer dimension
'num_heads': 4, # Number of attention heads
'num_stgat_blocks': 1, # Number of graph network blocks
'num_transformer_layers': 1, # Number of Transformer layers
'dropout': 0.1 # Dropout rate
}
Training Configuration
TRAIN_CONFIG = {
'learning_rate': 0.0001, # Learning rate
'num_epochs': 100, # Number of training epochs
'patience': 10, # Early stopping patience value
'x_sum_loss_weight': 0.1 # Node sum loss weight
}
RF.py is a powerful time series prediction tool based on Random Forest regression model, capable of handling multivariate time series data and making accurate predictions.
Basic Usage Example
predictor = TimeSeriesPredictor(n_features=3, test_size=0.4)
predictor.load_and_prepare_data('your_data.xlsx')
X, y = predictor.create_features_target()
X_train, X_test, y_train, y_test = predictor.split_data()
X_train_scaled, X_test_scaled = predictor.preprocess_data()
predictor.train_model()
y_pred = predictor.predict()
metrics = predictor.evaluate_model()
predictor.visualize_input_vs_prediction()
predictor.visualize_continuous_prediction()
importance_df = predictor.feature_importance()
Parameter Description
n_features: Number of feature columns to use (default: 3)
test_size: Test set ratio
random_state: Random seed
Main Dependencies
pandas - Data processing
numpy - Numerical computations
matplotlib - Data visualization
seaborn - Statistical plots
scikit-learn - Machine learning algorithms
openpyxl - Excel file reading