File size: 17,539 Bytes
5e9bd47
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
import json
import copy
import random
import numpy as np


PAD = '<pad>'
SOS = '<sos>'
EOS = '<eos>'
UNK = '<unk>'
MASK = '<mask>'

Rxn = '[Rxn]'  # Reaction
Rct = '[Rct]'  # Reactant
Prd = '[Prd]'  # Product
Cnd = '[Cnd]'  # Condition
Idt = '[Idt]'  # Identifier
Mol = '[Mol]'  # Molecule
Txt = '[Txt]'  # Text
Sup = '[Sup]'  # Supplement
Noise = '[Nos]'


class ReactionTokenizer(object):

    def __init__(self, input_size=100, sep_xy=True, pix2seq=False):
        self.stoi = {}
        self.itos = {}
        self.pix2seq = pix2seq
        self.maxx = input_size  # height
        self.maxy = input_size  # width
        self.sep_xy = sep_xy
        self.special_tokens = [PAD, SOS, EOS, UNK, MASK]
        self.tokens = [Rxn, Rct, Prd, Cnd, Idt, Mol, Txt, Sup, Noise]
        self.fit_tokens(self.tokens)

    def __len__(self):
        if self.pix2seq:
            return 2094
        if self.sep_xy:
            return self.offset + self.maxx + self.maxy
        else:
            return self.offset + max(self.maxx, self.maxy)

    @property
    def max_len(self):
        return 256

    @property
    def PAD_ID(self):
        return self.stoi[PAD]

    @property
    def SOS_ID(self):
        return self.stoi[SOS]

    @property
    def EOS_ID(self):
        return self.stoi[EOS]

    @property
    def UNK_ID(self):
        return self.stoi[UNK]

    @property
    def NOISE_ID(self):
        return self.stoi[Noise]

    @property
    def offset(self):
        return 0 if self.pix2seq else len(self.stoi)

    @property
    def output_constraint(self):
        return True

    def fit_tokens(self, tokens):
        vocab = self.special_tokens + tokens
        if self.pix2seq:
            for i, s in enumerate(vocab):
                self.stoi[s] = 2001 + i
            self.stoi[EOS] = len(self) - 2
            # self.stoi[Noise] = len(self) - 1
        else:
            for i, s in enumerate(vocab):
                self.stoi[s] = i
        self.itos = {item[1]: item[0] for item in self.stoi.items()}
        self.bbox_category_to_token = {1: Mol, 2: Txt, 3: Idt, 4: Sup}
        self.token_to_bbox_category = {item[1]: item[0] for item in self.bbox_category_to_token.items()}

    def is_x(self, x):
        return 0 <= x - self.offset < self.maxx

    def is_y(self, y):
        if self.sep_xy:
            return self.maxx <= y - self.offset < self.maxx + self.maxy
        return 0 <= y - self.offset < self.maxy

    def x_to_id(self, x):
        if x < -0.001 or x > 1.001:
            print(x)
        else:
            x = min(max(x, 0), 1)
        assert 0 <= x <= 1
        return self.offset + round(x * (self.maxx - 1))

    def y_to_id(self, y):
        if y < -0.001 or y > 1.001:
            print(y)
        else:
            y = min(max(y, 0), 1)
        assert 0 <= y <= 1
        if self.sep_xy:
            return self.offset + self.maxx + round(y * (self.maxy - 1))
        return self.offset + round(y * (self.maxy - 1))

    def id_to_x(self, id, scale=1):
        if not self.is_x(id):
            return -1
        return (id - self.offset) / (self.maxx - 1) / scale

    def id_to_y(self, id, scale=1):
        if not self.is_y(id):
            return -1
        if self.sep_xy:
            return (id - self.offset - self.maxx) / (self.maxy - 1) * scale
        return (id - self.offset) / (self.maxy - 1) / scale

    def update_state(self, state, idx):
        if state is None:
            new_state = (Rxn, 'e')
        else:
            if state[1] == 'x1':
                new_state = (state[0], 'y1')
            elif state[1] == 'y1':
                new_state = (state[0], 'x2')
            elif state[1] == 'x2':
                new_state = (state[0], 'y2')
            elif state[1] == 'y2':
                new_state = (state[0], 'c')
            elif state[1] == 'c':
                if self.is_x(idx):
                    new_state = (state[0], 'x1')
                else:
                    new_state = (state[0], 'e')
            else:
                if state[0] == Rct:
                    if self.is_x(idx):
                        new_state = (Cnd, 'x1')
                    else:
                        new_state = (Cnd, 'e')
                elif state[0] == Cnd:
                    new_state = (Prd, 'x1')
                elif state[0] == Prd:
                    new_state = (Rxn, 'e')
                elif state[0] == Rxn:
                    if self.is_x(idx):
                        new_state = (Rct, 'x1')
                    else:
                        new_state = (EOS, 'e')
                else:
                    new_state = (EOS, 'e')
        return new_state

    def output_mask(self, state):
        # mask: True means forbidden
        mask = np.array([True] * len(self))
        if state[1] in ['y1', 'c']:
            mask[self.offset:self.offset+self.maxx] = False
        if state[1] in ['x1', 'x2']:
            if self.sep_xy:
                mask[self.offset+self.maxx:self.offset+self.maxx+self.maxy] = False
            else:
                mask[self.offset:self.offset+self.maxy] = False
        if state[1] == 'y2':
            for token in [Idt, Mol, Txt, Sup]:
                mask[self.stoi[token]] = False
        if state[1] == 'c':
            mask[self.stoi[state[0]]] = False
        if state[1] == 'e':
            if state[0] in [Rct, Cnd, Rxn]:
                mask[self.offset:self.offset + self.maxx] = False
            if state[0] == Rct:
                mask[self.stoi[Cnd]] = False
            if state[0] == Prd:
                mask[self.stoi[Rxn]] = False
                mask[self.stoi[Noise]] = False
            if state[0] in [Rxn, EOS]:
                mask[self.EOS_ID] = False
        return mask

    def update_states_and_masks(self, states, ids):
        new_states = [self.update_state(state, idx) for state, idx in zip(states, ids)]
        masks = np.array([self.output_mask(state) for state in new_states])
        return new_states, masks

    def bbox_to_sequence(self, bbox, category):
        sequence = []
        x1, y1, x2, y2 = bbox
        if x1 >= x2 or y1 >= y2:
            return []
        sequence.append(self.x_to_id(x1))
        sequence.append(self.y_to_id(y1))
        sequence.append(self.x_to_id(x2))
        sequence.append(self.y_to_id(y2))
        if category in self.bbox_category_to_token:
            sequence.append(self.stoi[self.bbox_category_to_token[category]])
        else:
            sequence.append(self.stoi[Noise])
        return sequence

    def sequence_to_bbox(self, sequence, scale=[1, 1]):
        if len(sequence) < 5:
            return None
        x1, y1 = self.id_to_x(sequence[0], scale[0]), self.id_to_y(sequence[1], scale[1])
        x2, y2 = self.id_to_x(sequence[2], scale[0]), self.id_to_y(sequence[3], scale[1])
        if x1 == -1 or y1 == -1 or x2 == -1 or y2 == -1 or x1 >= x2 or y1 >= y2 or sequence[4] not in self.itos:
            return None
        category = self.itos[sequence[4]]
        if category not in [Mol, Txt, Idt, Sup]:
            return None
        return {'category': category, 'bbox': (x1, y1, x2, y2), 'category_id': self.token_to_bbox_category[category]}

    def perturb_reaction(self, reaction, boxes):
        reaction = copy.deepcopy(reaction)
        options = []
        options.append(0)  # Option 0: add
        if not(len(reaction['reactants']) == 1 and len(reaction['conditions']) == 0 and len(reaction['products']) == 1):
            options.append(1)  # Option 1: delete
            options.append(2)  # Option 2: move
        choice = random.choice(options)
        if choice == 0:
            key = random.choice(['reactants', 'conditions', 'products'])
            # TODO: insert to a random position
            # We simply add a random box, which may be a duplicate box in this reaction
            reaction[key].append(random.randrange(len(boxes)))
        if choice == 1 or choice == 2:
            options = []
            for key, val in [('reactants', 1), ('conditions', 0), ('products', 1)]:
                if len(reaction[key]) > val:
                    options.append(key)
            key = random.choice(options)
            idx = random.randrange(len(reaction[key]))
            del_box = reaction[key][idx]
            reaction[key] = reaction[key][:idx] + reaction[key][idx+1:]
            if choice == 2:
                options = ['reactants', 'conditions', 'products']
                options.remove(key)
                newkey = random.choice(options)
                reaction[newkey].append(del_box)
        return reaction

    def augment_reaction(self, reactions, data):
        area, boxes, labels = data['area'], data['boxes'], data['labels']
        nonempty_boxes = [i for i in range(len(area)) if area[i] > 0]
        if len(nonempty_boxes) == 0:
            return None
        if len(reactions) == 0 or random.randrange(100) < 20:
            num_reactants = random.randint(1, 3)
            num_conditions = random.randint(0, 3)
            num_products = random.randint(1, 3)
            reaction = {
                'reactants': random.choices(nonempty_boxes, k=num_reactants),
                'conditions': random.choices(nonempty_boxes, k=num_conditions),
                'products': random.choices(nonempty_boxes, k=num_products)
            }
        else:
            assert len(reactions) > 0
            reaction = self.perturb_reaction(random.choice(reactions), boxes)
        return reaction

    def reaction_to_sequence(self, reaction, data, shuffle_bbox=False):
        reaction = copy.deepcopy(reaction)
        area, boxes, labels = data['area'], data['boxes'], data['labels']
        # If reactants or products are empty (because of image cropping), skip the reaction
        if all([area[i] == 0 for i in reaction['reactants']]) or all([area[i] == 0 for i in reaction['products']]):
            return []
        if shuffle_bbox:
            random.shuffle(reaction['reactants'])
            random.shuffle(reaction['conditions'])
            random.shuffle(reaction['products'])
        sequence = []
        for idx in reaction['reactants']:
            if area[idx] == 0:
                continue
            sequence += self.bbox_to_sequence(boxes[idx].tolist(), labels[idx].item())
        sequence.append(self.stoi[Rct])
        for idx in reaction['conditions']:
            if area[idx] == 0:
                continue
            sequence += self.bbox_to_sequence(boxes[idx].tolist(), labels[idx].item())
        sequence.append(self.stoi[Cnd])
        for idx in reaction['products']:
            if area[idx] == 0:
                continue
            sequence += self.bbox_to_sequence(boxes[idx].tolist(), labels[idx].item())
        sequence.append(self.stoi[Prd])
        sequence.append(self.stoi[Rxn])
        return sequence

    def data_to_sequence(self, data, rand_order=False, shuffle_bbox=False, add_noise=False, mix_noise=False):
        sequence = [self.SOS_ID]
        sequence_out = [self.SOS_ID]
        reactions = copy.deepcopy(data['reactions'])
        reactions_seqs = []
        for reaction in reactions:
            seq = self.reaction_to_sequence(reaction, data, shuffle_bbox=shuffle_bbox)
            reactions_seqs.append([seq, seq])
        noise_seqs = []
        if add_noise:
            total_len = sum(len(seq) for seq, seq_out in reactions_seqs)
            while total_len < self.max_len:
                reaction = self.augment_reaction(reactions, data)
                if reaction is None:
                    break
                seq = self.reaction_to_sequence(reaction, data)
                if len(seq) == 0:
                    continue
                if mix_noise:
                    seq[-1] = self.NOISE_ID
                    seq_out = [self.PAD_ID] * (len(seq) - 1) + [self.NOISE_ID]
                else:
                    seq_out = [self.PAD_ID] * (len(seq) - 1) + [self.NOISE_ID]
                noise_seqs.append([seq, seq_out])
                total_len += len(seq)
        if rand_order:
            random.shuffle(reactions_seqs)
        reactions_seqs += noise_seqs
        if mix_noise:
            random.shuffle(reactions_seqs)
        for seq, seq_out in reactions_seqs:
            sequence += seq
            sequence_out += seq_out
        sequence.append(self.EOS_ID)
        sequence_out.append(self.EOS_ID)
        return sequence, sequence_out

    def sequence_to_data(self, sequence, scores=None, scale=None):
        reactions = []
        i = 0
        cur_reaction = {'reactants': [], 'conditions': [], 'products': []}
        flag = 'reactants'
        if len(sequence) > 0 and sequence[0] == self.SOS_ID:
            i += 1
        while i < len(sequence):
            if sequence[i] == self.EOS_ID:
                break
            if sequence[i] in self.itos:
                if self.itos[sequence[i]] in [Rxn, Noise]:
                    cur_reaction['label'] = self.itos[sequence[i]]
                    if len(cur_reaction['reactants']) > 0 and len(cur_reaction['products']) > 0:
                        reactions.append(cur_reaction)
                    cur_reaction = {'reactants': [], 'conditions': [], 'products': []}
                    flag = 'reactants'
                elif self.itos[sequence[i]] == Rct:
                    flag = 'conditions'
                elif self.itos[sequence[i]] == Cnd:
                    flag = 'products'
                elif self.itos[sequence[i]] == Prd:
                    flag = None
            elif i+5 <= len(sequence) and flag is not None:
                bbox = self.sequence_to_bbox(sequence[i:i+5], scale)
                if bbox is not None:
                    cur_reaction[flag].append(bbox)
                    i += 4
            i += 1
        return reactions

    def sequence_to_tokens(self, sequence):
        return [self.itos[x] if x in self.itos else x for x in sequence]


class BboxTokenizer(ReactionTokenizer):

    def __init__(self, input_size=100, sep_xy=True, pix2seq=False):
        super(BboxTokenizer, self).__init__(input_size, sep_xy, pix2seq)

    @property
    def max_len(self):
        return 500

    @property
    def output_constraint(self):
        return False

    def random_category(self):
        return random.choice(list(self.bbox_category_to_token.keys()))
        # return random.choice([random.choice(list(self.bbox_category_to_token.keys())), self.NOISE_ID])

    def random_bbox(self):
        _x1, _y1, _x2, _y2 = random.random(), random.random(), random.random(), random.random()
        x1, y1, x2, y2 = min(_x1, _x2), min(_y1, _y2), max(_x1, _x2), max(_y1, _y2)
        category = self.random_category()
        return [x1, y1, x2, y2], category

    def jitter_bbox(self, bbox, ratio=0.2):
        x1, y1, x2, y2 = bbox
        w, h = x2 - x1, y2 - y1
        _x1 = x1 + random.uniform(-w*ratio, w*ratio)
        _y1 = y1 + random.uniform(-h*ratio, h*ratio)
        _x2 = x2 + random.uniform(-w * ratio, w * ratio)
        _y2 = y2 + random.uniform(-h * ratio, h * ratio)
        x1, y1, x2, y2 = min(_x1, _x2), min(_y1, _y2), max(_x1, _x2), max(_y1, _y2)
        category = self.random_category()
        return np.clip([x1, y1, x2, y2], 0, 1), category

    def augment_box(self, bboxes):
        if len(bboxes) == 0:
            return self.random_bbox()
        if random.random() < 0.5:
            return self.random_bbox()
        else:
            return self.jitter_bbox(random.choice(bboxes))

    def data_to_sequence(self, data, add_noise=False, rand_order=False):
        sequence = [self.SOS_ID]
        sequence_out = [self.SOS_ID]
        if rand_order:
            perm = np.random.permutation(len(data['boxes']))
            boxes = data['boxes'][perm].tolist()
            labels = data['labels'][perm].tolist()
        else:
            boxes = data['boxes'].tolist()
            labels = data['labels'].tolist()
        for bbox, category in zip(boxes, labels):
            seq = self.bbox_to_sequence(bbox, category)
            sequence += seq
            # sequence[-1] = self.random_category()
            sequence_out += seq
        if add_noise:
            while len(sequence) < self.max_len:
                bbox, category = self.augment_box(boxes)
                sequence += self.bbox_to_sequence(bbox, category)
                sequence_out += [self.PAD_ID] * 4 + [self.NOISE_ID]
        sequence.append(self.EOS_ID)
        sequence_out.append(self.EOS_ID)
        return sequence, sequence_out

    def sequence_to_data(self, sequence, scores=None, scale=None):
        bboxes = []
        i = 0
        if len(sequence) > 0 and sequence[0] == self.SOS_ID:
            i += 1
        while i < len(sequence):
            if sequence[i] == self.EOS_ID:
                break
            if i+4 < len(sequence):
                bbox = self.sequence_to_bbox(sequence[i:i+5], scale)
                if bbox is not None:
                    if scores is not None:
                        bbox['score'] = scores[i + 4]
                    bboxes.append(bbox)
                    i += 4
            i += 1
        return bboxes


def get_tokenizer(args):
    tokenizer = {}
    if args.pix2seq:
        args.coord_bins = 2000
        args.sep_xy = False
    format = args.format
    if format == 'reaction':
        tokenizer[format] = ReactionTokenizer(args.coord_bins, args.sep_xy, args.pix2seq)
    if format == 'bbox':
        tokenizer[format] = BboxTokenizer(args.coord_bins, args.sep_xy, args.pix2seq)
    return tokenizer