app.py

import streamlit as st
from transformers import AutoTokenizer, AutoModel, AutoModelForSequenceClassification, pipeline
import torch
import plotly.express as px
import numpy as np
from sklearn.decomposition import PCA
from utils import visualize_attention, list_supported_models, plot_token_embeddings

st.set_page_config(page_title="Transformer Visualizer", layout="wide")
st.title("🧠 Transformer Visualizer")
st.markdown("Explore how Transformer models process and understand language.")

task = st.sidebar.selectbox("Select Task", ["Text Classification", "Text Generation", "Question Answering"])
model_name = st.sidebar.selectbox("Select Model", list_supported_models(task))

text_input = st.text_area("Enter input text", "The quick brown fox jumps over the lazy dog.")

if st.button("Run"):
    st.info(f"Loading model: `{model_name}`...")
    
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    if task == "Text Classification":
        model = AutoModelForSequenceClassification.from_pretrained(model_name, output_attentions=True)
    else:
        model = AutoModel.from_pretrained(model_name, output_attentions=True)

    inputs = tokenizer(text_input, return_tensors="pt", return_token_type_ids=False)
    outputs = model(**inputs)
    attentions = outputs.attentions

    st.success("Model inference complete!")

    # Tokenization Visualization
    st.subheader("🔠 Tokenization")
    tokens = tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])
    token_ids = inputs["input_ids"][0].tolist()
    st.write(list(zip(tokens, token_ids)))

    # Token Embeddings Visualization
    st.subheader("🌐 Token Embedding Space (PCA)")
    with torch.no_grad():
        hidden_states = model.base_model.embeddings.word_embeddings(inputs["input_ids"]).squeeze(0)
    fig_embed = plot_token_embeddings(hidden_states, tokens)
    st.plotly_chart(fig_embed, use_container_width=True)

    # Attention Visualization
    if attentions:
        st.subheader("👁️ Attention Visualization")
        fig = visualize_attention(attentions, tokenizer, inputs)
        st.plotly_chart(fig, use_container_width=True)
    else:
        st.warning("This model does not return attention weights.")

    if task == "Text Classification":
        st.subheader("✅ Prediction")
        pipe = pipeline("text-classification", model=model, tokenizer=tokenizer)
        prediction = pipe(text_input)
        st.write(prediction)

st.sidebar.markdown("---")
st.sidebar.write("App by Rahiya Esar 💖")