File size: 6,393 Bytes
bc12901
 
 
 
2359223
 
 
ab36703
bc12901
 
2359223
bc6a638
bc12901
 
 
 
 
 
 
bc6a638
225fcc2
 
 
 
 
2359223
 
225fcc2
 
2359223
 
 
 
8171e8e
225fcc2
2359223
8171e8e
 
bc6a638
2359223
225fcc2
 
 
1af0b6d
 
 
 
 
 
 
 
fcfd908
1af0b6d
 
 
 
0b2b653
 
1af0b6d
 
 
fcfd908
 
 
 
 
 
 
 
bc12901
bc6a638
2359223
 
 
 
 
 
 
 
 
 
 
 
 
 
bc6a638
d229b67
2359223
 
 
 
 
 
 
 
bc6a638
87ad231
2359223
 
 
 
 
225fcc2
bc6a638
0b2b653
bc6a638
 
2359223
 
 
bc6a638
2359223
 
0b2b653
2359223
 
 
 
 
bc6a638
2359223
 
 
 
 
 
0b2b653
2359223
 
 
 
 
 
 
 
 
 
27d0a44
 
 
 
 
 
 
 
 
 
2359223
 
 
 
 
 
 
 
27d0a44
 
 
2359223
 
 
 
 
 
 
 
 
 
177edb5
 
 
 
27d0a44
2359223
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bc6a638
177edb5
 
 
 
 
 
2359223
 
 
 
 
bc6a638
177edb5
 
 
d229b67
2359223
 
 
 
bc6a638
2359223
 
177edb5
 
 
 
 
 
2359223
 
 
2b1c83d
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
229
230
231
import os

os.environ["TOKENIZERS_PARALLELISM"] = "false"

import functools
from PIL import Image, ImageDraw
import gradio as gr

import torch
from docquery.pipeline import get_pipeline
from docquery.document import load_bytes, load_document, ImageDocument


def ensure_list(x):
    if isinstance(x, list):
        return x
    else:
        return [x]


CHECKPOINTS = {
    "LayoutLMv1 🦉": "impira/layoutlm-document-qa",
    "Donut 🍩": "naver-clova-ix/donut-base-finetuned-docvqa",
}

PIPELINES = {}


def construct_pipeline(model):
    global PIPELINES
    if model in PIPELINES:
        return PIPELINES[model]

    device = "cuda" if torch.cuda.is_available() else "cpu"
    ret = get_pipeline(checkpoint=CHECKPOINTS[model], device=device)
    PIPELINES[model] = ret
    return ret


@functools.lru_cache(1024)
def run_pipeline(model, question, document, top_k):
    pipeline = construct_pipeline(model)
    return pipeline(question=question, **document.context, top_k=top_k)


# TODO: Move into docquery
# TODO: Support words past the first page (or window?)
def lift_word_boxes(document):
    return document.context["image"][0][1]


def expand_bbox(word_boxes):
    if len(word_boxes) == 0:
        return None

    min_x, min_y, max_x, max_y = zip(*[x[1] for x in word_boxes])
    min_x, min_y, max_x, max_y = [min(min_x), min(min_y), max(max_x), max(max_y)]
    return [min_x, min_y, max_x, max_y]


# LayoutLM boxes are normalized to 0, 1000
def normalize_bbox(box, width, height, padding=0.005):
    min_x, min_y, max_x, max_y = [c / 1000 for c in box]
    if padding != 0:
        min_x = max(0, min_x - padding)
        min_y = max(0, min_y - padding)
        max_x = min(max_x + padding, 1)
        max_y = min(max_y + padding, 1)
    return [min_x * width, min_y * height, max_x * width, max_y * height]


examples = [
    [
        "invoice.png",
        "What is the invoice number?",
    ],
    [
        "contract.jpeg",
        "What is the purchase amount?",
    ],
    [
        "statement.png",
        "What are net sales for 2020?",
    ],
]


def process_path(path):
    if path:
        try:
            document = load_document(path)
            return document, document.preview, None
        except Exception:
            pass
    return None, None, None


def process_upload(file):
    if file:
        return process_path(file.name)
    else:
        return None, None, None


colors = ["#64A087", "green", "black"]


def process_question(question, document, model=list(CHECKPOINTS.keys())[0]):
    if document is None:
        return None, None

    predictions = run_pipeline(model, question, document, 3)
    image = document.preview.copy()
    draw = ImageDraw.Draw(image, "RGBA")
    for i, p in enumerate(ensure_list(predictions)):
        if i > 0:
            # Keep the code around to produce multiple boxes, but only show the top
            # prediction for now
            break

        if "start" in p and "end" in p:
            x1, y1, x2, y2 = normalize_bbox(
                expand_bbox(lift_word_boxes(document)[p["start"] : p["end"] + 1]),
                image.width,
                image.height,
            )
            draw.rectangle(((x1, y1), (x2, y2)), fill=(0, 255, 0, int(0.4 * 255)))

    return image, predictions


def load_example_document(img, question, model):
    document = ImageDocument(Image.fromarray(img))
    preview, answer = process_question(question, document, model)
    return document, question, preview, answer


CSS = """
#short-upload-box .w-full {
    min-height: 10rem !important;
}
#question input {
    font-size: 16px;
}
"""

with gr.Blocks(css=CSS) as demo:
    gr.Markdown("# DocQuery: Query Documents w/ NLP")
    document = gr.Variable()
    example_question = gr.Textbox(visible=False)
    example_image = gr.Image(visible=False)

    gr.Markdown("## 1. Upload a file or select an example")
    with gr.Row(equal_height=True):
        with gr.Column():
            upload = gr.File(
                label="Upload a file", interactive=True, elem_id="short-upload-box"
            )
            url = gr.Textbox(label="... or a URL", interactive=True)
        gr.Examples(
            examples=examples,
            inputs=[example_image, example_question],
        )

    gr.Markdown("## 2. Ask a question")

    with gr.Row(equal_height=True):
        question = gr.Textbox(
            label="Question",
            placeholder="e.g. What is the invoice number?",
            lines=1,
            max_lines=1,
            elem_id="question",
        )
        model = gr.Radio(
            choices=list(CHECKPOINTS.keys()),
            value=list(CHECKPOINTS.keys())[0],
            label="Model",
        )

    with gr.Row():
        clear_button = gr.Button("Clear", variant="secondary")
        submit_button = gr.Button("Submit", variant="primary", elem_id="submit-button")

    with gr.Row():
        image = gr.Image(visible=True)
        with gr.Column():
            output = gr.JSON(label="Output")

    clear_button.click(
        lambda _: (None, None, None, None),
        inputs=clear_button,
        outputs=[image, document, question, output],
    )
    upload.change(fn=process_upload, inputs=[upload], outputs=[document, image, output])
    url.change(fn=process_path, inputs=[url], outputs=[document, image, output])

    question.submit(
        fn=process_question,
        inputs=[question, document, model],
        outputs=[image, output],
    )

    submit_button.click(
        process_question,
        inputs=[question, document, model],
        outputs=[image, output],
    )

    model.change(
        process_question, inputs=[question, document, model], outputs=[image, output]
    )

    example_image.change(
        fn=load_example_document,
        inputs=[example_image, example_question, model],
        outputs=[document, question, image, output],
    )

    gr.Markdown("### More Info")
    gr.Markdown(
        "DocQuery uses LayoutLMv1 fine-tuned on DocVQA, a document visual question"
        " answering dataset, as well as SQuAD, which boosts its English-language comprehension."
        " To use it, simply upload an image or PDF, type a question, and click 'submit', or "
        " click one of the examples to load them."
    )
    gr.Markdown("[Github Repo](https://github.com/impira/docquery)")

if __name__ == "__main__":
    demo.launch()