Skip to main content
The data preparation process extracts facial landmarks from training images and generates a structured dataset that the model can learn from.

​
Overview

The prepare_data.py script processes images organized in emotion-specific folders, extracts facial landmark coordinates using OpenCV, and saves the normalized data to data.txt for training.

​
Directory Structure

Organize your training images in the following structure: 
data/
├── happy/
│   ├── image1.jpg
│   ├── image2.jpg
│   └── ...
└── sad/
    ├── image1.jpg
    ├── image2.jpg
    └── ...
The data/ directory should be located at the root of your project, parallel to the source/ folder.

​
Supported Emotions

Currently, EmoChat processes only two emotions:
  • happy (label: 0)
  • sad (label: 1)
Folder names are case-insensitive, but must match these exact emotion names.
Only images in happy/ and sad/ folders will be processed. Other emotion folders will be ignored.

​
Data Preparation Workflow

1

Create the data directory

Create a data/ folder at your project root with happy/ and sad/ subfolders:
mkdir -p data/happy data/sad
2

Add training images

Place facial images into the respective emotion folders. Ensure:
  • Images contain clearly visible faces
  • Good lighting conditions
  • Various angles and expressions
  • At least 20-30 images per emotion for better results
3

Run the preparation script

Execute the data preparation script:
cd source
python prepare_data.py
4

Verify output

Check that data.txt was created in the source/ directory. This file contains the extracted features and labels.

​
How It Works

​
Loading Emotion Folders

The script identifies valid emotion directories: 
ALLOWED_EMOTIONS = {"happy", "sad"}

emotion_folders = [
    e for e in sorted(os.listdir(data_dir))
    if os.path.isdir(os.path.join(data_dir, e)) and e.lower() in ALLOWED_EMOTIONS
]
# Sorted to ensure consistent indices: happy=0, sad=1
emotion_folders = sorted(emotion_folders, key=str.lower)

​
Extracting Facial Landmarks

For each image, the script:
  1. Reads the image using OpenCV
  2. Extracts facial landmarks (68 points × 2 coordinates = 136 features)
  3. Normalizes the coordinates
  4. Associates the data with an emotion label
for emotion_indx, emotion in enumerate(emotion_folders):
    emotion_path = os.path.join(data_dir, emotion)

    for image_path_ in os.listdir(emotion_path):
        image_path = os.path.join(emotion_path, image_path_)
        
        image = cv2.imread(image_path)
        if image is None:
            continue
        
        face_landmarks = get_face_landmarks(image)
        
        if len(face_landmarks) > 0:
            sample = face_landmarks + [int(emotion_indx)]
            output.append(sample)

​
Generating data.txt

The extracted features are saved to a text file: 
np.savetxt("data.txt", np.asarray(output))
Each row in data.txt contains:
  • 136 facial landmark coordinates (features)
  • 1 emotion label (0 for happy, 1 for sad)

​
Troubleshooting

​
”No se encontró el directorio de datos”

Problem: The data/ directory doesn’t exist. Solution: Create the directory structure: 
mkdir -p data/happy data/sad

​
“No se encontraron carpetas ‘happy’ o ‘sad’”

Problem: The emotion folders are missing or misnamed. Solution: Ensure folders are named exactly happy and sad (case-insensitive).

​
”No se generaron muestras”

Problem: No faces were detected in the images. Solution:
  • Verify images contain visible faces
  • Ensure images are not corrupted
  • Try images with better lighting and frontal faces
  • Check that OpenCV can read your image formats

​
Low Sample Count

Problem: Only a few images were processed successfully. Solution:
  • Add more high-quality training images
  • Ensure faces are clearly visible and well-lit
  • Remove blurry or low-resolution images

​
Best Practices

Image Quality

Use clear, well-lit images with visible facial features for better landmark detection.

Dataset Size

Aim for at least 50-100 images per emotion for reliable model performance.

Diversity

Include multiple people, angles, and lighting conditions to improve generalization.

Balance

Keep roughly equal numbers of images for each emotion to avoid bias.

​
Next Steps

Once you’ve successfully generated data.txt, proceed to Model Training to train your emotion recognition classifier.

Build docs developers (and LLMs) love

Get started for freeTalk to us