Having trouble with VOC 2012 segmentation with the void = 255 label

vision
1

Hello, I’m trying to implement a segmentation model using the VOC 2012 data, and I was having trouble with calculating the loss.

I am using cross entropy loss, but then I get the error

RuntimeError: Assertion `cur_target >= 0 && cur_target < n_classes' failed.  at ..\aten\src\THNN/generic/ClassNLLCriterion.c:92

when I try to set the number of classes to 22 for the output. This is because the data has a label of 255 for void, and thus pytorch thinks there are 256 classes. I was wondering how to rectify this? I loaded the data using torchvision.datasets.VOCSegmentation. Another thing I was not clear about was do people actually try to label the void pixels in practice? Thanks so much

2

You could pass ignore_index=255 to the initialization of your criterion, if your model does not predict the void class.

2 Likes
3

That worked! Thanks so much

4

Oh another issue. How would you recommend computing the confusion matrix to calculate IoU? Would sklearn’s confusion_matrix function take care of this if we feed in a list of all the classes?

5

sklearn's confusion matrix should most likely work.
If you would like to fill the confusion matrix in your evaluation loop, have a look at this post.

6

I’ve noticed both the scitkit method and the for loop method is a bit slow. Is there a vectorized way of implementing this or is this just the nature of constructing the confusion matrix? Also the confusion matrix is updated for
every minibatch correct?

7

Yes, the confusion matrix is updated for each mini batch so avoid storing the values.
Also there is a vectorized (and slightly more complicated way) to fill the confusion matrix using .put_:

# Create dummy data
n_samples = 1000
n_classes = 5

conf_matrix = torch.zeros(n_classes, n_classes)

preds = torch.randn(n_samples, n_classes)
preds = torch.argmax(preds, 1)
labels = torch.randint(0, n_classes, (n_samples,))

# For loop
t0 = time.perf_counter()
for p, t in zip(preds, labels):
    conf_matrix[p, t] += 1
print('Took {}s'.format(time.perf_counter() - t0))

# put_ method
conf_matrix2 = torch.zeros(n_classes, n_classes)

t0 = time.perf_counter()
lin_index = preds * n_classes + labels  # You need to calculate the linear index here
conf_matrix.put_(lin_index, torch.tensor(1.).expand_as(lin_index), accumulate=True)
print('Took {}s'.format(time.perf_counter() - t0))
1 Like