Model Training Notes
Validation Accuracy: train0
Note: "v1" = IMAGENET1K_V1, "v2" = V2
| Model | Run 0 |
|---|---|
| ResNet50 v1 | 0.764273 |
| ResNet50 v2 | 0.729282 |
| ResNet101 v1 | 0.775936 |
| ResNet101 v2 | 0.790055 |
Validation Accuracy: train1
Utilizes new labeled test set from Stanford Cars for more training data!
| Model | Run 0 |
|---|---|
| ResNet50 v1 | 0.848023 |
| ResNet50 v2 | 0.833607 |
| ResNet101 v1 | 0.867381 |
| ResNet101 v2 | 0.861614 |
Hyperparameterization: ResNet101v1 (train1 best model)
Hyperparameters changed: optimizer and learning rate
| Description | Run 0 |
|---|---|
| Adam, lr=1e-4 | 0.867381 (baseline) ⭐ |
| Adam, lr=3e-4 | 0.717875 |
| Adam, lr=5e-5 | 0.841050 |
| SGD, lr=1e-2 | 0.691104 |
| SGD, lr=5e-3 | 0.417627 |
Observations & Conclusions
- More data improves accuracy: All models saw substantial gains in
train1compared totrain0. - Deeper models help: ResNet101 generally outperforms ResNet50.
- Optimizer matters: Adam (
lr=1e-4) yielded the highest accuracy; both lower/higher learning rates and SGD performed worse. - IMAGENET v1 vs v2: The difference between v1 and v2 initializations is minor compared to the effect of data volume and model size.
- Performance margins: The right optimizer and learning rate can more than double validation accuracy for the same architecture.