This README file provides an overview of the code in the "Multi Task Learning.ipynb" Jupyter Notebook, which demonstrates multi-task learning using an enhanced ResNet50 architecture. Multi-task learning involves training a single neural network to perform both fine-grained and coarse-grained image classification tasks simultaneously.
The Jupyter Notebook contains code for multi-task learning using an enhanced ResNet50 architecture. It showcases multiple steps to improve model performance, making it suitable for real-world multi-task learning scenarios.
The code uses the CIFAR-100 dataset for multi-task learning. Images are loaded and preprocessed for both fine-grained and coarse-grained classification tasks.
An initial custom neural network model is presented, but the accuracy results are relatively low.
To enhance performance, a pre-trained ResNet50 model is employed and fine-tuned for multi-task learning. This approach leads to a significant improvement in accuracy:
- Test Fine-grained Accuracy: 36.35%
- Test Coarse-grained Accuracy: 49.72%
The code showcases further adjustments made to the ResNet50-based model. These enhancements include introducing more dense layers, fine-tuning the learning rate, and modifying loss weights. These adjustments result in even better results:
- Test Fine-grained Accuracy: 43.43%
- Test Coarse-grained Accuracy: 60.93%
Here are the performance results for the different models:
| Model Description | Fine-grained Accuracy | Coarse-grained Accuracy |
|---|---|---|
| Initial Model | 26.79% | 38.64% |
| ResNet50 Model (before adjustments) | 36.35% | 49.72% |
| Enhanced ResNet50 Model (after adjustments) | 43.43% | 60.93% |
.png)
The table summarizes the performance of the models, and the accompanying image shows accuracy and loss plots for the enhanced ResNet50 model. These results demonstrate the effectiveness of multi-task learning with the enhanced architecture. Further adjustments and fine-tuning can be explored to achieve even better performance.