How to use torch to achieve a search algorithm？

vision
1

Give two lists, P=[p_1,p_2,p_3,..,p_i] with i positive value, N=[n_1,n_2,...,n_j] with j negative value, give a target T, I want to select some elements from P and N so that their sum is T.

I want to use Torch to achieve this, So, the Goal is to minimize the function:
Loss= w_1*P + w_2*N - T,
where w_1 and w_2 are trainable parameters.
But the loss does not converge but decreases linearly, from positive to negative values.

data=torch.load('use_data.pkl')

delta = torch.zeros(data[0].shape, requires_grad=True, device="cuda") 
`#P and delta shape: 382 `

delta2 = torch.zeros(data[2].shape, requires_grad=True, device="cuda") # N
`#N and delta2 shape: 386 `
P=data[0]
N=data[2]

T=data[4][0]

opt = torch.optim.Adam([delta,delta2], lr=0.0001)

for s in range(10000):
    opt.zero_grad()
    main_loss=torch.sum(delta*P)+torch.sum(delta2*N)-T
    #loss = main_loss + (delta.norm(2) + delta2.norm(2))*0.01
    main_loss.backward()
    opt.step()
2

It sounds a bit as if your goal might be to get to zero rather than minimize.
In that case, you could, for example minimize the square of what you currently have, i.e.
loss = (w_1 * P + w_ 2 * N - T) ** 2

Best regards

Thomas

1 Like
3

Thanks for your reply, and there has another problem: why are the element values of w_1 and w_2 the same? Each element in w_1 is close to 0.6, and w_2 is close to -0.6 ? :sweat_smile: