File size: 14,615 Bytes
8019be0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import abc
from typing import Optional
import torch
from torch import Tensor
from dataclasses import dataclass
from .schedule import Schedule
import torch.nn.functional as F


@dataclass
class ModelPrediction:
    token_logits: Tensor
    length_posterior: Optional[Tensor]
    expected_gaps: Tensor

    def __init__(
        self,
        token_logits: Tensor,
        length_posterior: Optional[Tensor] = None,
        expected_gaps: Optional[Tensor] = None,
    ):
        assert length_posterior is not None or expected_gaps is not None
        self.token_logits = token_logits
        self.length_posterior = length_posterior
        self.expected_gaps = expected_gaps
        if self.expected_gaps is None:
            _, _, L = self.length_posterior.shape
            index = torch.arange(0, L, device=token_logits.device).view(1, 1, -1)
            self.expected_gaps = (F.softmax(self.length_posterior, dim=-1) * index).sum(dim=-1)


@dataclass
class Rate:
    unmask_rate: Tensor  # Shape [Batch, Length, Vocab]
    length_rate: Tensor  # Shape [Batch]


@dataclass
class HittingTime:
    insertion_time: Tensor  # Shape [Batch, Length]
    unmasking_time: Tensor  # Shape [Batch, Length]

    def __iter__(self):
        yield from [self.insertion_time, self.unmasking_time]


@dataclass
class JointInterpolantResult:
    # Joint Interpolant
    xt: Tensor  # Shape [Batch, Length]
    st: Tensor  # Shape [Batch, Length]
    _x1: Tensor
    _pad_token: int
    _mask_token: int

    @property
    def mask_indices(self) -> Tensor:
        return self.xt == self._mask_token

    @property
    def unmasked(self) -> Tensor:
        return torch.gather(self._x1, 1, self.st)

    @property
    def xt_length(self) -> Tensor:
        # Calculate length of xt
        return (self.xt != self._pad_token).sum(dim=1)

    @property
    def x1_length(self) -> Tensor:
        # Calculate length of x1
        return (self._x1 != self._pad_token).sum(dim=1)

    @property
    def gaps_and_mask(self) -> tuple[Tensor, Tensor]:
        x1_len = self.x1_length
        gaps = self.st.clone()

        pad_front = gaps.new_zeros((gaps.shape[0], 1)) - 1  # -1 for the front padding
        pad_back = gaps.new_zeros((gaps.shape[0], 1))
        gaps = torch.cat([pad_front, gaps, pad_back], dim=1)  # Add a leading zero

        gaps.scatter_(
            1, self.xt_length.unsqueeze(1) + 1, x1_len.unsqueeze(1)
        )  # Fill the last position with x1_len

        gaps = gaps[:, 1:] - gaps[:, :-1] - 1
        gaps = torch.clamp(gaps, min=0)

        idx = torch.arange(gaps.size(1), device=self.xt.device).unsqueeze(
            0
        )  # shape [1, max_gap]
        mask = idx <= self.xt_length.unsqueeze(1)
        gaps[~mask] = 0

        return gaps, mask


class JointInterpolant(abc.ABC):
    def __init__(
        self,
        vocab_size: int,
        mask_token: int,
        pad_token: int,
        max_length: int,
    ):
        """
        TODO: Add knobs
        """
        self.mask_token = mask_token
        self.pad_token = pad_token
        self.max_length = max_length
        self.vocab_size = vocab_size

    @abc.abstractmethod
    def elbo_weight(self, t: Tensor, x1: Tensor):
        """
        Return the ELBO weight for the training, can be changed depends on the empirical results
        Shape:
            t: [B]
        Returns:
            weight_unmask: [B, L]
            weight_delete: [B, L+1]
        """
        raise NotImplementedError

    @abc.abstractmethod
    def to_actual_rate(self, prediction: ModelPrediction, t: Tensor) -> Rate:
        raise NotImplementedError

    @abc.abstractmethod
    def sample_interpolant(self, t: Tensor, x1: Tensor) -> JointInterpolantResult:
        """
        Sample the interpolant xt from x1 at time t
        Shapes:
            x1: [B, L]
            t: [B]
        Returns:
            xt: [B, L]
            st: [B, L] boolean mask of positions that corresponds to xt
            xt_mask_indices: [B, L] boolean mask of positions that are masked at xt
            x1_remained: [B, L] tokens that are not deleted, used for the training target
            gap_counts: [B, L+1] the number of deleted tokens between xt slots
        """
        raise NotImplementedError


class AnyOrderMaskInsertionInterpolant(JointInterpolant):
    def __init__(
        self,
        insertion_schedule: Schedule,
        unmask_schedule: Schedule,
        vocab_size: int,
        mask_token: int,
        pad_token: int,
        max_length: int,
    ):
        super().__init__(vocab_size, mask_token, pad_token, max_length)
        self.insertion_schedule = insertion_schedule
        self.unmask_schedule = unmask_schedule
        #self.max_length = 500

    def expected_mask_fraction(self, t: Tensor, xt: Tensor) -> Tensor:
        """
        Compute the expected fraction of tokens that should be masked at time t.
        For AnyOrderMaskInsertionInterpolant, tokens are:
        - Deleted (pad) if t < insertion_time  
        - Masked if insertion_time <= t < unmasking_time
        - Unmasked if t >= unmasking_time
        
        We approximate: E[fraction masked] ≈ max(0, insertion_schedule.at(t) - unmask_schedule.at(t))
        
        Args:
            t: [B] current time
            xt: [B, L] current sequence (to get current length)
        Returns:
            [B] expected number of masked tokens per sequence
        """
        # Get schedule values at time t
        insertion_progress = self.insertion_schedule.at(t)  # [B]
        unmask_progress = self.unmask_schedule.at(t)  # [B]
        
        # Expected fraction of tokens that are inserted but not yet unmasked
        # Clamp to ensure non-negative
        expected_mask_frac = torch.clamp(insertion_progress - unmask_progress, min=0.0, max=1.0)
        
        # Get current sequence length (non-pad tokens)
        current_length = (xt != self.pad_token).sum(dim=1).float()  # [B]
        
        # Expected number of masked tokens
        expected_num_masked = expected_mask_frac * current_length  # [B]
        
        return expected_num_masked

    def hitting_time(self, t: Tensor, x1: Tensor) -> tuple[Tensor, Tensor]:
        """
        t1 is sampled from a uniform distribution over [0, 1]. when t1 < self.mask_schedule.at(t)
        t2 is sampled from a uniform distribution over [t1, 1]
        """
        B, L = x1.shape
        eps = 1e-6

        insert_time = self.insertion_schedule.sample((B, L), device=x1.device)
        insert_time = eps + (1 - eps) * insert_time  # ensure t1 is not 0
        unmask_time = self.unmask_schedule.sample_truncated(
            insert_time, (B, L), device=x1.device
        )

        return insert_time, unmask_time

    def elbo_weight(self, t: Tensor, x1: Tensor):
        """
        Return the ELBO weight for the training, can be changed depends on the empirical results
        """
        insert_weight = self.insertion_schedule.rate_scale_factor(t)
        insert_weight = insert_weight[:, None].expand(-1, x1.shape[1] + 1)

        unmask_weight = self.unmask_schedule.rate_scale_factor(t)
        unmask_weight = unmask_weight.unsqueeze(1).expand(-1, x1.shape[1])

        return unmask_weight, insert_weight

    def to_actual_rate(
        self, xt: Tensor, prediction: ModelPrediction, t: Tensor
    ) -> Rate:
        """
        Return the actual rate for the sampling
        Args:
            xt: [B, L] the sampled tokens
            prediction: ModelPrediction object containing token_posterior and expected_gaps
            t: [B] the time parameter
        """
        token_posterior = F.softmax(prediction.token_logits, dim=-1)  # (B, L, V)
        unmask_rate = token_posterior * self.unmask_schedule.rate_scale_factor(t).view(
            -1, 1, 1
        )
        
        length_rate = (
            prediction.expected_gaps
            * self.insertion_schedule.rate_scale_factor(t).view(-1, 1)
        )
        #print("expected_gaps:", prediction.expected_gaps, "length_rate:", length_rate)

        return Rate(
            unmask_rate=unmask_rate,  # (B, L, V)
            length_rate=length_rate,  # (B, L+1)
        )

    def sample_interpolant(self, t: Tensor, x1: Tensor) -> JointInterpolantResult:
        """
        Shapes:
            x1: [B, L]
            t: [B]
        Returns:
            xt: [B, L]
            st: [B, L] boolean mask of positions that corresponds to xt
            xt_mask_indices: [B, L] boolean mask of positions that are masked at xt
            x1_remained: [B, L] tokens that are not deleted, used for the training target
            gap_counts: [B, L+1] the number of deleted tokens between xt slots
        """
        # sample the stopping time (B, L, 2)
        insertion_time, unmasking_time = self.hitting_time(t, x1)

        clean_tokens = x1.ne(self.pad_token)
        deleted_tokens = clean_tokens & (t[:, None] < insertion_time)
        masked_tokens = (
            clean_tokens
            & (t[:, None] >= insertion_time)
            & (t[:, None] < unmasking_time)
        )

        xt = torch.where(
            deleted_tokens,
            self.pad_token,  # for deletion, change to pad token
            torch.where(
                masked_tokens,
                self.mask_token,  # for masking, change to mask token
                x1,
            ),
        )

        st = xt.ne(self.pad_token).to(torch.int32).argsort(dim=1, descending=True, stable=True) # edited to sort integers
        xt = torch.gather(xt, 1, st)
        st[xt == self.pad_token] = 0

        return JointInterpolantResult(
            xt=xt, st=st, _x1=x1, _pad_token=self.pad_token, _mask_token=self.mask_token
        )
        
    def sample_interpolant_plan(self, t: Tensor, x1: Tensor) -> JointInterpolantResult:
        """
        Shapes:
            x1: [B, L]
            t: [B]
        Returns:
            xt: [B, L]
            st: [B, L] boolean mask of positions that corresponds to xt
            xt_mask_indices: [B, L] boolean mask of positions that are masked at xt
            x1_remained: [B, L] tokens that are not deleted, used for the training target
            gap_counts: [B, L+1] the number of deleted tokens between xt slots
        """
        # sample the stopping time (B, L, 2)
        insertion_time, unmasking_time = self.hitting_time(t, x1)

        clean_tokens = x1.ne(self.pad_token)
        deleted_tokens = clean_tokens & (t[:, None] < insertion_time)
        masked_tokens = (
            clean_tokens
            & (t[:, None] >= insertion_time)
            & (t[:, None] < unmasking_time)
        )

        xt = torch.where(
            deleted_tokens,
            self.pad_token,  # for deletion, change to pad token
            torch.where(
                masked_tokens,
                self.mask_token,  # for masking, change to mask token
                x1,
            ),
        )
        st = xt.ne(self.pad_token).to(torch.int32).argsort(dim=1, descending=True, stable=True) # edited to sort integers
        xt = torch.gather(xt, 1, st)
        st[xt == self.pad_token] = 0
        num_gaps = (st != 0).sum(dim=1) + 1  # [B]

        deleted_mask = deleted_tokens  # [B, L]
        
        # Create gap assignment tensor: gap_assignment[b, gap_idx, x1_pos] = 1 if x1_pos is in gap gap_idx
        B, L = x1.shape
        max_gaps = L + 1
        gap_assignment = torch.zeros(B, max_gaps, L, device=x1.device, dtype=torch.float)
        
        # For each deleted position in x1, determine which gap it belongs to
        # Gap index = number of non-deleted positions (st values) that come before it
        pos_indices = torch.arange(L, device=x1.device).view(1, L, 1)  # [1, L, 1]
        st_expanded = st.unsqueeze(1)  # [B, 1, L]
        st_valid_mask = (st != 0).unsqueeze(1)  # [B, 1, L]
        
        # Count how many valid st entries are less than each position
        # gap_indices[b, pos] = number of st values < pos for deleted positions
        gap_indices = ((st_expanded < pos_indices) & st_valid_mask).sum(dim=2)  # [B, L]
        
        # Set gap_assignment[b, gap_idx, pos] = 1 where pos is deleted and belongs to gap_idx
        batch_idx = torch.arange(B, device=x1.device).view(B, 1).expand(B, L)
        pos_idx = torch.arange(L, device=x1.device).view(1, L).expand(B, L)
        
        gap_assignment[batch_idx[deleted_mask], gap_indices[deleted_mask], pos_idx[deleted_mask]] = 1.0

        return JointInterpolantResult(
            xt=xt, st=st, _x1=x1, _pad_token=self.pad_token, _mask_token=self.mask_token
        ), deleted_mask, gap_assignment


class MDMInterpolant(JointInterpolant):
    def __init__(
        self,
        unmask_schedule: Schedule,
        vocab_size: int,
        mask_token: int,
        pad_token: int,
        max_length: int,
    ):
        super().__init__(vocab_size, mask_token, pad_token, max_length)
        self.unmask_schedule = unmask_schedule

    def elbo_weight(self, t: Tensor, x1: Tensor):
        """
        Return the ELBO weight for the training, can be changed depends on the empirical results
        there's no weight_delete for the vanilla MDM
        """
        weight_unmask = self.unmask_schedule.rate_scale_factor(t)
        weight_unmask_expanded = weight_unmask.unsqueeze(1).expand(
            -1, x1.shape[1]
        )  # (B,L)
        return weight_unmask_expanded

    def to_actual_rate(self, xt: Tensor, prediction: Tensor, t: Tensor) -> Rate:
        """
        Return the actual rate for the sampling
        """
        token_posterior = F.softmax(prediction, dim=-1)  # (B, L, V)
        unmask_rate = token_posterior * self.unmask_schedule.rate_scale_factor(t).view(
            -1, 1, 1
        )

        return Rate(
            unmask_rate=unmask_rate,  # (B, L, V)
            length_rate=None,  # (B, L+1)
        )

    def sample_interpolant(self, t: Tensor, x1: Tensor) -> JointInterpolantResult:
        # sample the stopping time (B, L, 2)
        eps = 1e-6
        unmask_time = self.unmask_schedule.sample(
            (x1.shape[0], x1.shape[1]), device=x1.device
        )
        unmask_time = unmask_time * (1 - eps) + eps

        xt = torch.where(
            t[:, None] < unmask_time,
            self.mask_token,  # for masking, change to mask token
            x1,
        )
        st = torch.arange(xt.shape[1], device=xt.device, dtype=torch.long).repeat(
            xt.shape[0], 1
        )

        return JointInterpolantResult(
            xt=xt, st=st, _x1=x1, _pad_token=self.pad_token, _mask_token=self.mask_token
        )