For me, it seems to require CMD-OPTION-C on macOS (Safari 12.0) not CMD-OPTION-J to pull up the JavaScript Console.
How to address issues with imbalanced data? aka some classes have very few photos compared to others?
Aka data augmentation that is weighted by ratios of class imbalances or something like that?
thanks
When there are unsuitable images (e.g. drawings instead of photos) in the training dataset. How can I best remove them? Should I?
What’s the metrics=error_rate line for?
Yes, I’d like to know how to handle images that are inherently not squared, say all of them will be very rectangular
When doing a lr_find(), is it actually training the model?
Is the size of validation-set always 20% or does it depend upon your data size ?
This is going to be explained in a few minutes
It should be removed manually.
No, it’s trying out different LRs to help find the best via visualization.
It’s a mock training with a various range of learning rates. But the original model is loaded after, so it doesn’t change the weights.
looks like 3e-3 would have been better
what if curve is seen flat for many iterations unlike this one where it goes high in just few iterations
what is the y axis in the lr_find graph?
Question : Is training loss and error rate same thing computed on training data and test data ?
Error rate
Karpathy said validation sets should be made carefully, Rachel also has an article about it. when Is it ok to randomly split data ?
Just on the training set.
Not sure, but @william has an intersting approach to curating scraped datasets. Have a look here:
You can send a ‘size’ parameter to ImageDataBunch that will crop and pad your images to get them to be of your desired size.