ImageDataBunch.from_df doesn't take classes

fastai
1

I’m using fastai for inference, so I have few labels for each image. The method ‘from_df’ doesn’t take the number of classes so a lot of the images/labels get ignored when I train. Is there a way to pass in the set of classes?

This is the warning I get:

/usr/local/lib/python3.6/dist-packages/fastai/data_block.py:525: UserWarning: You are labelling your items with CategoryList.
Your valid set contained the following unknown labels, the corresponding items have been discarded.

2

hey, you have to create imageitemlist before doing imagedatabunch. Since your problem is a multi-label image classification problem. Your image item list will probably look like this.

src = (ImageItemList.from_csv(path, ‘train_v2.csv’, folder=‘train-jpg’, suffix=’.jpg’)
.random_split_by_pct(0.2)
.label_from_df(label_delim=’ ')#if your images labels are seperated by space
)
Then do the databunch

data = (src.transform(tfms, size=128)
.databunch(bs=64).normalize(imagenet_stats)
)
Hope it fixes the problem. If not, please share your notebook. I will look into it.

3

Thank you, I think the issue is that I have few images for each label, yet thousands of labels. I’m training in a distributed fashion, so it would be helpful if I could pass in the set of labels, instead of extracting them from a csv/dataframe, as the chunk that each machine gets will most likely get only a small subset of the labels. I’m not sure whether this is fine, but I feel like this could affect the embeddings I extract.