Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/Xander44-4/traffic_reducer/llms.txt

Use this file to discover all available pages before exploring further.

The scikit-learn phase-classifier model (modelo_semaforo_ia.pkl) was trained on a fully synthetic dataset that maps four-direction vehicle counts to the expected winning signal phase. The model is loaded at server startup by app.py; its presence is required for the /predict endpoint to respond — the endpoint returns 500 if the model failed to load.

​
Dataset file

traffic_reducer_dataset/modelo_entrenado/dataset_sintetico_entrenamiento.csv
The CSV contains 5,000 rows (plus header). Each row represents one simulated traffic snapshot at a four-way intersection.

​
Schema

ColumnTypeRangeDescription
Norteint0–59Vehicle count in the North lane
Surint0–59Vehicle count in the South lane
Esteint0–59Vehicle count in the East lane
Oesteint0–59Vehicle count in the West lane
GANADOR_ESPERADOint0–3Expected winning phase (0 = Norte, 1 = Sur, 2 = Este, 3 = Oeste)

​
Sample rows

Norte,Sur,Este,Oeste,GANADOR_ESPERADO
49,38,53,13,2
4,58,46,9,1
7,12,23,39,3
7,3,38,18,2
49,45,14,24,0
48,12,57,6,2
8,46,58,23,2
54,1,59,50,2
17,24,22,50,3
31,39,9,13,1

​
Label generation logic

GANADOR_ESPERADO is always the index of the maximum count across the four directions: 
GANADOR_ESPERADO = argmax([Norte, Sur, Este, Oeste])
For example, row 49,38,53,13 → argmax([49,38,53,13]) = index 2 (Este). The dataset encodes a strict majority-rule policy: whichever direction has the most vehicles gets the green phase. There are no tie-breaking rules in the synthetic data; ties are avoided by construction during generation.
Because the dataset is purely synthetic and labels are always argmax, any reasonable classifier achieves near-100% accuracy. The real intelligence of Traffic Reducer lies in the YOLOv8 vehicle detection pipeline, not the phase classifier. The classifier’s job is simply to formalise the argmax rule as a trained artifact that can be swapped for a more sophisticated policy in the future.

​
Loading and testing the model

import pickle
import numpy as np

with open('traffic_reducer_dataset/modelo_entrenado/modelo_semaforo_ia.pkl', 'rb') as f:
    model = pickle.load(f)

# Predict: Norte=10, Sur=50, Este=10, Oeste=10 → Sur wins (index 1)
prediction = model.predict([[10, 50, 10, 10]])
print(prediction)  # [1]
If pickle.load raises an error (e.g., the file was saved with joblib), use joblib.load instead — app.py tries both automatically: 
import joblib

model = joblib.load('traffic_reducer_dataset/modelo_entrenado/modelo_semaforo_ia.pkl')
prediction = model.predict([[10, 50, 10, 10]])
print(prediction)  # [1]

​
Retraining with scikit-learn

To retrain the classifier from scratch — for example after extending the dataset or switching algorithms — run the following script from the project root: 
import pandas as pd
from sklearn.ensemble import RandomForestClassifier
import joblib

df = pd.read_csv('traffic_reducer_dataset/modelo_entrenado/dataset_sintetico_entrenamiento.csv')
X = df[['Norte', 'Sur', 'Este', 'Oeste']]
y = df['GANADOR_ESPERADO']

clf = RandomForestClassifier(n_estimators=100, random_state=42)
clf.fit(X, y)

joblib.dump(clf, 'traffic_reducer_dataset/modelo_entrenado/modelo_semaforo_ia.pkl')
print("Model saved.")
After saving, restart traffic_app/app.py so the new model is picked up at startup. The MODEL_PATH variable in app.py must point to the correct absolute path on your machine.
To test a different policy — for example, one that weights pedestrian counts or time-of-day — extend the CSV with additional columns and update the X feature matrix in the training script. The /predict endpoint passes only the four zone counts to the model, so you would also need to update the prediction logic in app.py to include the new features.

Build docs developers (and LLMs) love

Get started for freeTalk to us