Simple multi-laber classification example with Pytorch and MultiLabelSoftMarginLoss (https://en.wikipedia.org/wiki/Multi-label_classification)

multilabel_example.py

import torch
import torch.nn as nn
import numpy as np
import torch.optim as optim
from torch.autograd import Variable

# (1, 0) => target labels 0+2
# (0, 1) => target labels 1
# (1, 1) => target labels 3
train = []
labels = []
for i in range(10000):
    category = (np.random.choice([0, 1]), np.random.choice([0, 1]))
    if category == (1, 0):
        train.append([np.random.uniform(0.1, 1), 0])
        labels.append([1, 0, 1])
    if category == (0, 1):
        train.append([0, np.random.uniform(0.1, 1)])
        labels.append([0, 1, 0])
    if category == (0, 0):
        train.append([np.random.uniform(0.1, 1), np.random.uniform(0.1, 1)])
        labels.append([0, 0, 1])

class _classifier(nn.Module):
    def __init__(self, nlabel):
        super(_classifier, self).__init__()
        self.main = nn.Sequential(
            nn.Linear(2, 64),
            nn.ReLU(),
            nn.Linear(64, nlabel),
        )

    def forward(self, input):
        return self.main(input)

nlabel = len(labels[0]) # => 3
classifier = _classifier(nlabel)

optimizer = optim.Adam(classifier.parameters())
criterion = nn.MultiLabelSoftMarginLoss()

epochs = 5
for epoch in range(epochs):
    losses = []
    for i, sample in enumerate(train):
        inputv = Variable(torch.FloatTensor(sample)).view(1, -1)
        labelsv = Variable(torch.FloatTensor(labels[i])).view(1, -1)
        
        output = classifier(inputv)
        loss = criterion(output, labelsv)

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        losses.append(loss.data.mean())
    print('[%d/%d] Loss: %.3f' % (epoch+1, epochs, np.mean(losses)))

output.txt

$ python multilabel.py
[1/5] Loss: 0.092
[2/5] Loss: 0.005
[3/5] Loss: 0.001
[4/5] Loss: 0.000
[5/5] Loss: 0.000

I think this code does not work as you'd expect.

As per PyTorch documentation https://pytorch.org/docs/stable/nn.html#multilabelmarginloss the target vector is NOT a multi-hot encoding:

(v.0.1.12) The criterion only considers the first non zero y[j] targets.
(v.0.4.1) The criterion only considers a contiguous block of non-negative targets that starts at the front.

And this can also be verified here https://github.com/pytorch/pytorch/blob/949559552004db317bc5ca53d67f2c62a54383f5/aten/src/THNN/generic/MultiLabelMarginCriterion.c at lines 57 and 65 for example (also please have a look at line 39 and 40 where the range of the target is checked).

In fact, the correct way of denoting a target for class 0+2 (example from line 7) should be to replace line 16:

labels.append([1, 0, 1])

with

labels.append([0,2,-1])

(as a side note, line 20 should have if category == (1, 1): to match the description at line 9)

I have trouble coding out the accuracy since the prediction variable for normal one label classification requires the max. How do we work our way around this?

Is 0.092 equivalent to 92% or 9.2% for the first iterative loss

Thank you @Renthal. I just wasted 2 hours on this and finally read your comment.

The code in this gist is incorrect. As @Renthal said, the leftmost columns for each example should be the ground truth class indices. The remaining columns should be filled with -1. Of course, each example may belong to different number of classes.

@erobic @Renthal note that he is using MultiLabelSoftMarginLoss, not MultiLabelMarginLoss.

When I change label format with -1 padded. As is shown in below:

for i in range(10000):
    category = (np.random.choice([0, 1]), np.random.choice([0, 1]))

    if category == (1, 1):
        train.append([np.random.uniform(0.1, 1), np.random.uniform(0.1, 1)])
        # labels.append([1, 0, 1])
        labels.append([0, 2, -1])

    if category == (1, 0):
        train.append([np.random.uniform(0.1, 1), 0])
        # labels.append([0, 1, 0])
        labels.append([1, -1, -1])

    if category == (0, 1):
        train.append([0, np.random.uniform(0.1, 1)])
        # labels.append([0, 0, 1])
        labels.append([2, -1, -1])

    if category == (0, 0):
        train.append([np.random.uniform(0.1, 1), np.random.uniform(0.1, 1)])
        # labels.append([1, 0, 0])
        labels.append([0, -1, -1])

But, I get amazing loss value:

[1/5] Loss: -1262.730
[2/5] Loss: -7461.019
[3/5] Loss: -18611.219
[4/5] Loss: -34584.168
[5/5] Loss: -55333.562

Final Problems: how to decode output logits of multi-class model?

I create custom multi-class loss function, but trained too slowly.

class MultilabelCrossEntropyLoss(nn.Module):
    def __init__(self):
        super(MultilabelCrossEntropyLoss, self).__init__()

    def forward(self, source: torch.Tensor, target: torch.Tensor) -> torch.Tensor:

        source = source.sigmoid()

        score = -1. * target * source.log() - (1 - target) * torch.log(1-source)
        return score.sum()

I got the result:

[1/500] Loss: 1.067
[2/500] Loss: 0.815
[3/500] Loss: 0.722
[4/500] Loss: 0.664
[5/500] Loss: 0.622
[6/500] Loss: 0.591
[7/500] Loss: 0.566
[8/500] Loss: 0.546
[9/500] Loss: 0.529
[10/500] Loss: 0.515
[11/500] Loss: 0.503
[12/500] Loss: 0.492
[13/500] Loss: 0.483
[14/500] Loss: 0.475
[15/500] Loss: 0.468
[16/500] Loss: 0.461
[17/500] Loss: 0.456
[18/500] Loss: 0.450

Why ?