- network structure:
generator.layers = {
struct('type', 'input', 'output_shape', [100, batch_size])
struct('type', 'fully_connect', 'output_shape', [3136, batch_size], 'activation', 'leaky_relu')
struct('type', 'reshape', 'output_shape', [7,7,64, batch_size])
struct('type', 'conv2d_transpose', 'output_shape', [14, 14, 32, batch_size], 'kernel_size', 5, 'stride', 2, 'padding', 'same', 'activation', 'leaky_relu')
struct('type', 'conv2d_transpose', 'output_shape', [28, 28, 1, batch_size], 'kernel_size', 5, 'stride', 2, 'padding', 'same', 'activation', 'sigmoid')
};
discriminator.layers = {
struct('type', 'input', 'output_shape', [28, 28, 1, batch_size])
struct('type', 'conv2d', 'output_maps', 32, 'kernel_size', 5, 'padding', 'same', 'activation', 'leaky_relu')
struct('type', 'sub_sampling', 'scale', 2)
struct('type', 'conv2d', 'output_maps', 64, 'kernel_size', 5, 'padding', 'same', 'activation', 'leaky_relu')
struct('type', 'sub_sampling', 'scale', 2)
struct('type', 'reshape', 'output_shape', [3136, batch_size])
struct('type', 'fully_connect', 'output_shape', [1, batch_size], 'activation', 'sigmoid')
};
- result:
- network structure:
generator.layers = {
struct('type', 'input', 'output_shape', [100, batch_size])
struct('type', 'fully_connect', 'output_shape', [1024, batch_size], 'activation', 'relu')
struct('type', 'fully_connect', 'output_shape', [28*28, batch_size], 'activation', 'sigmoid')
struct('type', 'reshape', 'output_shape', [28, 28, 1, batch_size])
};
discriminator.layers = {
struct('type', 'input', 'output_shape', [28,28,1, batch_size])
struct('type', 'reshape', 'output_shape', [28*28, batch_size])
struct('type', 'fully_connect', 'output_shape', [1024, batch_size], 'activation', 'relu')
struct('type', 'fully_connect', 'output_shape', [1, batch_size], 'activation', 'sigmoid')
};
- result:
https://grzegorzgwardys.wordpress.com/2016/04/22/8/
Dumoulin V, Visin F. A guide to convolution arithmetic for deep learning[J]. 2016.
https://github.com/rasmusbergpalm/DeepLearnToolbox/tree/master/CNN
http://neuralnetworksanddeeplearning.com/index.html