app.py

from __future__ import annotations
import psutil
import pandas as pd
import streamlit as st
import plotly.express as px
from models import NLI_MODEL_OPTIONS, NSP_MODEL_OPTIONS, METHOD_OPTIONS
from zeroshot_classification.classifiers import NSPZeroshotClassifier, NLIZeroshotClassifier

print(f"Total mem: {psutil.virtual_memory().total}")

def init_state(key: str):
    if key not in st.session_state:
        st.session_state[key] = None


for k in [
    "current_model",
    "current_model_option",
    "current_method_option",
    "current_prediction",
    "current_chart",
]:
    init_state(k)


def load_model(model_option: str, method_option: str, random_state: int = 0):
    with st.spinner("Loading selected model..."):
        if method_option == "Natural Language Inference":
            st.session_state.current_model = NLIZeroshotClassifier(
                model_name=model_option, random_state=random_state
            )
        else:
            st.session_state.current_model = NSPZeroshotClassifier(
                model_name=model_option, random_state=random_state
            )
        st.success("Model loaded!")


def visualize_output(labels: list[str], probabilities: list[float]):
    data = pd.DataFrame({"labels": labels, "probability": probabilities}).sort_values(
        by="probability", ascending=False
    )
    chart = px.bar(
        data,
        x="probability",
        y="labels",
        color="labels",
        orientation="h",
        height=290,
        width=500,
    ).update_layout(
        {
            "xaxis": {"title": "probability", "visible": True, "showticklabels": True},
            "yaxis": {"title": None, "visible": True, "showticklabels": True},
            "margin": dict(
                l=10,  # left
                r=10,  # right
                t=50,  # top
                b=10,  # bottom
            ),
            "showlegend": False,
        }
    )
    return chart


st.title("Zero-shot Turkish Text Classification")
method_option = st.radio(
    "Select a zero-shot classification method.",
    [
        METHOD_OPTIONS["nli"],
        METHOD_OPTIONS["nsp"],
    ],
)
if method_option == METHOD_OPTIONS["nli"]:
    model_option = st.selectbox(
        "Select a natural language inference model.", NLI_MODEL_OPTIONS, index=3
    )
if method_option == METHOD_OPTIONS["nsp"]:
    model_option = st.selectbox(
        "Select a BERT model for next sentence prediction.", NSP_MODEL_OPTIONS, index=0
    )

if model_option != st.session_state.current_model_option:
    st.session_state.current_model_option = model_option
    st.session_state.current_method_option = method_option
    load_model(
        st.session_state.current_model_option, st.session_state.current_method_option
    )


st.header("Configure prompts and labels")
col1, col2 = st.columns(2)
col1.subheader("Candidate labels")
labels = col1.text_area(
    label="These are the labels that the model will try to predict for the given text input. Your input labels should be comma separated and meaningful.",
    value="spor,dünya,siyaset,ekonomi,sanat",
    key="current_labels",
)

col1.header("Make predictions")
text = col1.text_area(
    "Enter a sentence or a paragraph to classify.",
    value="Ian Anderson, Jethro Tull konserinde yan flüt çalarak zeybek oynadı.",
    key="current_text",
)
col2.subheader("Prompt template")
prompt_template = col2.text_area(
    label="Prompt template is used to transform NLI and NSP tasks into a general-use zero-shot classifier. Models replace {} with the labels that you have given.",
    value="Bu metin {} kategorisine aittir",
    key="current_template",
)
col2.header("")


make_pred = col1.button("Predict")
if make_pred:
    st.session_state.current_prediction = (
        st.session_state.current_model.predict_on_texts(
            [st.session_state.current_text],
            candidate_labels=st.session_state.current_labels.split(","),
            prompt_template=st.session_state.current_template,
        )
    )
    if "scores" in st.session_state.current_prediction[0]:
        st.session_state.current_chart = visualize_output(
            st.session_state.current_prediction[0]["labels"],
            st.session_state.current_prediction[0]["scores"],
        )
    elif "probabilities" in st.session_state.current_prediction[0]:
        st.session_state.current_chart = visualize_output(
            st.session_state.current_prediction[0]["labels"],
            st.session_state.current_prediction[0]["probabilities"],
        )
    col2.plotly_chart(st.session_state.current_chart, use_container_width=True)