Spaces:
Running
Running
Pclanglais
commited on
Commit
•
e98a756
1
Parent(s):
e2b4df4
Update app.py
Browse files
app.py
CHANGED
@@ -1,28 +1,18 @@
|
|
1 |
-
import spaces
|
2 |
import transformers
|
3 |
import re
|
4 |
-
from transformers import
|
5 |
import torch
|
6 |
import gradio as gr
|
7 |
-
import json
|
8 |
-
import os
|
9 |
-
import shutil
|
10 |
-
import requests
|
11 |
-
import pandas as pd
|
12 |
import difflib
|
13 |
from concurrent.futures import ThreadPoolExecutor
|
|
|
14 |
|
15 |
# OCR Correction Model
|
16 |
-
|
17 |
-
|
18 |
-
import torch
|
19 |
-
from transformers import GPT2LMHeadModel, GPT2Tokenizer
|
20 |
-
|
21 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
22 |
|
23 |
# Load pre-trained model and tokenizer
|
24 |
-
|
25 |
-
model = GPT2LMHeadModel.from_pretrained(model_name)
|
26 |
tokenizer = GPT2Tokenizer.from_pretrained(model_name)
|
27 |
|
28 |
# CSS for formatting
|
@@ -33,78 +23,12 @@ css = """
|
|
33 |
margin-right: 2em;
|
34 |
font-size: 1.2em;
|
35 |
}
|
36 |
-
:target {
|
37 |
-
background-color: #CCF3DF;
|
38 |
-
}
|
39 |
-
.source {
|
40 |
-
float: left;
|
41 |
-
max-width: 17%;
|
42 |
-
margin-left: 2%;
|
43 |
-
}
|
44 |
-
.tooltip {
|
45 |
-
position: relative;
|
46 |
-
cursor: pointer;
|
47 |
-
font-variant-position: super;
|
48 |
-
color: #97999b;
|
49 |
-
}
|
50 |
-
.tooltip:hover::after {
|
51 |
-
content: attr(data-text);
|
52 |
-
position: absolute;
|
53 |
-
left: 0;
|
54 |
-
top: 120%;
|
55 |
-
white-space: pre-wrap;
|
56 |
-
width: 500px;
|
57 |
-
max-width: 500px;
|
58 |
-
z-index: 1;
|
59 |
-
background-color: #f9f9f9;
|
60 |
-
color: #000;
|
61 |
-
border: 1px solid #ddd;
|
62 |
-
border-radius: 5px;
|
63 |
-
padding: 5px;
|
64 |
-
display: block;
|
65 |
-
box-shadow: 0 4px 8px rgba(0,0,0,0.1);
|
66 |
-
}
|
67 |
-
.deleted {
|
68 |
-
background-color: #ffcccb;
|
69 |
-
text-decoration: line-through;
|
70 |
-
}
|
71 |
.inserted {
|
72 |
background-color: #90EE90;
|
73 |
}
|
74 |
-
.manuscript {
|
75 |
-
display: flex;
|
76 |
-
margin-bottom: 10px;
|
77 |
-
align-items: baseline;
|
78 |
-
}
|
79 |
-
.annotation {
|
80 |
-
width: 15%;
|
81 |
-
padding-right: 20px;
|
82 |
-
color: grey !important;
|
83 |
-
font-style: italic;
|
84 |
-
text-align: right;
|
85 |
-
}
|
86 |
-
.content {
|
87 |
-
width: 80%;
|
88 |
-
}
|
89 |
-
h2 {
|
90 |
-
margin: 0;
|
91 |
-
font-size: 1.5em;
|
92 |
-
}
|
93 |
-
.title-content h2 {
|
94 |
-
font-weight: bold;
|
95 |
-
}
|
96 |
-
.bibliography-content {
|
97 |
-
color: darkgreen !important;
|
98 |
-
margin-top: -5px;
|
99 |
-
}
|
100 |
-
.paratext-content {
|
101 |
-
color: #a4a4a4 !important;
|
102 |
-
margin-top: -5px;
|
103 |
-
}
|
104 |
</style>
|
105 |
"""
|
106 |
|
107 |
-
# Helper functions
|
108 |
def generate_html_diff(old_text, new_text):
|
109 |
d = difflib.Differ()
|
110 |
diff = list(d.compare(old_text.split(), new_text.split()))
|
@@ -113,64 +37,31 @@ def generate_html_diff(old_text, new_text):
|
|
113 |
if word.startswith(' '):
|
114 |
html_diff.append(word[2:])
|
115 |
elif word.startswith('+ '):
|
116 |
-
html_diff.append(f'<span
|
117 |
return ' '.join(html_diff)
|
118 |
|
119 |
-
def
|
120 |
-
|
121 |
-
text = re.sub(r'\n', ' ', text)
|
122 |
-
text = re.sub(r'\s+', ' ', text)
|
123 |
-
return text.strip()
|
124 |
-
|
125 |
-
def split_text(text, max_tokens=500):
|
126 |
-
parts = text.split("\n")
|
127 |
chunks = []
|
128 |
-
current_chunk =
|
129 |
-
|
130 |
-
for part in parts:
|
131 |
-
if current_chunk:
|
132 |
-
temp_chunk = current_chunk + "\n" + part
|
133 |
-
else:
|
134 |
-
temp_chunk = part
|
135 |
|
136 |
-
|
137 |
-
|
138 |
-
if
|
139 |
-
current_chunk
|
140 |
-
|
141 |
-
if current_chunk:
|
142 |
-
chunks.append(current_chunk)
|
143 |
-
current_chunk = part
|
144 |
|
145 |
if current_chunk:
|
146 |
-
chunks.append(current_chunk)
|
147 |
-
|
148 |
-
if len(chunks) == 1 and len(tokenizer.tokenize(chunks[0])) > max_tokens:
|
149 |
-
long_text = chunks[0]
|
150 |
-
chunks = []
|
151 |
-
while len(tokenizer.tokenize(long_text)) > max_tokens:
|
152 |
-
split_point = len(long_text) // 2
|
153 |
-
while split_point < len(long_text) and not re.match(r'\s', long_text[split_point]):
|
154 |
-
split_point += 1
|
155 |
-
if split_point >= len(long_text):
|
156 |
-
split_point = len(long_text) - 1
|
157 |
-
chunks.append(long_text[:split_point].strip())
|
158 |
-
long_text = long_text[split_point:].strip()
|
159 |
-
if long_text:
|
160 |
-
chunks.append(long_text)
|
161 |
|
162 |
return chunks
|
163 |
|
164 |
-
|
165 |
-
# Function to generate text
|
166 |
def ocr_correction(prompt, max_new_tokens=600, num_threads=os.cpu_count()):
|
167 |
prompt = f"""### Text ###\n{prompt}\n\n\n### Correction ###\n"""
|
168 |
input_ids = tokenizer.encode(prompt, return_tensors="pt").to(device)
|
169 |
|
170 |
-
# Set the number of threads for PyTorch
|
171 |
torch.set_num_threads(num_threads)
|
172 |
|
173 |
-
# Generate text
|
174 |
with ThreadPoolExecutor(max_workers=num_threads) as executor:
|
175 |
future = executor.submit(
|
176 |
model.generate,
|
@@ -183,41 +74,23 @@ def ocr_correction(prompt, max_new_tokens=600, num_threads=os.cpu_count()):
|
|
183 |
)
|
184 |
output = future.result()
|
185 |
|
186 |
-
# Decode and return the generated text
|
187 |
result = tokenizer.decode(output[0], skip_special_tokens=True)
|
188 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
189 |
|
190 |
-
|
191 |
-
|
192 |
-
|
193 |
-
# OCR Correction Class
|
194 |
-
class OCRCorrector:
|
195 |
-
def __init__(self, system_prompt="Le dialogue suivant est une conversation"):
|
196 |
-
self.system_prompt = system_prompt
|
197 |
-
|
198 |
-
def correct(self, user_message):
|
199 |
-
generated_text = ocr_correction(user_message)
|
200 |
-
html_diff = generate_html_diff(user_message, generated_text)
|
201 |
-
return generated_text, html_diff
|
202 |
-
|
203 |
-
# Combined Processing Class
|
204 |
-
class TextProcessor:
|
205 |
-
def __init__(self):
|
206 |
-
self.ocr_corrector = OCRCorrector()
|
207 |
-
|
208 |
-
@spaces.GPU(duration=120)
|
209 |
-
def process(self, user_message):
|
210 |
-
#OCR Correction
|
211 |
-
corrected_text, html_diff = self.ocr_corrector.correct(user_message)
|
212 |
-
|
213 |
-
# Combine results
|
214 |
-
ocr_result = f'<h2 style="text-align:center">OCR Correction</h2>\n<div class="generation">{html_diff}</div>'
|
215 |
-
|
216 |
-
final_output = f"{css}{ocr_result}"
|
217 |
-
return final_output
|
218 |
-
|
219 |
-
# Create the TextProcessor instance
|
220 |
-
text_processor = TextProcessor()
|
221 |
|
222 |
# Define the Gradio interface
|
223 |
with gr.Blocks(theme='JohnSmith9982/small_and_pretty') as demo:
|
@@ -225,7 +98,7 @@ with gr.Blocks(theme='JohnSmith9982/small_and_pretty') as demo:
|
|
225 |
text_input = gr.Textbox(label="Your (bad?) text", type="text", lines=5)
|
226 |
process_button = gr.Button("Process Text")
|
227 |
text_output = gr.HTML(label="Processed text")
|
228 |
-
process_button.click(
|
229 |
|
230 |
if __name__ == "__main__":
|
231 |
demo.queue().launch()
|
|
|
|
|
1 |
import transformers
|
2 |
import re
|
3 |
+
from transformers import GPT2LMHeadModel, GPT2Tokenizer
|
4 |
import torch
|
5 |
import gradio as gr
|
|
|
|
|
|
|
|
|
|
|
6 |
import difflib
|
7 |
from concurrent.futures import ThreadPoolExecutor
|
8 |
+
import os
|
9 |
|
10 |
# OCR Correction Model
|
11 |
+
model_name = "PleIAs/OCRonos-Vintage"
|
|
|
|
|
|
|
|
|
12 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
13 |
|
14 |
# Load pre-trained model and tokenizer
|
15 |
+
model = GPT2LMHeadModel.from_pretrained(model_name).to(device)
|
|
|
16 |
tokenizer = GPT2Tokenizer.from_pretrained(model_name)
|
17 |
|
18 |
# CSS for formatting
|
|
|
23 |
margin-right: 2em;
|
24 |
font-size: 1.2em;
|
25 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
.inserted {
|
27 |
background-color: #90EE90;
|
28 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
</style>
|
30 |
"""
|
31 |
|
|
|
32 |
def generate_html_diff(old_text, new_text):
|
33 |
d = difflib.Differ()
|
34 |
diff = list(d.compare(old_text.split(), new_text.split()))
|
|
|
37 |
if word.startswith(' '):
|
38 |
html_diff.append(word[2:])
|
39 |
elif word.startswith('+ '):
|
40 |
+
html_diff.append(f'<span class="inserted">{word[2:]}</span>')
|
41 |
return ' '.join(html_diff)
|
42 |
|
43 |
+
def split_text(text, max_tokens=400):
|
44 |
+
tokens = tokenizer.tokenize(text)
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
chunks = []
|
46 |
+
current_chunk = []
|
|
|
|
|
|
|
|
|
|
|
|
|
47 |
|
48 |
+
for token in tokens:
|
49 |
+
current_chunk.append(token)
|
50 |
+
if len(current_chunk) >= max_tokens:
|
51 |
+
chunks.append(tokenizer.convert_tokens_to_string(current_chunk))
|
52 |
+
current_chunk = []
|
|
|
|
|
|
|
53 |
|
54 |
if current_chunk:
|
55 |
+
chunks.append(tokenizer.convert_tokens_to_string(current_chunk))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
56 |
|
57 |
return chunks
|
58 |
|
|
|
|
|
59 |
def ocr_correction(prompt, max_new_tokens=600, num_threads=os.cpu_count()):
|
60 |
prompt = f"""### Text ###\n{prompt}\n\n\n### Correction ###\n"""
|
61 |
input_ids = tokenizer.encode(prompt, return_tensors="pt").to(device)
|
62 |
|
|
|
63 |
torch.set_num_threads(num_threads)
|
64 |
|
|
|
65 |
with ThreadPoolExecutor(max_workers=num_threads) as executor:
|
66 |
future = executor.submit(
|
67 |
model.generate,
|
|
|
74 |
)
|
75 |
output = future.result()
|
76 |
|
|
|
77 |
result = tokenizer.decode(output[0], skip_special_tokens=True)
|
78 |
+
return result.split("### Correction ###")[1].strip()
|
79 |
+
|
80 |
+
def process_text(user_message):
|
81 |
+
chunks = split_text(user_message)
|
82 |
+
corrected_chunks = []
|
83 |
+
|
84 |
+
for chunk in chunks:
|
85 |
+
corrected_chunk = ocr_correction(chunk)
|
86 |
+
corrected_chunks.append(corrected_chunk)
|
87 |
+
|
88 |
+
corrected_text = ' '.join(corrected_chunks)
|
89 |
+
html_diff = generate_html_diff(user_message, corrected_text)
|
90 |
|
91 |
+
ocr_result = f'<h2 style="text-align:center">OCR Correction</h2>\n<div class="generation">{html_diff}</div>'
|
92 |
+
final_output = f"{css}{ocr_result}"
|
93 |
+
return final_output
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
94 |
|
95 |
# Define the Gradio interface
|
96 |
with gr.Blocks(theme='JohnSmith9982/small_and_pretty') as demo:
|
|
|
98 |
text_input = gr.Textbox(label="Your (bad?) text", type="text", lines=5)
|
99 |
process_button = gr.Button("Process Text")
|
100 |
text_output = gr.HTML(label="Processed text")
|
101 |
+
process_button.click(process_text, inputs=text_input, outputs=[text_output])
|
102 |
|
103 |
if __name__ == "__main__":
|
104 |
demo.queue().launch()
|