app.py

import pandas as pd
import altair as alt
import streamlit as st
from sklearn.decomposition import PCA
from sklearn.manifold import MDS
from sentence_transformers import SentenceTransformer

st.set_page_config(layout='wide')

# constants
LIST_SIZE = 5


# cached resources
@st.cache_resource
def get_model() -> SentenceTransformer:
    return SentenceTransformer('panalexeu/xlm-roberta-ua-distilled')


@st.cache_resource
def get_pca() -> PCA:
    return PCA(n_components=2)


@st.cache_resource
def get_mds() -> MDS:
    return MDS(n_components=2)


# define state
if 'sentences' not in st.session_state:
    st.session_state['s_i'] = 0
    st.session_state['sentences'] = []

# layout
st.title('Playground')

left_column, right_column = st.columns([1, 1], gap='large', border=True)

# sentences
with left_column:
    st.subheader('Sentences')

    sub_left, sub_right = st.columns(2, gap='small', vertical_alignment='bottom')

    # read input
    with sub_left:
        if text_input := st.chat_input(placeholder='What is on your mind?', ):

            if len(st.session_state['sentences']) < LIST_SIZE:
                st.session_state['sentences'].append(text_input)
            else:  # Rewrite previous entries
                st.session_state['sentences'][st.session_state['s_i']] = text_input

                st.session_state['s_i'] += 1
                if st.session_state['s_i'] == LIST_SIZE:
                    st.session_state['s_i'] = 0

    with sub_right:
        if clear_btn := st.button(label='Clear', icon=':material/clear:', type='primary'):
            st.session_state['sentences'] = []

    # display sentences
    for i, sentence in enumerate(st.session_state['sentences']):
        st.write(i, sentence)

# embeds display
with (right_column):
    st.subheader('xlm-roberta-ua-distilled')

    if compute_btn := st.button('Compute', icon='🤖'):
        # calculating embeddings
        embeds = get_model().encode(st.session_state['sentences'])

        # projecting them
        proj_points = get_mds().fit_transform(embeds)

        # displaying projected embeddings
        points_df = pd.DataFrame(proj_points, columns=['x', 'y'])
        points_df['sentence'] = st.session_state['sentences']
        chart = alt.Chart(points_df) \
            .mark_circle(size=100) \
            .encode(x='x', y='y', tooltip='sentence') \
            .properties(title='2-d Projection of Sentence Embeddings (MDS)') \
            .interactive()

        st.altair_chart(chart)