使用shap库解读huggingface分类模型

需要安装transformers、shap、matplotlib、ipython等库。

Pycharm会显示 <IPython.core.display.HTML object>，只能在jupyter notebook中正常显示。

import shap
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import transformers

model_dir = "roberta/trained_model"

tokenizer = AutoTokenizer.from_pretrained(model_dir)
model = AutoModelForSequenceClassification.from_pretrained(model_dir, num_labels=2)

classifier = transformers.pipeline("text-classification", model=model, tokenizer=tokenizer, top_k=1)

explainer = shap.Explainer(classifier)
text = ["应用软件（Application software）是指专门用于满足用户特定需求的软件。"]
shap_values = explainer(text)
shap.plots.text(shap_values)

from matplotlib import pyplot as plt
plt.rcParams['font.sans-serif'] = 'AR PL UKai CN'
plt.rcParams['axes.unicode_minus'] = False
shap.plots.bar(shap_values[0, :, 1], order=shap.Explanation.argsort.flip)

其中，plt.rcParams['font.sans-serif'] = 'AR PL UKai CN'这句是为了在jupyter中显示中文。下面这段代码用于显示系统中所有字体。

from matplotlib.font_manager import FontManager
fm = FontManager()
my_fonts = set(f.name for f in fm.ttflist)
my_fonts

怎样用SHAP VALUE 解读深度学习 NLP 模型