app.py

import gradio as gr
import json
import matplotlib.pyplot as plt
import pandas as pd
import io
import base64
import math
import ast
import logging
import numpy as np
import plotly.graph_objects as go

# Set up logging
logging.basicConfig(level=logging.DEBUG)
logger = logging.getLogger(__name__)

# Function to safely parse JSON or Python dictionary input
def parse_input(json_input):
    logger.debug("Attempting to parse input: %s", json_input)
    try:
        # Try to parse as JSON first
        data = json.loads(json_input)
        logger.debug("Successfully parsed as JSON")
        return data
    except json.JSONDecodeError as e:
        logger.error("JSON parsing failed: %s", str(e))
        try:
            # If JSON fails, try to parse as Python literal (e.g., with single quotes)
            data = ast.literal_eval(json_input)
            logger.debug("Successfully parsed as Python literal")
            # Convert Python dictionary to JSON-compatible format (replace single quotes with double quotes)
            def dict_to_json(obj):
                if isinstance(obj, dict):
                    return {str(k): dict_to_json(v) for k, v in obj.items()}
                elif isinstance(obj, list):
                    return [dict_to_json(item) for item in obj]
                else:
                    return obj
            converted_data = dict_to_json(data)
            logger.debug("Converted to JSON-compatible format")
            return converted_data
        except (SyntaxError, ValueError) as e:
            logger.error("Python literal parsing failed: %s", str(e))
            raise ValueError(f"Malformed input: {str(e)}. Ensure property names are in double quotes (e.g., \"content\") or correct Python dictionary format.")

# Function to ensure a value is a float, converting from string if necessary
def ensure_float(value):
    if value is None:
        logger.debug("Replacing None logprob with 0.0")
        return 0.0  # Default to 0.0 for None to ensure visualization
    if isinstance(value, str):
        try:
            return float(value)
        except ValueError:
            logger.error("Failed to convert string '%s' to float", value)
            return 0.0  # Default to 0.0 for invalid strings
    if isinstance(value, (int, float)):
        return float(value)
    return 0.0  # Default for any other type

# Function to get or generate a token value (default to "Unknown" if missing)
def get_token(entry):
    token = entry.get("token", "Unknown")
    if token == "Unknown":
        logger.warning("Missing 'token' key for entry: %s, using 'Unknown'", entry)
    return token

# Function to create an empty Plotly figure
def create_empty_figure(title):
    return go.Figure().update_layout(title=title, xaxis_title="", yaxis_title="", showlegend=False)

# Function to process and visualize the full log probs with dynamic top_logprobs, handling missing tokens
def visualize_logprobs(json_input):
    try:
        # Parse the input (handles both JSON and Python dictionaries)
        data = parse_input(json_input)
        
        # Ensure data is a list or dictionary with 'content'
        if isinstance(data, dict) and "content" in data:
            content = data["content"]
        elif isinstance(data, list):
            content = data
        else:
            raise ValueError("Input must be a list or dictionary with 'content' key")

        # Extract tokens, log probs, and top alternatives, skipping non-finite values with fixed filter of -100000
        tokens = []
        logprobs = []
        top_alternatives = []  # List to store all top_logprobs (dynamic length)
        for entry in content:
            logprob = ensure_float(entry.get("logprob", None))
            if math.isfinite(logprob) and logprob >= -100000:
                token = get_token(entry)  # Safely get token, defaulting to "Unknown" if missing
                tokens.append(token)
                logprobs.append(logprob)
                # Get top_logprobs, default to empty dict if None
                top_probs = entry.get("top_logprobs", {})
                if top_probs is None:
                    logger.debug("top_logprobs is None for token: %s, using empty dict", token)
                    top_probs = {}  # Default to empty dict for None
                # Ensure all values in top_logprobs are floats and create a list of tuples
                finite_top_probs = []
                for key, value in top_probs.items():
                    float_value = ensure_float(value)
                    if float_value is not None and math.isfinite(float_value):
                        finite_top_probs.append((key, float_value))
                # Sort by log probability (descending) to get all alternatives
                sorted_probs = sorted(finite_top_probs, key=lambda x: x[1], reverse=True)
                top_alternatives.append(sorted_probs)  # Store all alternatives, dynamic length
            else:
                logger.debug("Skipping entry with logprob: %s (type: %s)", entry.get("logprob"), type(entry.get("logprob", None)))

        # Check if there's valid data after filtering
        if not logprobs or not tokens:
            return (create_empty_figure("Log Probabilities of Generated Tokens"), None, "No finite log probabilities to display.", create_empty_figure("Top Token Log Probabilities"), create_empty_figure("Significant Probability Drops"))

        # 1. Main Log Probability Plot (Interactive Plotly)
        main_fig = go.Figure()
        main_fig.add_trace(go.Scatter(x=list(range(len(logprobs))), y=logprobs, mode='markers+lines', name='Log Prob', marker=dict(color='blue')))
        main_fig.update_layout(
            title="Log Probabilities of Generated Tokens",
            xaxis_title="Token Position",
            yaxis_title="Log Probability",
            hovermode="closest",
            clickmode='event+select'
        )
        main_fig.update_traces(
            customdata=[f"Token: {tok}, Log Prob: {prob:.4f}, Position: {i}" for i, (tok, prob) in enumerate(zip(tokens, logprobs))],
            hovertemplate='<b>%{customdata}</b><extra></extra>'
        )

        # 2. Probability Drop Analysis (Interactive Plotly)
        if len(logprobs) < 2:
            drops_fig = create_empty_figure("Significant Probability Drops")
        else:
            drops = [logprobs[i+1] - logprobs[i] for i in range(len(logprobs)-1)]
            drops_fig = go.Figure()
            drops_fig.add_trace(go.Bar(x=list(range(len(drops))), y=drops, name='Drop', marker_color='red'))
            drops_fig.update_layout(
                title="Significant Probability Drops",
                xaxis_title="Token Position",
                yaxis_title="Log Probability Drop",
                hovermode="closest",
                clickmode='event+select'
            )
            drops_fig.update_traces(
                customdata=[f"Drop: {drop:.4f}, From: {tokens[i]} to {tokens[i+1]}, Position: {i}" for i, drop in enumerate(drops)],
                hovertemplate='<b>%{customdata}</b><extra></extra>'
            )

        # Create DataFrame for the table with dynamic top_logprobs
        table_data = []
        max_alternatives = max(len(alts) for alts in top_alternatives) if top_alternatives else 0
        for i, entry in enumerate(content):
            logprob = ensure_float(entry.get("logprob", None))
            if math.isfinite(logprob) and logprob >= -100000 and "top_logprobs" in entry:
                token = get_token(entry)  # Safely get token, defaulting to "Unknown" if missing
                top_logprobs = entry.get("top_logprobs", {})
                if top_logprobs is None:
                    logger.debug("top_logprobs is None for token: %s, using empty dict", token)
                    top_logprobs = {}  # Default to empty dict for None
                # Ensure all values in top_logprobs are floats
                finite_top_logprobs = []
                for key, value in top_logprobs.items():
                    float_value = ensure_float(value)
                    if float_value is not None and math.isfinite(float_value):
                        finite_top_logprobs.append((key, float_value))
                # Sort by log probability (descending)
                sorted_probs = sorted(finite_top_logprobs, key=lambda x: x[1], reverse=True)
                row = [token, f"{logprob:.4f}"]
                for alt_token, alt_logprob in sorted_probs[:max_alternatives]:  # Use max number of alternatives
                    row.append(f"{alt_token}: {alt_logprob:.4f}")
                # Pad with empty strings if fewer alternatives than max
                while len(row) < 2 + max_alternatives:
                    row.append("")
                table_data.append(row)

        df = (
            pd.DataFrame(
                table_data,
                columns=["Token", "Log Prob"] + [f"Alt {i+1}" for i in range(max_alternatives)],
            )
            if table_data
            else None
        )

        # Generate colored text
        if logprobs:
            min_logprob = min(logprobs)
            max_logprob = max(logprobs)
            if max_logprob == min_logprob:
                normalized_probs = [0.5] * len(logprobs)
            else:
                normalized_probs = [
                    (lp - min_logprob) / (max_logprob - min_logprob) for lp in logprobs
                ]

            colored_text = ""
            for i, (token, norm_prob) in enumerate(zip(tokens, normalized_probs)):
                r = int(255 * (1 - norm_prob))  # Red for low confidence
                g = int(255 * norm_prob)        # Green for high confidence
                b = 0
                color = f"rgb({r}, {g}, {b})"
                colored_text += f'<span style="color: {color}; font-weight: bold;">{token}</span>'
                if i < len(tokens) - 1:
                    colored_text += " "
            colored_text_html = f"<p>{colored_text}</p>"
        else:
            colored_text_html = "No finite log probabilities to display."

        # Top Token Log Probabilities (Interactive Plotly, dynamic length)
        alt_viz_fig = create_empty_figure("Top Token Log Probabilities") if not logprobs or not top_alternatives else go.Figure()
        if logprobs and top_alternatives:
            for i, (token, probs) in enumerate(zip(tokens, top_alternatives)):
                for j, (alt_tok, prob) in enumerate(probs):
                    alt_viz_fig.add_trace(go.Bar(x=[f"{token} (Pos {i})"], y=[prob], name=f"{alt_tok}", marker_color=['blue', 'green', 'red', 'purple', 'orange'][:len(probs)]))
            alt_viz_fig.update_layout(
                title="Top Token Log Probabilities",
                xaxis_title="Token (Position)",
                yaxis_title="Log Probability",
                barmode='stack',
                hovermode="closest",
                clickmode='event+select'
            )
            alt_viz_fig.update_traces(
                customdata=[f"Token: {tok}, Alt: {alt}, Log Prob: {prob:.4f}, Position: {i}" for i, (tok, alts) in enumerate(zip(tokens, top_alternatives)) for alt, prob in alts],
                hovertemplate='<b>%{customdata}</b><extra></extra>'
            )

        return (main_fig, df, colored_text_html, alt_viz_fig, drops_fig)

    except Exception as e:
        logger.error("Visualization failed: %s", str(e))
        return (create_empty_figure("Log Probabilities of Generated Tokens"), None, "No finite log probabilities to display.", create_empty_figure("Top Token Log Probabilities"), create_empty_figure("Significant Probability Drops"))

# Gradio interface with full dataset visualization, dynamic top_logprobs, and handling missing tokens
with gr.Blocks(title="Log Probability Visualizer") as app:
    gr.Markdown("# Log Probability Visualizer")
    gr.Markdown(
        "Paste your JSON or Python dictionary log prob data below to visualize all tokens at once. Fixed filter ≥ -100000, dynamic number of top_logprobs, handles missing 'token'."
    )

    with gr.Row():
        json_input = gr.Textbox(
            label="JSON Input",
            lines=10,
            placeholder="Paste your JSON (e.g., {\"content\": [...]}) or Python dict (e.g., {'content': [...]}) here...",
        )

    with gr.Row():
        plot_output = gr.Plot(label="Log Probability Plot (Click for Tokens)")
        drops_output = gr.Plot(label="Probability Drops (Click for Details)")

    with gr.Row():
        table_output = gr.Dataframe(label="Token Log Probabilities and Top Alternatives")
        alt_viz_output = gr.Plot(label="Top Token Log Probabilities (Click for Details)")

    with gr.Row():
        text_output = gr.HTML(label="Colored Text (Confidence Visualization)")

    btn = gr.Button("Visualize")
    btn.click(
        fn=visualize_logprobs,
        inputs=[json_input],
        outputs=[plot_output, table_output, text_output, alt_viz_output, drops_output],
    )

app.launch()