Upload app.py
Browse files
app.py
CHANGED
|
@@ -1,5 +1,6 @@
|
|
| 1 |
import gradio as gr
|
| 2 |
import spaces
|
|
|
|
| 3 |
from inflections_funcs import start_model, make_beams, get_beam_tokens, calculate_score_vectors
|
| 4 |
|
| 5 |
# Global model initialization
|
|
@@ -10,7 +11,7 @@ print("Model loaded.")
|
|
| 10 |
@spaces.GPU
|
| 11 |
def predict(prompt, dark_mode):
|
| 12 |
"""
|
| 13 |
-
Generates a response and returns
|
| 14 |
"""
|
| 15 |
# Generate beams
|
| 16 |
generated_dicts, transcription = make_beams(model, processor, prompt)
|
|
@@ -20,17 +21,19 @@ def predict(prompt, dark_mode):
|
|
| 20 |
score_vectors = calculate_score_vectors(model, generated_dicts)
|
| 21 |
|
| 22 |
if not beam_tokens or not score_vectors:
|
| 23 |
-
return "<span style='color: grey'>No tokens generated.</span>"
|
| 24 |
|
| 25 |
first_beam_tokens = beam_tokens[0]
|
| 26 |
first_beam_scores = score_vectors[0]
|
|
|
|
| 27 |
|
| 28 |
-
# Construct HTML output
|
| 29 |
-
|
|
|
|
|
|
|
| 30 |
|
| 31 |
# Adjust colors based on dark mode
|
| 32 |
normal_color = "white" if dark_mode else "black"
|
| 33 |
-
border_color = "#444" if dark_mode else "#eee"
|
| 34 |
|
| 35 |
for token, score in zip(first_beam_tokens, first_beam_scores):
|
| 36 |
# Replace SentencePiece space character (U+2581) with a regular space
|
|
@@ -38,8 +41,7 @@ def predict(prompt, dark_mode):
|
|
| 38 |
|
| 39 |
if score < 0.6:
|
| 40 |
color = "red"
|
| 41 |
-
|
| 42 |
-
bg_color = "#ffe6e6" if not dark_mode else "#441111"
|
| 43 |
style = f"color: {color}; background-color: {bg_color};"
|
| 44 |
else:
|
| 45 |
color = normal_color
|
|
@@ -48,23 +50,59 @@ def predict(prompt, dark_mode):
|
|
| 48 |
html_output += f'<span style="{style}" title="Score: {score:.4f}">{display_token}</span>'
|
| 49 |
|
| 50 |
html_output += '</div>'
|
| 51 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 52 |
|
| 53 |
# Gradio Interface
|
| 54 |
with gr.Blocks() as demo:
|
| 55 |
gr.Markdown("# LLM Token Probability Visualizer")
|
| 56 |
gr.Markdown("Input a prompt and see the first beam's tokens. Tokens with a probability score < 0.6 are highlighted in red.")
|
| 57 |
|
| 58 |
-
with gr.
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 63 |
|
| 64 |
-
|
| 65 |
-
|
| 66 |
|
| 67 |
-
|
| 68 |
|
| 69 |
if __name__ == "__main__":
|
| 70 |
demo.launch()
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
import spaces
|
| 3 |
+
from torch import exp
|
| 4 |
from inflections_funcs import start_model, make_beams, get_beam_tokens, calculate_score_vectors
|
| 5 |
|
| 6 |
# Global model initialization
|
|
|
|
| 11 |
@spaces.GPU
|
| 12 |
def predict(prompt, dark_mode):
|
| 13 |
"""
|
| 14 |
+
Generates a response and returns a main highlighted HTML string and a detailed token list HTML string.
|
| 15 |
"""
|
| 16 |
# Generate beams
|
| 17 |
generated_dicts, transcription = make_beams(model, processor, prompt)
|
|
|
|
| 21 |
score_vectors = calculate_score_vectors(model, generated_dicts)
|
| 22 |
|
| 23 |
if not beam_tokens or not score_vectors:
|
| 24 |
+
return "<span style='color: grey'>No tokens generated.</span>", ""
|
| 25 |
|
| 26 |
first_beam_tokens = beam_tokens[0]
|
| 27 |
first_beam_scores = score_vectors[0]
|
| 28 |
+
first_beam_score = exp(generated_dicts.sequences_scores[0]).item()
|
| 29 |
|
| 30 |
+
# Construct Main HTML output
|
| 31 |
+
# We use a <h3> tag to simulate a Markdown header for the beam title and score
|
| 32 |
+
html_output = f'<h3 style="font-family: sans-serif;">Beam 1 | Score: {first_beam_score:.4f}</h3>'
|
| 33 |
+
html_output += '<div style="font-family: monospace; white-space: pre-wrap; line-height: 1.5;">'
|
| 34 |
|
| 35 |
# Adjust colors based on dark mode
|
| 36 |
normal_color = "white" if dark_mode else "black"
|
|
|
|
| 37 |
|
| 38 |
for token, score in zip(first_beam_tokens, first_beam_scores):
|
| 39 |
# Replace SentencePiece space character (U+2581) with a regular space
|
|
|
|
| 41 |
|
| 42 |
if score < 0.6:
|
| 43 |
color = "red"
|
| 44 |
+
bg_color = "#441111" if dark_mode else "#ffe6e6"
|
|
|
|
| 45 |
style = f"color: {color}; background-color: {bg_color};"
|
| 46 |
else:
|
| 47 |
color = normal_color
|
|
|
|
| 50 |
html_output += f'<span style="{style}" title="Score: {score:.4f}">{display_token}</span>'
|
| 51 |
|
| 52 |
html_output += '</div>'
|
| 53 |
+
|
| 54 |
+
# Construct Detailed HTML output
|
| 55 |
+
# We create a simple table to list tokens and their scores
|
| 56 |
+
detail_table_style = "border-collapse: collapse; width: 100%; max-width: 500px; font-family: monospace;"
|
| 57 |
+
th_style = "border: 1px solid #ccc; padding: 8px; text-align: left; background-color: #f2f2f2;" if not dark_mode else "border: 1px solid #444; padding: 8px; text-align: left; background-color: #333;"
|
| 58 |
+
td_style = "border: 1px solid #ccc; padding: 8px; text-align: left;" if not dark_mode else "border: 1px solid #444; padding: 8px; text-align: left;"
|
| 59 |
+
|
| 60 |
+
detail_output = f'<table style="{detail_table_style}"><thead><tr><th style="{th_style}">Token</th><th style="{th_style}">Score</th></tr></thead><tbody>'
|
| 61 |
+
for token, score in zip(first_beam_tokens, first_beam_scores):
|
| 62 |
+
display_token = token.replace(' ', ' ')
|
| 63 |
+
detail_output += f'<tr><td style="{td_style}">{display_token}</td><td style="{td_style}">{score:.4f}</td></tr>'
|
| 64 |
+
detail_output += '</tbody></table>'
|
| 65 |
+
|
| 66 |
+
return html_output, detail_output
|
| 67 |
|
| 68 |
# Gradio Interface
|
| 69 |
with gr.Blocks() as demo:
|
| 70 |
gr.Markdown("# LLM Token Probability Visualizer")
|
| 71 |
gr.Markdown("Input a prompt and see the first beam's tokens. Tokens with a probability score < 0.6 are highlighted in red.")
|
| 72 |
|
| 73 |
+
with gr.Column():
|
| 74 |
+
prompt_input = gr.Textbox(label="Prompt", placeholder="Enter your prompt here...", lines=3)
|
| 75 |
+
dark_mode_toggle = gr.Checkbox(label="Dark Mode", value=True)
|
| 76 |
+
submit_btn = gr.Button("Generate")
|
| 77 |
+
|
| 78 |
+
output_html = gr.HTML(label="Highlighted Beam")
|
| 79 |
+
|
| 80 |
+
toggle_btn = gr.Button("Show Token Details")
|
| 81 |
+
detail_html = gr.HTML(label="Detailed Token Probabilities", visible=False)
|
| 82 |
+
|
| 83 |
+
def toggle_details(visible):
|
| 84 |
+
return gr.update(visible=not visible, value="") if visible else gr.update(visible=True)
|
| 85 |
+
|
| 86 |
+
# This is slightly wrong. toggle_details should just toggle visibility.
|
| 87 |
+
# Let's use a better way.
|
| 88 |
+
def update_visibility(visible):
|
| 89 |
+
return gr.update(visible=not visible)
|
| 90 |
+
|
| 91 |
+
submit_btn.click(fn=predict, inputs=[prompt_input, dark_mode_toggle], outputs=[output_html, detail_html])
|
| 92 |
+
|
| 93 |
+
# We need to track the visibility state. Since we can't easily,
|
| 94 |
+
# we can use a gr.State or just a simple toggle function.
|
| 95 |
+
|
| 96 |
+
def on_toggle_click(current_visible):
|
| 97 |
+
return gr.update(visible=not current_visible)
|
| 98 |
+
|
| 99 |
+
# To track visibility, we can use a gr.State
|
| 100 |
+
visibility_state = gr.State(value=False)
|
| 101 |
|
| 102 |
+
def toggle_view(visible):
|
| 103 |
+
return not visible, gr.update(visible=not visible)
|
| 104 |
|
| 105 |
+
toggle_btn.click(fn=toggle_view, inputs=visibility_state, outputs=[visibility_state, detail_html])
|
| 106 |
|
| 107 |
if __name__ == "__main__":
|
| 108 |
demo.launch()
|