perform NMF processing on input data A: A≈UV

initialize parameters: Conv 1 (3, 16), Conv 2 (16, 32), Conv 3(32, 64)

define activation function σ(x)=max(0, x)

define pooling operations P^(l)(i, j)=maxp.q∈[1∶s]F^(l)(i+p, j+q)

define dropout regularization and dropout probability p=0.5

initialization optimizer, learning rate η=0.0005

execute for each training round:

forward propagation:

compute convolution output, apply ReLU, pooling and dropout

map features to full connection layer

calculate predicted value ŷ

calculate cross entropy loss function 

backpropagation:

updating model parameters using the Adam optimizer

validation:

monitoring validation set losses

if the loss is not improved for several consecutive rounds

stop the training in advance

return to model