The matrix above * the first row of the matrix = B00 a00*w00 + a01*w01 + a02*w02 + a10*w10 + a11*w11 + a12*w12 + a20*w20 + a21*w21 + a22*w22
the matrix above* The second row
of the matrix = the matrix above B01 * the third row
of the matrix = the matrix above B10 * the fourth row of the matrix = B11
Model structure
- Pooling layer is replaced with convolutional layer
- Use strided covolutions on D
- Use fractional-strided convolutions on G
- Both G and D use batch normalization to help solve the problem of poor initialization, and the gradient is propagated to each layer
- BN should not be used in the input and output layers
- Remove the fully connected layer and use global pooling
- Relu is used on G except for the output layer using tanh
- Use LeakyReLU on D