VGG style convolution neural network for the classic MNIST dataset on Kaggle. Currently gets 99.6% on the Kaggle leaderboards.
The input are 28 x 28 greyscale images 4 convolution layers with filter size 3x3 and ReLU activations. Max pooling layers after every other convolution layer. 2 hidden layers with dropout. Softmax output.
Layer Type | Parameters |
---|---|
Input | size: 28x28, channel: 1 |
convolution | kernel: 3x3, channel: 128 |
ReLU | |
convolution | kernel: 3x3, channel: 128 |
ReLU | |
max pool | kernel: 2x2 |
dropout | 0.2 |
convolution | kernel: 3x3, channel: 256 |
ReLU | |
convolution | kernel: 3x3, channel: 256 |
ReLU | |
max pool | kernel: 2x2 |
dropout | 0.2 |
fully connected | units: 1024 |
ReLU | |
dropout | 0.5 |
fully connected | units: 1024 |
ReLU | |
dropout | 0.5 |
softmax | units: 10 |
Images are randomly transformed 'on the fly' while they are being prepared in each batch. The CPU will prepare each batch while the GPU will run the previous batch through the network.
Stream data from SSD instead of holding all images in memory (need to install SSD first). Try different network archetectures and data pre-processing.