app.py

import random

import numpy as np
import streamlit as st
import nltk
import torch
import umap
from nltk.tokenize import word_tokenize
from transformers import AutoModel, AutoTokenizer

from aligner import Aligner
from plotools import (
    plot_align_matrix_heatmap_plotly,
    plot_similarity_matrix_heatmap_plotly,
    show_assignments_plotly,
)
from utils import centering, convert_to_word_embeddings, encode_sentence

device = "cuda" if torch.cuda.is_available() else "cpu"
torch.manual_seed(42)
np.random.seed(42)
random.seed(42)
nltk.download("punkt")


@st.cache_resource
def init_model(model: str):
    tokenizer = AutoTokenizer.from_pretrained(model)
    model = (
        AutoModel.from_pretrained(model, output_hidden_states=True).to(device).eval()
    )
    return tokenizer, model


@st.cache_resource(max_entries=100)
def init_aligner(
    ot_type: str, sinkhorn: bool, distortion: float, threshhold: float, tau: float
):
    return Aligner(
        ot_type=ot_type,
        sinkhorn=sinkhorn,
        dist_type="cos",
        weight_type="uniform",
        distortion=distortion,
        thresh=threshhold,
        tau=tau,
        div_type="--",
    )


def main():
    st.set_page_config(layout="wide")

    # Sidebar
    st.sidebar.markdown("## Settings & Parameters")
    model = st.sidebar.selectbox(
        "model", ["microsoft/deberta-v3-base", "bert-base-uncased"]
    )
    layer = st.sidebar.slider(
        "layer number for embeddings",
        0,
        11,
        value=9,
    )
    is_centering = st.sidebar.checkbox("centering embeddings", value=True)
    ot_type = st.sidebar.radio(
        "type",
        ["OT", "POT", "UOT"],
        index=1,
        horizontal=True,
        help="optimal transport algorithm to be used",
    )
    ot_type = ot_type.lower()
    sinkhorn = st.sidebar.checkbox(
        "sinkhorn", value=True, help="use sinkhorn algorithm"
    )
    distortion = st.sidebar.slider(
        "distortion: $\kappa$",
        0.0,
        1.0,
        value=0.20,
        help="suppression of off-diagonal alignments",
    )
    tau = st.sidebar.slider(
        "m / $\\tau$",
        0.0,
        1.0,
        value=0.98,
        help="fraction of fertility to be aligned (fraction of mass to be transported) / penalties",
    )
    threshhold = st.sidebar.slider(
        "threshhold: $\lambda$",
        0.0,
        1.0,
        value=0.22,
        help="sparsity of alignment matrix",
    )
    show_similarity = st.sidebar.checkbox("show similarity matrix", value=True)
    show_assignments = st.sidebar.checkbox("show assignments", value=True)
    if show_assignments:
        n_neighbors = st.sidebar.slider(
            "n_neighbors (see [details](https://umap-learn.readthedocs.io/en/latest/parameters.html#n-neighbors).)",
            2, 15, value=8,
            help="number of nearest neighbors for umap balancing between the preservation of local and global structures"
        )

    # Content
    st.markdown(
        "## Playground: Unbalanced Optimal Transport for Unbalanced Word Alignment"
    )

    col1, col2 = st.columns(2)

    with col1:
        sent1 = st.text_area(
            "sentence 1",
            "By one estimate, fewer than 20,000 lions exist in the wild, a drop of about 40 percent in the past two decades.",
            help="Initial text",
        )
    with col2:
        sent2 = st.text_area(
            "sentence 2",
            "Today there are only around 20,000 wild lions left in the world.",
            help="Text to compare",
        )

    tokenizer, model = init_model(model)
    aligner = init_aligner(ot_type, sinkhorn, distortion, threshhold, tau)
    
    with st.container():
      if sent1 != '' and sent2 != '':
        sent1 = word_tokenize(sent1.lower())
        sent2 = word_tokenize(sent2.lower())
        print(sent1)
        print(sent2)
        hidden_output, input_id, offset_map = encode_sentence(sent1, sent2, tokenizer, model, layer=layer)
        if is_centering:
            hidden_output = centering(hidden_output)
        s1_vec, s2_vec = convert_to_word_embeddings(offset_map, input_id, hidden_output, tokenizer, pair=True)
        align_matrix, cost_matrix, loss, similarity_matrix = aligner.compute_alignment_matrixes(s1_vec, s2_vec)
        print(align_matrix.shape, cost_matrix.shape)
        
        st.write(f"**Word alignment matrix** (loss: :blue[{loss}])")
        fig = plot_align_matrix_heatmap_plotly(align_matrix.T, sent1, sent2, threshhold, cost_matrix.T)
        st.plotly_chart(fig, use_container_width=True)

        if show_similarity:
            st.write(f"**Word similarity matrix**")
            fig2 = plot_similarity_matrix_heatmap_plotly(similarity_matrix.T, sent1, sent2, cost_matrix.T)
            st.plotly_chart(fig2, use_container_width=True)

        if show_assignments:
            st.write(f"**Alignments after UMAP**")
            word_embeddings = torch.vstack([s1_vec, s2_vec])
            umap_embeddings = umap.UMAP(
                n_neighbors=n_neighbors,
                n_components=2,
                random_state=42,
                metric="cosine",
            ).fit_transform(word_embeddings.detach().numpy())
            print(umap_embeddings.shape)
            fig3 = show_assignments_plotly(
                align_matrix, umap_embeddings, sent1, sent2, thr=threshhold
            )
            st.plotly_chart(fig3, use_container_width=True)
    
    st.divider()
    st.subheader('Refs')
    st.write("Yuki Arase, Han Bao, Sho Yokoi, [Unbalanced Optimal Transport for Unbalanced Word Alignment](https://arxiv.org/abs/2306.04116), ACL2023 [[github](https://github.com/yukiar/OTAlign/tree/main)]")

if __name__ == '__main__':
    main()