Sports Classifier App (trained using PyTorch)

Important Note

• Pre-trained model used was EfficientNet_B0 (model pre-trained on ImageNet dataset) : https://github.com/lukemelas/EfficientNet-PyTorch
• Custom model was trained using the smallest EfficientNet, model imported using torch image model (timm) : https://huggingface.co/timm
• Model Trained on 100 different sports, refer to classes.txt for the list of different sports
• For the purpose of using Streamlit cloud, the smallest model is chosen for a faster inference speed, with less accuracy
• Accuracy of the model is 93%

Link to app

• https://sports-classifier.streamlit.app/

How to use the app?

Step 1: Upload the image of a sport

Step 2: Click on the classify button

General steps for fine-tuning a pre-trained CNN model:

1. Label images:

• For image classification tasks, the labeling process involves assigning a class label to each image and organizing them into folders based on their classes. Tools like RoboFlow or CVAT simplify this process.

2. Split Data:

• Divide dataset into training and validation sets. The training set is used to train the model, while the validation set is used to evaluate its performance during training.

3. Pre-process the images: (Resize images and batching dataset)

• Resize images into appropriate sizes, each pre-trained model has a different image size. For EfficientNet, the image size used was 128x128 for the smallest model.
• Batch size for image classification can vary, a good starting point is 32 for batch size.

4. Augment Images (if needed):

• Data augmentation techniques can be applied to artificially increase the diversity of your dataset. This helps the model generalize better.

5. Train the model:

• The models were trained using PyTorch, feel free to use tensorflow if you want. Monitor metrics such as train loss and val loss during training.

6. Validate the Model:

• Evaluate the trained model on the validation set to ensure it generalizes well to new, unseen data.

7. Inference on Test Set:

• After training, perform inference on a separate test set to assess the model's performance on completely unseen data.