Quick question: resize of target image?

fastai
21

I am so sorry, I followed everything even went through the docs and I couldn’t understand why it was behaving strangely to me.
So I went to 2019 Lesson7-superres.ipynb and it was written this way too and I couldn’t find out why it works there but in my notebook doesn’t.

I want to resize my targets only, how do I do it?
Input size = 96 and target size = 96 * 2

def get_data(bs, size):
data = (src.label_from_func(lambda x: path_hr/x.name)
       .transform(get_transforms(), size = size)
       .databunch(bs=bs, no_check = True).normalize(imagenet_stats, do_y=True))

data.c = 3
return data
22

There is no way to do this directly without writing a custom transform, so pre-resize your targets to the right size and don’t pass any size to .transform is probably the easiest way.

23

Fantastic, thank you very much @sgugger.

This is what was confusing to me, because in the 2019 Lesson7-superres.ipynb and 2018 lesson 14 enhance.ipynb suggest that both fastai v0.4 and fastai v1 had a way of doing what I was attempting.

Please give Lesson7-superres.ipynb a look and let me know what you found.

Cuz this helps when doing progressive resizing. It’s how @jeremy did it so quick in the video lesson 7

24

I got it @sgugger it’s working, thank you, you got me thinking so went to check the docs.
I just needed to change transform to transform_y. Dunno how @Jeremy did it in the video

def get_data(bs, size):
    data = (src.label_from_func(lambda x: path_hr/x.name)
           .transform_y(tfms = get_transforms(), size = size)
           .databunch(bs=bs, no_check = True).normalize(imagenet_stats, do_y=True))

data.c = 2
return data