Multi label classification in pytorch

Question

I have a multi-label classification problem. I have 11 classes, around 4k examples. Each example can have from 1 to 4-5 label. At the moment, i'm training a classifier separately for each class with log_loss. As you can expect, it is taking quite some time to train 11 classifier, and i would like to try another approach and to train only 1 classifier. The idea is that the last layer of this classifer would have 11 nodes, and would output a real number by classes which would be converted to a proba by a sigmoid. The loss I want to optimize is the mean of the log_loss on all classes.

Unfortunately, i'm some kind of noob with pytorch, and even by reading the source code of the losses, i can't figure out if one of the already existing losses does exactly what i want, or if I should create a new loss, and if that's the case, i don't really know how to do it.

To be very specific, i want to give for each element of the batch one vector of size 11(which contains a real number for each label (the closer to infinity, the closer this class is predicted to be 1), and 1 vector of size 11 (which contains a 1 at every true label), and be able to compute the mean log_loss on all 11 labels, and optimize my classifier based on that loss.

Any help would be greatly appreciated :)

Answer 1

You are looking for torch.nn.BCELoss. Here's example code:

import torch

batch_size = 2
num_classes = 11

loss_fn = torch.nn.BCELoss()

outputs_before_sigmoid = torch.randn(batch_size, num_classes)
sigmoid_outputs = torch.sigmoid(outputs_before_sigmoid)
target_classes = torch.randint(0, 2, (batch_size, num_classes))  # randints in [0, 2).

loss = loss_fn(sigmoid_outputs, target_classes)

# alternatively, use BCE with logits, on outputs before sigmoid.
loss_fn_2 = torch.nn.BCEWithLogitsLoss()
loss2 = loss_fn_2(outputs_before_sigmoid, target_classes)
assert loss == loss2