ImageClassifierData.from_csv() error

Part 1 (2018) Beginner (2018)
1

(I also posted this in the Deep Learning category but got no responses so I’m going to post it here as well.)

Hi everyone.

I am trying to load bounding boxes that I saved to a csv file with pandas. I’ve manually opened the CSV file after saving it and ensured the contents are correct. However, when I call
md = ImageClassifierData.from_csv(PATH, 'imgs', BBOX_CSV, tfms=tfms, bs=bs, continuous=True) and then view the data with
next(iter(md.val_dl))
I see that a lot of my data (bounding box coordinates) have been turned to 0s. About half of the bounding boxes were turned into 0s. Does anyone have any ideas for what could be going wrong? Also, let me know if I can be more specific about anything.

Thanks!
Zane

2

Did you figure this out? I am seeing something similar and I am trying to figure this out.

Thanks!

-Dimitris

3

I didn’t necessarily figure this out, but I could work around the issue by removing the data transformations I was using. (Basically I didn’t use the normalize and the later denorm steps).

I also tried messing around with feeding the data in as numpy arrays (if removing the transforms themselves doesn’t fix it) and that worked as well.

Let me know if this helps,
Zane

4

Thanks Zane>> It works…I tried removing the tfms_y argument and y-Boundary Boxes of the valid data are intact and not 0.

5

The problem there was --we try to resize the Image to 224 and the boundary boxes also needs to be resized…Maybe image was resized to 1200 to 224, So the Boundary_box also needs to be resized here…Instead try Original Image in its original size and do not transform y-tfm_y-which resizes the boundary Box.