File size: 10,522 Bytes
bc12901
 
 
 
2359223
 
 
ab36703
bc12901
 
2359223
bc6a638
bc12901
 
 
 
 
 
 
bc6a638
225fcc2
 
 
 
 
2359223
 
225fcc2
 
2359223
 
 
 
8171e8e
225fcc2
2359223
8171e8e
 
bc6a638
225fcc2
 
 
1af0b6d
 
 
 
87500f1
 
1af0b6d
 
fcfd908
1af0b6d
 
 
 
0b2b653
 
1af0b6d
 
 
fcfd908
 
 
 
 
 
 
 
bc12901
bc6a638
2359223
 
 
 
 
 
 
 
 
 
 
 
 
 
bc6a638
d229b67
2359223
 
 
 
15fad86
 
 
d207d63
15fad86
194858a
15fad86
2359223
 
15fad86
 
 
d207d63
15fad86
194858a
15fad86
bc6a638
87ad231
2359223
 
 
 
15fad86
 
 
d207d63
15fad86
194858a
15fad86
225fcc2
bc6a638
0b2b653
bc6a638
 
2359223
 
194858a
bc6a638
d1e1ea7
2359223
99d94a6
2359223
d1e1ea7
 
 
2359223
 
 
bc6a638
87500f1
 
 
2359223
87500f1
 
 
2359223
 
 
0b2b653
2359223
194858a
 
 
 
 
d1e1ea7
194858a
15fad86
2359223
 
 
d207d63
 
194858a
 
d207d63
 
2359223
 
27d0a44
d207d63
 
 
 
 
 
27d0a44
 
 
d207d63
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
37a2f41
d207d63
 
 
 
 
 
 
 
 
 
27d0a44
37a2f41
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
27d0a44
 
 
194858a
d207d63
 
 
 
 
 
 
 
2359223
 
 
 
 
 
d207d63
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
27d0a44
37a2f41
d207d63
 
 
 
27d0a44
2359223
d207d63
 
 
 
 
 
 
 
 
 
 
 
 
2359223
d207d63
 
 
 
 
 
194858a
d207d63
2359223
d207d63
 
 
 
 
194858a
d207d63
 
60bdd81
 
d207d63
 
194858a
 
 
 
 
 
 
 
 
 
d207d63
2359223
15fad86
 
 
 
 
194858a
d207d63
60bdd81
 
15fad86
2359223
194858a
 
 
 
 
 
 
 
 
 
2359223
15fad86
6d0b7db
 
 
194858a
6d0b7db
66c9f11
6d0b7db
 
194858a
6d0b7db
bc6a638
177edb5
 
 
194858a
177edb5
 
2359223
 
 
194858a
2359223
bc6a638
177edb5
194858a
 
 
177edb5
d229b67
2359223
 
 
194858a
177edb5
2359223
 
6d0b7db
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
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
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


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, page):
    return document.context["image"][page][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,
                gr.update(visible=True, value=document.preview),
                gr.update(visible=True),
                gr.update(visible=False, value=None),
                gr.update(visible=False, value=None),
            )
        except Exception:
            pass
    return (
        None,
        gr.update(visible=False, value=None),
        gr.update(visible=False),
        gr.update(visible=False, value=None),
        gr.update(visible=False, value=None),
    )


def process_upload(file):
    if file:
        return process_path(file.name)
    else:
        return (
            None,
            gr.update(visible=False, value=None),
            gr.update(visible=False),
            gr.update(visible=False, value=None),
            gr.update(visible=False, value=None),
        )


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


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

    text_value = None
    predictions = run_pipeline(model, question, document, 3)
    pages = [x.copy().convert("RGB") for x in document.preview]
    for i, p in enumerate(ensure_list(predictions)):
        if i == 0:
            text_value = p["answer"]
        else:
            # 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:
            image = pages[p["page"]]
            draw = ImageDraw.Draw(image, "RGBA")
            x1, y1, x2, y2 = normalize_bbox(
                expand_bbox(
                    lift_word_boxes(document, p["page"])[p["start"] : p["end"] + 1]
                ),
                image.width,
                image.height,
            )
            draw.rectangle(((x1, y1), (x2, y2)), fill=(0, 255, 0, int(0.4 * 255)))

    return (
        gr.update(visible=True, value=pages),
        gr.update(visible=True, value=predictions),
        gr.update(
            visible=True,
            value=text_value,
        ),
    )


def load_example_document(img, question, model):
    if img is not None:
        document = ImageDocument(Image.fromarray(img))
        preview, answer, answer_text = process_question(question, document, model)
        return document, question, preview, gr.update(visible=True), answer, answer_text
    else:
        return None, None, None, gr.update(visible=False), None


CSS = """
#question input {
    font-size: 16px;
}
#url-textbox {
    padding: 0 !important;
}
#short-upload-box .w-full {
    min-height: 10rem !important;
}
/* I think something like this can be used to re-shape
 * the table
 */
/*
.gr-samples-table tr {
    display: inline;
}
.gr-samples-table .p-2 {
    width: 100px;
}
*/
#select-a-file {
    width: 100%;
}
#file-clear {
    padding-top: 2px !important;
    padding-bottom: 2px !important;
    padding-left: 8px !important;
    padding-right: 8px !important;
}
.gradio-container .gr-button-primary {
    background: linear-gradient(180deg, #CDF9BE 0%, #AFF497 100%);
    border: 1px solid #B0DCCC;
    border-radius: 8px;
    color: #1B8700;
}
.gradio-container.dark button#submit-button {
    background: linear-gradient(180deg, #CDF9BE 0%, #AFF497 100%);
    border: 1px solid #B0DCCC;
    border-radius: 8px;
    color: #1B8700
}

table.gr-samples-table tr td {
    border: none;
    outline: none;
}

table.gr-samples-table tr td:first-of-type {
    width: 0%;
}

div#short-upload-box div.absolute {
    display: none !important;
}

gradio-app > div > div > div > div.w-full > div, .gradio-app > div > div > div > div.w-full > div {
    gap: 0px 2%;
}

gradio-app div div div div.w-full, .gradio-app div div div div.w-full {
    gap: 0px;
}

gradio-app h2, .gradio-app h2 {
    padding-top: 10px;
}
"""

with gr.Blocks(css=CSS) as demo:
    gr.Markdown("# DocQuery: Document Query Engine")
    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."
        " [Github Repo](https://github.com/impira/docquery)"
    )

    document = gr.Variable()
    example_question = gr.Textbox(visible=False)
    example_image = gr.Image(visible=False)

    with gr.Row(equal_height=True):
        with gr.Column():
            with gr.Row():
                gr.Markdown("## 1. Select a file", elem_id="select-a-file")
                img_clear_button = gr.Button(
                    "Clear", variant="secondary", elem_id="file-clear", visible=False
                )
            image = gr.Gallery(visible=False)
            with gr.Row(equal_height=True):
                url = gr.Textbox(
                    show_label=False,
                    placeholder="URL",
                    lines=1,
                    max_lines=1,
                    elem_id="url-textbox",
                )
                submit = gr.Button("Get")
            gr.Markdown("— or —")
            upload = gr.File(
                label=None, interactive=True, elem_id="short-upload-box"
            )
            gr.Examples(
                examples=examples,
                inputs=[example_image, example_question],
            )

        with gr.Column() as col:
            gr.Markdown("## 2. Ask a question")
            question = gr.Textbox(
                label="Question",
                placeholder="e.g. What is the invoice number?",
                lines=1,
                max_lines=1,
            )
            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.Column():
                output_text = gr.Textbox(label="Top Answer", visible=False)
                output = gr.JSON(label="Output", visible=False)

    img_clear_button.click(
        lambda _: (
            gr.update(visible=False, value=None),
            None,
            gr.update(visible=False, value=None),
            gr.update(visible=False, value=None),
            gr.update(visible=False),
            None,
            None,
            None,
        ),
        inputs=img_clear_button,
        outputs=[
            image,
            document,
            output,
            output_text,
            img_clear_button,
            example_image,
            upload,
            url,
        ],
    )
    clear_button.click(
        lambda _: (
            gr.update(visible=False, value=None),
            None,
            None,
            gr.update(visible=False, value=None),
            gr.update(visible=False, value=None),
            None,
            None,
            None,
        ),
        inputs=clear_button,
        outputs=[
            image,
            document,
            question,
            output,
            output_text,
            example_image,
            upload,
            url,
        ],
    )

    upload.change(
        fn=process_upload,
        inputs=[upload],
        outputs=[document, image, img_clear_button, output, output_text],
    )
    submit.click(
        fn=process_path,
        inputs=[url],
        outputs=[document, image, img_clear_button, output, output_text],
    )

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

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

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

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

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