Classe Order in Multi prediction

fastai fastai dev
1

How to impose a class order in multi label modeling.
I am getting different class order sometimes, and it bugs me. How I can impose a specific label order when creating the data?
Just to clarify, I created a model, trained it and predict, the class order was:
['a', 'b', 'c']
Now I loaded more data, and when I do

data.classes
>>]'b', 'a', 'c']

I get a different order.
So my old initial trained model (and weights) don’t “work” with this new data order, how I can create the Databunch object with the order I want in the output?

In previous versions, the class order was obtained from the first appearance of each label in the training dataset. This is not practical for me.

2

It’s weird, they’re coded in order though.
You can always pass your classes when you create the labels.

3

I am using the block api, and a model that was trained before with the old ImageDataBunch.from_df.
Where I should pass the label order? in .label_from_df ?

4

There, and you can pass classes=bla to use this specific list of class (in that order).

1 Like
5

thanks, it works for me.