I made a model which predicts how significantly the pupil is dilated

(Michael Woodburn) #1

I call it Eyetrain. I would love feedback.

It uses float labels, not discrete categories.
I’m waiting for a larger dataset which is coming within a month’s time.

Two issues:

  1. The images appear distorted when displayed in the Image Bunch… I hope they are not being read that way. Perhaps a change to get_transforms() will fix it?

  2. log_preds = learn.predict(img)
    This is the last line. It gives me a tuple, which I do not know what to do with. e^log_preds[0] = PLR seems logical but… it usually gives me a number >0.8 which is NOT close to the labels the model was taught off.

1 Like

(eduarda lacerda) #2

Hey! I am having the same problem when the data bunch is displayed. Since this topic was created months ago, I suppose you have already fixed the problem. Are the images really being read that way? How did you managed to fix it?

0 Likes

(Michael Woodburn) #3

tfms = [[],[]]

data = (ImageList(fnames_nomeds)
.split_by_rand_pct(0.2)
.label_from_func(get_subject_label)
.transform(tfms, size=224)
.databunch())

Set transforms to empty

1 Like