app.py

import torch
from transformers import AutoModel, AutoConfig
import joblib
import pickle
import pandas as pd
import numpy as np


excel_path = "./Dict.xlsx"
keywords = pd.read_excel(excel_path)

keywords = keywords.assign(Words = keywords.Words.str.lower())
keywords = keywords.assign(Words = keywords.Words.str.split(",")).explode("Words")
keywords = keywords.assign(Words = keywords.Words.str.strip())
keywords = keywords.assign(Fields = keywords.Fields.str.strip())

# Load the model (joblib file) and vectorizer (pickle file) from Hugging Face
model_path = "./classifier.joblib"
vectorizer_path = "./vectorizer"

classifier = joblib.load(model_path)
with open(vectorizer_path, "rb") as f:
    tfidf_vectorizer = pickle.load(f)


import gradio as gr
def classify_text(text):

  import pandas as pd
  import numpy as np
  from snowballstemmer import stemmer
  import re
  import nltk

  Sample = pd.DataFrame({"Description": [text]})
  Sample = Sample.assign(Description = Sample.Description.str.lower())
  Sample = Sample.assign(Description = Sample.Description.str.lower())
  Sample = Sample.reset_index(names = "n").reset_index(names = "Doc").drop("n",axis = 1)
  Sample = Sample.assign(Description = Sample.Description.astype("str"))
  Sample = Sample.groupby(["Doc"]).Description.apply(" ".join).reset_index()
  Sample['Description'] = Sample['Description'].str.strip()
  keyword_mapping = {}
  for keyword, field in zip(keywords['Words'], keywords['Fields']):
    keyword_mapping.setdefault(keyword, []).append(field)
    # Function to find matching keywords in a description and return the first corresponding field
  def find_field(description):
    matching_fields = set()
    for keyword, fields in keyword_mapping.items():
        if re.search(r'\b{}\b'.format(re.escape(keyword)), description):
            matching_fields.update(fields)
    return list(matching_fields)[0] if matching_fields else "None"

  # Apply the function to the 'Description' column of df_sample
  Sample['AssignedField'] = Sample['Description'].apply(find_field)
  done = Sample.loc[~Sample.AssignedField.isna()]
  pos = Sample.loc[Sample.AssignedField.isna()].Doc

  nltk.download('stopwords')
  stopword_es = nltk.corpus.stopwords.words('spanish')
  stemmer_es = stemmer("spanish")

  Sample = pd.DataFrame({"Description": [text]})
  Sample = Sample.assign(Description = Sample.Description.str.lower())
  Sample = Sample.reset_index(names = "n").reset_index(names = "Doc").drop("n",axis = 1)
      # Split into tokens and reshape the data to one row per token
  Sample = Sample.assign(Description = Sample.Description.str.split()).explode("Description")
  Sample = Sample.assign(Description = Sample.Description.str.replace('([&%—●•:–>.€ª,;()-/])',''))
  Sample = Sample.assign(Description = Sample.Description.str.replace('"',''))
  Sample = Sample.assign(Description = Sample.Description.str.replace('([“”\[\]\d+])',''))
  Sample = Sample.reset_index()
  # Filter Stopwords
  positions = [i for i, desc in enumerate(Sample.Description) if desc in stopword_es]
  Sample.drop(positions, inplace=True)

  stemmer_es = stemmer("spanish")
  Sample["Description"] = Sample["Description"].apply(lambda word: stemmer_es.stemWord(word))

  Sample = Sample.assign(Description = Sample.Description.astype("str"))
  Sample = Sample.groupby(["Doc"]).Description.apply(" ".join).reset_index()

  new_text_tfidf = tfidf_vectorizer.transform(Sample["Description"])

  # Make probability predictions for the new text documents
  predicted_probabilities = classifier.predict_proba(new_text_tfidf)

  # Get the top four predicted class labels for each text document
  top_four_predictions = []
  for probs in predicted_probabilities:
    # Sort the probabilities and get the indices of the top four
    top_four_indices = sorted(range(len(probs)), key=lambda i: probs[i], reverse=True)[:3]

    # Get the corresponding class labels for the top four indices
    top_four_classes = [classifier.classes_[i] for i in top_four_indices]

    top_four_predictions.append(top_four_classes)

  preds = pd.DataFrame(top_four_predictions).rename(columns = {0:"Pred_1",
                                                    1:"Pred_2",
                                                            2: "Pred_3"})

  results_1 = pd.concat([Sample.reset_index(),preds],axis = 1).drop("index",axis = 1).assign(AssignedField = "None")[["Doc","Description","AssignedField","Pred_1","Pred_2","Pred_3"]]
  done[["Pred_1","Pred_2","Pred_3"]] = ["None","None","None"]
  final_results = results_1.merge(done, on = ["Description"], how = "left")
  return done["AssignedField"][0],results_1["Pred_1"][0],results_1["Pred_2"][0],results_1["Pred_3"][0]

# Create a Gradio interface
iface = gr.Interface(
    fn=classify_text,
    inputs=gr.Textbox(label = "Description", placeholder= "Enter your Service Description", lines = 3),
    outputs=([gr.Textbox() for _ in range(4)]),
    title="Text Classification App",
    theme = gr.themes.Default(primary_hue="blue"),
    examples = [["ilaap"],["finrep"],["liquidez"],["Data governance"]])  # Use a compact theme
# Launch the Gradio interface
iface.launch()