[FIXED] How to explicitly include decision variables in pytorch

Issue

I’m implementing a neural network from a paper in PyTorch. Here is the screenshot of the paper: Here N_Psi is a neural network, and K is a decision matrix. The way I came up with is to include an extra linear layer for K, but wondering if there's any chance to explicitly define K as decision variables in a more direct way?

Any hint would be very helpful. Thanks in advance!

Solution

You can define your decision matrix as a fully connected layer with no bias, using nn.Linear. Then you have to add this additional layer to your optimizer parameter list. Given N your neural network, K you linear layer, and optim your torch.optim.Optimizer class, you can:

optimizer = optim(list(N.parameters()) + list(K.parameters()))

Then in the inference stage, given x_n+1 and x_n, do something like:

mse = F.mse_loss(N(x_n+1), K(N(x_n)))
reg_1 = K.weight.pow(2).sum()
reg_2 = p2v(N.parameters()).sum()
loss = mse + lamb_1*reg_1 + lamb_2*reg_2

Where we imported:

• torch.nn.functional as F
• torch.nn.utils.parameters_to_vector as p2v

Answered By - Ivan