ResNet Structure Used:

1. Convolution 1: (3, 3, 32)
2. Two building blocks: (3, 3, 32), no down sample
2. Two building blocks: (3, 3, 64), (2, 2) stride down sample
3. Two building blocks: (3, 3, 128), (2, 2) stride down sample
4. Average pooling
5. Flatten
6. FC 1: (128 -> 10)
7. Softmax

Hyper parameters:

1. Learning rate: 0.01
2. momentum: 0.9
3. Batch size: 128
4. Max epochs: 70

Training results:

Test accuracy: 0.9924