Coursework for EEEM066 – Fundamentals of Machine Learning, University of Surrey
This project tackles the problem of fine-grained knife classification across 192 classes using deep learning models. It explores architectural choices, hyperparameter tuning, and augmentation strategies for improving classification performance.
- Dataset: 9,928 real-world knife images (192 classes) + 351 test samples
- Metric: mean Average Precision (mAP)
- Goal: Optimize classification performance using CNNs and training pipeline enhancements
Model | mAP | Training Time |
---|---|---|
EfficientNet B0 | 0.192 | 16.4 min |
ResNet-34 | 0.063 | 7.8 min |
Inception V3 | 0.408 | 7.9 min |
Default config: Color Jitter
Additional tested: Horizontal Flip
, Random Rotation
, and combinations.
Augmentation Combo | mAP |
---|---|
Default | 0.192 |
Default + Flip | 0.238 |
Default + Rotation | 0.255 |
Flip + Rotation + Default | 0.294 |
Learning Rates:
LR | mAP |
---|---|
0.00001 | 0.015 |
0.0003 | 0.192 |
0.001 | 0.507 |
0.003 | 0.586 |
Batch Sizes with LR=0.003:
BS | mAP |
---|---|
32 | 0.534 |
64 | 0.192 |
128 | 0.555 |
256 | 0.600 |
File | Description |
---|---|
EEEM066_CW1_TAMAN_BACHANI_6846172.ipynb |
Full code with model training and tuning |
EEEM066_CW1_Report_TAMAN_BACHANI_6846172.pdf |
Final coursework report (optional to view) |
assets/ |
Figures and visual results for this project |
- Inception V3 outperformed others in base architecture tests.
- Combined augmentations noticeably improved model robustness.
- Careful tuning of learning rate and batch size significantly boosted performance (from 0.192 to 0.600 mAP).
Taman Bachani
MSc Artificial Intelligence – University of Surrey
GitHub | LinkedIn