app.py

import gradio as gr
import regex as re
import torch
import nltk
import pandas as pd
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from nltk.tokenize import sent_tokenize
import plotly.express as px
import time
import tqdm

nltk.download('punkt')

# Define the device (GPU or CPU)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

# Define the model and tokenizer
checkpoint = "ieq/IEQ-BERT"
tokenizer = AutoTokenizer.from_pretrained(checkpoint).to(device)
model = AutoModelForSequenceClassification.from_pretrained(checkpoint).to(device)


# Define the function for preprocessing text
def prep_text(text):
    clean_sents = []  # append clean con sentences
    sent_tokens = str(text).split('.')
    for sent_token in sent_tokens:
        word_tokens = [str(word_token).strip().lower() for word_token in sent_token.split()]
        word_tokens = [word_token for word_token in word_tokens if word_token not in punctuations]
        clean_sents.append(' '.join((word_tokens)))
    joined_clean_sents = '. '.join(clean_sents).strip(' ')
    return joined_clean_sents


# APP INFO
def app_info():
    check = """
    Please go to either the "Single-Text-Prediction" or "Multi-Text-Prediction" tab to analyse your text. 
    """

    return check


# Create Gradio interface for app info
iface1 = gr.Interface(
    fn=app_info, inputs=None, outputs=['text'], title="General-Infomation",
    description='''
    This app, powered by the IEQ-BERT model (sadickam/sdg-classification-bert), is for automating the classification of text concerning 
    with respect to indoor environmetal quality (IEQ). IEQ refers to the quality of the indoor air, lighting, 
    temperature, and acoustics within a building, as well as the overall comfort and well-being of its occupants. It encompasses various 
    factors that can impact the health, productivity, and satisfaction of people who spend time indoors, such as office workers, students, 
    patients, and residents. This app assigns five labels to any given text and a text may be assigned one or more labels. The five labels include
    the following:
    - Acoustic
    - Indoor air quality (IAQ)
    - No IEQ (label assigned when no IEQ is defected)
    - Thermal
    - Visual

    Because IEQ-BERT is capable of assigning one or more labels to a text, it is possible that the returned prediction like 
    (Acoustic_No IEQ) or (NO IEQ_Thermal). These multiple predictions that include "No IEQ" may suggest lack of contextual 
    clarity in the text and need manual review to affirm label. 

    This app has two analysis modules summarised below:
    - Single-Text-Prediction - Analyses text pasted in a text box and return IEQ prediction.
    - Multi-Text-Prediction - Analyses multiple rows of texts in an uploaded CSV file and returns a downloadable CSV file with IEQ prediction for each row of text.

    This app runs on a free server and may therefore not be suitable for analysing large CSV files. 
    If you need assistance with analysing large CSV, do get in touch using the contact information in the Contact section.

    <h3>Contact</h3>
    <p>We would be happy to receive your feedback regarding this app. If you would also like to collaborate with us to explore some use cases for the model 
    powering this app, we are happy to hear from you.</p>

    Dr Abdul-Manan Sadick - [email protected]
    Dr Giorgia Chinazzo - [email protected]
    ''')


# SINGLE TEXT
# Define the prediction function
def predict_single_text(text):
    """
    Predicts the IEQ labels for a single text.

    Args:
        text (str): The text to be analyzed.

    Returns:
        top_prediction (dict): A dictionary containing the top predicted IEQ labels and their corresponding probabilities.
        fig (plotly.graph_objs.Figure): A bar chart showing the likelihood of each IEQ label.
    """
    # Preprocess the input text
    cleaned_text = prep_text(text)

    # Check if the text is empty after preprocessing
    if cleaned_text == "":
        raise gr.Error('This model needs some text input to return a prediction')

    # Tokenize the preprocessed text
    tokenized_text = tokenizer(cleaned_text, return_tensors="pt", truncation=True, max_length=512, padding=True).to(
        device)

    # Make predictions
    with torch.no_grad():
        outputs = model(**tokenized_text)
    logits = outputs.logits

    # Calculate the probabilities
    probabilities = torch.sigmoid(logits).squeeze()

    # Define the threshold for prediction
    threshold = 0.3

    # Get the predicted labels
    predicted_labels_ = (probabilities.cpu().numpy() > threshold).tolist()

    # Define the list of IEQ labels
    label_list = [
        'Acoustic',
        'Indoor air quality',
        'No IEQ',
        'Thermal',
        'Visual'
    ]

    # Map the predicted labels to their corresponding names
    predicted_labels = [label_list[i] for i in range(len(label_list)) if predicted_labels_[i] == 1]

    # Get the probabilities of the predicted labels
    predicted_prob = [round(a_, 3) for a_ in probabilities.cpu().numpy().tolist() if a_ > threshold]

    # Create a dictionary containing the top predicted IEQ labels and their corresponding probabilities
    top_prediction = (dict(zip(predicted_labels, predicted_prob)))

    # Create a bar chart showing the likelihood of each IEQ label
    # Make dataframe for plotly bar chart
    u, v = zip(*dict(zip(label_list, probabilities.cpu().numpy().tolist())).items())
    m = list(u)
    n = list(v)
    df2 = pd.DataFrame()
    df2['IEQ'] = m
    df2['Likelihood'] = n

    # plot graph of predictions
    fig = px.bar(df2, x="Likelihood", y="IEQ", orientation="h")

    fig.update_layout(
        # barmode='stack',
        template='seaborn', font=dict(family="Arial", size=12, color="black"),
        autosize=True,
        # width=800,
        # height=500,
        xaxis_title="Likelihood of IEQ",
        yaxis_title="Indoor environmental quality (IEQ)",
        # legend_title="Topics"
    )

    fig.update_xaxes(tickangle=0, tickfont=dict(family='Arial', color='black', size=12))
    fig.update_yaxes(tickangle=0, tickfont=dict(family='Arial', color='black', size=12))
    fig.update_annotations(font_size=12)

    return top_prediction, fig

    # Create Gradio interface for single text


iface2 = gr.Interface(fn=predict_single_text,
                      inputs=gr.Textbox(lines=7, label="Paste or type text here"),
                      outputs=[gr.Label(label="Top Prediction", show_label=True),
                               gr.Plot(label="Likelihood of all labels", show_label=True)],
                      title="Single Text Prediction",
                      article="**Note:** The quality of model predictions may depend on the quality of information provided."
                      )


# UPLOAD CSV
# Define the prediction function
def predict_from_csv(file, column_name, progress=gr.Progress()):
    """
    Predicts the IEQ labels for a list of texts in a CSV file.

    Args:
        file (str): The path to the CSV file.
        column_name (str): The name of the column containing the text to be analyzed.
        progress (gr.Progress): A progress bar to display the analysis progress.

    Returns:
        fig (plotly.graph_objs.Figure): A histogram showing the frequency of each IEQ label.
        output_csv (gr.File): A downloadable CSV file containing the predictions.
    """
    # Read the CSV file
    df_docs = pd.read_csv(file)

    # Check if the specified column exists
    if column_name not in df_docs.columns:
        raise gr.Error(f"The column '{column_name}' does not exist in the uploaded CSV file.")

    # Extract the text list from the specified column
    text_list = df_docs[column_name].tolist()

    # Define the list of IEQ labels
    label_list = [
        'Acoustic',
        'Indoor air quality',
        'No IEQ',
        'Thermal',
        'Visual'
    ]

    # Initialize lists to store the predictions
    labels_predicted = []
    prediction_scores = []

    # Preprocess text and make predictions
    for text_input in progress.tqdm(text_list, desc="Analysing data"):
        # Sleep to avoid rate limiting
        time.sleep(0.02)

        # Preprocess the text
        cleaned_text = prep_text(text_input)

        # Tokenize the text
        tokenized_text = tokenizer(cleaned_text, return_tensors="pt", truncation=True, max_length=512, padding=True).to(
            device)

        # Make predictions
        with torch.no_grad():
            outputs = model(**tokenized_text)
        logits = outputs.logits

        # Calculate the probabilities
        predictions = torch.sigmoid(logits).squeeze()

        # Define the threshold for prediction
        threshold = 0.3

        # Get the predicted labels
        predicted_labels_ = (predictions.cpu().numpy() > threshold).tolist()

        # Map the predicted labels to their corresponding names
        predicted_labels = [label_list[i] for i in range(len(label_list)) if predicted_labels_[i] == 1]

        # Get the probabilities of the predicted labels
        prediction_score = [round(a_, 3) for a_ in predictions.cpu().numpy().tolist() if a_ > threshold]

        # Append the predictions to the lists
        labels_predicted.append(predicted_labels)
        prediction_scores.append(prediction_score)

        # Append the predictions to the DataFrame
    df_docs['IEQ_predicted'] = labels_predicted
    df_docs['prediction_scores'] = prediction_scores

    # Save the predictions to a CSV file
    df_docs.to_csv('IEQ_predictions.csv')

    # Create a downloadable CSV file
    output_csv = gr.File(value='IEQ_predictions.csv', visible=True)

    # Create a histogram showing the frequency of each IEQ label
    fig = px.histogram(df_docs, y="IEQ_predicted")
    fig.update_layout(
        template='seaborn',
        font=dict(family="Arial", size=12, color="black"),
        autosize=True,
        # width=800,
        # height=500,
        xaxis_title="IEQ counts",
        yaxis_title="Indoor environmetal quality (IEQ)",
    )
    fig.update_xaxes(tickangle=0, tickfont=dict(family='Arial', color='black', size=12))
    fig.update_yaxes(tickangle=0, tickfont=dict(family='Arial', color='black', size=12))
    fig.update_annotations(font_size=12)

    return fig, output_csv


# Define the input component
file_input = gr.File(label="Upload CSV file here", show_label=True, file_types=[".csv"])
column_name_input = gr.Textbox(label="Enter the column name containing the text to be analyzed", show_label=True)

# Create the Gradio interface
iface3 = gr.Interface(fn=predict_from_csv,
                      inputs=[file_input, column_name_input],
                      outputs=[gr.Plot(label='Frequency of IEQs', show_label=True),
                               gr.File(label='Download output CSV', show_label=True)],
                      title="Multi-text Prediction (CVS)",
                      description='**NOTE:** Please enter the column name containing the text to be analyzed.')

# Create a tabbed interface
demo = gr.TabbedInterface(interface_list=[iface1, iface2, iface3],
                          tab_names=["General-App-Info", "Single-Text-Prediction", "Multi-Text-Prediction (CSV)"],
                          title="Indoor Environmetal Quality (IEQ) Text Classifier App",
                          theme='soft'
                          )

# Launch the interface
demo.queue().launch()