Request: allow val_idxs provision to image_data_from_csv


(RobG) #1

I’m enjoying migrating my notebooks to v1 and can now load/train/predict vision projects.

I’m not sure if feature requests / observations should be posted in the forum or git, so my apologies in advance for any breach of etiquette here.

In v0 we can pass an explicit list of val_idxs to ImageClassifierData.from_csv but this capability doesn’t appear to exist with the image_data_from_csv of v1 which only takes a valid_pct. This is useful functionality when using cross validation. I was able to work around this and build my own loaders with explicit validation data but it would be nice if the previous capability was preserved.

No doubt I’ll have other ‘observations’ so clarity on where to post would be welcome :slight_smile:


#2

Yes, this is indeed a useful feature. Note that you can already decide on your validation set by separately creating your train_ds and valid_ds as you wish.
Will implement this at some point, note that there is a dedicated thread for features request, but creating a topic is also fine.