Spaces:
Running
Running
Synced repo using 'sync_with_huggingface' Github Action
Browse files- gradio_app.py +51 -81
gradio_app.py
CHANGED
@@ -9,59 +9,40 @@ if "APP_PATH" in os.environ:
|
|
9 |
if app_path not in sys.path:
|
10 |
sys.path.append(app_path)
|
11 |
|
12 |
-
import
|
13 |
-
|
14 |
from typing import List
|
15 |
|
16 |
import pypdfium2
|
17 |
-
|
18 |
-
|
19 |
-
from surya.detection import batch_text_detection
|
20 |
-
from surya.input.pdflines import get_page_text_lines, get_table_blocks
|
21 |
-
from surya.layout import batch_layout_detection
|
22 |
-
from surya.model.detection.model import load_model, load_processor
|
23 |
-
from surya.model.layout.model import load_model as load_layout_model
|
24 |
-
from surya.model.layout.processor import load_processor as load_layout_processor
|
25 |
-
from surya.model.recognition.model import load_model as load_rec_model
|
26 |
-
from surya.model.recognition.processor import load_processor as load_rec_processor
|
27 |
-
from surya.model.table_rec.model import load_model as load_table_model
|
28 |
-
from surya.model.table_rec.processor import load_processor as load_table_processor
|
29 |
-
from surya.model.ocr_error.model import load_model as load_ocr_error_model, load_tokenizer as load_ocr_error_processor
|
30 |
-
from surya.postprocessing.heatmap import draw_polys_on_image, draw_bboxes_on_image
|
31 |
-
from surya.ocr import run_ocr
|
32 |
-
from surya.postprocessing.text import draw_text_on_image
|
33 |
-
from PIL import Image
|
34 |
-
from surya.languages import CODE_TO_LANGUAGE
|
35 |
-
from surya.input.langs import replace_lang_with_code
|
36 |
-
from surya.schema import OCRResult, TextDetectionResult, LayoutResult, TableResult
|
37 |
-
from surya.settings import settings
|
38 |
-
from surya.tables import batch_table_recognition
|
39 |
-
from surya.postprocessing.util import rescale_bbox
|
40 |
-
from pdftext.extraction import plain_text_output
|
41 |
-
from surya.ocr_error import batch_ocr_error_detection
|
42 |
-
|
43 |
|
44 |
-
|
45 |
-
return load_model(), load_processor()
|
46 |
|
47 |
-
|
48 |
-
return load_rec_model(), load_rec_processor()
|
49 |
|
50 |
-
|
51 |
-
return load_layout_model(), load_layout_processor()
|
52 |
|
53 |
-
|
54 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
55 |
|
56 |
-
def load_ocr_error_cached():
|
57 |
-
return load_ocr_error_model(), load_ocr_error_processor()
|
58 |
|
59 |
-
#
|
60 |
def run_ocr_errors(pdf_file, page_count, sample_len=512, max_samples=10, max_pages=15):
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
|
|
|
|
|
|
|
|
|
|
65 |
|
66 |
sample_gap = len(text) // max_samples
|
67 |
if len(text) == 0 or sample_gap == 0:
|
@@ -75,29 +56,29 @@ def run_ocr_errors(pdf_file, page_count, sample_len=512, max_samples=10, max_pag
|
|
75 |
for i in range(0, len(text), sample_gap):
|
76 |
samples.append(text[i:i + sample_len])
|
77 |
|
78 |
-
results =
|
79 |
label = "This PDF has good text."
|
80 |
if results.labels.count("bad") / len(results.labels) > .2:
|
81 |
label = "This PDF may have garbled or bad OCR text."
|
82 |
return label, results.labels
|
83 |
|
84 |
-
#
|
85 |
def text_detection(img) -> (Image.Image, TextDetectionResult):
|
86 |
-
pred =
|
87 |
polygons = [p.polygon for p in pred.bboxes]
|
88 |
det_img = draw_polys_on_image(polygons, img.copy())
|
89 |
return det_img, pred
|
90 |
|
91 |
-
#
|
92 |
def layout_detection(img) -> (Image.Image, LayoutResult):
|
93 |
-
pred =
|
94 |
polygons = [p.polygon for p in pred.bboxes]
|
95 |
labels = [f"{p.label}-{p.position}" for p in pred.bboxes]
|
96 |
layout_img = draw_polys_on_image(polygons, img.copy(), labels=labels, label_font_size=18)
|
97 |
return layout_img, pred
|
98 |
|
99 |
-
#
|
100 |
-
def table_recognition(img, highres_img,
|
101 |
if skip_table_detection:
|
102 |
layout_tables = [(0, 0, highres_img.size[0], highres_img.size[1])]
|
103 |
table_imgs = [highres_img]
|
@@ -108,23 +89,14 @@ def table_recognition(img, highres_img, filepath, page_idx: int, use_pdf_boxes:
|
|
108 |
layout_tables = []
|
109 |
for tb in layout_tables_lowres:
|
110 |
highres_bbox = rescale_bbox(tb, img.size, highres_img.size)
|
|
|
|
|
111 |
table_imgs.append(
|
112 |
highres_img.crop(highres_bbox)
|
113 |
)
|
114 |
layout_tables.append(highres_bbox)
|
115 |
|
116 |
-
|
117 |
-
page_text = get_page_text_lines(filepath, [page_idx], [highres_img.size])[0]
|
118 |
-
table_bboxes = get_table_blocks(layout_tables, page_text, highres_img.size)
|
119 |
-
except PdfiumError:
|
120 |
-
# This happens when we try to get text from an image
|
121 |
-
table_bboxes = [[] for _ in layout_tables]
|
122 |
-
|
123 |
-
if not use_pdf_boxes or any(len(tb) == 0 for tb in table_bboxes):
|
124 |
-
det_results = batch_text_detection(table_imgs, det_model, det_processor)
|
125 |
-
table_bboxes = [[{"bbox": tb.bbox, "text": None} for tb in det_result.bboxes] for det_result in det_results]
|
126 |
-
|
127 |
-
table_preds = batch_table_recognition(table_imgs, table_bboxes, table_model, table_processor)
|
128 |
table_img = img.copy()
|
129 |
|
130 |
for results, table_bbox in zip(table_preds, layout_tables):
|
@@ -132,7 +104,7 @@ def table_recognition(img, highres_img, filepath, page_idx: int, use_pdf_boxes:
|
|
132 |
labels = []
|
133 |
colors = []
|
134 |
|
135 |
-
for item in results.
|
136 |
adjusted_bboxes.append([
|
137 |
(item.bbox[0] + table_bbox[0]),
|
138 |
(item.bbox[1] + table_bbox[1]),
|
@@ -140,31 +112,33 @@ def table_recognition(img, highres_img, filepath, page_idx: int, use_pdf_boxes:
|
|
140 |
(item.bbox[3] + table_bbox[1])
|
141 |
])
|
142 |
labels.append(item.label)
|
143 |
-
if
|
144 |
colors.append("blue")
|
145 |
else:
|
146 |
colors.append("red")
|
147 |
table_img = draw_bboxes_on_image(adjusted_bboxes, highres_img, labels=labels, label_font_size=18, color=colors)
|
148 |
return table_img, table_preds
|
149 |
|
150 |
-
#
|
151 |
def ocr(img, highres_img, langs: List[str]) -> (Image.Image, OCRResult):
|
152 |
replace_lang_with_code(langs)
|
153 |
-
img_pred =
|
154 |
|
155 |
bboxes = [l.bbox for l in img_pred.text_lines]
|
156 |
text = [l.text for l in img_pred.text_lines]
|
157 |
-
rec_img = draw_text_on_image(bboxes, text, img.size, langs
|
158 |
return rec_img, img_pred
|
159 |
|
160 |
def open_pdf(pdf_file):
|
161 |
return pypdfium2.PdfDocument(pdf_file)
|
162 |
|
163 |
-
def
|
164 |
doc = open_pdf(pdf_file)
|
165 |
-
|
|
|
|
|
166 |
|
167 |
-
def get_page_image(pdf_file, page_num, dpi=
|
168 |
doc = open_pdf(pdf_file)
|
169 |
renderer = doc.render(
|
170 |
pypdfium2.PdfBitmap.to_pil,
|
@@ -173,18 +147,14 @@ def get_page_image(pdf_file, page_num, dpi=96):
|
|
173 |
)
|
174 |
png = list(renderer)[0]
|
175 |
png_image = png.convert("RGB")
|
|
|
176 |
return png_image
|
177 |
|
178 |
def get_uploaded_image(in_file):
|
179 |
return Image.open(in_file).convert("RGB")
|
180 |
|
181 |
# Load models if not already loaded in reload mode
|
182 |
-
|
183 |
-
det_model, det_processor = load_det_cached()
|
184 |
-
rec_model, rec_processor = load_rec_cached()
|
185 |
-
layout_model, layout_processor = load_layout_cached()
|
186 |
-
table_model, table_processor = load_table_cached()
|
187 |
-
ocr_error_model, ocr_error_processor = load_ocr_error_cached()
|
188 |
|
189 |
with gr.Blocks(title="Surya") as demo:
|
190 |
gr.Markdown("""
|
@@ -224,8 +194,8 @@ with gr.Blocks(title="Surya") as demo:
|
|
224 |
|
225 |
def show_image(file, num=1):
|
226 |
if file.endswith('.pdf'):
|
227 |
-
count =
|
228 |
-
img = get_page_image(file, num)
|
229 |
return [
|
230 |
gr.update(visible=True, maximum=count),
|
231 |
gr.update(value=img)]
|
@@ -283,7 +253,7 @@ with gr.Blocks(title="Surya") as demo:
|
|
283 |
pil_image_highres = get_page_image(in_file, page_number, dpi=settings.IMAGE_DPI_HIGHRES)
|
284 |
else:
|
285 |
pil_image_highres = pil_image
|
286 |
-
table_img, pred = table_recognition(pil_image, pil_image_highres,
|
287 |
return table_img, [p.model_dump() for p in pred]
|
288 |
table_rec_btn.click(
|
289 |
fn=table_rec_img,
|
@@ -293,10 +263,10 @@ with gr.Blocks(title="Surya") as demo:
|
|
293 |
# Run bad PDF text detection
|
294 |
def ocr_errors_pdf(file, page_count, sample_len=512, max_samples=10, max_pages=15):
|
295 |
if file.endswith('.pdf'):
|
296 |
-
count =
|
297 |
else:
|
298 |
raise gr.Error("This feature only works with PDFs.", duration=5)
|
299 |
-
label, results = run_ocr_errors(file, count)
|
300 |
return gr.update(label="Result json:" + label, value=results)
|
301 |
ocr_errors_btn.click(
|
302 |
fn=ocr_errors_pdf,
|
|
|
9 |
if app_path not in sys.path:
|
10 |
sys.path.append(app_path)
|
11 |
|
12 |
+
import io
|
13 |
+
import tempfile
|
14 |
from typing import List
|
15 |
|
16 |
import pypdfium2
|
17 |
+
import gradio as gr
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
18 |
|
19 |
+
from surya.models import load_predictors
|
|
|
20 |
|
21 |
+
from surya.debug.draw import draw_polys_on_image, draw_bboxes_on_image
|
|
|
22 |
|
23 |
+
from surya.debug.text import draw_text_on_image
|
|
|
24 |
|
25 |
+
from PIL import Image
|
26 |
+
from surya.recognition.languages import CODE_TO_LANGUAGE, replace_lang_with_code
|
27 |
+
from surya.table_rec import TableResult
|
28 |
+
from surya.detection import TextDetectionResult
|
29 |
+
from surya.recognition import OCRResult
|
30 |
+
from surya.layout import LayoutResult
|
31 |
+
from surya.settings import settings
|
32 |
+
from surya.common.util import rescale_bbox, expand_bbox
|
33 |
|
|
|
|
|
34 |
|
35 |
+
# just copy from streamlit_app.py
|
36 |
def run_ocr_errors(pdf_file, page_count, sample_len=512, max_samples=10, max_pages=15):
|
37 |
+
from pdftext.extraction import plain_text_output
|
38 |
+
with tempfile.NamedTemporaryFile(suffix=".pdf") as f:
|
39 |
+
f.write(pdf_file.getvalue())
|
40 |
+
f.seek(0)
|
41 |
+
|
42 |
+
# Sample the text from the middle of the PDF
|
43 |
+
page_middle = page_count // 2
|
44 |
+
page_range = range(max(page_middle - max_pages, 0), min(page_middle + max_pages, page_count))
|
45 |
+
text = plain_text_output(f.name, page_range=page_range)
|
46 |
|
47 |
sample_gap = len(text) // max_samples
|
48 |
if len(text) == 0 or sample_gap == 0:
|
|
|
56 |
for i in range(0, len(text), sample_gap):
|
57 |
samples.append(text[i:i + sample_len])
|
58 |
|
59 |
+
results = predictors["ocr_error"](samples)
|
60 |
label = "This PDF has good text."
|
61 |
if results.labels.count("bad") / len(results.labels) > .2:
|
62 |
label = "This PDF may have garbled or bad OCR text."
|
63 |
return label, results.labels
|
64 |
|
65 |
+
# just copy from streamlit_app.py
|
66 |
def text_detection(img) -> (Image.Image, TextDetectionResult):
|
67 |
+
pred = predictors["detection"]([img])[0]
|
68 |
polygons = [p.polygon for p in pred.bboxes]
|
69 |
det_img = draw_polys_on_image(polygons, img.copy())
|
70 |
return det_img, pred
|
71 |
|
72 |
+
# just copy from streamlit_app.py
|
73 |
def layout_detection(img) -> (Image.Image, LayoutResult):
|
74 |
+
pred = predictors["layout"]([img])[0]
|
75 |
polygons = [p.polygon for p in pred.bboxes]
|
76 |
labels = [f"{p.label}-{p.position}" for p in pred.bboxes]
|
77 |
layout_img = draw_polys_on_image(polygons, img.copy(), labels=labels, label_font_size=18)
|
78 |
return layout_img, pred
|
79 |
|
80 |
+
# just copy from streamlit_app.py
|
81 |
+
def table_recognition(img, highres_img, skip_table_detection: bool) -> (Image.Image, List[TableResult]):
|
82 |
if skip_table_detection:
|
83 |
layout_tables = [(0, 0, highres_img.size[0], highres_img.size[1])]
|
84 |
table_imgs = [highres_img]
|
|
|
89 |
layout_tables = []
|
90 |
for tb in layout_tables_lowres:
|
91 |
highres_bbox = rescale_bbox(tb, img.size, highres_img.size)
|
92 |
+
# Slightly expand the box
|
93 |
+
highres_bbox = expand_bbox(highres_bbox)
|
94 |
table_imgs.append(
|
95 |
highres_img.crop(highres_bbox)
|
96 |
)
|
97 |
layout_tables.append(highres_bbox)
|
98 |
|
99 |
+
table_preds = predictors["table_rec"](table_imgs)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
100 |
table_img = img.copy()
|
101 |
|
102 |
for results, table_bbox in zip(table_preds, layout_tables):
|
|
|
104 |
labels = []
|
105 |
colors = []
|
106 |
|
107 |
+
for item in results.cells:
|
108 |
adjusted_bboxes.append([
|
109 |
(item.bbox[0] + table_bbox[0]),
|
110 |
(item.bbox[1] + table_bbox[1]),
|
|
|
112 |
(item.bbox[3] + table_bbox[1])
|
113 |
])
|
114 |
labels.append(item.label)
|
115 |
+
if "Row" in item.label:
|
116 |
colors.append("blue")
|
117 |
else:
|
118 |
colors.append("red")
|
119 |
table_img = draw_bboxes_on_image(adjusted_bboxes, highres_img, labels=labels, label_font_size=18, color=colors)
|
120 |
return table_img, table_preds
|
121 |
|
122 |
+
# just copy from streamlit_app.py
|
123 |
def ocr(img, highres_img, langs: List[str]) -> (Image.Image, OCRResult):
|
124 |
replace_lang_with_code(langs)
|
125 |
+
img_pred = predictors["recognition"]([img], [langs], predictors["detection"], highres_images=[highres_img])[0]
|
126 |
|
127 |
bboxes = [l.bbox for l in img_pred.text_lines]
|
128 |
text = [l.text for l in img_pred.text_lines]
|
129 |
+
rec_img = draw_text_on_image(bboxes, text, img.size, langs)
|
130 |
return rec_img, img_pred
|
131 |
|
132 |
def open_pdf(pdf_file):
|
133 |
return pypdfium2.PdfDocument(pdf_file)
|
134 |
|
135 |
+
def page_counter(pdf_file):
|
136 |
doc = open_pdf(pdf_file)
|
137 |
+
doc_len = len(doc)
|
138 |
+
doc.close()
|
139 |
+
return doc_len
|
140 |
|
141 |
+
def get_page_image(pdf_file, page_num, dpi=settings.IMAGE_DPI):
|
142 |
doc = open_pdf(pdf_file)
|
143 |
renderer = doc.render(
|
144 |
pypdfium2.PdfBitmap.to_pil,
|
|
|
147 |
)
|
148 |
png = list(renderer)[0]
|
149 |
png_image = png.convert("RGB")
|
150 |
+
doc.close()
|
151 |
return png_image
|
152 |
|
153 |
def get_uploaded_image(in_file):
|
154 |
return Image.open(in_file).convert("RGB")
|
155 |
|
156 |
# Load models if not already loaded in reload mode
|
157 |
+
predictors = load_predictors()
|
|
|
|
|
|
|
|
|
|
|
158 |
|
159 |
with gr.Blocks(title="Surya") as demo:
|
160 |
gr.Markdown("""
|
|
|
194 |
|
195 |
def show_image(file, num=1):
|
196 |
if file.endswith('.pdf'):
|
197 |
+
count = page_counter(file)
|
198 |
+
img = get_page_image(file, num, settings.IMAGE_DPI)
|
199 |
return [
|
200 |
gr.update(visible=True, maximum=count),
|
201 |
gr.update(value=img)]
|
|
|
253 |
pil_image_highres = get_page_image(in_file, page_number, dpi=settings.IMAGE_DPI_HIGHRES)
|
254 |
else:
|
255 |
pil_image_highres = pil_image
|
256 |
+
table_img, pred = table_recognition(pil_image, pil_image_highres, skip_table_detection)
|
257 |
return table_img, [p.model_dump() for p in pred]
|
258 |
table_rec_btn.click(
|
259 |
fn=table_rec_img,
|
|
|
263 |
# Run bad PDF text detection
|
264 |
def ocr_errors_pdf(file, page_count, sample_len=512, max_samples=10, max_pages=15):
|
265 |
if file.endswith('.pdf'):
|
266 |
+
count = page_counter(file)
|
267 |
else:
|
268 |
raise gr.Error("This feature only works with PDFs.", duration=5)
|
269 |
+
label, results = run_ocr_errors(io.BytesIO(open(file.name, "rb").read()), count)
|
270 |
return gr.update(label="Result json:" + label, value=results)
|
271 |
ocr_errors_btn.click(
|
272 |
fn=ocr_errors_pdf,
|