A simplified MobileNetV2 implementation built from scratch in TensorFlow for CIFAR-10 classification.
- Depthwise Separable Convolutions
- Inverted Residual Blocks
- Linear Bottlenecks
- Residual Connections
- Batch Normalization
- Data Augmentation
- CIFAR-10
- 10 classes
- 32x32 RGB images
| Experiment | Test Accuracy |
|---|---|
| Baseline | 65.5% |
| + Data Augmentation | 66.8% |
- Dropout
- Learning Rate Scheduling
- Transfer Learning with MobileNetV2
✅ Completed
This project was built to implement and understand the core ideas behind MobileNetV2 from scratch on CIFAR-10.
Key learnings:
- Depthwise separable convolutions
- Inverted residual bottlenecks
- Residual connections
- Data augmentation
- Bias-variance analysis
Future work:
- Transfer learning with pretrained MobileNetV2