File size: 8,831 Bytes
8026e91
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import numpy as np
import torch as t
import torch.nn as nn

from jukebox.vqvae.encdec import Encoder, Decoder, assert_shape
from jukebox.vqvae.bottleneck import NoBottleneck, Bottleneck
from jukebox.utils.logger import average_metrics
from jukebox.utils.audio_utils import spectral_convergence, spectral_loss, multispectral_loss, audio_postprocess

def dont_update(params):
    for param in params:
        param.requires_grad = False

def update(params):
    for param in params:
        param.requires_grad = True

def calculate_strides(strides, downs):
    return [stride ** down for stride, down in zip(strides, downs)]

def _loss_fn(loss_fn, x_target, x_pred, hps):
    if loss_fn == 'l1':
        return t.mean(t.abs(x_pred - x_target)) / hps.bandwidth['l1']
    elif loss_fn == 'l2':
        return t.mean((x_pred - x_target) ** 2) / hps.bandwidth['l2']
    elif loss_fn == 'linf':
        residual = ((x_pred - x_target) ** 2).reshape(x_target.shape[0], -1)
        values, _ = t.topk(residual, hps.linf_k, dim=1)
        return t.mean(values) / hps.bandwidth['l2']
    elif loss_fn == 'lmix':
        loss = 0.0
        if hps.lmix_l1:
            loss += hps.lmix_l1 * _loss_fn('l1', x_target, x_pred, hps)
        if hps.lmix_l2:
            loss += hps.lmix_l2 * _loss_fn('l2', x_target, x_pred, hps)
        if hps.lmix_linf:
            loss += hps.lmix_linf * _loss_fn('linf', x_target, x_pred, hps)
        return loss
    else:
        assert False, f"Unknown loss_fn {loss_fn}"

class VQVAE(nn.Module):
    def __init__(self, input_shape, levels, downs_t, strides_t,
                 emb_width, l_bins, mu, commit, spectral, multispectral,
                 multipliers=None, use_bottleneck=True, **block_kwargs):
        super().__init__()

        self.sample_length = input_shape[0]
        x_shape, x_channels = input_shape[:-1], input_shape[-1]
        self.x_shape = x_shape

        self.downsamples = calculate_strides(strides_t, downs_t)
        self.hop_lengths = np.cumprod(self.downsamples)
        self.z_shapes = z_shapes = [(x_shape[0] // self.hop_lengths[level],) for level in range(levels)]
        self.levels = levels

        if multipliers is None:
            self.multipliers = [1] * levels
        else:
            assert len(multipliers) == levels, "Invalid number of multipliers"
            self.multipliers = multipliers
        def _block_kwargs(level):
            this_block_kwargs = dict(block_kwargs)
            this_block_kwargs["width"] *= self.multipliers[level]
            this_block_kwargs["depth"] *= self.multipliers[level]
            return this_block_kwargs

        encoder = lambda level: Encoder(x_channels, emb_width, level + 1,
                                        downs_t[:level+1], strides_t[:level+1], **_block_kwargs(level))
        decoder = lambda level: Decoder(x_channels, emb_width, level + 1,
                                        downs_t[:level+1], strides_t[:level+1], **_block_kwargs(level))
        self.encoders = nn.ModuleList()
        self.decoders = nn.ModuleList()
        for level in range(levels):
            self.encoders.append(encoder(level))
            self.decoders.append(decoder(level))

        if use_bottleneck:
            self.bottleneck = Bottleneck(l_bins, emb_width, mu, levels)
        else:
            self.bottleneck = NoBottleneck(levels)

        self.downs_t = downs_t
        self.strides_t = strides_t
        self.l_bins = l_bins
        self.commit = commit
        self.spectral = spectral
        self.multispectral = multispectral

    def preprocess(self, x):
        # x: NTC [-1,1] -> NCT [-1,1]
        assert len(x.shape) == 3
        x = x.permute(0,2,1).float()
        return x

    def postprocess(self, x):
        # x: NTC [-1,1] <- NCT [-1,1]
        x = x.permute(0,2,1)
        return x

    def _decode(self, zs, start_level=0, end_level=None):
        # Decode
        if end_level is None:
            end_level = self.levels
        assert len(zs) == end_level - start_level
        xs_quantised = self.bottleneck.decode(zs, start_level=start_level, end_level=end_level)
        assert len(xs_quantised) == end_level - start_level

        # Use only lowest level
        decoder, x_quantised = self.decoders[start_level], xs_quantised[0:1]
        x_out = decoder(x_quantised, all_levels=False)
        x_out = self.postprocess(x_out)
        return x_out

    def decode(self, zs, start_level=0, end_level=None, bs_chunks=1):
        z_chunks = [t.chunk(z, bs_chunks, dim=0) for z in zs]
        x_outs = []
        for i in range(bs_chunks):
            zs_i = [z_chunk[i] for z_chunk in z_chunks]
            x_out = self._decode(zs_i, start_level=start_level, end_level=end_level)
            x_outs.append(x_out)
        return t.cat(x_outs, dim=0)

    def _encode(self, x, start_level=0, end_level=None):
        # Encode
        if end_level is None:
            end_level = self.levels
        x_in = self.preprocess(x)
        xs = []
        for level in range(self.levels):
            encoder = self.encoders[level]
            x_out = encoder(x_in)
            xs.append(x_out[-1])
        zs = self.bottleneck.encode(xs)
        return zs[start_level:end_level]

    def encode(self, x, start_level=0, end_level=None, bs_chunks=1):
        x_chunks = t.chunk(x, bs_chunks, dim=0)
        zs_list = []
        for x_i in x_chunks:
            zs_i = self._encode(x_i, start_level=start_level, end_level=end_level)
            zs_list.append(zs_i)
        zs = [t.cat(zs_level_list, dim=0) for zs_level_list in zip(*zs_list)]
        return zs

    def sample(self, n_samples):
        zs = [t.randint(0, self.l_bins, size=(n_samples, *z_shape), device='cuda') for z_shape in self.z_shapes]
        return self.decode(zs)

    def forward(self, x, hps, loss_fn='l1'):
        metrics = {}

        N = x.shape[0]

        # Encode/Decode
        x_in = self.preprocess(x)
        xs = []
        for level in range(self.levels):
            encoder = self.encoders[level]
            x_out = encoder(x_in)
            xs.append(x_out[-1])

        zs, xs_quantised, commit_losses, quantiser_metrics = self.bottleneck(xs)
        x_outs = []
        for level in range(self.levels):
            decoder = self.decoders[level]
            x_out = decoder(xs_quantised[level:level+1], all_levels=False)
            assert_shape(x_out, x_in.shape)
            x_outs.append(x_out)

        # Loss
        def _spectral_loss(x_target, x_out, hps):
            if hps.use_nonrelative_specloss:
                sl = spectral_loss(x_target, x_out, hps) / hps.bandwidth['spec']
            else:
                sl = spectral_convergence(x_target, x_out, hps)
            sl = t.mean(sl)
            return sl

        def _multispectral_loss(x_target, x_out, hps):
            sl = multispectral_loss(x_target, x_out, hps) / hps.bandwidth['spec']
            sl = t.mean(sl)
            return sl

        recons_loss = t.zeros(()).to(x.device)
        spec_loss = t.zeros(()).to(x.device)
        multispec_loss = t.zeros(()).to(x.device)
        x_target = audio_postprocess(x.float(), hps)

        for level in reversed(range(self.levels)):
            x_out = self.postprocess(x_outs[level])
            x_out = audio_postprocess(x_out, hps)
            this_recons_loss = _loss_fn(loss_fn, x_target, x_out, hps)
            this_spec_loss = _spectral_loss(x_target, x_out, hps)
            this_multispec_loss = _multispectral_loss(x_target, x_out, hps)
            metrics[f'recons_loss_l{level + 1}'] = this_recons_loss
            metrics[f'spectral_loss_l{level + 1}'] = this_spec_loss
            metrics[f'multispectral_loss_l{level + 1}'] = this_multispec_loss
            recons_loss += this_recons_loss
            spec_loss += this_spec_loss
            multispec_loss += this_multispec_loss

        commit_loss = sum(commit_losses)
        loss = recons_loss + self.spectral * spec_loss + self.multispectral * multispec_loss + self.commit * commit_loss

        with t.no_grad():
            sc = t.mean(spectral_convergence(x_target, x_out, hps))
            l2_loss = _loss_fn("l2", x_target, x_out, hps)
            l1_loss = _loss_fn("l1", x_target, x_out, hps)
            linf_loss = _loss_fn("linf", x_target, x_out, hps)

        quantiser_metrics = average_metrics(quantiser_metrics)

        metrics.update(dict(
            recons_loss=recons_loss,
            spectral_loss=spec_loss,
            multispectral_loss=multispec_loss,
            spectral_convergence=sc,
            l2_loss=l2_loss,
            l1_loss=l1_loss,
            linf_loss=linf_loss,
            commit_loss=commit_loss,
            **quantiser_metrics))

        for key, val in metrics.items():
            metrics[key] = val.detach()

        return x_out, loss, metrics