Modified version of LeNet-5 to recognize handwritten digit
Architecture: Convolutional Neural Network (Coursera)
Dataset: MNIST (Subset)
Here are the results (training) I got for different configurations
| Epochs | Batch Size | Learning Rate | Loss | Accuracy |
|---|---|---|---|---|
| 5 | 512 | 0.0001 | 0.313 | 90.947 |
| 5 | 512 | 0.0005 | 0.811 | 76.036 |
| 5 | 256 | 0.0001 | 0.6852 | 75.724 |
| 5 | 256 | 0.0005 | 0.5342 | 87.793 |
| 5 | 128 | 0.0001 | 0.4005 | 85.771 |
| 5 | 128 | 0.0005 | 0.5240 | 78.402 |
| 5 | 64 | 0.0001 | 0.7928 | 68.419 |
| 5 | 64 | 0.0005 | 0.04527 | 98.62 |
Test Accuracy
| Epochs | Batch Size | Learning Rate | Loss | Accuracy |
|---|---|---|---|---|
| 5 | 512 | 0.0001 | 0.263 | 92.423 |
| 5 | 512 | 0.0005 | 0.794 | 76.387 |
| 5 | 256 | 0.0001 | 0.6551 | 76.181 |
| 5 | 256 | 0.0005 | 0.5298 | 87.690 |
| 5 | 128 | 0.0001 | 0.3781 | 86.167 |
| 5 | 128 | 0.0005 | 0.5240 | 79.066 |
| 5 | 64 | 0.0001 | 0.7707 | 68.970 |
| 5 | 64 | 0.0005 | 0.04527 | 98.377 |
training: python model.py --train true <MNIST-path> <model-save-path>
training: python model.py --test true <MNIST-path> <model-save-path>