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