Cat Emotion Classifier

This project trains a deep learning model to classify cat emotions from images. The final model uses a ConvNeXt-Large architecture and incorporates modern training techniques to achieve high accuracy.

Project Evolution

The project started with a custom CNN with residual connections, which yielded a validation accuracy of ~26%. To improve performance, the following iterative steps were taken:

1. Adopted Transfer Learning (ResNet-50): The model was upgraded to a standard ResNet-50 architecture with fine-tuning of all layers and more aggressive data augmentation. This significantly improved validation accuracy to ~60%.
2. Switched to ConvNeXt-Large: To push performance further, the architecture was switched to ConvNeXt-Large.
3. Advanced Training Techniques: Several modern techniques were introduced, including TrivialAugmentWide for data augmentation, label smoothing, the AdamW optimizer, a cosine annealing learning rate scheduler, gradient clipping, and Automatic Mixed Precision (AMP) to manage memory.

This iterative process resulted in a final validation accuracy of ~70.4% with smaller training and validation gap.

Running the Project

1. Clone the repository:

   git clone https://github.com/parsany/CatRecognition.git
   cd CatRecognition

2. Install the required dependencies. It is highly recommended to use a virtual environment.

   pip install torch torchvision scikit-learn matplotlib

   Ensure you have a version of PyTorch that supports your available hardware (CUDA for NVIDIA GPUs).

Dataset

The model is trained on the "Cat Emotions" dataset, which is publicly available on Roboflow. Link to Dataset

Final Model & Training

The best-performing model uses the following configuration:

• Model: convnext_large pretrained on ImageNet.
• Data Augmentation: TrivialAugmentWide for robust training.
• Optimizer: AdamW with a learning rate of 1e-5 and weight decay of 1e-4.
• Scheduler: Cosine Annealing learning rate scheduler.
• Loss Function: Cross-Entropy Loss with Label Smoothing (smoothing factor of 0.1).
• Training Features:
  • Automatic Mixed Precision (AMP): To reduce memory footprint and speed up training.
  • Gradient Clipping: To prevent exploding gradients and stabilize training.
• Hyperparameters:
  • Epochs: 30
  • Batch Size: 8 (optimized for an ~8GB VRAM GPU)
  • Dropout Rate: 0.5

Evaluation

The model is evaluated based on its validation accuracy and loss throughout training. The final confusion matrix is generated from the best-saved model checkpoint (best_convnext_model.pth) to provide a clear view of its classification performance across the different cat emotions. The final model achieved a validation accuracy of 70.41%.

Update:

tested on the dog emotion generally and got 85%%.