app.py

import gradio as gr
import torch
import joblib
import numpy as np
import pandas as pd
from transformers import AutoTokenizer, AutoModel

# Load IndoBERT tokenizer
tokenizer = AutoTokenizer.from_pretrained("indolem/indobert-base-uncased")

# Load IndoBERT model
model = AutoModel.from_pretrained("indolem/indobert-base-uncased")

# Mapping dictionaries for labels
priority_score_mapping = {1: "LOW", 2: "MEDIUM", 3: "HIGH"}
problem_domain_mapping = {0: "OPERATIONAL", 1: "TECHNICAL"}

# Load the trained Random Forest models
best_classifier1 = joblib.load('best_classifier1_optimized.pkl')
best_classifier2 = joblib.load('best_classifier2_optimized.pkl')

markdown_text = '''
## Label Description

### Priority Score
* **Low** label, means that the temporary/corrective solution can solve the problem. A permanent solution will be provided later because the impact on the business can still be handled.
* **Medium** label, means that there's a need to determine the time constraint to solve the problem. If it remains too long, it will impact the business side.
* **High** label, means that the problem is urgent and must be solved immediately.

### Problem Domain
* **Operational** label, means that the scope of the problem is on the business or daily operational.
* **Technical** label, means that the scope of the problem is on the technical (technology) side like the mobile/web application.
'''

description="Write the feedback about the capsule hotel that you've ever visited or stayed there. The machine learning model will predict the priority score and problem domain of the feedback."

# Function to perform predictions
def predict(text):
    # Convert the sentences into input features
    encoded_inputs = tokenizer(text, padding=True, truncation=True, return_tensors="pt", max_length=128)

    # Perform word embedding using IndoBERT model
    with torch.no_grad():
      outputs = model(**encoded_inputs)
      embeddings = outputs.last_hidden_state

    # Convert the embeddings to numpy array
    embeddings = embeddings.numpy()

    embeddings_custom_flat = embeddings.reshape(embeddings.shape[0], -1)

    # Ensure mean_pooled_embeddings has exactly 768 features
    num_features_expected = 768
    if embeddings_custom_flat.shape[1] < num_features_expected:
      # If the number of features is less than 768, pad the embeddings
      pad_width = num_features_expected - embeddings_custom_flat.shape[1]
      embeddings_custom_flat = np.pad(embeddings_custom_flat, ((0, 0), (0, pad_width)), mode='constant')

    elif embeddings_custom_flat.shape[1] > num_features_expected:
      # If the number of features is more than 768, truncate the embeddings
      embeddings_custom_flat = embeddings_custom_flat[:, :num_features_expected]

    # Predict the priority_score for the custom input
    custom_priority_score = best_classifier1.predict(embeddings_custom_flat)

    # Predict the problem_domain for the custom input
    custom_problem_domain = best_classifier2.predict(embeddings_custom_flat)

    # Map numerical labels to human-readable labels
    mapped_priority_score = priority_score_mapping.get(custom_priority_score[0], "unknown")
    mapped_problem_domain = problem_domain_mapping.get(custom_problem_domain[0], "unknown")

    return f"Predicted Priority Score: {mapped_priority_score}, Predicted Problem Domain: {mapped_problem_domain}"


# Create a Gradio interface
gr.Interface(fn=predict, inputs="text", outputs="text", title="Simple Risk Classifier Demo (Case Study: Capsule Hotel)", description=description, article=markdown_text).launch(debug=True)