app.py

import gradio as gr
import re
from transformers import AutoModelForSequenceClassification, AutoTokenizer, pipeline
import torch
from keybert import KeyBERT
from datasets import load_dataset
import shap
from transformers_interpret import SequenceClassificationExplainer
from ferret import Benchmark

#model_identifier = "karalif/myTestModel"
#model = AutoModelForSequenceClassification.from_pretrained(model_identifier)
#tokenizer = AutoTokenizer.from_pretrained(model_identifier)

name = "karalif/myTestModel"
model = AutoModelForSequenceClassification.from_pretrained(name)
tokenizer = AutoTokenizer.from_pretrained(name, normalization=True)

bench = Benchmark(model, tokenizer)
#text = "hvað er maðurinn eiginlega að pæla ég fatta ekki??????????"

def get_prediction(text):

    explanations_formality = bench.explain(text, target=0)
    explanations_sentiment = bench.explain(text, target=1)
    explanations_politeness = bench.explain(text, target=2)
    explanations_toxicity = bench.explain(text, target=3)

    encoding = tokenizer(text, return_tensors="pt", padding="max_length", truncation=True, max_length=200)
    encoding = {k: v.to(model.device) for k, v in encoding.items()}

    with torch.no_grad():
        outputs = model(**encoding)

    logits = outputs.logits
    sigmoid = torch.nn.Sigmoid()
    probs = sigmoid(logits.squeeze().cpu()).numpy()

    kw_model = KeyBERT()
    keywords = kw_model.extract_keywords(text, keyphrase_ngram_range=(1, 1), stop_words='english', use_maxsum=True, nr_candidates=20, top_n=5)

    response = ""
    labels = ['Politeness', 'Toxicity', 'Sentiment', 'Formality']
    colors = ['#b8e994', '#f8d7da', '#fff3cd', '#bee5eb']  # Corresponding colors for labels

    for i, label in enumerate(labels):
        response += f"<span style='background-color:{colors[i]}; color:black;'>{label}</span>: {probs[i]*100:.1f}%<br>"
    
    influential_keywords = "INFLUENTIAL KEYWORDS:<br>"
    for keyword, score in keywords:
        influential_keywords += f"{keyword} (Score: {score:.2f})<br>"
    
    return response, keywords, influential_keywords

def replace_encoding(tokens):
    return [token.replace('Ġ', ' ')
                 .replace('ð', 'ð')
                 .replace('é', 'é')
                 .replace('æ', 'æ')
                 .replace('ý', 'ý')
                 .replace('á', 'á')
                 .replace('ú', 'ú')
                 .replace('ÃŃ', 'í')
                 .replace('Ãö', 'ö')
                 .replace('þ', 'þ')
                 .replace('Ãģ', 'Á')
                 .replace('Ãį', 'Ú')
                 .replace('Ãĵ', 'Ó')
                 .replace('ÃĨ', 'Æ')
                 .replace('ÃIJ', 'Ð')
                 .replace('Ãĸ', 'Ö')
                 .replace('Ãī', 'É')
                 .replace('Ãļ', 'ý')
            for token in tokens[1:-1]]

def predict(text):
    greeting_pattern = r"^(Halló|Hæ|Sæl|Góðan dag|Kær kveðja|Daginn|Kvöldið|Ágætis|Elsku)"
    
    prediction_output, keywords, influential_keywords = get_prediction(text)
    greeting_feedback = ""
    
    modified_input = text
    for keyword, _ in keywords:
        modified_input = modified_input.replace(keyword, f"<span style='color:green;'>{keyword}</span>")
    
    if not re.match(greeting_pattern, text, re.IGNORECASE):
        greeting_feedback = "OTHER FEEDBACK:<br>Heilsaðu dóninn þinn<br>"

    response = f"INPUT:<br>{modified_input}<br><br>MY PREDICTION:<br>{prediction_output}<br>{influential_keywords}<br>{greeting_feedback}"

    # Include influential words in the response
    explanation_lists = [explanations_toxicity, explanations_formality, explanations_sentiment, explanations_politeness]
    labels = ['Toxicity', 'Formality', 'Sentiment', 'Politeness']

    response += "<br>MOST INFLUENTIAL WORDS FOR EACH LABEL:<br>"
    for i, explanations in enumerate(explanation_lists):
        label = labels[i]
        for explanation in explanations:
            if explanation.explainer == 'Partition SHAP':
                tokens = replace_encoding(explanation.tokens)
                token_score_pairs = zip(tokens, explanation.scores)
                formatted_output = ' '.join([f"{token} ({score})" for token, score in token_score_pairs])
                response += f"{label}: {formatted_output}<br>"

    response += "<br>TOP 2 MOST INFLUENTIAL WORDS FOR EACH LABEL:<br>"
    for i, explanations in enumerate(explanation_lists):
        label = labels[i]
        response += f"{label}:<br>"
        for explanation in explanations:
            if explanation.explainer == 'Partition SHAP':
                sorted_scores = sorted(enumerate(explanation.scores), key=lambda x: abs(x[1]), reverse=True)[:2]
                tokens = replace_encoding(explanation.tokens)
                tokens = [tokens[idx] for idx, _ in sorted_scores]
                formatted_output = ' '.join(tokens)
                response += f"{formatted_output}<br>"

    return response

description_html = """
    <center>
        <img src='http://www.ru.is/media/HR_logo_vinstri_transparent.png' width='250' height='auto'>
    </center>
    """

demo = gr.Interface(
    fn=predict,
    inputs=gr.TextArea(label="Enter text here:"),
    outputs=gr.HTML(label="Leiðrétt"),
    description=description_html,
    examples=[
        ["Það voru vitni að árásinni sem tilkynntu málið til lögreglu sem kom skjótt á vettvang."],
        ["Ég held þetta sé ekki góður tími fara heimsókn."],
        ["Sæl og blessuð Kristín, hvað er að frella af þér gamla??"],
        ["Hver á þenan bússtað? já eða nei."],
        ["Hafi þau svo látið gólfið þorna vel og síðan flotað það til lagfæringar eftir motturnar."],
    ],
    theme=gr.themes.Default(primary_hue="red", secondary_hue="pink")
)

demo.launch()