app.py

import os
import streamlit as st
import pandas as pd
import numpy as np
import plotly.express as px
import plotly.graph_objects as go
from sklearn.metrics.pairwise import paired_cosine_distances
from sklearn.preprocessing import normalize
from rolaser import RoLaserEncoder

@st.cache_resource(show_spinner=False)
def load_models():
    laser_checkpoint = f"{os.environ['LASER']}/models/laser2.pt"
    laser_vocab = f"{os.environ['LASER']}/models/laser2.cvocab"
    laser_tokenizer = 'spm'
    laser_model = RoLaserEncoder(model_path=laser_checkpoint, vocab=laser_vocab, tokenizer=laser_tokenizer)

    rolaser_checkpoint = f"{os.environ['ROLASER']}/models/rolaser.pt"
    rolaser_vocab = f"{os.environ['ROLASER']}/models/rolaser.cvocab"
    rolaser_tokenizer = 'roberta'
    rolaser_model = RoLaserEncoder(model_path=rolaser_checkpoint, vocab=rolaser_vocab, tokenizer=rolaser_tokenizer)

    c_rolaser_checkpoint = f"{os.environ['ROLASER']}/models/c-rolaser.pt"
    c_rolaser_vocab = f"{os.environ['ROLASER']}/models/c-rolaser.cvocab"
    c_rolaser_tokenizer = 'char'
    c_rolaser_model = RoLaserEncoder(model_path=c_rolaser_checkpoint, vocab=c_rolaser_vocab, tokenizer=c_rolaser_tokenizer)
    return laser_model, rolaser_model, c_rolaser_model

@st.cache_data(show_spinner=False)
def load_sample_data():
    STD_SENTENCES = ['See you tomorrow.'] * 10
    UGC_SENTENCES = [
        'See you t03orro3.',
        'C. U. tomorrow.',
        'sea you tomorrow.',
        'See yo utomorrow.',
        'Cu 2moro.',
        'See you tkmoerow.',
        'See yow tomorrow.',
        'See you tmrw.',
        'C. Yew tomorrow.',
        'c ya 2morrow.'
    ]
    return STD_SENTENCES, UGC_SENTENCES

def main():
    sample_std, sample_ugc = load_sample_data()
    laser_model, rolaser_model, c_rolaser_model = load_models()

    st.title('Pairwise Cosine Distance Calculator')

    info = '''
    :bookmark: **Paper:** [Making Sentence Embeddings Robust to User-Generated Content (Nishimwe et al., 2024)](https://arxiv.org/abs/2403.17220)  
    :link: **Github:** [https://github.com/lydianish/RoLASER](https://github.com/lydianish/RoLASER)  
    :computer: **Demo:** This app computes the cosine distance between standard and non-standard text input pairs using LASER, RoLASER, and c-RoLASER models.
    '''
    st.markdown(info)
            
    st.header('Standard and Non-standard Text Input Pairs')

    cols = st.columns(3)
    num_pairs = cols[1].number_input('Number of Text Input Pairs (1-10):', min_value=1, max_value=10, value=5)

    with st.form('text_input_form'):

        std_text_inputs = []
        ugc_text_inputs = []

        for i in range(num_pairs):
            col1, col2 = st.columns(2)
            with col1:
                text_input1 = st.text_input(f'Standard text {i+1}:', key=f'std{i}', value=sample_std[i])
                std_text_inputs.append(text_input1)
            with col2:
                text_input2 = st.text_input(f'Non-standard text {i+1}:', key=f'ugc{i}', value=sample_ugc[i])
                ugc_text_inputs.append(text_input2)

        st.caption('*The models are case-insensitive: all texts will be lowercased.*')

        st.form_submit_button('Compute')
    
    X_std_laser = normalize(laser_model.encode(std_text_inputs))
    X_ugc_laser = normalize(laser_model.encode(ugc_text_inputs))
    X_cos_laser = paired_cosine_distances(X_std_laser, X_ugc_laser)

    X_std_rolaser = normalize(rolaser_model.encode(std_text_inputs))
    X_ugc_rolaser = normalize(rolaser_model.encode(ugc_text_inputs))
    X_cos_rolaser = paired_cosine_distances(X_std_rolaser, X_ugc_rolaser)

    X_std_c_rolaser = normalize(c_rolaser_model.encode(std_text_inputs))
    X_ugc_c_rolaser = normalize(c_rolaser_model.encode(ugc_text_inputs))
    X_cos_c_rolaser = paired_cosine_distances(X_std_c_rolaser, X_ugc_c_rolaser)

    outputs = pd.DataFrame(columns=[ 'model', 'pair', 'ugc', 'std', 'cos'])
    outputs['model'] = np.repeat(['LASER', 'RoLASER', 'c-RoLASER'], num_pairs)   
    outputs['pair'] = np.tile(np.arange(1,num_pairs+1), 3)
    outputs['std'] = np.tile(std_text_inputs, 3)
    outputs['ugc'] = np.tile(ugc_text_inputs, 3)
    outputs['cos'] = np.concatenate([X_cos_laser, X_cos_rolaser, X_cos_c_rolaser])

    st.header('Cosine Distance Scores')
    st.caption('*This bar plot is interactive: Hover on the bars to display values. Click on the legend items to filter models.*')
    fig = px.bar(outputs, x='pair', y='cos', color='model', barmode='group', hover_data=['ugc', 'std'])
    fig.update_xaxes(title_text='Text Input Pair')
    fig.update_yaxes(title_text='Cosine Distance')
    st.plotly_chart(fig, use_container_width=True)

    if num_pairs > 1:
        st.header('Cosine Distance Statistics')
        st.caption('*This box plot is interactive: Hover on the boxes to display values. Click on the legend items to filter models.*')
        fig = go.Figure()
        fig.add_trace(go.Box(
            y=outputs[outputs['model']=='LASER']['cos'],
            name='LASER',
            boxmean='sd'
        ))
        fig.add_trace(go.Box(
            y=outputs[outputs['model']=='RoLASER']['cos'],
            name='RoLASER',
            boxmean='sd'
        ))
        fig.add_trace(go.Box(
            y=outputs[outputs['model']=='c-RoLASER']['cos'],
            name='c-RoLASER',
            boxmean='sd'
        ))
        fig.update_xaxes(title_text='Model')
        fig.update_yaxes(title_text='Cosine Distance')
        st.plotly_chart(fig, use_container_width=True)
    
if __name__ == "__main__":
    main()