Spaces:
Running
Running
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,132 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
3 |
+
import torch
|
4 |
+
from typing import Tuple, List, Dict
|
5 |
+
import numpy as np
|
6 |
+
|
7 |
+
# Select smaller models that are suitable for this task
|
8 |
+
AVAILABLE_MODELS = {
|
9 |
+
"distilgpt2": "distilgpt2",
|
10 |
+
"bloomz-560m": "bigscience/bloomz-560m",
|
11 |
+
"gpt2-medium": "gpt2-medium",
|
12 |
+
"opt-350m": "facebook/opt-350m",
|
13 |
+
"pythia-160m": "EleutherAI/pythia-160m"
|
14 |
+
}
|
15 |
+
|
16 |
+
class TextGenerator:
|
17 |
+
def __init__(self):
|
18 |
+
self.model = None
|
19 |
+
self.tokenizer = None
|
20 |
+
|
21 |
+
def load_model(self, model_name: str) -> str:
|
22 |
+
"""Load the selected model and tokenizer"""
|
23 |
+
try:
|
24 |
+
self.model = AutoModelForCausalLM.from_pretrained(AVAILABLE_MODELS[model_name])
|
25 |
+
self.tokenizer = AutoTokenizer.from_pretrained(AVAILABLE_MODELS[model_name])
|
26 |
+
return f"Successfully loaded {model_name}"
|
27 |
+
except Exception as e:
|
28 |
+
return f"Error loading model: {str(e)}"
|
29 |
+
|
30 |
+
def get_next_token_predictions(self, text: str, top_k: int = 10) -> Tuple[List[str], List[float]]:
|
31 |
+
"""Get predictions for the next token"""
|
32 |
+
if not self.model or not self.tokenizer:
|
33 |
+
return [], []
|
34 |
+
|
35 |
+
inputs = self.tokenizer(text, return_tensors="pt")
|
36 |
+
with torch.no_grad():
|
37 |
+
outputs = self.model(**inputs)
|
38 |
+
logits = outputs.logits[0, -1, :]
|
39 |
+
probs = torch.nn.functional.softmax(logits, dim=-1)
|
40 |
+
|
41 |
+
top_k_probs, top_k_indices = torch.topk(probs, top_k)
|
42 |
+
top_k_tokens = [self.tokenizer.decode([idx.item()]) for idx in top_k_indices]
|
43 |
+
top_k_probs = top_k_probs.tolist()
|
44 |
+
|
45 |
+
return top_k_tokens, top_k_probs
|
46 |
+
|
47 |
+
def format_predictions(tokens: List[str], probs: List[float]) -> str:
|
48 |
+
"""Format the predictions for display"""
|
49 |
+
if not tokens or not probs:
|
50 |
+
return "No predictions available"
|
51 |
+
|
52 |
+
formatted = "Predicted next tokens:\n\n"
|
53 |
+
for token, prob in zip(tokens, probs):
|
54 |
+
formatted += f"'{token}' : {prob:.4f}\n"
|
55 |
+
return formatted
|
56 |
+
|
57 |
+
generator = TextGenerator()
|
58 |
+
|
59 |
+
def update_output(model_name: str, text: str, custom_token: str, selected_token: str) -> Tuple[str, str, str, Dict, str]:
|
60 |
+
"""Update the interface based on user interactions"""
|
61 |
+
output = text
|
62 |
+
|
63 |
+
# Load model if it changed
|
64 |
+
if not generator.model or generator.model.name_or_path != AVAILABLE_MODELS[model_name]:
|
65 |
+
load_message = generator.load_model(model_name)
|
66 |
+
if "Error" in load_message:
|
67 |
+
return text, "", "", gr.update(choices=[]), load_message
|
68 |
+
|
69 |
+
# Add custom token or selected token
|
70 |
+
if custom_token:
|
71 |
+
output += custom_token
|
72 |
+
elif selected_token:
|
73 |
+
output += selected_token.strip("'")
|
74 |
+
|
75 |
+
# Get new predictions
|
76 |
+
tokens, probs = generator.get_next_token_predictions(output)
|
77 |
+
predictions = format_predictions(tokens, probs)
|
78 |
+
|
79 |
+
# Update dropdown choices
|
80 |
+
token_choices = [f"'{token}'" for token in tokens]
|
81 |
+
|
82 |
+
return output, "", "", gr.update(choices=token_choices), predictions
|
83 |
+
|
84 |
+
with gr.Blocks() as app:
|
85 |
+
gr.Markdown("# Interactive Text Generation")
|
86 |
+
|
87 |
+
with gr.Row():
|
88 |
+
model_dropdown = gr.Dropdown(
|
89 |
+
choices=list(AVAILABLE_MODELS.keys()),
|
90 |
+
value="distilgpt2",
|
91 |
+
label="Select Model"
|
92 |
+
)
|
93 |
+
|
94 |
+
with gr.Row():
|
95 |
+
text_input = gr.Textbox(
|
96 |
+
lines=5,
|
97 |
+
label="Generated Text",
|
98 |
+
placeholder="Start typing or select a token..."
|
99 |
+
)
|
100 |
+
|
101 |
+
with gr.Row():
|
102 |
+
custom_token = gr.Textbox(
|
103 |
+
label="Custom Token",
|
104 |
+
placeholder="Type your own token..."
|
105 |
+
)
|
106 |
+
token_dropdown = gr.Dropdown(
|
107 |
+
choices=[],
|
108 |
+
label="Select from predicted tokens"
|
109 |
+
)
|
110 |
+
|
111 |
+
with gr.Row():
|
112 |
+
predictions_output = gr.Textbox(
|
113 |
+
label="Predictions",
|
114 |
+
lines=12
|
115 |
+
)
|
116 |
+
|
117 |
+
with gr.Row():
|
118 |
+
status_output = gr.Textbox(
|
119 |
+
label="Status",
|
120 |
+
lines=1
|
121 |
+
)
|
122 |
+
|
123 |
+
# Update when model changes or token is added
|
124 |
+
for trigger in [model_dropdown, custom_token, token_dropdown]:
|
125 |
+
trigger.change(
|
126 |
+
fn=update_output,
|
127 |
+
inputs=[model_dropdown, text_input, custom_token, token_dropdown],
|
128 |
+
outputs=[text_input, custom_token, token_dropdown, token_dropdown, predictions_output]
|
129 |
+
)
|
130 |
+
|
131 |
+
if __name__ == "__main__":
|
132 |
+
app.launch()
|