Hello everyone, I’m having a “weighted” problem with this particular loss function.
I’m training a capsule network model with two classes, but the classes are unbalanced(class 0 has 32% of the training samples of the dataset and class 1 has 68%).
I’m using inversed weights to force the network to learn from the class 0 data, but the model gets 68% of accuracy all the time, it’s just learning from the class 1 (just in the capsule network model).
I have tried with an InceptionV3 with weighted cross entropy loss as criterion, to see if the weighted works:
criterion = nn.CrossEntropyLoss(torch.FloatTensor([0.68, 0.32]).cuda())
And the model gets 98% accuracy, it works.
But when I try to weight the capsule network loss function like this:
def margin_loss(self, x, labels, size_average=True): batch_size = x.size(0) v_c = torch.sqrt((x**2).sum(dim=2, keepdim=True)) left = F.relu(0.9 - v_c).view(batch_size, -1) right = F.relu(v_c - 0.1).view(batch_size, -1) loss = labels * left + 0.5 * (1.0 - labels) * right # This line is not included in the original code, it's the "weighted" that i want to apply loss = loss * self.weights loss = loss.sum(dim=1).mean() return loss
Nothing happen, is still the same accuracy.
I have tried with the following weights:
self.weights = torch.Tensor([0.68, 0.32]).cuda() # 68% accuracy in all epochs self.weights = torch.Tensor([0.99, 0.01]).cuda() # Check the attached image
But the model is still stuck on 68% accuracy(class 1).
Look at this image, on the right are the inceptionv3 results, on the left the capsule network results in [loss, accuracy] format, every record is an epoch:
The code is based on this capsule networks repo
You can check the complete model in there.
Any idea how to weight that particular loss function? I will be reading your answers! Thanks!