Spaces:
Running
Running
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,98 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import torch
|
3 |
+
from transformers import GPT2LMHeadModel, GPT2Tokenizer
|
4 |
+
import numpy as np
|
5 |
+
import matplotlib.pyplot as plt
|
6 |
+
import seaborn as sns
|
7 |
+
|
8 |
+
# Load model and tokenizer
|
9 |
+
model_name = "gpt2"
|
10 |
+
model = GPT2LMHeadModel.from_pretrained(model_name)
|
11 |
+
tokenizer = GPT2Tokenizer.from_pretrained(model_name)
|
12 |
+
|
13 |
+
def get_token_probabilities(text, top_k=10):
|
14 |
+
# Tokenize the input text
|
15 |
+
input_ids = tokenizer.encode(text, return_tensors="pt")
|
16 |
+
|
17 |
+
# Get the last token's position
|
18 |
+
last_token_position = input_ids.shape[1] - 1
|
19 |
+
|
20 |
+
# Get model predictions
|
21 |
+
with torch.no_grad():
|
22 |
+
outputs = model(input_ids)
|
23 |
+
logits = outputs.logits
|
24 |
+
|
25 |
+
# Get probabilities for the next token after the last token
|
26 |
+
next_token_logits = logits[0, last_token_position, :]
|
27 |
+
next_token_probs = torch.softmax(next_token_logits, dim=0)
|
28 |
+
|
29 |
+
# Get top k most likely tokens
|
30 |
+
topk_probs, topk_indices = torch.topk(next_token_probs, top_k)
|
31 |
+
|
32 |
+
# Convert to numpy for easier handling
|
33 |
+
topk_probs = topk_probs.numpy()
|
34 |
+
topk_indices = topk_indices.numpy()
|
35 |
+
|
36 |
+
# Decode tokens
|
37 |
+
topk_tokens = [tokenizer.decode([idx]) for idx in topk_indices]
|
38 |
+
|
39 |
+
# Create a plot
|
40 |
+
plt.figure(figsize=(10, 6))
|
41 |
+
sns.barplot(x=topk_probs, y=topk_tokens)
|
42 |
+
plt.title(f"Top {top_k} token probabilities after: '{text}'")
|
43 |
+
plt.xlabel("Probability")
|
44 |
+
plt.ylabel("Tokens")
|
45 |
+
plt.tight_layout()
|
46 |
+
|
47 |
+
# Save the plot to a file
|
48 |
+
plt.savefig("token_probabilities.png")
|
49 |
+
plt.close()
|
50 |
+
|
51 |
+
return "token_probabilities.png", dict(zip(topk_tokens, topk_probs.tolist()))
|
52 |
+
|
53 |
+
def interface():
|
54 |
+
with gr.Blocks() as demo:
|
55 |
+
gr.Markdown("# GPT-2 Next Token Probability Visualizer")
|
56 |
+
gr.Markdown("Enter text and see the probabilities of possible next tokens.")
|
57 |
+
|
58 |
+
with gr.Row():
|
59 |
+
with gr.Column():
|
60 |
+
input_text = gr.Textbox(
|
61 |
+
label="Input Text",
|
62 |
+
placeholder="Type some text here...",
|
63 |
+
value="Hello, my name is"
|
64 |
+
)
|
65 |
+
top_k = gr.Slider(
|
66 |
+
minimum=5,
|
67 |
+
maximum=20,
|
68 |
+
value=10,
|
69 |
+
step=1,
|
70 |
+
label="Number of top tokens to show"
|
71 |
+
)
|
72 |
+
btn = gr.Button("Generate Probabilities")
|
73 |
+
|
74 |
+
with gr.Column():
|
75 |
+
output_image = gr.Image(label="Probability Distribution")
|
76 |
+
output_table = gr.JSON(label="Token Probabilities")
|
77 |
+
|
78 |
+
btn.click(
|
79 |
+
fn=get_token_probabilities,
|
80 |
+
inputs=[input_text, top_k],
|
81 |
+
outputs=[output_image, output_table]
|
82 |
+
)
|
83 |
+
|
84 |
+
gr.Examples(
|
85 |
+
examples=[
|
86 |
+
["Hello, my name is", 10],
|
87 |
+
["The capital of France is", 10],
|
88 |
+
["Once upon a time", 10],
|
89 |
+
["The best way to learn is to", 10]
|
90 |
+
],
|
91 |
+
inputs=[input_text, top_k],
|
92 |
+
)
|
93 |
+
|
94 |
+
return demo
|
95 |
+
|
96 |
+
if __name__ == "__main__":
|
97 |
+
demo = interface()
|
98 |
+
demo.launch()
|