File size: 10,739 Bytes
e12111a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Module for sampling t, t_index for diffusion."""
from __future__ import annotations

from dataclasses import dataclass

import chex
import jax.numpy as jnp
import jax.random
from jax import lax


def scatter_add(x: jnp.ndarray, indices: jnp.ndarray, updates: jnp.ndarray) -> jnp.ndarray:
    """Use lax.scatter_add to add value at indices.

    Args:
        x: array to be updated, (n, ).
        indices: index values < n, (batch, ).
        updates: value to be added, (batch, ).

    Returns:
        Updated array x.
    """
    dim_num = lax.ScatterDimensionNumbers(
        update_window_dims=(),
        inserted_window_dims=(0,),
        scatter_dims_to_operand_dims=(0,),
    )
    return lax.scatter_add(
        x,
        scatter_indices=indices[:, None],
        updates=updates,
        dimension_numbers=dim_num,
    )


def scatter_set(x: jnp.ndarray, indices: jnp.ndarray, updates: jnp.ndarray) -> jnp.ndarray:
    """Use lax.scatte to set value at indices.

    When indices are not unique, values at the same index is overwritten
    by the last value.

    Args:
        x: array to be updated, (n, ).
        indices: index values < n, (batch, ).
        updates: value to be added, (batch, ).

    Returns:
        Updated array x.
    """
    dim_num = lax.ScatterDimensionNumbers(
        update_window_dims=(),
        inserted_window_dims=(0,),
        scatter_dims_to_operand_dims=(0,),
    )
    return lax.scatter(
        x,
        scatter_indices=indices[:, None],
        updates=updates,
        dimension_numbers=dim_num,
    )


@dataclass
class TimeSampler:
    """Time sampler for diffusion training.

    Each device has its own time sampler.
    """

    num_timesteps: int  # number of diffusion steps
    uniform_time_sampling: bool  # sample time uniformly
    warmup_steps: int = 10
    decay: float = 0.9
    uniform_prob: float = 0.01

    def t_index_to_t(self, t_index: jnp.ndarray) -> jnp.ndarray:
        """Convert t_index to t.

        t_index = 0 corresponds to t = 1 / num_timesteps.
        t_index = num_timesteps - 1 corresponds to t = 1.

        Args:
            t_index: t_index, shape (batch, ).

        Returns:
            t: t, shape (batch, ).
        """
        return jnp.asarray(t_index + 1, jnp.float32) / self.num_timesteps

    def sample_uniformly(
        self,
        key: jax.Array,
        t_index_minval: jnp.ndarray,
        t_index_maxval: jnp.ndarray,
    ) -> tuple[jnp.ndarray, jnp.ndarray, jnp.ndarray]:
        """Sample noise and time uniformly.

        Args:
            key: random key.
            t_index_minval: minium value of t_index, in [0, num_timesteps),
                shape (batch, ), the values are inclusive.
            t_index_maxval: maximum value of t_index, in [0, num_timesteps),
                shape (batch, ), the values are exclusive.

        Returns:
            t, values between [0, 1), shape (batch, ).
            t_index, values between [0, num_timesteps), shape (batch, ).
            probs_t, probability of sampled t, shape (batch, ).
        """
        t_index = jax.random.randint(
            key=key,
            shape=t_index_minval.shape,
            minval=t_index_minval,  # inclusive
            maxval=t_index_maxval,  # exclusive
        )
        t = self.t_index_to_t(t_index)
        probs_t = jnp.ones_like(t) / self.num_timesteps
        return t, t_index, probs_t

    def t_probs_from_loss_sq(self, loss_sq_hist: jnp.ndarray) -> jnp.ndarray:
        """Get probability of sampling t from loss_sq_hist.

        Args:
            loss_sq_hist: loss_sq_hist, shape (num_timesteps, ).

        Returns:
            probs: probability of sampling each t, shape (num_timesteps, ).
        """
        probs = jnp.sqrt(loss_sq_hist)
        probs /= probs.sum()
        probs = jnp.nan_to_num(probs, copy=False, nan=1.0 / self.num_timesteps)
        probs *= 1 - self.uniform_prob
        probs += self.uniform_prob / self.num_timesteps
        return probs

    def t_probs_from_loss_count(
        self,
        loss_count_hist: jnp.ndarray,
    ) -> jnp.ndarray:
        """Get probability of sampling t from loss_count_hist.

        Args:
            loss_count_hist: loss_count_hist, shape (num_timesteps, ).

        Returns:
            probs: probability of sampling each t, shape (num_timesteps, ).
        """
        probs = self.warmup_steps - loss_count_hist
        probs = jnp.maximum(probs, 0)
        probs /= probs.sum()
        probs = jnp.nan_to_num(probs, copy=False, nan=1.0 / self.num_timesteps)
        return probs

    def sample_with_importance(
        self,
        key: jax.Array,
        t_index_minval: jnp.ndarray,
        t_index_maxval: jnp.ndarray,
        probs: jnp.ndarray,
    ) -> tuple[jnp.ndarray, jnp.ndarray, jnp.ndarray]:
        """Sample noise and time via importance sampling.

        https://arxiv.org/abs/2102.09672
        https://github.com/microsoft/muzic/blob/61e4436516c61b6b358ef709b446de539817decf/getmusic/getmusic/modeling/roformer/diffusion_roformer.py#L376
        https://github.com/openai/improved-diffusion/blob/783b6740edb79fdb7d063250db2c51cc9545dcd1/improved_diffusion/resample.py#L70

        Args:
            key: random key.
            t_index_minval: minium value of t_index, in [0, num_timesteps),
                shape (batch, ), the values are inclusive.
            t_index_maxval: maximum value of t_index, in [0, num_timesteps),
                shape (batch, ), the values are exclusive.
            probs: probability of sampling each t, shape (num_timesteps, ).

        Returns:
            t, values between [0, 1), shape (batch, ).
            t_index, values between [0, num_timesteps), shape (batch, ).
            probs_t, probability of sampled t, shape (batch, ).
        """
        batch_size = t_index_minval.shape[0]

        # extend the probs to the batch size
        # (batch_size, num_timesteps)
        probs_batch = jnp.tile(probs, (batch_size, 1))

        # mask out the timesteps out of range
        # and renomalize the probs
        mask = jnp.tile(jnp.arange(self.num_timesteps), (batch_size, 1))
        mask = jnp.logical_and(
            mask >= t_index_minval[:, None],
            mask < t_index_maxval[:, None],
        )
        probs_batch = jnp.where(mask, probs_batch, 0.0)
        probs_batch /= probs_batch.sum(axis=1, keepdims=True)

        # sample timesteps
        logits = jnp.log(probs_batch)
        t_index = jax.random.categorical(
            key=key,
            logits=logits,
            shape=(batch_size,),
        )
        t = self.t_index_to_t(t_index)
        probs_t = probs[t_index]

        return t, t_index, probs_t

    def sample(
        self,
        key: jax.Array,
        batch_size: int,
        t_index_min: int,
        t_index_max: int,
        loss_count_hist: jnp.ndarray,
        loss_sq_hist: jnp.ndarray,
    ) -> tuple[jnp.ndarray, jnp.ndarray, jnp.ndarray]:
        """Sample noise and time.

        Each time step in consideration should be sampled enough.

        Args:
            key: random key.
            batch_size: batch size.
            t_index_min: minium value of t_index, in [0, num_timesteps),
                the values are inclusive.
            t_index_max: maximum value of t_index, in [0, num_timesteps),
                the values are exclusive.
            loss_count_hist: count of time steps, shape (num_timesteps, ).
            loss_sq_hist: weighted average of squared loss, shape (num_timesteps, ).

        Returns:
            t, values between [0, 1), shape (batch, ).
            t_index, values between [0, num_timesteps), shape (batch, ).
            probs_t, probability of sampled t, shape (batch, ).
        """
        if t_index_min > t_index_max:
            raise ValueError(f"t_index_min {t_index_min} > t_index_max {t_index_max}.")
        if t_index_min < 0:
            raise ValueError(f"t_index_min {t_index_min} < 0.")
        if t_index_max > self.num_timesteps:
            raise ValueError(f"t_index_max {t_index_max} > num_timesteps {self.num_timesteps}.")

        t_index_minval = jnp.full((batch_size,), t_index_min, dtype=jnp.int32)
        t_index_maxval = jnp.full((batch_size,), t_index_max, dtype=jnp.int32)
        if self.uniform_time_sampling:
            return self.sample_uniformly(
                key=key,
                t_index_minval=t_index_minval,
                t_index_maxval=t_index_maxval,
            )

        chex.assert_shape(loss_count_hist, (self.num_timesteps,))
        chex.assert_shape(loss_sq_hist, (self.num_timesteps,))
        min_loss_count_hist = jnp.min(loss_count_hist[t_index_min:t_index_max])

        probs = lax.select(
            min_loss_count_hist < self.warmup_steps,
            self.t_probs_from_loss_count(loss_count_hist=loss_count_hist),
            self.t_probs_from_loss_sq(loss_sq_hist=loss_sq_hist),
        )
        t, t_index, probs_t = self.sample_with_importance(
            key, t_index_minval, t_index_maxval, probs
        )
        # during warmup, although we sample with non-uniform probability,
        # this is to ensure to sample every timestep enough times
        # so the probs_t will be overwritten by uniform probability
        probs_t = lax.select(
            min_loss_count_hist < self.warmup_steps,
            jnp.ones_like(probs_t) / self.num_timesteps,
            probs_t,
        )
        return t, t_index, probs_t

    def update_stats(
        self,
        loss_batch: jnp.ndarray,
        t_index: jnp.ndarray,
        loss_count_hist: jnp.ndarray,
        loss_sq_hist: jnp.ndarray,
    ) -> tuple[jnp.ndarray, jnp.ndarray]:
        """Update the loss_sq_hist and loss_count_hist.

        Args:
            loss_batch: loss of the current batch, shape (batch, ).
            t_index: t_index of the current batch, shape (batch, ).
            loss_count_hist: count of time steps, shape (num_timesteps, ).
            loss_sq_hist: weighted average of squared loss, shape (num_timesteps, ).

        Returns:
            loss_count_hist: updated loss_count_hist, shape (num_timesteps, ).
            loss_sq_hist: updated loss_sq_hist, shape (num_timesteps, ).
        """
        # (batch, )
        loss_sq_batch = loss_batch**2
        # (batch, )
        loss_sq_prev = loss_sq_hist[t_index]
        loss_sq_updated = (1 - self.decay) * loss_sq_batch + self.decay * loss_sq_prev
        loss_sq_hist = scatter_set(x=loss_sq_hist, indices=t_index, updates=loss_sq_updated)
        loss_count_hist = scatter_add(
            x=loss_count_hist,
            indices=t_index,
            updates=jnp.ones_like(t_index),
        )
        return loss_count_hist, loss_sq_hist