app.py

import streamlit as st
from streamlit_mic_recorder import mic_recorder
from transformers import pipeline
import torch
from transformers import BertTokenizer, BertForSequenceClassification
# from transformers import AutoModelForSequenceClassification, AutoTokenizer
import numpy as np
import pandas as pd
import time
import altair as alt


def callback():
    if st.session_state.my_recorder_output:
        audio_bytes = st.session_state.my_recorder_output['bytes']
        st.audio(audio_bytes)


@st.cache_resource
def load_text_to_speech_model(model="openai/whisper-base"):
    pipe = pipeline("automatic-speech-recognition", model=model)
    return pipe


def translate(inputs, model="openai/whisper-base"):
    pipe = pipeline("automatic-speech-recognition", model=model)
    translate_result = pipe(inputs, generate_kwargs={'task': 'translate'})
    return translate_result['text']


# def encode_depracated(docs, tokenizer):
#     '''
#     This function takes list of texts and returns input_ids and attention_mask of texts
#     '''
#     encoded_dict = tokenizer.batch_encode_plus(docs, add_special_tokens=True, max_length=128, padding='max_length',
#                             return_attention_mask=True, truncation=True, return_tensors='pt')
#     input_ids = encoded_dict['input_ids']
#     attention_masks = encoded_dict['attention_mask']
#     return input_ids, attention_masks


# def load_classification_model():
#     CUSTOMMODEL_PATH = "./bert-itserviceclassification"
#     PRETRAINED_LM = "bert-base-uncased"
#     tokenizer = BertTokenizer.from_pretrained(PRETRAINED_LM, do_lower_case=True)
#     model = BertForSequenceClassification.from_pretrained(PRETRAINED_LM,
#                                                         num_labels=8,
#                                                         output_attentions=False,
#                                                         output_hidden_states=False)
#     model.load_state_dict(torch.load(CUSTOMMODEL_PATH, map_location ='cpu'))
#     return model, tokenizer

@st.cache_resource
def load_classification_model():
    PRETRAINED_LM = "kkngan/bert-base-uncased-it-service-classification"
    # model = AutoModelForSequenceClassification.from_pretrained(PRETRAINED_LM, num_labels=8)
    # tokenizer = AutoTokenizer.from_pretrained(PRETRAINED_LM)
    tokenizer = BertTokenizer.from_pretrained(PRETRAINED_LM, do_lower_case=True)
    model = BertForSequenceClassification.from_pretrained(PRETRAINED_LM,
                                                         num_labels=8)
    return model, tokenizer


def predict(text, model, tokenizer):
    lookup_key ={0: 'Hardware',
    1: 'Access',
    2: 'Miscellaneous',
    3: 'HR Support',
    4: 'Purchase',
    5: 'Administrative rights',
    6: 'Storage',
    7: 'Internal Project'}
    # with torch.no_grad():
    #     input_ids, att_mask = encode([text], tokenizer)
    #     logits = model(input_ids = input_ids, attention_mask=att_mask).logits
    inputs = tokenizer(text,
                   padding = True,
                   truncation = True,
                   return_tensors='pt')
    outputs = model(**inputs)
    predicted_class_id = outputs.logits.argmax().item()
    predicted_label = lookup_key.get(predicted_class_id)
    probability = torch.nn.functional.softmax(outputs.logits, dim=-1).cpu().detach().numpy()
    return predicted_label, predicted_class_id, probability


def display_result(translate_text, prediction, predicted_class_id, probability):
    st.markdown('<font color="purple"><b>Text:</b></font>', unsafe_allow_html=True)
    st.write(f'{translate_text}')
    st.write(f'\n')
    st.write(f'\n')

    # st.markdown(f'<font color="green"><b>Predicted Class: (Probability: {(probability[0][predicted_class_id] * 100):.2f}%) </b></font>', unsafe_allow_html=True)
    st.markdown('<font color="green"><b>Predicted Class:</b></font>', unsafe_allow_html=True)
    st.write(f'{prediction}')

    # Convert probability to bar cart
    st.write(f'\n')
    st.write(f'\n')

    # Show Probability of each Service Category
    category = ('Hardware', 'Access', 'Miscellaneous', 'HR Support', 'Purchase', 'Administrative rights', 'Storage', 'Internal Project')
    probability = np.array(probability[0])
    df = pd.DataFrame({'Category': category, 'Probability (%)': probability * 100})
    df['Probability (%)'] = df['Probability (%)'].apply(lambda x: round(x, 2))

    base = alt.Chart(df).encode(
                x='Probability (%)',
                y=alt.Y('Category').sort('-x'),
                
                # color='b:O',
                tooltip=['Category',alt.Tooltip('Probability (%)', format=",.2f")],
                text='Probability (%)'
                ).properties(title="Probability of each Service Category")
    chart = base.mark_bar() + base.mark_text(align='left', dx=2)
    st.altair_chart(chart, use_container_width=True)


def main():
    # st.cache_resource.clear()
    st.set_page_config(layout="wide", page_title="NLP IT Service Classification", page_icon="🤖",)
    st.markdown('<b>🤖 Welcome to IT Service Classification Assistant!!! 🤖</b>', unsafe_allow_html=True)
    st.write(f'\n')
    st.write(f'\n')

    with st.sidebar:
        st.image('front_page_image.jpg' , use_column_width=True)
        text_to_speech_model = st.selectbox("Pick select a speech to text model", ["openai/whisper-base", "openai/whisper-large-v3"])
        options = st.selectbox("Pick select an input method", ["Start a recording", "Upload an audio", "Enter a transcript"])
        if options == "Start a recording":
            audio = mic_recorder(key='my_recorder', callback=callback)
        elif options == "Upload an audio":
            audio = st.file_uploader("Please upload an audio", type=["wav", "mp3"])
        else:
            text = st.text_area("Please input the transcript (Only support English)")
        button = st.button('Submit')

    if button:        
        with st.spinner(text="Loading... It may take a while if you are running the app for the first time."):
            start_time = time.time()
            if options == "Start a recording":
                # transcibe_text, translate_text = transcribe_and_translate(upload=audio["bytes"])
                translate_text = translate(inputs=audio["bytes"], model=text_to_speech_model)
            elif options == "Upload an audio":
                # transcibe_text, translate_text = transcribe_and_translate(upload=audio.getvalue())
                translate_text = translate(inputs=audio.getvalue(), model=text_to_speech_model)
            else:
                translate_text = text
            model, tokenizer = load_classification_model()
            prediction, predicted_class_id, probability = predict(text=translate_text, model=model, tokenizer=tokenizer)
            end_time = time.time()

        display_result(translate_text, prediction, predicted_class_id, probability)

        st.write(f'\n')
        st.write(f'\n')
        st.markdown(f'*It took {(end_time-start_time):.2f} sec to process the input.', unsafe_allow_html=True)


if __name__ == '__main__':
    main()