Spaces:
Sleeping
Sleeping
Update sroie_inference.py
Browse files- sroie_inference.py +114 -114
sroie_inference.py
CHANGED
@@ -1,114 +1,114 @@
|
|
1 |
-
import torch
|
2 |
-
import cv2
|
3 |
-
import numpy as np
|
4 |
-
from PIL import Image, ImageDraw, ImageFont
|
5 |
-
from transformers import LayoutLMv3TokenizerFast, LayoutLMv3Processor, LayoutLMv3ForTokenClassification
|
6 |
-
from utils import OCR, unnormalize_box
|
7 |
-
|
8 |
-
|
9 |
-
labels = ["O", "B-COMPANY", "I-COMPANY", "B-DATE", "I-DATE", "B-ADDRESS", "I-ADDRESS", "B-TOTAL", "I-TOTAL"]
|
10 |
-
id2label = {v: k for v, k in enumerate(labels)}
|
11 |
-
label2id = {k: v for v, k in enumerate(labels)}
|
12 |
-
|
13 |
-
tokenizer = LayoutLMv3TokenizerFast.from_pretrained("Theivaprakasham/layoutlmv3-finetuned-sroie", apply_ocr=False)
|
14 |
-
processor = LayoutLMv3Processor.from_pretrained("Theivaprakasham/layoutlmv3-finetuned-sroie", apply_ocr=False)
|
15 |
-
model = LayoutLMv3ForTokenClassification.from_pretrained("Theivaprakasham/layoutlmv3-finetuned-sroie")
|
16 |
-
|
17 |
-
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
18 |
-
model.to(device)
|
19 |
-
|
20 |
-
|
21 |
-
def blur(image, boxes):
|
22 |
-
image = np.array(image)
|
23 |
-
for box in boxes:
|
24 |
-
|
25 |
-
blur_x = int(box[0])
|
26 |
-
blur_y = int(box[1])
|
27 |
-
blur_width = int(box[2]-box[0])
|
28 |
-
blur_height = int(box[3]-box[1])
|
29 |
-
|
30 |
-
roi = image[blur_y:blur_y + blur_height, blur_x:blur_x + blur_width]
|
31 |
-
blur_image = cv2.GaussianBlur(roi, (201, 201), 0)
|
32 |
-
image[blur_y:blur_y + blur_height, blur_x:blur_x + blur_width] = blur_image
|
33 |
-
|
34 |
-
return Image.fromarray(image, 'RGB')
|
35 |
-
|
36 |
-
|
37 |
-
def prediction(image):
|
38 |
-
boxes, words = OCR(image)
|
39 |
-
encoding = processor(image, words, boxes=boxes, return_offsets_mapping=True, return_tensors="pt", truncation=True)
|
40 |
-
offset_mapping = encoding.pop('offset_mapping')
|
41 |
-
|
42 |
-
for k, v in encoding.items():
|
43 |
-
encoding[k] = v.to(device)
|
44 |
-
|
45 |
-
outputs = model(**encoding)
|
46 |
-
|
47 |
-
predictions = outputs.logits.argmax(-1).squeeze().tolist()
|
48 |
-
token_boxes = encoding.bbox.squeeze().tolist()
|
49 |
-
|
50 |
-
inp_ids = encoding.input_ids.squeeze().tolist()
|
51 |
-
inp_words = [tokenizer.decode(i) for i in inp_ids]
|
52 |
-
|
53 |
-
width, height = image.size
|
54 |
-
is_subword = np.array(offset_mapping.squeeze().tolist())[:, 0] != 0
|
55 |
-
|
56 |
-
true_predictions = [id2label[pred] for idx, pred in enumerate(predictions) if not is_subword[idx]]
|
57 |
-
true_boxes = [unnormalize_box(box, width, height) for idx, box in enumerate(token_boxes) if not is_subword[idx]]
|
58 |
-
true_words = []
|
59 |
-
|
60 |
-
for id, i in enumerate(inp_words):
|
61 |
-
if not is_subword[id]:
|
62 |
-
true_words.append(i)
|
63 |
-
else:
|
64 |
-
true_words[-1] = true_words[-1]+i
|
65 |
-
|
66 |
-
true_predictions = true_predictions[1:-1]
|
67 |
-
true_boxes = true_boxes[1:-1]
|
68 |
-
true_words = true_words[1:-1]
|
69 |
-
|
70 |
-
preds = []
|
71 |
-
l_words = []
|
72 |
-
bboxes = []
|
73 |
-
|
74 |
-
for i, j in enumerate(true_predictions):
|
75 |
-
if j != 'others':
|
76 |
-
preds.append(true_predictions[i])
|
77 |
-
l_words.append(true_words[i])
|
78 |
-
bboxes.append(true_boxes[i])
|
79 |
-
|
80 |
-
d = {}
|
81 |
-
for id, i in enumerate(preds):
|
82 |
-
if i not in d.keys():
|
83 |
-
d[i] = l_words[id]
|
84 |
-
else:
|
85 |
-
d[i] = d[i] + ", " + l_words[id]
|
86 |
-
|
87 |
-
d = {k: v.strip() for (k, v) in d.items()}
|
88 |
-
|
89 |
-
keys_to_pop = []
|
90 |
-
for k, v in d.items():
|
91 |
-
if k[:2] == "I-":
|
92 |
-
d["B-" + k[2:]] = d["B-" + k[2:]] + ", " + v
|
93 |
-
keys_to_pop.append(k)
|
94 |
-
|
95 |
-
if "O" in d: d.pop("O")
|
96 |
-
if "B-TOTAL" in d: d.pop("B-TOTAL")
|
97 |
-
for k in keys_to_pop: d.pop(k)
|
98 |
-
|
99 |
-
blur_boxes = []
|
100 |
-
for prediction, box in zip(preds, bboxes):
|
101 |
-
if prediction != 'O' and prediction[2:] != 'TOTAL':
|
102 |
-
blur_boxes.append(box)
|
103 |
-
|
104 |
-
image = (blur(image, blur_boxes))
|
105 |
-
|
106 |
-
draw = ImageDraw.Draw(image, "RGBA")
|
107 |
-
font = ImageFont.load_default()
|
108 |
-
|
109 |
-
for prediction, box in zip(preds, bboxes):
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
return d, image
|
114 |
-
|
|
|
1 |
+
import torch
|
2 |
+
import cv2
|
3 |
+
import numpy as np
|
4 |
+
from PIL import Image, ImageDraw, ImageFont
|
5 |
+
from transformers import LayoutLMv3TokenizerFast, LayoutLMv3Processor, LayoutLMv3ForTokenClassification
|
6 |
+
from utils import OCR, unnormalize_box
|
7 |
+
|
8 |
+
|
9 |
+
labels = ["O", "B-COMPANY", "I-COMPANY", "B-DATE", "I-DATE", "B-ADDRESS", "I-ADDRESS", "B-TOTAL", "I-TOTAL"]
|
10 |
+
id2label = {v: k for v, k in enumerate(labels)}
|
11 |
+
label2id = {k: v for v, k in enumerate(labels)}
|
12 |
+
|
13 |
+
tokenizer = LayoutLMv3TokenizerFast.from_pretrained("Theivaprakasham/layoutlmv3-finetuned-sroie", apply_ocr=False)
|
14 |
+
processor = LayoutLMv3Processor.from_pretrained("Theivaprakasham/layoutlmv3-finetuned-sroie", apply_ocr=False)
|
15 |
+
model = LayoutLMv3ForTokenClassification.from_pretrained("Theivaprakasham/layoutlmv3-finetuned-sroie")
|
16 |
+
|
17 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
18 |
+
model.to(device)
|
19 |
+
|
20 |
+
|
21 |
+
def blur(image, boxes):
|
22 |
+
image = np.array(image)
|
23 |
+
for box in boxes:
|
24 |
+
|
25 |
+
blur_x = int(box[0])
|
26 |
+
blur_y = int(box[1])
|
27 |
+
blur_width = int(box[2]-box[0])
|
28 |
+
blur_height = int(box[3]-box[1])
|
29 |
+
|
30 |
+
roi = image[blur_y:blur_y + blur_height, blur_x:blur_x + blur_width]
|
31 |
+
blur_image = cv2.GaussianBlur(roi, (201, 201), 0)
|
32 |
+
image[blur_y:blur_y + blur_height, blur_x:blur_x + blur_width] = blur_image
|
33 |
+
|
34 |
+
return Image.fromarray(image, 'RGB')
|
35 |
+
|
36 |
+
|
37 |
+
def prediction(image):
|
38 |
+
boxes, words = OCR(image)
|
39 |
+
encoding = processor(image, words, boxes=boxes, return_offsets_mapping=True, return_tensors="pt", truncation=True)
|
40 |
+
offset_mapping = encoding.pop('offset_mapping')
|
41 |
+
|
42 |
+
for k, v in encoding.items():
|
43 |
+
encoding[k] = v.to(device)
|
44 |
+
|
45 |
+
outputs = model(**encoding)
|
46 |
+
|
47 |
+
predictions = outputs.logits.argmax(-1).squeeze().tolist()
|
48 |
+
token_boxes = encoding.bbox.squeeze().tolist()
|
49 |
+
|
50 |
+
inp_ids = encoding.input_ids.squeeze().tolist()
|
51 |
+
inp_words = [tokenizer.decode(i) for i in inp_ids]
|
52 |
+
|
53 |
+
width, height = image.size
|
54 |
+
is_subword = np.array(offset_mapping.squeeze().tolist())[:, 0] != 0
|
55 |
+
|
56 |
+
true_predictions = [id2label[pred] for idx, pred in enumerate(predictions) if not is_subword[idx]]
|
57 |
+
true_boxes = [unnormalize_box(box, width, height) for idx, box in enumerate(token_boxes) if not is_subword[idx]]
|
58 |
+
true_words = []
|
59 |
+
|
60 |
+
for id, i in enumerate(inp_words):
|
61 |
+
if not is_subword[id]:
|
62 |
+
true_words.append(i)
|
63 |
+
else:
|
64 |
+
true_words[-1] = true_words[-1]+i
|
65 |
+
|
66 |
+
true_predictions = true_predictions[1:-1]
|
67 |
+
true_boxes = true_boxes[1:-1]
|
68 |
+
true_words = true_words[1:-1]
|
69 |
+
|
70 |
+
preds = []
|
71 |
+
l_words = []
|
72 |
+
bboxes = []
|
73 |
+
|
74 |
+
for i, j in enumerate(true_predictions):
|
75 |
+
if j != 'others':
|
76 |
+
preds.append(true_predictions[i])
|
77 |
+
l_words.append(true_words[i])
|
78 |
+
bboxes.append(true_boxes[i])
|
79 |
+
|
80 |
+
d = {}
|
81 |
+
for id, i in enumerate(preds):
|
82 |
+
if i not in d.keys():
|
83 |
+
d[i] = l_words[id]
|
84 |
+
else:
|
85 |
+
d[i] = d[i] + ", " + l_words[id]
|
86 |
+
|
87 |
+
d = {k: v.strip() for (k, v) in d.items()}
|
88 |
+
|
89 |
+
keys_to_pop = []
|
90 |
+
for k, v in d.items():
|
91 |
+
if k[:2] == "I-":
|
92 |
+
d["B-" + k[2:]] = d["B-" + k[2:]] + ", " + v
|
93 |
+
keys_to_pop.append(k)
|
94 |
+
|
95 |
+
if "O" in d: d.pop("O")
|
96 |
+
if "B-TOTAL" in d: d.pop("B-TOTAL")
|
97 |
+
for k in keys_to_pop: d.pop(k)
|
98 |
+
|
99 |
+
blur_boxes = []
|
100 |
+
for prediction, box in zip(preds, bboxes):
|
101 |
+
if prediction != 'O' and prediction[2:] != 'TOTAL':
|
102 |
+
blur_boxes.append(box)
|
103 |
+
|
104 |
+
image = (blur(image, blur_boxes))
|
105 |
+
|
106 |
+
#draw = ImageDraw.Draw(image, "RGBA")
|
107 |
+
#font = ImageFont.load_default()
|
108 |
+
|
109 |
+
#for prediction, box in zip(preds, bboxes):
|
110 |
+
# draw.rectangle(box)
|
111 |
+
# draw.text((box[0]+10, box[1]-10), text=prediction, font=font, fill="black", font_size="8")
|
112 |
+
|
113 |
+
return d, image
|
114 |
+
|