summaryrefslogtreecommitdiff
path: root/src/text_recognizer/networks/transducer/transducer.py
blob: d7e3d08d0eee683e65e0ab54f28c22bfe93e22e8 (plain)
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
"""Transducer and the transducer loss function.py

Stolen from:
    https://github.com/facebookresearch/gtn_applications/blob/master/transducer.py

"""
from pathlib import Path
import itertools
from typing import Dict, List, Optional, Union, Tuple

from loguru import logger
import gtn
import torch
from torch import nn
from torch import Tensor

from text_recognizer.datasets.iam_preprocessor import Preprocessor


def make_scalar_graph(weight) -> gtn.Graph:
    scalar = gtn.Graph()
    scalar.add_node(True)
    scalar.add_node(False, True)
    scalar.add_arc(0, 1, 0, 0, weight)
    return scalar


def make_chain_graph(sequence) -> gtn.Graph:
    graph = gtn.Graph(False)
    graph.add_node(True)
    for i, s in enumerate(sequence):
        graph.add_node(False, i == (len(sequence) - 1))
        graph.add_arc(i, i + 1, s)
    return graph


def make_transitions_graph(
    ngram: int, num_tokens: int, calc_grad: bool = False
) -> gtn.Graph:
    transitions = gtn.Graph(calc_grad)
    transitions.add_node(True, ngram == 1)

    state_map = {(): 0}

    # First build transitions which include <s>:
    for n in range(1, ngram):
        for state in itertools.product(range(num_tokens), repeat=n):
            in_idx = state_map[state[:-1]]
            out_idx = transitions.add_node(False, ngram == 1)
            state_map[state] = out_idx
            transitions.add_arc(in_idx, out_idx, state[-1])

    for state in itertools.product(range(num_tokens), repeat=ngram):
        state_idx = state_map[state[:-1]]
        new_state_idx = state_map[state[1:]]
        # p(state[-1] | state[:-1])
        transitions.add_arc(state_idx, new_state_idx, state[-1])

    if ngram > 1:
        # Build transitions which include </s>:
        end_idx = transitions.add_node(False, True)
        for in_idx in range(end_idx):
            transitions.add_arc(in_idx, end_idx, gtn.epsilon)

    return transitions


def make_lexicon_graph(word_pieces: List, graphemes_to_idx: Dict) -> gtn.Graph:
    """Constructs a graph which transduces letters to word pieces."""
    graph = gtn.Graph(False)
    graph.add_node(True, True)
    for i, wp in enumerate(word_pieces):
        prev = 0
        for l in wp[:-1]:
            n = graph.add_node()
            graph.add_arc(prev, n, graphemes_to_idx[l], gtn.epsilon)
            prev = n
        graph.add_arc(prev, 0, graphemes_to_idx[wp[-1]], i)
    graph.arc_sort()
    return graph


def make_token_graph(
    token_list: List, blank: str = "none", allow_repeats: bool = True
) -> gtn.Graph:
    """Constructs a graph with all the individual token transition models."""
    if not allow_repeats and blank != "optional":
        raise ValueError("Must use blank='optional' if disallowing repeats.")

    ntoks = len(token_list)
    graph = gtn.Graph(False)

    # Creating nodes
    graph.add_node(True, True)
    for i in range(ntoks):
        # We can consume one or more consecutive word
        # pieces for each emission:
        # E.g. [ab, ab, ab] transduces to [ab]
        graph.add_node(False, blank != "forced")

    if blank != "none":
        graph.add_node()

    # Creating arcs
    if blank != "none":
        # Blank index is assumed to be last (ntoks)
        graph.add_arc(0, ntoks + 1, ntoks, gtn.epsilon)
        graph.add_arc(ntoks + 1, 0, gtn.epsilon)

    for i in range(ntoks):
        graph.add_arc((ntoks + 1) if blank == "forced" else 0, i + 1, i)
        graph.add_arc(i + 1, i + 1, i, gtn.epsilon)

        if allow_repeats:
            if blank == "forced":
                # Allow transitions from token to blank only
                graph.add_arc(i + 1, ntoks + 1, ntoks, gtn.epsilon)
            else:
                # Allow transition from token to blank and all other tokens
                graph.add_arc(i + 1, 0, gtn.epsilon)

        else:
            # allow transitions to blank and all other tokens except the same token
            graph.add_arc(i + 1, ntoks + 1, ntoks, gtn.epsilon)
            for j in range(ntoks):
                if i != j:
                    graph.add_arc(i + 1, j + 1, j, j)

    return graph


class TransducerLossFunction(torch.autograd.Function):
    @staticmethod
    def forward(
        ctx,
        inputs,
        targets,
        tokens,
        lexicon,
        transition_params=None,
        transitions=None,
        reduction="none",
    ) -> Tensor:
        B, T, C = inputs.shape

        losses = [None] * B
        emissions_graphs = [None] * B

        if transitions is not None:
            if transition_params is None:
                raise ValueError("Specified transitions, but not transition params.")

            cpu_data = transition_params.cpu().contiguous()
            transitions.set_weights(cpu_data.data_ptr())
            transitions.calc_grad = transition_params.requires_grad
            transitions.zero_grad()

        def process(b: int) -> None:
            # Create emission graph:
            emissions = gtn.linear_graph(T, C, inputs.requires_grad)
            cpu_data = inputs[b].cpu().contiguous()
            emissions.set_weights(cpu_data.data_ptr())
            target = make_chain_graph(targets[b])
            target.arc_sort(True)

            # Create token tot grapheme decomposition graph
            tokens_target = gtn.remove(gtn.project_output(gtn.compose(target, lexicon)))
            tokens_target.arc_sort()

            # Create alignment graph:
            aligments = gtn.project_input(
                gtn.remove(gtn.compose(tokens, tokens_target))
            )
            aligments.arc_sort()

            # Add transitions scores:
            if transitions is not None:
                aligments = gtn.intersect(transitions, aligments)
                aligments.arc_sort()

            loss = gtn.forward_score(gtn.intersect(emissions, aligments))

            # Normalize if needed:
            if transitions is not None:
                norm = gtn.forward_score(gtn.intersect(emissions, transitions))
                loss = gtn.subtract(loss, norm)

            losses[b] = gtn.negate(loss)

            # Save for backward:
            if emissions.calc_grad:
                emissions_graphs[b] = emissions

        gtn.parallel_for(process, range(B))

        ctx.graphs = (losses, emissions_graphs, transitions)
        ctx.input_shape = inputs.shape

        # Optionally reduce by target length
        if reduction == "mean":
            scales = [(1 / len(t) if len(t) > 0 else 1.0) for t in targets]
        else:
            scales = [1.0] * B

        ctx.scales = scales

        loss = torch.tensor([l.item() * s for l, s in zip(losses, scales)])
        return torch.mean(loss.to(inputs.device))

    @staticmethod
    def backward(ctx, grad_output) -> Tuple:
        losses, emissions_graphs, transitions = ctx.graphs
        scales = ctx.scales

        B, T, C = ctx.input_shape
        calc_emissions = ctx.needs_input_grad[0]
        input_grad = torch.empty((B, T, C)) if calc_emissions else None

        def process(b: int) -> None:
            scale = make_scalar_graph(scales[b])
            gtn.backward(losses[b], scale)
            emissions = emissions_graphs[b]
            if calc_emissions:
                grad = emissions.grad().weights_to_numpy()
                input_grad[b] = torch.tensor(grad).view(1, T, C)

        gtn.parallel_for(process, range(B))

        if calc_emissions:
            input_grad = input_grad.to(grad_output.device)
            input_grad *= grad_output / B

        if ctx.needs_input_grad[4]:
            grad = transitions.grad().weights_to_numpy()
            transition_grad = torch.tensor(grad).to(grad_output.device)
            transition_grad *= grad_output / B
        else:
            transition_grad = None

        return (
            input_grad,
            None,  # target
            None,  # tokens
            None,  # lexicon
            transition_grad,  # transition params
            None,  # transitions graph
            None,
        )


TransducerLoss = TransducerLossFunction.apply


class Transducer(nn.Module):
    def __init__(
        self,
        tokens: List,
        graphemes_to_idx: Dict,
        ngram: int = 0,
        transitions: str = None,
        blank: str = "none",
        allow_repeats: bool = True,
        reduction: str = "none",
    ) -> None:
        """A generic transducer loss function.

        Args:
            tokens (List) : A list of iterable objects (e.g. strings, tuples, etc)
                representing the output tokens of the model (e.g. letters,
                word-pieces, words). For example ["a", "b", "ab", "ba", "aba"]
                could be a list of sub-word tokens.
            graphemes_to_idx (dict) : A dictionary mapping grapheme units (e.g.
                "a", "b", ..) to their corresponding integer index.
            ngram (int) : Order of the token-level transition model. If `ngram=0`
                then no transition model is used.
            blank (string) : Specifies the usage of blank token
                'none' - do not use blank token
                'optional' - allow an optional blank inbetween tokens
                'forced' - force a blank inbetween tokens (also referred to as garbage token)
            allow_repeats (boolean) : If false, then we don't allow paths with
                consecutive tokens in the alignment graph. This keeps the graph
                unambiguous in the sense that the same input cannot transduce to
                different outputs.
        """
        super().__init__()
        if blank not in ["optional", "forced", "none"]:
            raise ValueError(
                "Invalid value specified for blank. Must be in ['optional', 'forced', 'none']"
            )
        self.tokens = make_token_graph(tokens, blank=blank, allow_repeats=allow_repeats)
        self.lexicon = make_lexicon_graph(tokens, graphemes_to_idx)
        self.ngram = ngram
        if ngram > 0 and transitions is not None:
            raise ValueError("Only one of ngram and transitions may be specified")

        if ngram > 0:
            transitions = make_transitions_graph(
                ngram, len(tokens) + int(blank != "none"), True
            )

        if transitions is not None:
            self.transitions = transitions
            self.transitions.arc_sort()
            self.transitions_params = nn.Parameter(
                torch.zeros(self.transitions.num_arcs())
            )
        else:
            self.transitions = None
            self.transitions_params = None
        self.reduction = reduction

    def forward(self, inputs: Tensor, targets: Tensor) -> TransducerLoss:
        TransducerLoss(
            inputs,
            targets,
            self.tokens,
            self.lexicon,
            self.transitions_params,
            self.transitions,
            self.reduction,
        )

    def viterbi(self, outputs: Tensor) -> List[Tensor]:
        B, T, C = outputs.shape

        if self.transitions is not None:
            cpu_data = self.transition_params.cpu().contiguous()
            self.transitions.set_weights(cpu_data.data_ptr())
            self.transitions.calc_grad = False

        self.tokens.arc_sort()

        paths = [None] * B

        def process(b: int) -> None:
            emissions = gtn.linear_graph(T, C, False)
            cpu_data = outputs[b].cpu().contiguous()
            emissions.set_weights(cpu_data.data_ptr())

            if self.transitions is not None:
                full_graph = gtn.intersect(emissions, self.transitions)
            else:
                full_graph = emissions

            # Find the best path and remove back-off arcs:
            path = gtn.remove(gtn.viterbi_path(full_graph))

            # Left compose the viterbi path with the "aligment to token"
            # transducer to get the outputs:
            path = gtn.compose(path, self.tokens)

            # When there are ambiguous paths (allow_repeats is true), we take
            # the shortest:
            path = gtn.viterbi_path(path)
            path = gtn.remove(gtn.project_output(path))
            paths[b] = path.labels_to_list()

        gtn.parallel_for(process, range(B))
        predictions = [torch.IntTensor(path) for path in paths]
        return predictions


def load_transducer_loss(
    num_features: int,
    ngram: int,
    tokens: str,
    lexicon: str,
    transitions: str,
    blank: str,
    allow_repeats: bool,
    prepend_wordsep: bool = False,
    use_words: bool = False,
    data_dir: Optional[Union[str, Path]] = None,
    reduction: str = "mean",
) -> Tuple[Transducer, int]:
    if data_dir is None:
        data_dir = (
            Path(__file__).resolve().parents[4] / "data" / "raw" / "iam" / "iamdb"
        )
        logger.debug(f"Using data dir: {data_dir}")
        if not data_dir.exists():
            raise RuntimeError(f"Could not locate iamdb directory at {data_dir}")
    else:
        data_dir = Path(data_dir)
    processed_path = (
        Path(__file__).resolve().parents[4] / "data" / "processed" / "iam_lines"
    )
    tokens_path = processed_path / tokens
    lexicon_path = processed_path / lexicon

    if transitions is not None:
        transitions = gtn.load(str(processed_path / transitions))

    preprocessor = Preprocessor(
        data_dir, num_features, tokens_path, lexicon_path, use_words, prepend_wordsep,
    )

    num_tokens = preprocessor.num_tokens

    criterion = Transducer(
        preprocessor.tokens,
        preprocessor.graphemes_to_index,
        ngram=ngram,
        transitions=transitions,
        blank=blank,
        allow_repeats=allow_repeats,
        reduction=reduction,
    )

    return criterion, num_tokens + int(blank != "none")