File size: 7,854 Bytes
f9740ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import abc
import torch
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
from torch.cuda.amp import custom_fwd, custom_bwd


from catsample import sample_categorical

def get_graph(config, device):
    if config.graph.type == "uniform":
        return Uniform(config.tokens)
    elif config.graph.type == "absorb":
        return Absorbing(config.tokens)
    else:
        raise ValueError(f"Graph {config.graph.type} not valid")


def unsqueeze_as(x, y, back=True):
    if back:
        return x.view(*x.shape, *((1,) * (len(y.shape) - len(x.shape))))
    else:
        return x.view(*((1,) * (len(y.shape) - len(x.shape))), *x.shape)


class Graph(abc.ABC):

    @property
    def dim(self):
        pass

    @property
    def absorb(self):
        """
        Whether input {dim - 1} is an absorbing state (used for denoising to always remove the mask).
        """
        pass


    @abc.abstractmethod
    def rate(self, i):
        """
        Computes the i-th column of the rate matrix Q, where i is [B_1, ..., B_n].

        This is intended to compute the "forward" rate of p(X_t | X_0 = i).
        """
        pass


    @abc.abstractmethod
    def transp_rate(self, i):
        """
        Computes the i-th row of the rate matrix Q.

        Can be used to compute the reverse rate.
        """
        pass


    @abc.abstractmethod
    def transition(self, i, sigma):
        """
        Computes the i-th column of the transition matrix e^{sigma Q}.
        """
        pass


    def sample_transition(self, i, sigma):
        """
        Samples the transition vector.
        """
        transition_vector = self.transition(i, sigma)
        return sample_categorical(transition_vector, method="hard")
    

    def reverse_rate(self, i, score):
        """
        Constructs the reverse rate. Which is score * transp_rate
        """
        normalized_rate = self.transp_rate(i) * score

        normalized_rate.scatter_(-1, i[..., None], torch.zeros_like(normalized_rate))
        normalized_rate.scatter_(-1, i[..., None], -normalized_rate.sum(dim=-1, keepdim=True))
        return normalized_rate

    def sample_rate(self, i, rate):
        return sample_categorical(F.one_hot(i, num_classes=self.dim).to(rate) + rate)

    
    @abc.abstractmethod
    def staggered_score(self, score, dsigma):
        """
        Computes p_{sigma - dsigma}(z) / p_{sigma}(x), which is approximated with
        e^{-{dsigma} E} score
        """
        pass
    

    @abc.abstractmethod
    def sample_limit(self, *batch_dims):
        """
        Sample the limiting distribution. Returns the probability vector as well.
        """
        pass


    @abc.abstractmethod
    def score_entropy(self, score, sigma, x, x0):
        """
        Computes the score entropy function (with requisite constant normalization)
        """
        pass


class Uniform(Graph):
    """
    Everything goes to everything else. Normalized down by dimension to avoid blowup.
    """
    def __init__(self, dim):
        self._dim = dim

    @property
    def dim(self):
        return self._dim
    
    @property
    def absorb(self):
        return False

    def rate(self, i):
        edge = torch.ones(*i.shape, self.dim, device=i.device) / self.dim
        edge = edge.scatter(-1, i[..., None], - (self.dim - 1) / self.dim)
        return edge

    def transp_rate(self, i):
        return self.rate(i)

    def transition(self, i, sigma):
        trans = torch.ones(*i.shape, self.dim, device=i.device) * (1 - (-sigma[..., None]).exp()) / self.dim
        trans = trans.scatter(-1, i[..., None], torch.zeros_like(trans))
        trans = trans.scatter(-1, i[..., None], 1 - trans.sum(dim=-1, keepdim=True))
        return trans
    
    def transp_transition(self, i, sigma):
        return self.transition(i, sigma)

    def sample_transition(self, i, sigma):
        move_chance = 1 - (-sigma).exp()
        move_indices = torch.rand(*i.shape, device=i.device) < move_chance
        i_pert = torch.where(move_indices, torch.randint_like(i, self.dim), i)
        return i_pert

    def staggered_score(self, score, dsigma):
        dim = score.shape[-1]
        epow = (-dsigma).exp()[..., None]
        return ((epow - 1) / (dim * epow)) * score.sum(dim=-1, keepdim=True) + score / epow

    def sample_limit(self, *batch_dims):
        return torch.randint(0, self.dim, batch_dims)

    def score_entropy(self, score, sigma, x, x0):
        esigm1 = torch.where(
            sigma < 0.5,
            torch.expm1(sigma),
            torch.exp(sigma) - 1
        )
        ratio = 1 - self.dim / (esigm1 + self.dim)

        # negative term
        neg_term = score.mean(dim=-1) - torch.gather(score, -1, x[..., None]).squeeze(-1) / self.dim
        # no move means scaling by the uniform ratio. move means alter only one ratio away from 1
        neg_term = torch.where(
            x == x0,
            ratio * neg_term,
            torch.gather(score, -1, x0[..., None]).squeeze(-1) / esigm1 + neg_term
        )

        # constant factor
        const = torch.where(
            x == x0,
            (self.dim - 1) / self.dim * ratio * (ratio.log() - 1),
            ((-ratio.log() - 1) / ratio - (self.dim - 2)) / self.dim 
        )

        #positive term
        sexp = score.exp()
        pos_term = sexp.mean(dim=-1) - torch.gather(sexp, -1, x[..., None]).squeeze(-1) / self.dim
        return pos_term - neg_term + const


class Absorbing(Graph):
    def __init__(self, dim):
        super().__init__()
        self._dim = dim

    @property
    def dim(self):
        return self._dim + 1
    
    @property
    def absorb(self):
        return True

    def rate(self, i):
        # edge = - F.one_hot(i, num_classes=self.dim)
        # edge.scatter_add_(-1, i[..., None], torch.ones_like(edge[..., :1]))
        return F.one_hot((self.dim - 1) * torch.ones_like(i), num_classes=self.dim) - F.one_hot(i, num_classes=self.dim)        

    def transp_rate(self, i):
        edge = -F.one_hot(i, num_classes=self.dim)
        edge[i == self.dim - 1] += 1
        return edge

    def transition(self, i, sigma):
        pass
    
    def transp_transition(self, i, sigma):
        sigma = unsqueeze_as(sigma, i[..., None])
        edge = (-sigma).exp() * F.one_hot(i, num_classes=self.dim)
        edge += torch.where(
            i == self.dim - 1,
            1 - (-sigma).squeeze(-1).exp(),
            0
        )[..., None]
        return edge

    def sample_transition(self, i, sigma):
        move_chance = 1 - (-sigma).exp()
        move_indices = torch.rand(*i.shape, device=i.device) < move_chance
        i_pert = torch.where(move_indices, self.dim - 1, i)
        return i_pert
    
    def staggered_score(self, score, dsigma):
        score = score.clone() # yeah yeah whatever we should probably do this
        extra_const = (1 - (dsigma).exp()) * score.sum(dim=-1)
        score *= dsigma.exp()[:, None]
        score[..., -1] += extra_const
        return score

    def sample_limit(self, *batch_dims):
        return (self.dim - 1) * torch.ones(*batch_dims, dtype=torch.int64)

    def score_entropy(self, score, sigma, x, x0):
        rel_ind = x == self.dim - 1
        esigm1 = torch.where(
            sigma < 0.5,
            torch.expm1(sigma),
            torch.exp(sigma) - 1
        )

        ratio = 1 / esigm1.expand_as(x)[rel_ind]
        other_ind = x0[rel_ind]

        # negative_term
        neg_term = ratio * torch.gather(score[rel_ind], -1, other_ind[..., None]).squeeze(-1)

        #positive term
        pos_term = score[rel_ind][:, :-1].exp().sum(dim=-1)

        # constant term
        const = ratio * (ratio.log() - 1)

        entropy = torch.zeros(*x.shape, device=x.device)
        entropy[rel_ind] += pos_term - neg_term + const
        return entropy