Home > Software engineering >  How can I create a loss function that will push the actual NN weights to move?
How can I create a loss function that will push the actual NN weights to move?

Time:10-23

I have a simple NN:

import torch
import torch.nn as nn
import torch.optim as optim

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.fc1 = nn.Linear(1, 5)
        self.fc2 = nn.Linear(5, 10)
        self.fc3 = nn.Linear(10, 1)

    def forward(self, x):
        x = self.fc1(x)
        x = torch.relu(x)        
        x = torch.relu(self.fc2(x))
        x = self.fc3(x)
        return x

net = Model()

opt = optim.Adam(net.parameters())

I also have some input features:

features = torch.rand((3,1)) 

I can train it normally with a simple MSE loss function:

for i in range(10):
    opt.zero_grad()
    out = net(features)
    loss = torch.mean(torch.square(torch.tensor(5) - torch.sum(out)))
    print('loss:', loss)
    loss.backward()
    opt.step()

However, I'm trying to create a loss function that takes the actual weight values into account:

loss = 1 - torch.mean(torch.tensor([torch.sum(w_arr) for w_arr in net.parameters()]))

But I'm getting an error:

RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

The goal here is to get each weight's value closest to 1 (or any other value) as possible.

CodePudding user response:

A quick error fix will be to include requires_grad = True while creating tensor. This way -

loss = 1 - torch.mean(torch.tensor([torch.sum(w_arr) for w_arr in net.parameters()], requires_grad=True))

But when converting list of weights to tensor, torch doesn't know origin of that tensor so the loss doesn't decrease. One way to do it is

for i in range(500):
    opt.zero_grad()
    out = net(features)
    loss = torch.mean(torch.square(torch.tensor(5) - torch.sum(out)))
    len_w = 0
    for w_arr in net.parameters():
        loss  = torch.mean(torch.abs(1 - w_arr))
        len_w  = 1
    loss /= len_w
    print('loss:', loss)
    loss.backward()
    opt.step()

In this way of loss computation, it makes sure that all the weights are close to 1.

  • Related