Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -9,7 +9,6 @@ import ast
|
|
9 |
import logging
|
10 |
import numpy as np
|
11 |
import plotly.graph_objects as go
|
12 |
-
from plotly.subplots import make_subplots
|
13 |
|
14 |
# Set up logging
|
15 |
logging.basicConfig(level=logging.DEBUG)
|
@@ -58,8 +57,12 @@ def ensure_float(value):
|
|
58 |
return float(value)
|
59 |
return None
|
60 |
|
|
|
|
|
|
|
|
|
61 |
# Function to process and visualize log probs with interactive Plotly plots
|
62 |
-
def visualize_logprobs(json_input
|
63 |
try:
|
64 |
# Parse the input (handles both JSON and Python dictionaries)
|
65 |
data = parse_input(json_input)
|
@@ -72,7 +75,7 @@ def visualize_logprobs(json_input, page=0):
|
|
72 |
else:
|
73 |
raise ValueError("Input must be a list or dictionary with 'content' key")
|
74 |
|
75 |
-
# Extract tokens
|
76 |
tokens = []
|
77 |
logprobs = []
|
78 |
top_alternatives = [] # List to store top 3 log probs (selected token + 2 alternatives)
|
@@ -100,20 +103,11 @@ def visualize_logprobs(json_input, page=0):
|
|
100 |
|
101 |
# Check if there's valid data after filtering
|
102 |
if not logprobs or not tokens:
|
103 |
-
return (
|
104 |
-
|
105 |
-
# Paginate data for large inputs (fixed page size of 1000)
|
106 |
-
page_size = 1000
|
107 |
-
total_pages = max(1, (len(logprobs) + page_size - 1) // page_size)
|
108 |
-
start_idx = page * page_size
|
109 |
-
end_idx = min((page + 1) * page_size, len(logprobs))
|
110 |
-
paginated_tokens = tokens[start_idx:end_idx]
|
111 |
-
paginated_logprobs = logprobs[start_idx:end_idx]
|
112 |
-
paginated_alternatives = top_alternatives[start_idx:end_idx] if top_alternatives else []
|
113 |
|
114 |
# 1. Main Log Probability Plot (Interactive Plotly)
|
115 |
main_fig = go.Figure()
|
116 |
-
main_fig.add_trace(go.Scatter(x=list(range(len(
|
117 |
main_fig.update_layout(
|
118 |
title="Log Probabilities of Generated Tokens",
|
119 |
xaxis_title="Token Position",
|
@@ -122,16 +116,15 @@ def visualize_logprobs(json_input, page=0):
|
|
122 |
clickmode='event+select'
|
123 |
)
|
124 |
main_fig.update_traces(
|
125 |
-
customdata=[f"Token: {tok}, Log Prob: {prob:.4f}, Position: {i
|
126 |
hovertemplate='<b>%{customdata}</b><extra></extra>'
|
127 |
)
|
128 |
|
129 |
# 2. Probability Drop Analysis (Interactive Plotly)
|
130 |
-
if len(
|
131 |
-
drops_fig =
|
132 |
-
drops_fig.add_trace(go.Bar(x=list(range(len(paginated_logprobs)-1)), y=[0], name='Drop', marker_color='red'))
|
133 |
else:
|
134 |
-
drops = [
|
135 |
drops_fig = go.Figure()
|
136 |
drops_fig.add_trace(go.Bar(x=list(range(len(drops))), y=drops, name='Drop', marker_color='red'))
|
137 |
drops_fig.update_layout(
|
@@ -142,13 +135,13 @@ def visualize_logprobs(json_input, page=0):
|
|
142 |
clickmode='event+select'
|
143 |
)
|
144 |
drops_fig.update_traces(
|
145 |
-
customdata=[f"Drop: {drop:.4f}, From: {
|
146 |
hovertemplate='<b>%{customdata}</b><extra></extra>'
|
147 |
)
|
148 |
|
149 |
-
# Create DataFrame for the table
|
150 |
table_data = []
|
151 |
-
for i, entry in enumerate(content
|
152 |
logprob = ensure_float(entry.get("logprob", None))
|
153 |
if logprob is not None and math.isfinite(logprob) and logprob >= -100000 and "top_logprobs" in entry and entry["top_logprobs"] is not None:
|
154 |
token = entry["token"]
|
@@ -183,38 +176,38 @@ def visualize_logprobs(json_input, page=0):
|
|
183 |
else None
|
184 |
)
|
185 |
|
186 |
-
# Generate colored text
|
187 |
-
if
|
188 |
-
min_logprob = min(
|
189 |
-
max_logprob = max(
|
190 |
if max_logprob == min_logprob:
|
191 |
-
normalized_probs = [0.5] * len(
|
192 |
else:
|
193 |
normalized_probs = [
|
194 |
-
(lp - min_logprob) / (max_logprob - min_logprob) for lp in
|
195 |
]
|
196 |
|
197 |
colored_text = ""
|
198 |
-
for i, (token, norm_prob) in enumerate(zip(
|
199 |
r = int(255 * (1 - norm_prob)) # Red for low confidence
|
200 |
g = int(255 * norm_prob) # Green for high confidence
|
201 |
b = 0
|
202 |
color = f"rgb({r}, {g}, {b})"
|
203 |
colored_text += f'<span style="color: {color}; font-weight: bold;">{token}</span>'
|
204 |
-
if i < len(
|
205 |
colored_text += " "
|
206 |
colored_text_html = f"<p>{colored_text}</p>"
|
207 |
else:
|
208 |
colored_text_html = "No finite log probabilities to display."
|
209 |
|
210 |
-
# Top 3 Token Log Probabilities (
|
211 |
-
alt_viz_fig = go.Figure()
|
212 |
-
if
|
213 |
-
for i, (token, probs) in enumerate(zip(
|
214 |
for j, (alt_tok, prob) in enumerate(probs):
|
215 |
-
alt_viz_fig.add_trace(go.Bar(x=[f"{token} (Pos {i
|
216 |
alt_viz_fig.update_layout(
|
217 |
-
title="Top 3 Token Log Probabilities
|
218 |
xaxis_title="Token (Position)",
|
219 |
yaxis_title="Log Probability",
|
220 |
barmode='stack',
|
@@ -222,35 +215,29 @@ def visualize_logprobs(json_input, page=0):
|
|
222 |
clickmode='event+select'
|
223 |
)
|
224 |
alt_viz_fig.update_traces(
|
225 |
-
customdata=[f"Token: {tok}, Alt: {alt}, Log Prob: {prob:.4f}, Position: {i
|
226 |
hovertemplate='<b>%{customdata}</b><extra></extra>'
|
227 |
)
|
228 |
-
alt_viz_html = alt_viz_fig.to_html(include_plotlyjs='cdn', full_html=False)
|
229 |
-
else:
|
230 |
-
alt_viz_html = "No finite log probabilities to display."
|
231 |
|
232 |
-
return (main_fig, df, colored_text_html,
|
233 |
|
234 |
except Exception as e:
|
235 |
logger.error("Visualization failed: %s", str(e))
|
236 |
-
return (
|
237 |
|
238 |
-
# Gradio interface with
|
239 |
with gr.Blocks(title="Log Probability Visualizer") as app:
|
240 |
gr.Markdown("# Log Probability Visualizer")
|
241 |
gr.Markdown(
|
242 |
-
"Paste your JSON or Python dictionary log prob data below to visualize the tokens and their probabilities.
|
243 |
)
|
244 |
|
245 |
with gr.Row():
|
246 |
-
|
247 |
-
|
248 |
-
|
249 |
-
|
250 |
-
|
251 |
-
)
|
252 |
-
with gr.Column(scale=1):
|
253 |
-
page = gr.Number(value=0, label="Page Number", precision=0, minimum=0)
|
254 |
|
255 |
with gr.Row():
|
256 |
plot_output = gr.Plot(label="Log Probability Plot (Click for Tokens)")
|
@@ -266,36 +253,8 @@ with gr.Blocks(title="Log Probability Visualizer") as app:
|
|
266 |
btn = gr.Button("Visualize")
|
267 |
btn.click(
|
268 |
fn=visualize_logprobs,
|
269 |
-
inputs=[json_input
|
270 |
-
outputs=[plot_output, table_output, text_output, alt_viz_output, drops_output
|
271 |
-
)
|
272 |
-
|
273 |
-
# Pagination controls
|
274 |
-
with gr.Row():
|
275 |
-
prev_btn = gr.Button("Previous Page")
|
276 |
-
next_btn = gr.Button("Next Page")
|
277 |
-
total_pages_output = gr.Number(label="Total Pages", interactive=False)
|
278 |
-
current_page_output = gr.Number(label="Current Page", interactive=False)
|
279 |
-
|
280 |
-
def update_page(json_input, current_page, action):
|
281 |
-
if action == "prev" and current_page > 0:
|
282 |
-
current_page -= 1
|
283 |
-
elif action == "next":
|
284 |
-
total_pages = visualize_logprobs(json_input, 0)[5] # Get total pages
|
285 |
-
if current_page < total_pages - 1:
|
286 |
-
current_page += 1
|
287 |
-
return gr.update(value=current_page), gr.update(value=total_pages)
|
288 |
-
|
289 |
-
prev_btn.click(
|
290 |
-
fn=update_page,
|
291 |
-
inputs=[json_input, page, gr.State()],
|
292 |
-
outputs=[page, total_pages_output]
|
293 |
-
)
|
294 |
-
|
295 |
-
next_btn.click(
|
296 |
-
fn=update_page,
|
297 |
-
inputs=[json_input, page, gr.State()],
|
298 |
-
outputs=[page, total_pages_output]
|
299 |
)
|
300 |
|
301 |
app.launch()
|
|
|
9 |
import logging
|
10 |
import numpy as np
|
11 |
import plotly.graph_objects as go
|
|
|
12 |
|
13 |
# Set up logging
|
14 |
logging.basicConfig(level=logging.DEBUG)
|
|
|
57 |
return float(value)
|
58 |
return None
|
59 |
|
60 |
+
# Function to create an empty Plotly figure
|
61 |
+
def create_empty_figure(title):
|
62 |
+
return go.Figure().update_layout(title=title, xaxis_title="", yaxis_title="", showlegend=False)
|
63 |
+
|
64 |
# Function to process and visualize log probs with interactive Plotly plots
|
65 |
+
def visualize_logprobs(json_input):
|
66 |
try:
|
67 |
# Parse the input (handles both JSON and Python dictionaries)
|
68 |
data = parse_input(json_input)
|
|
|
75 |
else:
|
76 |
raise ValueError("Input must be a list or dictionary with 'content' key")
|
77 |
|
78 |
+
# Extract tokens and log probs, skipping None or non-finite values with fixed filter of -100000
|
79 |
tokens = []
|
80 |
logprobs = []
|
81 |
top_alternatives = [] # List to store top 3 log probs (selected token + 2 alternatives)
|
|
|
103 |
|
104 |
# Check if there's valid data after filtering
|
105 |
if not logprobs or not tokens:
|
106 |
+
return (create_empty_figure("Log Probabilities of Generated Tokens"), None, "No finite log probabilities to display.", create_empty_figure("Top 3 Token Log Probabilities"), create_empty_figure("Significant Probability Drops"))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
107 |
|
108 |
# 1. Main Log Probability Plot (Interactive Plotly)
|
109 |
main_fig = go.Figure()
|
110 |
+
main_fig.add_trace(go.Scatter(x=list(range(len(logprobs))), y=logprobs, mode='markers+lines', name='Log Prob', marker=dict(color='blue')))
|
111 |
main_fig.update_layout(
|
112 |
title="Log Probabilities of Generated Tokens",
|
113 |
xaxis_title="Token Position",
|
|
|
116 |
clickmode='event+select'
|
117 |
)
|
118 |
main_fig.update_traces(
|
119 |
+
customdata=[f"Token: {tok}, Log Prob: {prob:.4f}, Position: {i}" for i, (tok, prob) in enumerate(zip(tokens, logprobs))],
|
120 |
hovertemplate='<b>%{customdata}</b><extra></extra>'
|
121 |
)
|
122 |
|
123 |
# 2. Probability Drop Analysis (Interactive Plotly)
|
124 |
+
if len(logprobs) < 2:
|
125 |
+
drops_fig = create_empty_figure("Significant Probability Drops")
|
|
|
126 |
else:
|
127 |
+
drops = [logprobs[i+1] - logprobs[i] for i in range(len(logprobs)-1)]
|
128 |
drops_fig = go.Figure()
|
129 |
drops_fig.add_trace(go.Bar(x=list(range(len(drops))), y=drops, name='Drop', marker_color='red'))
|
130 |
drops_fig.update_layout(
|
|
|
135 |
clickmode='event+select'
|
136 |
)
|
137 |
drops_fig.update_traces(
|
138 |
+
customdata=[f"Drop: {drop:.4f}, From: {tokens[i]} to {tokens[i+1]}, Position: {i}" for i, drop in enumerate(drops)],
|
139 |
hovertemplate='<b>%{customdata}</b><extra></extra>'
|
140 |
)
|
141 |
|
142 |
+
# Create DataFrame for the table
|
143 |
table_data = []
|
144 |
+
for i, entry in enumerate(content):
|
145 |
logprob = ensure_float(entry.get("logprob", None))
|
146 |
if logprob is not None and math.isfinite(logprob) and logprob >= -100000 and "top_logprobs" in entry and entry["top_logprobs"] is not None:
|
147 |
token = entry["token"]
|
|
|
176 |
else None
|
177 |
)
|
178 |
|
179 |
+
# Generate colored text
|
180 |
+
if logprobs:
|
181 |
+
min_logprob = min(logprobs)
|
182 |
+
max_logprob = max(logprobs)
|
183 |
if max_logprob == min_logprob:
|
184 |
+
normalized_probs = [0.5] * len(logprobs)
|
185 |
else:
|
186 |
normalized_probs = [
|
187 |
+
(lp - min_logprob) / (max_logprob - min_logprob) for lp in logprobs
|
188 |
]
|
189 |
|
190 |
colored_text = ""
|
191 |
+
for i, (token, norm_prob) in enumerate(zip(tokens, normalized_probs)):
|
192 |
r = int(255 * (1 - norm_prob)) # Red for low confidence
|
193 |
g = int(255 * norm_prob) # Green for high confidence
|
194 |
b = 0
|
195 |
color = f"rgb({r}, {g}, {b})"
|
196 |
colored_text += f'<span style="color: {color}; font-weight: bold;">{token}</span>'
|
197 |
+
if i < len(tokens) - 1:
|
198 |
colored_text += " "
|
199 |
colored_text_html = f"<p>{colored_text}</p>"
|
200 |
else:
|
201 |
colored_text_html = "No finite log probabilities to display."
|
202 |
|
203 |
+
# Top 3 Token Log Probabilities (Interactive Plotly)
|
204 |
+
alt_viz_fig = create_empty_figure("Top 3 Token Log Probabilities") if not logprobs or not top_alternatives else go.Figure()
|
205 |
+
if logprobs and top_alternatives:
|
206 |
+
for i, (token, probs) in enumerate(zip(tokens, top_alternatives)):
|
207 |
for j, (alt_tok, prob) in enumerate(probs):
|
208 |
+
alt_viz_fig.add_trace(go.Bar(x=[f"{token} (Pos {i})"], y=[prob], name=f"{alt_tok}", marker_color=['blue', 'green', 'red'][j]))
|
209 |
alt_viz_fig.update_layout(
|
210 |
+
title="Top 3 Token Log Probabilities",
|
211 |
xaxis_title="Token (Position)",
|
212 |
yaxis_title="Log Probability",
|
213 |
barmode='stack',
|
|
|
215 |
clickmode='event+select'
|
216 |
)
|
217 |
alt_viz_fig.update_traces(
|
218 |
+
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],
|
219 |
hovertemplate='<b>%{customdata}</b><extra></extra>'
|
220 |
)
|
|
|
|
|
|
|
221 |
|
222 |
+
return (main_fig, df, colored_text_html, alt_viz_fig, drops_fig)
|
223 |
|
224 |
except Exception as e:
|
225 |
logger.error("Visualization failed: %s", str(e))
|
226 |
+
return (create_empty_figure("Log Probabilities of Generated Tokens"), None, "No finite log probabilities to display.", create_empty_figure("Top 3 Token Log Probabilities"), create_empty_figure("Significant Probability Drops"))
|
227 |
|
228 |
+
# Gradio interface with improved layout
|
229 |
with gr.Blocks(title="Log Probability Visualizer") as app:
|
230 |
gr.Markdown("# Log Probability Visualizer")
|
231 |
gr.Markdown(
|
232 |
+
"Paste your JSON or Python dictionary log prob data below to visualize the tokens and their probabilities. Fixed filter ≥ -100000, 1000 tokens per page."
|
233 |
)
|
234 |
|
235 |
with gr.Row():
|
236 |
+
json_input = gr.Textbox(
|
237 |
+
label="JSON Input",
|
238 |
+
lines=10,
|
239 |
+
placeholder="Paste your JSON (e.g., {\"content\": [...]}) or Python dict (e.g., {'content': [...]}) here...",
|
240 |
+
)
|
|
|
|
|
|
|
241 |
|
242 |
with gr.Row():
|
243 |
plot_output = gr.Plot(label="Log Probability Plot (Click for Tokens)")
|
|
|
253 |
btn = gr.Button("Visualize")
|
254 |
btn.click(
|
255 |
fn=visualize_logprobs,
|
256 |
+
inputs=[json_input],
|
257 |
+
outputs=[plot_output, table_output, text_output, alt_viz_output, drops_output],
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
258 |
)
|
259 |
|
260 |
app.launch()
|