An Image Classification project utilizing MobileNetV2 and DenseNet-121. This project leverages advanced techniques to enhance model performance and accuracy, including:
Index | Technique | Description | Specific Details |
---|---|---|---|
1 | Cross Validation | Utilized to ensure that the model generalizes well to new data. | 5-fold cross-validation applied. |
2 | Early Stopping | To prevent overfitting by stopping training when validation metrics stop improving. | Patience level set to 5 epochs. |
3 | Transfer Learning | Applying knowledge gained from one problem to a different but related problem. | Leveraging pre-trained weights. |
4 | Hyperparameter Tuning | Systematically searching for the optimal parameters of a model. | Tuning epoch, batch size, and learning rate. |
5 | Data Augmentation | Increasing the diversity of data available for training models without actually collecting new data. | Includes random rotation, Gaussian blur, resizing, and bicubic interpolation. |
6 | Changing Optimization Algorithms | Experimenting with different optimizers to improve training performance. | Switch from Adam to AdamW optimizer. |
7 | Weighted Class Training | Adjusting the importance of a class based on its weight to address class imbalance. | More emphasis on classes like cats, dogs, birds, and planes. |
8 | Ensemble Methods | Combining predictions from multiple models to improve accuracy. | Soft voting ensemble of MobileNetV2 and DenseNet-121 logits. |
This table provides an at-a-glance overview of the methodologies and specific adaptations made to optimize the CIFAR10 Image Classification project.
Overview of entire project: CIFAR-10 Image Classification Presentation
View the Full Report here: Full Report
Full project at: Full Source Codes (1.0 to 3.2)
Ablation Study:
7 iteration of MobileV2 Settings 📱:
Enhancements Approaches | 1.0 | 1.1 | 2.0 | 2.1 | 2.2 | 2.3 | 2.4 |
---|---|---|---|---|---|---|---|
Increasing Epochs | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ |
Bicubic Interpolation | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | |
Random Rotation | ✔ | ||||||
Gaussian Blur | ✔ | ||||||
Transfer Learning | ✔ | ✔ | ✔ | ✔ | ✔ | ||
Decreased Learning Rate | ✔ | ✔ | ✔ | ✔ | |||
Increased Batch Size | ✔ | ✔ | ✔ | ||||
AdamW Optimization | ✔ | ✔ | |||||
Weighted Class Training | ✔ |
2 iteration of DenseNet-121 Settings:
Enhancements Approaches | 3.0 | 3.1 |
---|---|---|
Increase Epochs | ✔ | ✔ |
Increase Batch Size | ✔ | ✔ |
Weighted Class Training | ✔ | ✔ |
Transfer Learning | ✔ | |
Decreased Learning Rate | ✔ |
Ensemble Results using Soft Voting (2.4 + 3.1) :
Model | F1-Scores | Accuracy | ||||
---|---|---|---|---|---|---|
Train | Test | Improvement | Train | Test | Improvement | |
2.4 | 0.9442 | 0.8874 | - | 0.9430 | 0.8842 | - |
3.1 | 0.9802 | 0.9235 | +0.0361 | 0.9802 | 0.9234 | +0.0392 |
3.2 (2.4+3.1) | 0.9859 | 0.9342 | +0.0107 | 0.9859 | 0.9341 | +0.0107 |
Ensemble Method Results across individual classes :
F1-Score: