Problems with model prediction (incorrect output dimensions)

Part 1 (2019)
1

I was trying out an image segmentation task. While trying to train the model, I was getting the error:

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

I read some threads here and in an attempt to narrow down the problem, I tried to make a prediction and I noticed a problem. The following is by DataBunch:

ImageDataBunch;

Train: LabelList (116 items)
x: SegmentationItemList
Image (3, 128, 128),Image (3, 128, 128),Image (3, 128, 128),Image (3, 128, 128),Image (3, 128, 128)
y: SegmentationLabelList
ImageSegment (1, 128, 128),ImageSegment (1, 128, 128),ImageSegment (1, 128, 128),ImageSegment (1, 128, 128),ImageSegment (1, 128, 128)
Path: /content/drive/My Drive/AutoBound Data/cleaned/originalImages;

Valid: LabelList (29 items)
x: SegmentationItemList
Image (3, 128, 128),Image (3, 128, 128),Image (3, 128, 128),Image (3, 128, 128),Image (3, 128, 128)
y: SegmentationLabelList
ImageSegment (1, 128, 128),ImageSegment (1, 128, 128),ImageSegment (1, 128, 128),ImageSegment (1, 128, 128),ImageSegment (1, 128, 128)
Path: /content/drive/My Drive/AutoBound Data/cleaned/originalImages;

Test: None

When I make a prediction on a batch (9 items per batch), I get an output of the following size:

torch.Size([9, 2, 128, 128])

The segmented image in my data bunch had the dimensions [1,128,128], but the output image has the above dimensions. I believe this was causing the problem. How do I fix this?

2

How many classes you have for your segmentation task?

3

Just two classes. I’m trying to segment satellite images into areas with and without images.
This is a sample segmented image

segmented
mED0SPr.png576×530

and this is a sample original image

original
FODVbvY.png576×530

4

The output dimension [9, 2, 128, 128] is giving you the probabilities. So you need to take the argmax along the second dimension. The 2 here is prob for every class. So if you had 10 classes this value would have been 10.

5

Okay. So, the training should happen properly. But I keep getting that error I mentioned before:

While trying to train the model, I was getting the error:

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

Some threads mentioned they might be because of index out of bounds.

6

The problem is probably due to wrong dimensions in your architecture. Can you check your final linear layers for correct dimensions?

7

Hi Vishal, have you solved this problem? Because I am having the same error :slight_smile:

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