app.py

import streamlit as st
import streamlit.components.v1 as components
from transformers import (AutoModelForSequenceClassification, AutoTokenizer,
                          pipeline)
import shap
from PIL import Image

st.set_option('deprecation.showPyplotGlobalUse', False)
output_width = 800
output_height = 300
rescale_logits = False



st.set_page_config(page_title='Text Classification with Shap')
st.title('Interpreting HF Pipeline Text Classification with Shap')

form = st.sidebar.form("Model Selection")
form.header('Model Selection')

model_name = form.text_input("Enter the name of the text classification LLM (note: model must be fine-tuned on a text classification task)", value = "Hate-speech-CNERG/bert-base-uncased-hatexplain")
form.form_submit_button("Submit")


@st.cache_data()
def load_model(model_name):
    tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
    model = AutoModelForSequenceClassification.from_pretrained(model_name)

    return tokenizer, model

tokenizer, model = load_model(model_name)
pred = pipeline("text-classification", model=model, tokenizer=tokenizer, top_k=None)
explainer = shap.Explainer(pred, rescale_to_logits = rescale_logits)

col1, col2 = st.columns(2)
text = col1.text_area("Enter text input", value = "Classify me.")

result = pred(text)
top_pred = result[0][0]['label']
col2.write('')
for label in result[0]:
    col2.write(f'**{label["label"]}**: {label["score"]: .2f}')

shap_values = explainer([text])

force_plot = shap.plots.text(shap_values, display=False)
bar_plot = shap.plots.bar(shap_values[0, :, top_pred], order=shap.Explanation.argsort.flip, show=False)

st.markdown("""
<style>
.big-font {
    font-size:35px !important;
}
</style>
""", unsafe_allow_html=True)
st.markdown(f'<center><p class="big-font">Shap Bar Plot for <i>{top_pred}</i> Prediction</p></center>', unsafe_allow_html=True)
st.pyplot(bar_plot, clear_figure=True)

st.markdown('<center><p class="big-font">Shap Interactive Force Plot</p></center>', unsafe_allow_html=True)
components.html(force_plot, height=output_height, width=output_width, scrolling=True)