What if I told you that you could pick up a library model and instantly classify text with arbitrary categories without any training or fine tuning?
That is exactly what we are going to do with Hugging Face’s zero-shot learning model. We will also be using libpython-clj to do this exploration without leaving the comfort of our trusty Clojure REPL.
What’s for breakfast?
We’ll start off by taking some text from a recipe description and trying to decide if it’s for breakfast, lunch or dinner:
"French Toast with egg and bacon in the center with maple syrup on top. Sprinkle with powdered sugar if desired."
Next we will need to install the required python deps:
pip install numpy torch transformers lime
Now we just need to set up the libpython clojure namespace to load the Hugging Face transformers library.
1 2 3 4 5 6 7
Setup is complete. We are now ready to classify with zeroshot.
Classify with Zero Shot
To create the classifier with zero shot, you need only create it with a handy pipeline function.
After that you need just the text you want to classify and category labels you want to use.
1 2 3
Classification is only a function call away with:
1 2 3 4
Breakfast is the winner. Notice that all the probabilities add up to 1. This is because the default mode for
softmax. We can change that so the categories are each considered independently with the
1 2 3
This is a really powerful technique for such an easy to use library. However, how can we do anything with it if we don’t understand how it is working and get a handle on how to debug it. We need some level of trust in it for utility.
This is where LIME enters.
Using LIME for Interpretable Models
One of the biggest problems holding back applying state of the art machine learning models to real life problems is that of interpretability and trust. The lime technique is a well designed tool to help with this. One of the reasons that I really like it is that it is model agnostic. This means that you can use it with whatever code you want to use with it as long as you adhere to it’s api. You need to provide it with the input and a function that will classify and return the probabilities in a numpy array.
The creation of the explainer is only a
1 2 3 4 5
We need to create a function that will take in some text and then return the probabilities for the labels. Since the zeroshot classifier will reorder the returning labels/probs by the value, we need to make sure that it will match up by index to the original labels.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
Finally we make an explanation for our text here. We are only using 6 features and 100 samples, to keep the cpus down, but in real life you would want to use closer to the default amount of
5000 samples. The samples are how the explainers work, it modifies the text over and over again and sees the difference in classification values. For example, one of the sample texts for our case is
' Toast with bacon in the center with syrup on . with sugar desired.'.
1 2 3 4 5 6 7
Now it becomes more clear. The model is using mainly the word
toast to classify it as breakfast with supporting words also being
syrup. The word
the is also in there too which could be an artifact of the low numbers of samples we used or not. But now at least we have the tools to dig in and understand.
Exciting advances are happening in Deep Learning and NLP. To make them truly useful, we will need to continue to consider how to make them interpretable and debuggable.
As always, keep your Clojure REPL handy.