Output n_class is less than len(class_name) in training?

[ymzlygw]

Hi, Thanks for your work. I used your code to train a multi-class segmentation task. I have 10 classes, but after training, the output mask just have 8 classes( using np.unique). Do you know how this happend? Thanks for your reply!

[ymzlygw]

I make sure that in dataset part, I input all 10 classes name. Note that the two classes( nose, eyebrow) that not be detected are more smaller than other class, is that the reason? My config is :

ENCODER = 'resnet34'
ENCODER_WEIGHTS = 'imagenet'
CLASSES = ['background','face', 'skin', 'clothes', 'hair', 'eye', 'eyeball', 'mouth', 'nose', 'eyebrow']
ACTIVATION = 'softmax2d' 

pr_mask = (pr_mask.squeeze().cpu().numpy().round()) # (10, 512, 512) But after np.argmax: vi_mask = np.argmax(pr_mask, axis=0) n = np.unique(vi_mask) # here just have 8 classes: [0,1,2,3,4,5,6,7] , not included 8,9

Or my config is not ok for segmentation of smaller part?

[Arunavameister]

I have a similar problem where my output has 4 classes whereas I expect 5. The training seems to be fine and the the IoU(during training) seems quite high as well. @ymzlygw Have you found a solution yet?

[Arunavameister]

@qubvel Have you ever encountered this problem?

[qubvel]

Your model does not predict some classes, don't see anything strange here suppose you have classes: car, road, sky, grass but on the picture you have only: road, grass and sky

in that case if model not predict "car" and number of predicted classes is less than number of total classes, and that's correct

[Arunavameister]

Thanks for the reply, what you are saying is obviously true but in my case I am doing inference only on the train set and I have a really low loss and a really high IoU which makes me think that I am obviously missing something. The model should atleast overfit. I dont expect you to debug the problem for me ofcourse, was just wondering if this has been encountered before.

[github-actions[bot]]

This issue is stale because it has been open 60 days with no activity. Remove stale label or comment or this will be closed in 7 days.

[github-actions[bot]]

This issue was closed because it has been stalled for 7 days with no activity.