Example of Shapley explanations for a zero-shot classifier

zero-shot-explained.py

from transformers import pipeline
from numpy import argmax
import shap
from transformers import RobertaConfig, RobertaModel
from transformers import RobertaTokenizer
import torch
import numpy as np
import scipy as sp
# import datasets
import pandas as pd

# dataset  = datasets.load_dataset("emotion", split = "train")
# data     = pd.DataFrame({'text':dataset['text'],'emotion':dataset['label']})

transformer_name = 'roberta-large-mnli'

tokenizer = RobertaTokenizer.from_pretrained(transformer_name)
model     = RobertaModel.from_pretrained(transformer_name)

dataset = pd.read_excel(r"C:\Me\custom_data.xlsx", engine='openpyxl')
data    = pd.DataFrame({'text':dataset['Text'],'topic':dataset['Topic']})

states_of_mind = ['happy','sad','angry','neutral']
hypothesis     = "I am feeling {}."

zero_shot_classifier = pipeline("zero-shot-classification",
                                tokenizer=tokenizer,
                                model=transformer_name, 
                                device=0)

def predict_zeroshot(text):
    results = zero_shot_classifier(text,
        states_of_mind,
        hypothesis_template=hypothesis
        )
    scores = results["scores"]
    # you may want to convert your scores (which are probabilities, right?) to logits
    return scores
    
explainer = shap.Explainer(predict_zeroshot, tokenizer, output_names=states_of_mind)

shap_values = explainer(data['text'][:3])
shap.plots.text(shap_values)