File size: 12,387 Bytes
e3061ad
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

from omegaconf import DictConfig
from . import builders, musicgen
from einops import rearrange
from torch.nn import functional as F
from ..modules.conditioners import SegmentWithAttributes

import torch
import numpy as np
import random
import typing as tp
import math
import flashy


class MagnetSolver(musicgen.MusicGenSolver):
    """Solver for MAGNeT - Masked Audio Generation using
        a single Non-autoregressive Transformer https://arxiv.org/abs/2401.04577.
    """
    def __init__(self, cfg: DictConfig):
        super().__init__(cfg)

        # initialize generation parameters by config
        self.generation_params = {
            'use_sampling': self.cfg.generate.lm.use_sampling,
            'temp': self.cfg.generate.lm.temp,
            'top_k': self.cfg.generate.lm.top_k,
            'top_p': self.cfg.generate.lm.top_p,
            'max_cfg_coef': self.cfg.generate.lm.max_cfg_coef,
            'min_cfg_coef': self.cfg.generate.lm.min_cfg_coef,
            'decoding_steps': list(self.cfg.generate.lm.decoding_steps),
            'anneal_temp': self.cfg.generate.lm.anneal_temp,
            'span_scoring': self.cfg.generate.lm.span_scoring,
            'span_arrangement': self.cfg.generate.lm.span_arrangement
        }

        sequence_len = int(cfg.dataset.segment_duration * self.compression_model.frame_rate)
        self.mean_maskrate_to_u = torch.tensor(self._calc_mean_maskrate_to_u_LUT(sequence_len), device=self.device)
        self.ce_per_codebook = [torch.log(torch.tensor(self.compression_model.cardinality, device=self.device))
                                for _ in range(cfg.transformer_lm.n_q)]

    def build_model(self) -> None:
        self.cfg.transformer_lm.segment_duration = self.cfg.dataset.segment_duration
        self.cfg.transformer_lm.span_len = self.cfg.masking.span_len
        assert self.cfg.efficient_attention_backend == "xformers", "MAGNeT v1 models support only xformers backend."
        super().build_model()

    def _calc_mean_maskrate_to_u_LUT(self, T: int):
        """ Create a Look Up Table (LUT) transforming a discrete masking percentage m in 0,1,...,100 to u,
            the number of overlapping spans of length L to place s.t. the masking rate is approximately m/float(100).
            It first creates the inverse transformation, of the masking rate as function of u,
            using the expression choose(T - L, u) / choose(T, u), where L is the atomic span length used
            during masking. See https://arxiv.org/abs/2401.04577,
            appendix C, for the mean mask rate derivation.

            We leverage the fact that:
                                choose(T - L, u) / choose(T, u) = Prod_{j = 0}^{u - 1}((T - L - j)/(T - j))
            in the provided implementation, in order to avoid overflow.
        Args:
            T (float): Sequence length.
        Returns:
            (List) A LUT transforming m in 0,1,...,100 to u,
            s.t. the masking rate of the span-L mask is approximately m/float(100).
        """

        L = self.cfg.masking.span_len

        u2mean = [0.0]  # mean mask rate is 0.0 for u = 0
        v = (T - L) / float(T)
        for u in range(1, T):
            u2mean.append(1 - v)
            v *= (T - L - u) / (T - u)  # Overflow-safe implementation of choose(T - L, u) / choose(T, u).

        mean2u = []
        for maskperc in range(101):
            maskrate = maskperc / float(100)
            u = int(np.searchsorted(u2mean, maskrate))
            mean2u.append(u)

        return mean2u

    def _non_spans_mask(self, mask_probs: torch.Tensor, B: int, T: int, device: torch.device) -> torch.Tensor:
        """ Construct a boolean mask of shape [B, T, 1], with masking rates defined by mask_probs.
            The masked tokens are singletons, placed uniformly at random.
        Args:
            mask_probs (torch.Tensor): The desired masking rate per sample, of shape [B,]
            B (int): Batch size.
            T (int): Sequence length.
            device (torch.device): device of the output tensor
        Returns:
            (torch.Tensor): A mask of shape [B, T]
        """
        num_token_masked = (T * mask_probs).round().clamp(min=1)
        batch_randperm = torch.rand((B, T), device=device).argsort(dim=-1)
        return batch_randperm < rearrange(num_token_masked, 'b -> b 1')

    def _spans_mask(self, mask_probs: torch.Tensor, B: int, T: int, device: torch.device) -> torch.Tensor:
        """ Construct a spans mask with masking rates defined by mask_probs,
            where the atomic span length ( > 1 ) is defined by cfg.masking.span_len.
        Args:
            mask_probs (torch.Tensor): The desired masking rate per sample, of shape [B,]
            B (int): Batch size.
            T (int): Sequence length.
            device (torch.device): device of the output tensor
        Returns:
            (torch.Tensor): A spans mask of shape [B, T]
        """
        rounded_probs = torch.round(100 * mask_probs).long()
        k = self.mean_maskrate_to_u[rounded_probs].clamp(min=1)  # k is the number of span starts

        # sample random span starts
        batch_randperm = torch.rand((B, T), device=device).argsort(dim=-1)
        mask = batch_randperm < rearrange(k, 'b -> b 1')
        B, T = mask.shape
        shifted_mask = mask.clone()
        for _ in range(self.cfg.masking.span_len - 1):
            shifted_mask = torch.concat((torch.full((B, 1), False, device=device), shifted_mask[:, :-1]), dim=1)
            mask = torch.logical_or(mask, shifted_mask)

        return mask

    def _get_mask(self, mask_probs: torch.Tensor, B: int, T: int, device: torch.device) -> torch.Tensor:
        """ Construct a boolean mask with masking rates defined by mask_probs, and atomic
            span length defined by cfg.masking.span_len.
        Args:
            mask_probs (torch.Tensor): The desired masking rate per sample, of shape [B,]
            B (int): Batch size.
            T (int): Sequence length.
            device (torch.device): device of the output tensor
        Returns:
            (torch.Tensor): A boolean tensor of shape [B, T]
        """
        if self.cfg.masking.span_len <= 1:
            return self._non_spans_mask(mask_probs, B, T, device)

        return self._spans_mask(mask_probs, B, T, device)

    def _compute_cross_entropy_magnet(self, logits: torch.Tensor,
                                      targets: torch.Tensor, mask: torch.Tensor, stage: torch.Tensor) -> torch.Tensor:
        """ Compute cross entropy between multi-codebook targets and model's logits.
        The cross entropy is computed only on a specific codebook, defined by the stage argument.
        Valid timesteps for each codebook are pulled from the mask, where invalid
        timesteps are set to 0.

        Args:
            logits (torch.Tensor): Model's logits of shape [B, K, T, card].
            targets (torch.Tensor): Target codes, of shape [B, K, T].
            mask (torch.Tensor): Mask for valid target codes, of shape [B, K, T].
            stage (torch.Tensor): The codebook (idx) that is being optimized, as a scalar tensor.
        Returns:
            ce (torch.Tensor): Cross entropy of the codebook that is being optimized.
        """
        assert logits.shape[:-1] == targets.shape
        assert mask.shape == targets.shape
        ce = torch.zeros([], device=targets.device)
        logits_k = logits[:, stage, ...].contiguous().view(-1, logits.size(-1))  # [B x T, card]
        targets_k = targets[:, stage, ...].contiguous().view(-1)  # [B x T]
        mask_k = mask[:, stage, ...].contiguous().view(-1)  # [B x T]

        IGNORE_IDX = -1
        targets_k[~mask_k] = IGNORE_IDX
        q_ce = F.cross_entropy(logits_k, targets_k, ignore_index=IGNORE_IDX)

        ce += q_ce
        return ce

    def run_step(self, idx: int, batch: tp.Tuple[torch.Tensor, tp.List[SegmentWithAttributes]], metrics: dict) -> dict:
        """Perform one training or valid step on a given batch."""
        check_synchronization_points = idx == 1 and self.device == 'cuda'

        condition_tensors, audio_tokens, padding_mask = self._prepare_tokens_and_attributes(
            batch, check_synchronization_points)

        self.deadlock_detect.update('tokens_and_conditions')

        if check_synchronization_points:
            torch.cuda.set_sync_debug_mode('warn')

        B, K, T = audio_tokens.shape
        device = self.device

        # Choose the stage (codebook idx) for update, uniformly at random.
        stage_ = random.randint(0, K - 1)
        stage = torch.full((1, ), stage_, device=device)

        # masking
        rand_time = torch.zeros((B,), device=device).float().uniform_(0, 1)
        rand_mask_probs = torch.cos(rand_time * math.pi * 0.5)

        # stage mask
        stage_mask = self._get_mask(rand_mask_probs, B, T, device)  # [B, T]
        stage_mask = stage_mask.unsqueeze(1)  # [B, 1, T]

        # Keep all preceding codebooks.
        mask = torch.full((B, K, T), False, device=device)
        mask[:, stage, :] = stage_mask

        # Mask all codebooks larger than stage_
        mask_id = self.model.special_token_id
        mask[:, (stage_+1):, :] = torch.full((B, K - stage_ - 1, T), True, device=device)
        input_tokens = torch.where(mask, mask_id, audio_tokens)

        # Take loss only on the chosen stage, and only on the masked tokens.
        loss_mask = torch.full((B, K, T), False, device=device)
        loss_mask[:, stage, :] = stage_mask

        with self.autocast:
            model_output = self.model.compute_predictions(input_tokens, [], condition_tensors, stage=stage_)
            logits = model_output.logits
            loss_mask &= padding_mask
            ce = self._compute_cross_entropy_magnet(logits, audio_tokens, loss_mask, stage)
            loss = ce
        self.deadlock_detect.update('loss')

        if check_synchronization_points:
            torch.cuda.set_sync_debug_mode('default')

        if self.is_training:
            metrics['lr'] = self.optimizer.param_groups[0]['lr']
            if self.scaler is not None:
                loss = self.scaler.scale(loss)
            self.deadlock_detect.update('scale')
            if self.cfg.fsdp.use:
                loss.backward()
                flashy.distrib.average_tensors(self.model.buffers())
            elif self.cfg.optim.eager_sync:
                with flashy.distrib.eager_sync_model(self.model):
                    loss.backward()
            else:
                # this should always be slower but can be useful
                # for weird use cases like multiple backwards.
                loss.backward()
                flashy.distrib.sync_model(self.model)
            self.deadlock_detect.update('backward')

            if self.scaler is not None:
                self.scaler.unscale_(self.optimizer)
            if self.cfg.optim.max_norm:
                if self.cfg.fsdp.use:
                    metrics['grad_norm'] = self.model.clip_grad_norm_(self.cfg.optim.max_norm)  # type: ignore
                else:
                    metrics['grad_norm'] = torch.nn.utils.clip_grad_norm_(
                        self.model.parameters(), self.cfg.optim.max_norm
                    )
            if self.scaler is None:
                self.optimizer.step()
            else:
                self.scaler.step(self.optimizer)
                self.scaler.update()
            if self.lr_scheduler:
                self.lr_scheduler.step()
            self.optimizer.zero_grad()
            self.deadlock_detect.update('optim')
            if self.scaler is not None:
                scale = self.scaler.get_scale()
                metrics['grad_scale'] = scale
            if not loss.isfinite().all():
                raise RuntimeError("Model probably diverged.")

        metrics['ce'] = ce
        metrics['ppl'] = torch.exp(ce)

        return metrics


class AudioMagnetSolver(MagnetSolver):
    """Solver for audio-MAGNeT. A MAGNeT model for sound generation.

    More information can be found in the MAGNeT model card.
    """
    DATASET_TYPE: builders.DatasetType = builders.DatasetType.SOUND