Flowers Dataset Labels Unavailable

Hi there, I’m trying to extend my knowledge learned on the pets dataset on classification by using the flowers dataset. Has anyone worked on this here? If so could you let me know how you got the image labels? The labels provided in the train and validation text files are numbered and aren’t labeled by the name of the flower and by the looks of it the labels available on the main oxford link are different…

If we mean the active competition, this is often done on purpose so people can’t easily go cheat and bring in more data (with the potential to get the test set)

I don’t quite know if it’s active :sweat_smile:. I just picked up one of classification datasets on fastai2 titled ‘FLOWERS’…

1 Like

Ah nevermind, that’s different then! (With the FastGarden makes it a bit confusing). I’d look at where the source is coming from for that dataset first as fastai simply hosts it, they didn’t make it

Did that as well, the main site provides a labels.mat file but it doesn’t seem to match with the number of images provided in the fastai dataset.

Also is there an easy way to read .mat files in fastai?

1 Like

You can read .mat’s into pandas

Were you looking at this site? https://www.robots.ox.ac.uk/~vgg/data/flowers/102/categories.html

Yup! https://www.robots.ox.ac.uk/~vgg/data/flowers/102/imagelabels.mat

I was using scipy’s loadmat method which converted it into a python dict. Could you let me know how it’s done in pandas?

After reading the .mat files using scipy i get the number instead of the names themselves…Am I doing something wrong here?

there are 8189 ordered labels so you should find the ordered names from matlab or image files from the folder and join or merge them with labels.

The image files are basically named with the numbers itself.

And reading the .mat files again gives me 8189 files which are numbered

1 Like

I had a brain fart first thing in the morning. Takes some converting but you can get there:

Haha! Kinda referred to that exact page and got this:

1 Like

@harish3110 I’d check a few first manually to make sure it’s #1=1, etc but here’s something to look at:

https://gist.github.com/JosephKJ/94c7728ed1a8e0cd87fe6a029769cde1

Ah! hopefully, these are rightly labeled! I guess I’ll work with these names for now and map them accordingly! Thanks! :slight_smile:

Update: The above link seems to be legit @muellerzr! The few types of flowers I can a decipher seem to match! :slight_smile:

1 Like

I actually found quite a few mistakes in the link above, here is a cleaned version: GitHub - bdevnani3/oxfordflowers102-label-name-mapping: The Oxford Flowers dataset provides numerical categorical labels and a list of flowers that these labels cover but no mapping between the flower names and the labels. This dictionary contains that mapping.