Using MSE loss on batch

1

I’m trying to use MSE loss on a batch the following way:
My CNN’s output is a vector of 32 samples. So, for example, if my batch size is 4, I’ll have an output of 4X32 samples. Each output vector needs to be loss- calculated with another vector. Then, I want to take each vector and apply on it backward function and so on.
the code as it is right now:

loss_criterion = nn.MSELoss(reduction= 'none')
...    
...
for batch_idx, data in enumerate(training_loader, 0):
        optimizerS.zero_grad()
        simulated_spec = netS(data['image_tensor'], batch_size)
        S_loss = loss_criterion(data['metadata_tensor'], simulated_spec)
        S_loss.backward()
        optimizerS.step()

S_loss is now a tensor of (batch size, 32) , and I get a run time error:
raise RuntimeError(“grad can be implicitly created only for scalar outputs”)

how can it be solved and how do i get the MSE loss per every 32 samples?

2

You would just take the sum or the mean over the batch to get the total loss. i.e. S_loss.mean() or S_loss.sum(). I’m guessing if your total loss is supposed to be MSE you would be taking the mean.