Error when train customer dataset .

[ThorPham]

I has a dataset with 6 class and when i trained bug appear like this
ValueError: size shape must match input shape. Input is 2D, size is 3
This is a my dataset config

I traced my bug and bug in line

mmsegmentation/mmseg/ops/wrappers.py

Line 27 in 7b09967

return F.interpolate(input, size, scale_factor, mode, align_corners)

and i print the "size" in F.interpolate

I don't know why the last shape is different other with 3D dimension
Here is full log bug

Can you help me fix ?

[MeowZheng]

I suggest you print the filename at

mmsegmentation/mmseg/datasets/pipelines/loading.py

Line 61 in 7b09967

img_bytes = self.file_client.get(filename)

and

mmsegmentation/mmseg/datasets/pipelines/loading.py

Line 132 in 7b09967

img_bytes = self.file_client.get(filename)

I think you might load wrong data

[ThorPham]

@MeowZheng here is log :

and here is custom dataset

My dataset with 6 class and annotation is format RGB space with extension png with 6 color.

[jinwonkim93]

Have you check those files?

[ThorPham]

@jinwonkim93 Yes I write code read image and annotation and check shape of them. It's ok. Can you help me some other method to test those files.

[ThorPham]

@jinwonkim93 @MeowZheng I know my problem. That is annotation have more than color in PALETTE

[ThorPham]

@MeowZheng I just check shape ofresults.get('gt_semantic_seg') in VOC or coco is 2D. But my dataset is 3D.
Let me know the format dataset for the annotation i should convert

[jinwonkim93]

@ThorPham
Check this issue.
#1719

[ThorPham]

@jinwonkim93 . I done with issue. Thank you so much

[RyanYip-Kat]

In fact, my annotation file is str_mask2, the dimension is also 2D, but after saving as an segmentation file, this error will still be reported? What's the matter, please? thank you

[ThorPham]

@RyanYip-Kat can you show log error ?

[LaBiXiaoChai]

@RyanYip-Kat You successfully saved a single-channel picture using opencv, but when you re-read, you use a three-channel reading method, so it will be displayed as a three-channel. you can use cv2.imread("test.png", 0)