README.md

Environment Setup

Activate your conda or virtual Environment with python=3.12 (or similar) installed and do the following:

pip install -r requirements.txt

EfficientNet Model

This folder uses the EfficientNetV2-M architecture for Script Identification. The base trained model is automatically downloaded from GitHub releases via the efficientnet_identifier.py script. Following are the 12 languages it is typically trained on: Hindi, English, Gujarati, Punjabi, Marathi, Kannada, Telugu, Tamil, Malayalam, Assamese, Bengali, and Odia.

Configuration

All hyperparameters and paths are defined in config.py (which acts as a JSON configuration file). You must ensure paths and hyperparameters are set appropriately:

{
    "paths": {
        "base_dir": "path/to/base",
        "train_dir": "train_data",
        "test_dir": "test_data",
        "save_model_name": "models/efficientnet/efficientnetv2_m_sliced.pth"
    },
    "hyperparameters": {
        "batch_size": 32,
        "epochs": 30,
        "learning_rate": 0.0001,
        "image_size": 480,
        ...
    },
    "classes": [ ... ]
}

Training

To start the training:

python train.py

During training, the model uses dynamic data augmentation (rotations, color jitter, affine transformations). The script automatically saves the checkpoint whenever the validation accuracy improves.

Testing

To test your trained model on the test_data directory:

python test.py

The code will output class-wise accuracy along with macro precision, recall, and F1-score for the standard 12 classes, as well as a merged 10-class report (where Devanagari and Eastern Nagari scripts are grouped). Additionally, confusion matrices will be saved as images.

Inference

Script detection can be done using infer.py on a single image as input.

python infer.py --image_path demo_images/D_image_149_9.jpg
# {'predicted_class': 'Assamese'}

Simply replace demo_images/D_image_149_9.jpg with your image path.

To process a batch of images:

python infer.py --image_dir demo_images/ --batch
# predictions.csv

Using the --image_dir and --batch arguments, predictions can be made on a batch of images, and the output is saved to prediction.csv.

Evaluation

After inference for a batch of images and obtaining the predictions in a csv file, use this for calculating the accuracy for the language predictions against the actual labels.

The test CSV should contain two columns: Filepath (the image file name/path) and Language_test (the ground truth language labels).

python evaluation.py <path/to/test.csv> <path/to/prediction.csv>

API

Start the localhost FastAPI server:

uvicorn app:app --reload

To get the prediction for an image using the EfficientNet model:

• METHOD: POST
• URL:

http://127.0.0.1:8000/predict_lang/

• Parameters: file: image_file to be classified

The model will predict the language and return a JSON response.