File size: 8,601 Bytes
4ee6673
 
 
 
 
 
 
 
 
7a13891
 
 
4ee6673
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7a13891
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4ee6673
7a13891
 
4ee6673
7a13891
4ee6673
7a13891
4ee6673
 
 
7a13891
 
4ee6673
 
 
 
 
 
 
 
 
 
 
 
 
7a13891
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4ee6673
 
 
 
 
 
 
 
 
 
 
 
7a13891
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4ee6673
 
 
7a13891
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4ee6673
7a13891
 
 
 
 
 
 
 
 
 
4ee6673
7a13891
 
 
 
 
4ee6673
 
7a13891
4ee6673
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
# -*- coding: utf-8 -*-
"""DocAI_DeploymentGradio.ipynb

Automatically generated by Colaboratory.

Original file is located at
    https://colab.research.google.com/drive/1USSEj7nHh2n2hUhTJTC0Iwhj6mSR7-mD
"""

import os
os.system('pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cpu')

os.system('pip install pyyaml==5.1')

os.system('pip install -q git+https://github.com/huggingface/transformers.git')

os.system('pip install -q datasets seqeval')

os.system('pip install torch==1.8.0+cu101 torchvision==0.9.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html')
os.system('pip install -q detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu101/torch1.8/index.html')
os.system('pip install -q pytesseract')

!pip install gradio

!pip install -q git+https://github.com/huggingface/transformers.git

!pip install h5py

!pip install -q datasets seqeval

import gradio as gr

import numpy as np
import tensorflow as tf

import torch
import json

from datasets.features import ClassLabel
from transformers import AutoProcessor

from datasets import Features, Sequence, ClassLabel, Value, Array2D, Array3D
from datasets import load_dataset # this dataset uses the new Image feature :) 

from transformers import LayoutLMv3ForTokenClassification
from transformers import AutoModelForTokenClassification

#import cv2
from PIL import Image, ImageDraw, ImageFont

dataset = load_dataset("nielsr/funsd-layoutlmv3")

example = dataset["test"][0]

#image_path = "/root/.cache/huggingface/datasets/nielsr___funsd-layoutlmv3/funsd/1.0.0/0e3f4efdfd59aa1c3b4952c517894f7b1fc4d75c12ef01bcc8626a69e41c1bb9/funsd-layoutlmv3-test.arrow"

image_path = '/root/.cache/huggingface/datasets/nielsr___funsd-layoutlmv3/funsd/1.0.0/0e3f4efdfd59aa1c3b4952c517894f7b1fc4d75c12ef01bcc8626a69e41c1bb9'

example = dataset["test"][0]
example["image"].save("example1.png")

example1 = dataset["test"][1]
example1["image"].save("example2.png")

example2 = dataset["test"][2]
example2["image"].save("example3.png")

example2["image"]

#Image.open(dataset[2][image_path]).convert("RGB").save("example1.png")
#Image.open(dataset[1]["image_path"]).convert("RGB").save("example2.png")
#Image.open(dataset[0]["image_path"]).convert("RGB").save("example3.png")

words, boxes, ner_tags = example["tokens"], example["bboxes"], example["ner_tags"]

features = dataset["test"].features

column_names = dataset["test"].column_names
image_column_name = "image"
text_column_name = "tokens"
boxes_column_name = "bboxes"
label_column_name = "ner_tags"

def get_label_list(labels):
    unique_labels = set()
    for label in labels:
        unique_labels = unique_labels | set(label)
    label_list = list(unique_labels)
    label_list.sort()
    return label_list

if isinstance(features[label_column_name].feature, ClassLabel):
    label_list = features[label_column_name].feature.names
    # No need to convert the labels since they are already ints.
    id2label = {k: v for k,v in enumerate(label_list)}
    label2id = {v: k for k,v in enumerate(label_list)}
else:
    label_list = get_label_list(dataset["train"][label_column_name])
    id2label = {k: v for k,v in enumerate(label_list)}
    label2id = {v: k for k,v in enumerate(label_list)}
num_labels = len(label_list)

label2color = {'question':'blue', 'answer':'green', 'header':'orange', 'other':'violet'}

def prepare_examples(examples):
  images = examples[image_column_name]
  words = examples[text_column_name]
  boxes = examples[boxes_column_name]
  word_labels = examples[label_column_name]

  encoding = processor(images, words, boxes=boxes, word_labels=word_labels,
                       truncation=True, padding="max_length")

  return encoding

processor = AutoProcessor.from_pretrained("microsoft/layoutlmv3-base", apply_ocr=False)

model = LayoutLMv3ForTokenClassification.from_pretrained("microsoft/layoutlmv3-base",
                                                         id2label=id2label,
                                                         label2id=label2id)

# we need to define custom features for `set_format` (used later on) to work properly
features = Features({
    'pixel_values': Array3D(dtype="float32", shape=(3, 224, 224)),
    'input_ids': Sequence(feature=Value(dtype='int64')),
    'attention_mask': Sequence(Value(dtype='int64')),
    'bbox': Array2D(dtype="int64", shape=(512, 4)),
    'labels': Sequence(feature=Value(dtype='int64')),
})

eval_dataset = dataset["test"].map(
    prepare_examples,
    batched=True,
    remove_columns=column_names,
    features=features,
)

def unnormalize_box(bbox, width, height):
     return [
         width * (bbox[0] / 1000),
         height * (bbox[1] / 1000),
         width * (bbox[2] / 1000),
         height * (bbox[3] / 1000),
     ]

def process_image(image):

    print(type(image))
    width, height = image.size

    image = example["image"]
    words = example["tokens"]
    boxes = example["bboxes"]
    word_labels = example["ner_tags"]

    for k,v in encoding.items():
        print(k,v.shape)

    # encode
    #encoding = processor(image, truncation=True, return_offsets_mapping=True, return_tensors="pt")
    #offset_mapping = encoding.pop('offset_mapping')
    
    #encoding = processor(image, words, truncation=True,boxes=boxes, word_labels=word_labels,return_offsets_mapping=True, return_tensors="pt")
    #offset_mapping = encoding.pop('offset_mapping')

    encoding = processor(image, truncation=True,boxes=boxes, word_labels=word_labels,return_offsets_mapping=True, return_tensors="pt")
    offset_mapping = encoding.pop('offset_mapping')

    

    # forward pass
    with torch.no_grad():
        outputs = model(**encoding)

    # get predictions

    # We take the highest score for each token, using argmax. 
    # This serves as the predicted label for each token.
    logits = outputs.logits
    #logits.shape
    predictions = logits.argmax(-1).squeeze().tolist()

    labels = encoding.labels.squeeze().tolist()

    token_boxes = encoding.bbox.squeeze().tolist()
    width, height = image.size

    #true_predictions = [model.config.id2label[pred] for pred, label in zip(predictions, labels) if label != - 100]
    #true_labels = [model.config.id2label[label] for prediction, label in zip(predictions, labels) if label != -100]
    #true_boxes = [unnormalize_box(box, width, height) for box, label in zip(token_boxes, labels) if label != -100]
    
    
    # only keep non-subword predictions
    is_subword = np.array(offset_mapping.squeeze().tolist())[:,0] != 0
    true_predictions = [id2label[pred] for idx, pred in enumerate(predictions) if not is_subword[idx]]
    true_boxes = [unnormalize_box(box, width, height) for idx, box in enumerate(token_boxes) if not is_subword[idx]]

    # draw predictions over the image
    draw = ImageDraw.Draw(image)
    font = ImageFont.load_default()
    for prediction, box in zip(true_predictions, true_boxes):
        predicted_label = id2label(prediction)
        draw.rectangle(box, outline=label2color[predicted_label])
        draw.text((box[0]+10, box[1]-10), text=predicted_label, fill=label2color[predicted_label], font=font)
    
    return image

title = "DocumentAI - Extraction of Key Information using LayoutLMv3 model"
description = "Extraction of Form or Invoice Extraction - We use Microsoft's LayoutLMv3 trained on Invoice Dataset to predict the Biller Name, Biller Address, Biller post_code, Due_date, GST, Invoice_date, Invoice_number, Subtotal and Total. To use it, simply upload an image or use the example image below. Results will show up in a few seconds."

article="<b>References</b><br>[1] Y. Xu et al., “LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking.” 2022. <a href='https://arxiv.org/abs/2204.08387'>Paper Link</a><br>[2]  <a href='https://github.com/NielsRogge/Transformers-Tutorials/tree/master/LayoutLMv3'>LayoutLMv3 training and inference</a>" 

examples =[['example1.png'],['example2.png'],['example3.png']]

css = """.output_image, .input_image {height: 600px !important}"""

iface = gr.Interface(fn=process_image, 
                     inputs=gr.inputs.Image(type="pil"), 
                     outputs=gr.outputs.Image(type="pil", label="annotated predict image"),
                     title=title,
                     description=description,
                     article=article,
                     examples=examples,
                     css=css,
                     analytics_enabled = True, enable_queue=True
                     )

iface.launch(inline=False, share=False, debug=False)