Loss in multiple outputs network

1

Hi all, i have a network composed by a RNN layer , followed by 3 separated fully connected portions that output each a single value:
Untitled Diagram.drawio

The net is defined by :

class RNN(nn.Module):
    def __init__(self, hidden_size, num_layers,fd_n):
        super(RNN, self).__init__()
        self.hidden_size = hidden_size
        self.num_layers = num_layers
        self.rnn = nn.RNN(input_size=510,hidden_size=hidden_size,num_layers=num_layers,batch_first=True,dropout=0.0)
        self.fcx   = nn.Linear(in_features=hidden_size,out_features=fd_n)
        self.fcx2  = nn.Linear(in_features=fd_n,out_features=1)
        self.fcy   = nn.Linear(in_features=hidden_size,out_features=fd_n)
        self.fcy2  = nn.Linear(in_features=fd_n,out_features=1)
        self.fcw   = nn.Linear(in_features=hidden_size,out_features=fd_n)
        self.fcw2  = nn.Linear(in_features=fd_n,out_features=1)

    def forward(self,x):
        h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(device) 
        out_rnn, _ = self.rnn(x,h0)
        outx     = self.fcx(F.relu(out_rnn[:, -1, :]))
        outx     = self.fcx2(F.relu(outx))
        outy     = self.fcy(F.relu(out_rnn[:, -1, :]))
        outy     = self.fcy2(F.relu(outy))
        outw     = self.fcw(F.relu(out_rnn[:, -1, :]))
        outw     = self.fcw2(F.relu(outw))

        return outx,outy,outw

while the training part is:

        outputs_x, outputs_y, outputs_w  = model(inputs)
        loss_x = criterion_x(outputs_x, labels)
        loss_y = criterion_y(outputs_y, labels)
        loss_w = criterion_w(outputs_w, labels)

        loss = loss_x + loss_y + loss_w

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

At the beginning I always get :

UserWarning: Using a target size (torch.Size([1520, 3])) that is different to the input size (torch.Size([1520, 1])). This will likely lead to incorrect results due to broadcasting. Please ensure they have the same size.

am I correctly backpropagating ?
Thanks

2

I would not ignore the warning as broadcasting might be used which could yield invalid loss values.
The warning is raised since the criterion_x/y/w calls are using a target with 3 features while each model output has 1 feature/logit only. Assuming the target contains the values for all 3 model outputs, you could torch.cat((outputs_x, outputs_y, outputs_w), dim=1) the outputs and calculate the loss afterwards.

3

Seems like the problem is solved.
thanks :slight_smile: