File size: 11,276 Bytes
15f37f2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# -*- coding: utf-8 -*-
import os
import pickle
from functools import lru_cache
import pytesseract
import numpy as np
from PIL import Image
import torch
from torchvision.transforms import ToTensor

PAD_TOKEN_BOX = [0, 0, 0, 0]
GRID_SIZE = 1000


def normalize_box(box, width, height, size=1000):
    """
    Takes a bounding box and normalizes it to a thousand pixels. If you notice it is
    just like calculating percentage except takes 1000 instead of 100.
    """
    return [
        int(size * (box[0] / width)),
        int(size * (box[1] / height)),
        int(size * (box[2] / width)),
        int(size * (box[3] / height)),
    ]


@lru_cache(maxsize=10)
def resize_align_bbox(bbox, orig_w, orig_h, target_w, target_h):
    x_scale = target_w / orig_w
    y_scale = target_h / orig_h
    orig_left, orig_top, orig_right, orig_bottom = bbox
    x = int(np.round(orig_left * x_scale))
    y = int(np.round(orig_top * y_scale))
    xmax = int(np.round(orig_right * x_scale))
    ymax = int(np.round(orig_bottom * y_scale))
    return [x, y, xmax, ymax]


def get_topleft_bottomright_coordinates(df_row):
    left, top, width, height = df_row["left"], df_row["top"], df_row["width"], df_row["height"]
    return [left, top, left + width, top + height]


def apply_ocr(image_fp):
    """
    Returns words and its bounding boxes from an image
    """
    image = Image.open(image_fp)
    width, height = image.size

    ocr_df = pytesseract.image_to_data(image, output_type="data.frame")
    ocr_df = ocr_df.dropna().reset_index(drop=True)
    float_cols = ocr_df.select_dtypes("float").columns
    ocr_df[float_cols] = ocr_df[float_cols].round(0).astype(int)
    ocr_df = ocr_df.replace(r"^\s*$", np.nan, regex=True)
    ocr_df = ocr_df.dropna().reset_index(drop=True)
    words = list(ocr_df.text.apply(lambda x: str(x).strip()))
    actual_bboxes = ocr_df.apply(get_topleft_bottomright_coordinates, axis=1).values.tolist()

    # add as extra columns
    assert len(words) == len(actual_bboxes)
    return {"words": words, "bbox": actual_bboxes}

def get_tokens_with_boxes(unnormalized_word_boxes, pad_token_box, word_ids,max_seq_len = 512):
    
    # assert len(unnormalized_word_boxes) == len(word_ids), this should not be applied, since word_ids may have higher 
    # length and the bbox corresponding to them may not exist
    
    unnormalized_token_boxes = []
    
    for i, word_idx in enumerate(word_ids):
        if word_idx is None:
            break
        unnormalized_token_boxes.append(unnormalized_word_boxes[word_idx])

    # all remaining are padding tokens so why add them in a loop one by one
    num_pad_tokens = len(word_ids) - i - 1
    if num_pad_tokens > 0:
        unnormalized_token_boxes.extend([pad_token_box] * num_pad_tokens)
        
        
    if len(unnormalized_token_boxes)<max_seq_len:
        unnormalized_token_boxes.extend([pad_token_box] * (max_seq_len-len(unnormalized_token_boxes)))
        
    return unnormalized_token_boxes


def get_centroid(actual_bbox):
    centroid = []
    for i in actual_bbox:
        width = i[2] - i[0]
        height = i[3] - i[1]
        centroid.append([i[0] + width / 2, i[1] + height / 2])
    return centroid


def get_pad_token_id_start_index(words, encoding, tokenizer): 
#     assert len(words) < len(encoding["input_ids"])  This condition, was creating errors on some sample images
    for idx in range(len(encoding["input_ids"])):
        if encoding["input_ids"][idx] == tokenizer.pad_token_id:
            break
    return idx


def get_relative_distance(bboxes, centroids, pad_tokens_start_idx):

    a_rel_x = []
    a_rel_y = []

    for i in range(0, len(bboxes)-1):
        if i >= pad_tokens_start_idx:
            a_rel_x.append([0] * 8)
            a_rel_y.append([0] * 8)
            continue

        curr = bboxes[i]
        next = bboxes[i+1]

        a_rel_x.append(
            [
                curr[0],  # top left x
                curr[2],  # bottom right x
                curr[2] - curr[0],  # width
                next[0] - curr[0],  # diff top left x
                next[0] - curr[0],  # diff bottom left x
                next[2] - curr[2],  # diff top right x
                next[2] - curr[2],  # diff bottom right x
                centroids[i+1][0] - centroids[i][0],
            ]
        )

        a_rel_y.append(
            [
                curr[1],  # top left y
                curr[3],  # bottom right y
                curr[3] - curr[1],  # height
                next[1] - curr[1],  # diff top left y
                next[3] - curr[3],  # diff bottom left y
                next[1] - curr[1],  # diff top right y
                next[3] - curr[3],  # diff bottom right y
                centroids[i+1][1] - centroids[i][1],
            ]
        )

    # For the last word
    
    a_rel_x.append([0]*8)  
    a_rel_y.append([0]*8)


    return a_rel_x, a_rel_y
     


def apply_mask(inputs, tokenizer):
    inputs = torch.as_tensor(inputs)
    rand = torch.rand(inputs.shape)
    # where the random array is less than 0.15, we set true
    mask_arr = (rand < 0.15) * (inputs != tokenizer.cls_token_id) * (inputs != tokenizer.pad_token_id)
    # create selection from mask_arr
    selection = torch.flatten(mask_arr.nonzero()).tolist()
    # apply selection pad_tokens_start_idx to inputs.input_ids, adding MASK tokens
    inputs[selection] = 103
    return inputs


def read_image_and_extract_text(image):
    original_image = Image.open(image).convert("RGB")
    return apply_ocr(image)


def create_features(
        image,
        tokenizer,
        add_batch_dim=False,
        target_size=(500,384),  # This was the resolution used by the authors
        max_seq_length=512,
        path_to_save=None,
        save_to_disk=False,
        apply_mask_for_mlm=False,
        extras_for_debugging=False,
        use_ocr = False,
        bounding_box = None,
        words = None
):

    # step 1: read original image and extract OCR entries
    original_image = Image.open(image).convert("RGB")

    if (use_ocr == False) and (bounding_box == None or words == None):
        raise Exception('Please provide the bounding box and words or pass the argument "use_ocr" = True')

    if use_ocr == True:
      entries = apply_ocr(image)
      bounding_box = entries["bbox"]
      words = entries["words"]

    CLS_TOKEN_BOX = [0, 0, *original_image.size]    # Can be variable, but as per the paper, they have mentioned that it covers the whole image
    # step 2: resize image
    resized_image = original_image.resize(target_size)

    # step 3: normalize image to a grid of 1000 x 1000 (to avoid the problem of differently sized images)
    width, height = original_image.size
    normalized_word_boxes = [
        normalize_box(bbox, width, height, GRID_SIZE) for bbox in bounding_box
    ]
    assert len(words) == len(normalized_word_boxes), "Length of words != Length of normalized words"

    # step 4: tokenize words and get their bounding boxes (one word may split into multiple tokens)
    encoding = tokenizer(words,
                         padding="max_length",
                         max_length=max_seq_length,
                         is_split_into_words=True,
                         truncation=True,
                         add_special_tokens=False)
    
    unnormalized_token_boxes = get_tokens_with_boxes(bounding_box,
                                                                  PAD_TOKEN_BOX,
                                                                  encoding.word_ids())

    # step 5: add special tokens and truncate seq. to maximum length
    unnormalized_token_boxes = [CLS_TOKEN_BOX] + unnormalized_token_boxes[:-1]
    # add CLS token manually to avoid autom. addition of SEP too (as in the paper)
    encoding["input_ids"] = [tokenizer.cls_token_id] + encoding["input_ids"][:-1]

    # step 6: Add bounding boxes to the encoding dict
    encoding["unnormalized_token_boxes"] = unnormalized_token_boxes
   
    # step 7: apply mask for the sake of pre-training
    if apply_mask_for_mlm:
        encoding["mlm_labels"] = encoding["input_ids"]
        encoding["input_ids"] = apply_mask(encoding["input_ids"], tokenizer)
        assert len(encoding["mlm_labels"]) == max_seq_length, "Length of mlm_labels != Length of max_seq_length"
       
    assert len(encoding["input_ids"]) == max_seq_length, "Length of input_ids != Length of max_seq_length"
    assert len(encoding["attention_mask"]) == max_seq_length, "Length of attention mask != Length of max_seq_length"
    assert len(encoding["token_type_ids"]) == max_seq_length, "Length of token type ids != Length of max_seq_length"

    # step 8: normalize the image
    encoding["resized_scaled_img"] = ToTensor()(resized_image)

    # step 9: apply mask for the sake of pre-training
    if apply_mask_for_mlm:
        encoding["mlm_labels"] = encoding["input_ids"]
        encoding["input_ids"] = apply_mask(encoding["input_ids"], tokenizer)

    # step 10: rescale and align the bounding boxes to match the resized image size (typically 224x224)
    resized_and_aligned_bboxes = []

    for bbox in unnormalized_token_boxes:
        # performing the normalization of the bounding box
        resized_and_aligned_bboxes.append(resize_align_bbox(tuple(bbox), *original_image.size, *target_size))

    encoding["resized_and_aligned_bounding_boxes"] = resized_and_aligned_bboxes
    
    # step 11: add the relative distances in the normalized grid
    bboxes_centroids = get_centroid(resized_and_aligned_bboxes)
    pad_token_start_index = get_pad_token_id_start_index(words, encoding, tokenizer)
    a_rel_x, a_rel_y = get_relative_distance(resized_and_aligned_bboxes, bboxes_centroids, pad_token_start_index)

    # step 12: convert all to tensors
    for k, v in encoding.items():
        encoding[k] = torch.as_tensor(encoding[k])

    encoding.update({
        "x_features": torch.as_tensor(a_rel_x, dtype=torch.int32),
        "y_features": torch.as_tensor(a_rel_y, dtype=torch.int32),
        })

    # step 13: add tokens for debugging
    if extras_for_debugging:
        input_ids = encoding["mlm_labels"] if apply_mask_for_mlm else encoding["input_ids"]
        encoding["tokens_without_padding"] = tokenizer.convert_ids_to_tokens(input_ids)
        encoding["words"] = words


    # step 14: add extra dim for batch
    if add_batch_dim:
        encoding["x_features"].unsqueeze_(0)
        encoding["y_features"].unsqueeze_(0)
        encoding["input_ids"].unsqueeze_(0)
        encoding["resized_scaled_img"].unsqueeze_(0)

    # step 15: save to disk
    if save_to_disk:
        os.makedirs(path_to_save, exist_ok=True)
        image_name = os.path.basename(image)
        with open(f"{path_to_save}{image_name}.pickle", "wb") as f:
            pickle.dump(encoding, f)

    # step 16: keys to keep, resized_and_aligned_bounding_boxes have been added for the purpose to test if the bounding boxes are drawn correctly or not, it maybe removed
    
    keys = ['resized_scaled_img', 'x_features','y_features','input_ids','resized_and_aligned_bounding_boxes']
    
    if apply_mask_for_mlm:
        keys.append('mlm_labels')
    
    final_encoding = {k:encoding[k] for k in keys}
    
    del encoding
    return final_encoding