cifar10-lenet

Modernized variation of LeNet trained on the CIFAR-10 dataset using MindSpore 2.8.0 + CANN 8.5.0 with the structure below.

1. 1st convolution layer:
   1. 3 input channels
   2. 32 output channels
   3. 3 by 3 kernel with unit stride
   4. Padding of 1px in all directions
   5. ReLU activation
2. 1st batch normalization layer with 32 channels, epsilon = 1e-5, momentum = 0.9
3. 1st max pooling layer with 2 by 2 window and a stride of 2
4. 2nd convolution layer:
   1. 32 input channels
   2. 64 output channels
   3. 3 by 3 kernel with unit stride
   4. Padding of 1px in all directions
   5. ReLU activation
5. 2nd batch normalization layer with 64 channels, epsilon = 1e-5, momentum = 0.9
6. 2nd max pooling layer with 2 by 2 window and a stride of 2
7. Flattening layer to convert the 64 x 8 x 8 feature maps to 4096 output channels
8. 1st hidden fully connected layer: 4096 input channels, 2048 output channels, ReLU activation
9. 1st dropout layer with p=0.5
10. 2nd hidden fully connected layer: 2048 input channels, 1024 output channels, ReLU activation
11. 2nd dropout layer with p=0.5
12. Final fully connected layer: 1024 input channels, 10 output channels

Usage

See the included notebook for details.