utils.py

import plotly.graph_objects as go
import numpy as np
from sklearn.decomposition import PCA

def list_supported_models(task):
    if task == "Text Classification":
        return ["distilbert-base-uncased", "bert-base-uncased", "roberta-base"]
    elif task == "Text Generation":
        return ["gpt2", "distilgpt2"]
    elif task == "Question Answering":
        return ["deepset/roberta-base-squad2", "distilbert-base-cased-distilled-squad"]
    return []

def visualize_attention(attentions, tokenizer, inputs):
    tokens = tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])
    last_layer_attention = attentions[-1][0]  # [heads, seq_len, seq_len]
    avg_attention = last_layer_attention.mean(dim=0).detach().numpy()

    fig = go.Figure(data=go.Heatmap(
        z=avg_attention,
        x=tokens,
        y=tokens,
        colorscale='Viridis'
    ))
    fig.update_layout(title="Average Attention - Last Layer", xaxis_nticks=len(tokens))
    return fig

def plot_token_embeddings(embeddings, tokens):
    pca = PCA(n_components=2)
    reduced = pca.fit_transform(embeddings.detach().numpy())
    
    fig = go.Figure()
    for i, token in enumerate(tokens):
        fig.add_trace(go.Scatter(
            x=[reduced[i][0]], y=[reduced[i][1]],
            text=[token],
            mode='markers+text',
            textposition='top center',
            marker=dict(size=10),
            name=token
        ))
    fig.update_layout(title="Token Embeddings (PCA)", xaxis_title="PC 1", yaxis_title="PC 2")
    return fig