app.py

import numpy as np
import plotly.graph_objects as go
import json
import gradio as gr
from nltk.corpus import words
import nltk


# load files w embeddings, attention scores, and tokens
vocab_embeddings = np.load('vocab_embeddings.npy')
with open('vocab_attention_scores.json', 'r') as f:
    vocab_attention_scores = json.load(f)
with open('vocab_tokens.json', 'r') as f:
    vocab_tokens = json.load(f)

# attention scores to numpy arrs
b_gen_attention = np.array([score['B-GEN'] for score in vocab_attention_scores])
i_gen_attention = np.array([score['I-GEN'] for score in vocab_attention_scores])
b_unfair_attention = np.array([score['B-UNFAIR'] for score in vocab_attention_scores])
i_unfair_attention = np.array([score['I-UNFAIR'] for score in vocab_attention_scores])
b_stereo_attention = np.array([score['B-STEREO'] for score in vocab_attention_scores])
i_stereo_attention = np.array([score['I-STEREO'] for score in vocab_attention_scores])
o_attention = np.array([score['O'] for score in vocab_attention_scores])  # Use actual O scores

# remove non-dict english words, but keep subwords ##
nltk.download('words')
english_words = set(words.words())

filtered_indices = [i for i, token in enumerate(vocab_tokens) if token in english_words or token.startswith("##")]
filtered_tokens = [vocab_tokens[i] for i in filtered_indices]

b_gen_attention_filtered = b_gen_attention[filtered_indices]
i_gen_attention_filtered = i_gen_attention[filtered_indices]
b_unfair_attention_filtered = b_unfair_attention[filtered_indices]
i_unfair_attention_filtered = i_unfair_attention[filtered_indices]
b_stereo_attention_filtered = b_stereo_attention[filtered_indices]
i_stereo_attention_filtered = i_stereo_attention[filtered_indices]
o_attention_filtered = o_attention[filtered_indices]

# plot top 500 O tokens for comparison
top_500_o_indices = np.argsort(o_attention_filtered)[-500:]
top_500_o_tokens = [filtered_tokens[i] for i in top_500_o_indices]
o_attention_filtered_top_500 = o_attention_filtered[top_500_o_indices]

# tool tip for tokens
def create_hover_text(tokens, b_gen, i_gen, b_unfair, i_unfair, b_stereo, i_stereo, o_val):
    hover_text = []
    for i in range(len(tokens)):
        hover_text.append(
            f"Token: {tokens[i]}<br>"
            f"B-GEN: {b_gen[i]:.3f}, I-GEN: {i_gen[i]:.3f}<br>"
            f"B-UNFAIR: {b_unfair[i]:.3f}, I-UNFAIR: {i_unfair[i]:.3f}<br>"
            f"B-STEREO: {b_stereo[i]:.3f}, I-STEREO: {i_stereo[i]:.3f}<br>"
            f"O: {o_val[i]:.3f}"
        )
    return hover_text

# ploting top 100 tokens for each entity
def select_top_100(*data_arrays):
    indices_list = []
    for data in data_arrays:
        if data is not None:
            top_indices = np.argsort(data)[-100:] 
            indices_list.append(top_indices)
    
    combined_indices = np.unique(np.concatenate(indices_list))

    # filter based on combined indices
    filtered_data = [data[combined_indices] if data is not None else None for data in data_arrays]
    tokens_filtered = [filtered_tokens[i] for i in combined_indices]

    return (*filtered_data, tokens_filtered)

# plots for 1 2 and 3 D
def create_plot(selected_dimensions):
    # plot data
    attention_map = {
        'Generalization': b_gen_attention_filtered + i_gen_attention_filtered,
        'Unfairness': b_unfair_attention_filtered + i_unfair_attention_filtered,
        'Stereotype': b_stereo_attention_filtered + i_stereo_attention_filtered,
    }

    # init x, y, z so they can be moved around
    x_data, y_data, z_data = None, None, None

    # use selected dimentsions to order dimensions
    if len(selected_dimensions) > 0:
        x_data = attention_map[selected_dimensions[0]]
    if len(selected_dimensions) > 1:
        y_data = attention_map[selected_dimensions[1]] 
    if len(selected_dimensions) > 2:
        z_data = attention_map[selected_dimensions[2]]

    # select top 100 dps for each selected dimension
    x_data, y_data, z_data, tokens_filtered = select_top_100(x_data, y_data, z_data)

    # filter the O tokens using the same dimensions
    o_x = attention_map[selected_dimensions[0]][top_500_o_indices]
    if len(selected_dimensions) > 1:
        o_y = attention_map[selected_dimensions[1]][top_500_o_indices]
    else:
        o_y = np.zeros_like(o_x) 
    if len(selected_dimensions) > 2:
        o_z = attention_map[selected_dimensions[2]][top_500_o_indices]
    else:
        o_z = np.zeros_like(o_x)

    # hover text for GUS tokens
    classified_hover_text = create_hover_text(
        tokens_filtered, 
        b_gen_attention_filtered, i_gen_attention_filtered, 
        b_unfair_attention_filtered, i_unfair_attention_filtered, 
        b_stereo_attention_filtered, i_stereo_attention_filtered, 
        o_attention_filtered
    )

    # hover text for O tokens
    o_hover_text = create_hover_text(
        top_500_o_tokens, 
        b_gen_attention_filtered[top_500_o_indices], i_gen_attention_filtered[top_500_o_indices], 
        b_unfair_attention_filtered[top_500_o_indices], i_unfair_attention_filtered[top_500_o_indices], 
        b_stereo_attention_filtered[top_500_o_indices], i_stereo_attention_filtered[top_500_o_indices], 
        o_attention_filtered_top_500
    )


    # plot
    fig = go.Figure()

    if x_data is not None and y_data is not None and z_data is not None:
        # 3d scatter plot
        fig.add_trace(go.Scatter3d(
            x=x_data,
            y=y_data,
            z=z_data,
            mode='markers',
            marker=dict(
                size=6,
                color=x_data,  # color based on the x-axis data
                colorscale='Viridis',
                opacity=0.85,
            ),
            text=classified_hover_text, 
            hoverinfo='text',
            name='Classified Tokens'
        ))
        # add top 500 O tags to the plot too
        fig.add_trace(go.Scatter3d(
            x=o_x,
            y=o_y,
            z=o_z,
            mode='markers',
            marker=dict(
                size=6,
                color='grey',
                opacity=0.5,
            ),
            text=o_hover_text,
            hoverinfo='text',
            name='O Tokens'
        ))
    elif x_data is not None and y_data is not None:
        # 2d scatter plot
        fig.add_trace(go.Scatter(
            x=x_data,
            y=y_data,
            mode='markers',
            marker=dict(
                size=6,
                color=x_data, # color based on the x-axis data
                colorscale='Viridis',
                opacity=0.85,
            ),
            text=classified_hover_text,
            hoverinfo='text',
            name='Classified Tokens'
        ))
        # add top 500 O tags to the plot too
        fig.add_trace(go.Scatter(
            x=o_x,
            y=o_y,
            mode='markers',
            marker=dict(
                size=6,
                color='grey',
                opacity=0.5,
            ),
            text=o_hover_text, 
            hoverinfo='text',
            name='O Tokens'
        ))
    elif x_data is not None:
        # 1D scatter plot
        fig.add_trace(go.Scatter(
            x=x_data,
            y=np.zeros_like(x_data),
            mode='markers',
            marker=dict(
                size=6,
                color=x_data, 
                colorscale='Viridis',
                opacity=0.85,
            ),
            text=classified_hover_text,
            hoverinfo='text',
            name='GUS Tokens'
        ))
        fig.add_trace(go.Scatter(
            x=o_x,
            y=np.zeros_like(o_x), 
            mode='markers',
            marker=dict(
                size=6,
                color='grey',
                opacity=0.5,
            ),
            text=o_hover_text,
            hoverinfo='text',
            name='O Tokens'
        ))

    # update layout dynamically
    if x_data is not None and y_data is not None and z_data is not None:
        # 3D
        fig.update_layout(
            title="GUS-Net Entity Attentions Visualization",
            scene=dict(
                xaxis=dict(title=f"{selected_dimensions[0]} Attention"),
                yaxis=dict(title=f"{selected_dimensions[1]} Attention"),
                zaxis=dict(title=f"{selected_dimensions[2]} Attention"),
            ),
            margin=dict(l=0, r=0, b=0, t=40),
        )
    elif x_data is not None and y_data is not None:
        # 2D
        fig.update_layout(
            title="GUS-Net Entity Attentions Visualization",
            xaxis_title=f"{selected_dimensions[0]} Attention",
            yaxis_title=f"{selected_dimensions[1]} Attention",
            margin=dict(l=0, r=0, b=0, t=40),
        )
    elif x_data is not None:
        # 1D
        fig.update_layout(
            title="GUS-Net Entity Attentions Visualization",
            xaxis_title=f"{selected_dimensions[0]} Attention",
            margin=dict(l=0, r=0, b=0, t=40),
        )

    return fig

def get_top_tokens_for_entities(selected_dimensions):
    entity_map = {
        'Generalization': b_gen_attention_filtered + i_gen_attention_filtered,
        'Unfairness': b_unfair_attention_filtered + i_unfair_attention_filtered,
        'Stereotype': b_stereo_attention_filtered + i_stereo_attention_filtered,
    }

    top_tokens_info = {}
    for dimension in selected_dimensions:
        if dimension in entity_map:
            attention_scores = entity_map[dimension]
            top_indices = np.argsort(attention_scores)[-10:]  # top 10 tokens
            top_tokens = [filtered_tokens[i] for i in top_indices]
            top_scores = attention_scores[top_indices]
            top_tokens_info[dimension] = list(zip(top_tokens, top_scores))

    return top_tokens_info

def update_gradio(selected_dimensions):
    fig = create_plot(selected_dimensions)

    top_tokens_info = get_top_tokens_for_entities(selected_dimensions)
    
    formatted_top_tokens = ""
    for entity, tokens_info in top_tokens_info.items():
        formatted_top_tokens += f"\nTop tokens for {entity}:\n"
        for token, score in tokens_info:
            formatted_top_tokens += f"Token: {token}, Attention Score: {score:.3f}\n"

    return fig, formatted_top_tokens


def render_gradio_interface():
    with gr.Blocks() as interface:
        with gr.Column():
            dimensions_input = gr.CheckboxGroup(
                choices=["Generalization", "Unfairness", "Stereotype"],
                label="Select Dimensions to Plot",
                value=["Generalization", "Unfairness", "Stereotype"]  # defaults to 3D
            )

            plot_output = gr.Plot(label="Token Attention Visualization")
            top_tokens_output = gr.Textbox(label="Top Tokens for Each Entity Class", lines=10)

            dimensions_input.change(
                fn=update_gradio,
                inputs=[dimensions_input],
                outputs=[plot_output, top_tokens_output]
            )

            interface.load(
                fn=lambda: update_gradio(["Generalization", "Unfairness", "Stereotype"]),
                inputs=None, 
                outputs=[plot_output, top_tokens_output]
            )

    return interface

interface = render_gradio_interface()
interface.launch()