TPnet-small

TPnet-small is a lightweight deep neural network (DNN) designed to predict traffic congestion using tabular smart mobility features. It serves as a compact yet powerful alternative to tree-based models.

Model Details

• Model type: Feedforward Deep Neural Network (3 hidden layers)
• Architecture: [128, 64, 32] with ReLU activations and Dropout
• Output: 3-class Softmax output (High, Medium, Low)
• Trained on: Smart Mobility Traffic Dataset from Kaggle

Training Details

• Optimizer: Adam
• Loss: Sparse Categorical Crossentropy
• Epochs: 50
• Batch size: 32
• Training time: ~6 seconds on CPU
• Train Accuracy: 99.1%
• Val Accuracy: 94.6%

Evaluation

Metric	Value
Accuracy	94.5%
F1 Score	0.944
Parameters	13,123
Model Size	~156 KB (.h5 format)

Includes training trajectory and confusion matrix plots.

How to Use

from tensorflow.keras.models import load_model

model = load_model("traffic_predictor_dnn.h5")
y_pred = model.predict(X_test)  # X_test must be scaled [n_samples, 20]

Limitations

• Performance limited by small dataset size and feature coverage

• Currently optimized for CPU inference, not edge deployment

Authors

• Created by @Yukin3