ASL Image Classification with EfficientNet

This project involves classifying American Sign Language (ASL) images using a pre-trained EfficientNet-B0 model. The project demonstrates transfer learning techniques and utilizes the PyTorch framework along with torchvision for model and data handling.

Project Structure

• ASL_Classification.ipynb: Jupyter notebook containing the entire workflow from data loading to model training and evaluation.
• models/: Directory where the trained models are saved.
• data/: Directory containing the dataset.
• README.md: Documentation for the project.

Dataset

The dataset contains images of ASL signs, divided into training and testing directories. Each subdirectory within train and test corresponds to a different ASL sign.

• data/train/: Training images
• data/test/: Testing images

Preprocessing

The preprocessing steps include:

1. Loading the dataset from the specified directories.
2. Applying the appropriate transformations for EfficientNet-B0 using the pre-trained weights.
3. Creating data loaders for training and testing.

Model

The model is based on EfficientNet-B0, a state-of-the-art architecture for image classification. The following steps are involved:

1. Loading the pre-trained EfficientNet-B0 model with default weights.
2. Freezing the feature extraction layers.
3. Replacing the classifier head with a custom classifier suited for the ASL dataset.

Training

The training process involves:

1. Defining the loss function (CrossEntropyLoss) and optimizer (SGD).
2. Training the model for a specified number of epochs (5 in this case).
3. Recording the training and testing losses and accuracies for each epoch.
4. Logging the training process using TensorBoard.

Evaluation

The model's performance is evaluated by printing the training and testing accuracies and losses over the epochs. The total training time is also recorded.

Dependencies

• torch
• torchvision
• matplotlib
• torchinfo
• TensorBoard

Results

The training and testing losses and accuracies are recorded and can be visualized using TensorBoard. The final trained model is saved in the models directory.

Conclusion

This project demonstrates how to use transfer learning with EfficientNet-B0 to classify ASL images. Further improvements can be made by tuning the model architecture, experimenting with different loss functions and optimizers, and using more advanced data augmentation techniques.