Skip to main content

​
Overview

The emotion recognition model is a scikit-learn classifier trained to predict emotions from facial landmark features. The model is serialized using Python’s pickle module and stored in the model file.

​
Loading the Model

The model is loaded at application startup in app.py:28-29: 
import pickle

model_path = "./model"
with open(model_path, "rb") as f:
    model = pickle.load(f)

​
Error Handling

If the model file is missing, the application raises a clear error: 
if not os.path.isfile(model_path):
    raise FileNotFoundError(
        f"No se encontró el modelo entrenado en '{model_path}'. "
        f"Ejecuta antes 'train_model.py'."
    )

​
Model Interface

​
predict() Method

The model follows the standard scikit-learn interface: 
output = model.predict([face_landmarks])
​
features
List[List[float]]
required
A 2D array where each row contains 136 facial landmark features (68 points × 2 coordinates). Typically passed as a single-element list: [face_landmarks].
​
predictions
numpy.ndarray
Returns an array of emotion indices. For a single prediction, access the result with output[0].Emotion Indices:
  • 0 = HAPPY
  • 1 = SAD

​
Emotion Mapping

The emotion labels are defined in app.py:19: 
emotions = ["HAPPY", "SAD"]
IndexEmotionDescription
0HAPPYPositive, joyful expression
1SADNegative, sorrowful expression

​
Usage Example

​
Complete Prediction Pipeline

import pickle
import cv2
from utils import get_face_landmarks

# 1. Load the model
with open("./model", "rb") as f:
    model = pickle.load(f)

# 2. Define emotion labels
emotions = ["HAPPY", "SAD"]

# 3. Load and process an image
image = cv2.imread("photo.jpg")
face_landmarks = get_face_landmarks(image, draw=False, static_image_mode=True)

# 4. Make prediction
if len(face_landmarks) > 0:
    output = model.predict([face_landmarks])
    emotion_index = int(output[0])
    
    # 5. Map to emotion label
    if 0 <= emotion_index < len(emotions):
        emotion = emotions[emotion_index]
        print(f"Detected emotion: {emotion}")
    else:
        print("Unknown emotion index")
else:
    print("No face detected")

​
Batch Prediction

import pickle
import cv2
from utils import get_face_landmarks

# Load model
with open("./model", "rb") as f:
    model = pickle.load(f)

emotions = ["HAPPY", "SAD"]

# Process multiple images
images = ["photo1.jpg", "photo2.jpg", "photo3.jpg"]
all_features = []

for img_path in images:
    image = cv2.imread(img_path)
    landmarks = get_face_landmarks(image)
    if len(landmarks) > 0:
        all_features.append(landmarks)

# Batch prediction
if len(all_features) > 0:
    predictions = model.predict(all_features)
    for i, pred in enumerate(predictions):
        emotion = emotions[int(pred)]
        print(f"Image {i+1}: {emotion}")

​
Integration with Flask Endpoint

This is how the model is used in the actual /predict endpoint (app.py:50-51): 
face_landmarks = get_face_landmarks(frame, draw=False, static_image_mode=True)

if len(face_landmarks) > 0:
    output = model.predict([face_landmarks])
    emotion = emotions[int(output[0])] if 0 <= int(output[0]) < len(emotions) else "UNKNOWN"
    return jsonify({'emotion': emotion})

​
Model Requirements

​
Input Format

  • Type: List of 136 float values
  • Structure: [x1, y1, x2, y2, ..., x68, y68]
  • Normalization: Coordinates normalized to 0.0-1.0 range
  • Source: Output from get_face_landmarks()

​
Output Format

  • Type: numpy.ndarray
  • Shape: (n_samples,) for n predictions
  • Values: Integer indices (0 or 1 for current model)

​
Model Training

To train or retrain the model, run: 
python train_model.py
This will generate a new model file with the trained classifier.

​
Model File Location

The model file is expected at: 
./model
Relative to the application root directory where app.py is located.

​
Technical Details

  • Format: Python pickle (protocol version depends on training environment)
  • Framework: scikit-learn (exact classifier type determined during training)
  • Feature dimension: 136 (68 facial landmarks × 2 coordinates)
  • Output classes: 2 (HAPPY, SAD)
  • Thread safety: Model predictions are thread-safe in scikit-learn

Build docs developers (and LLMs) love

Get started for freeTalk to us