Spaces:
Running
Running
File size: 8,558 Bytes
0244d3c 6934db6 7e141c2 c28bdaa 0244d3c f2687d8 0244d3c f2687d8 9825333 f2687d8 9825333 f2687d8 0244d3c c28bdaa f2687d8 c28bdaa 0244d3c c28bdaa f2687d8 0244d3c f2687d8 c28bdaa 0244d3c c28bdaa 6934db6 0244d3c c28bdaa 0244d3c 7e141c2 c28bdaa 7e141c2 c28bdaa 7e141c2 0244d3c 7e141c2 0244d3c 9825333 0244d3c 9825333 0244d3c 6934db6 0244d3c 7e141c2 0244d3c 7e141c2 0244d3c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 |
import gradio as gr
import json
import matplotlib.pyplot as plt
import pandas as pd
import io
import base64
import ast
import math
# Function to safely convert string representations of infinity
def parse_infinity(value):
if isinstance(value, str):
if value.lower() == '-infinity' or value.lower() == '-inf':
return float('-inf')
elif value.lower() == 'infinity' or value.lower() == 'inf':
return float('inf')
return value
# Function to process and visualize log probs
def visualize_logprobs(json_input):
try:
# Try to parse as JSON first, handling string representations of infinity
try:
# Attempt to load JSON, replacing -inf with "-Infinity" if needed
def replace_inf(s):
import re
return re.sub(r'-inf', '"-Infinity"', re.sub(r'inf', '"Infinity"', s))
data = json.loads(replace_inf(json_input))
# Convert string "Infinity" or "-Infinity" back to float if needed
if isinstance(data, dict) and 'content' in data:
for entry in data['content']:
if 'logprob' in entry:
entry['logprob'] = parse_infinity(entry['logprob'])
if 'top_logprobs' in entry:
entry['top_logprobs'] = {k: parse_infinity(v) for k, v in entry['top_logprobs'].items()}
elif isinstance(data, list):
for entry in data:
if 'logprob' in entry:
entry['logprob'] = parse_infinity(entry['logprob'])
if 'top_logprobs' in entry:
entry['top_logprobs'] = {k: parse_infinity(v) for k, v in entry['top_logprobs'].items()}
except json.JSONDecodeError:
# If JSON fails, try to parse as Python literal (e.g., with single quotes)
try:
data = ast.literal_eval(json_input)
# Ensure -inf is handled as float('-inf')
if isinstance(data, dict) and 'content' in data:
for entry in data['content']:
if 'logprob' in entry and isinstance(entry['logprob'], str):
entry['logprob'] = parse_infinity(entry['logprob'])
if 'top_logprobs' in entry:
entry['top_logprobs'] = {k: parse_infinity(v) for k, v in entry['top_logprobs'].items()}
elif isinstance(data, list):
for entry in data:
if 'logprob' in entry and isinstance(entry['logprob'], str):
entry['logprob'] = parse_infinity(entry['logprob'])
if 'top_logprobs' in entry:
entry['top_logprobs'] = {k: parse_infinity(v) for k, v in entry['top_logprobs'].items()}
except (SyntaxError, ValueError) as e:
raise ValueError(f"Malformed input: {str(e)}")
# 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 and log probs, skipping None values and handling non-finite values
tokens = []
logprobs = []
for entry in content:
if 'logprob' in entry and entry['logprob'] is not None and math.isfinite(entry['logprob']):
tokens.append(entry['token'])
logprobs.append(entry['logprob'])
# Prepare data for the table
table_data = []
for entry in content:
if 'logprob' in entry and entry['logprob'] is not None and math.isfinite(entry['logprob']):
token = entry['token']
logprob = entry['logprob']
top_logprobs = entry.get('top_logprobs', {})
# Filter out non-finite (e.g., -inf, inf, nan) log probs from top_logprobs
finite_top_logprobs = {k: v for k, v in top_logprobs.items() if math.isfinite(v)}
# Extract top 3 finite alternatives, sorted by log prob (most probable first)
top_3 = sorted(finite_top_logprobs.items(), key=lambda x: x[1], reverse=True)[:3]
row = [token, f"{logprob:.4f}"]
for alt_token, alt_logprob in top_3:
row.append(f"{alt_token}: {alt_logprob:.4f}")
# Pad with empty strings if fewer than 3 alternatives
while len(row) < 5:
row.append("")
table_data.append(row)
# Create the plot (only for finite log probs)
if logprobs:
plt.figure(figsize=(10, 5))
plt.plot(range(len(logprobs)), logprobs, marker='o', linestyle='-', color='b')
plt.title("Log Probabilities of Generated Tokens")
plt.xlabel("Token Position")
plt.ylabel("Log Probability")
plt.grid(True)
plt.xticks(range(len(logprobs)), tokens, rotation=45, ha='right')
plt.tight_layout()
# Save plot to a bytes buffer
buf = io.BytesIO()
plt.savefig(buf, format='png', bbox_inches='tight')
buf.seek(0)
plt.close()
# Convert buffer to base64 for Gradio
img_bytes = buf.getvalue()
img_base64 = base64.b64encode(img_bytes).decode('utf-8')
img_html = f'<img src="data:image/png;base64,{img_base64}" style="max-width: 100%; height: auto;">'
else:
img_html = "No finite log probabilities to plot."
# Create a DataFrame for the table
df = pd.DataFrame(
table_data,
columns=["Token", "Log Prob", "Top 1 Alternative", "Top 2 Alternative", "Top 3 Alternative"]
) if table_data else None
# Generate colored text based on log probabilities
if logprobs:
# Normalize log probs to [0, 1] for color scaling (0 = most uncertain, 1 = most confident)
min_logprob = min(logprobs)
max_logprob = max(logprobs)
if max_logprob == min_logprob:
normalized_probs = [0.5] * len(logprobs) # Avoid division by zero
else:
normalized_probs = [(lp - min_logprob) / (max_logprob - min_logprob) for lp in logprobs]
# Create HTML for colored text
colored_text = ""
for i, (token, norm_prob) in enumerate(zip(tokens, normalized_probs)):
# Map normalized probability to RGB color (green for high confidence, red for low)
r = int(255 * (1 - norm_prob)) # Red increases as uncertainty increases
g = int(255 * norm_prob) # Green decreases as uncertainty increases
b = 0 # Blue stays 0 for simplicity
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 += " " # Add space between tokens
colored_text_html = f'<p>{colored_text}</p>'
else:
colored_text_html = "No finite log probabilities to display."
return img_html, df, colored_text_html
except Exception as e:
return f"Error: {str(e)}", None, None
# Gradio interface
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 the tokens and their probabilities.")
# Input
json_input = gr.Textbox(label="JSON Input", lines=10, placeholder="Paste your JSON or Python dict here...")
# Outputs
plot_output = gr.HTML(label="Log Probability Plot")
table_output = gr.Dataframe(label="Token Log Probabilities and Top Alternatives")
text_output = gr.HTML(label="Colored Text (Confidence Visualization)")
# Button to trigger visualization
btn = gr.Button("Visualize")
btn.click(
fn=visualize_logprobs,
inputs=json_input,
outputs=[plot_output, table_output, text_output]
)
# Launch the app
app.launch() |