File size: 12,756 Bytes
1504958
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from typing import Optional, Union

import torch
import inspect
import math
import torch.nn as nn
from diffusers import ConfigMixin, ModelMixin
from diffusers.models.autoencoders.vae import (
    DecoderOutput,
    DiagonalGaussianDistribution,
)
from diffusers.models.modeling_outputs import AutoencoderKLOutput
from xora.models.autoencoders.conv_nd_factory import make_conv_nd


class AutoencoderKLWrapper(ModelMixin, ConfigMixin):
    """Variational Autoencoder (VAE) model with KL loss.

    VAE from the paper Auto-Encoding Variational Bayes by Diederik P. Kingma and Max Welling.
    This model is a wrapper around an encoder and a decoder, and it adds a KL loss term to the reconstruction loss.

    Args:
        encoder (`nn.Module`):
            Encoder module.
        decoder (`nn.Module`):
            Decoder module.
        latent_channels (`int`, *optional*, defaults to 4):
            Number of latent channels.
    """

    def __init__(
        self,
        encoder: nn.Module,
        decoder: nn.Module,
        latent_channels: int = 4,
        dims: int = 2,
        sample_size=512,
        use_quant_conv: bool = True,
    ):
        super().__init__()

        # pass init params to Encoder
        self.encoder = encoder
        self.use_quant_conv = use_quant_conv

        # pass init params to Decoder
        quant_dims = 2 if dims == 2 else 3
        self.decoder = decoder
        if use_quant_conv:
            self.quant_conv = make_conv_nd(
                quant_dims, 2 * latent_channels, 2 * latent_channels, 1
            )
            self.post_quant_conv = make_conv_nd(
                quant_dims, latent_channels, latent_channels, 1
            )
        else:
            self.quant_conv = nn.Identity()
            self.post_quant_conv = nn.Identity()
        self.use_z_tiling = False
        self.use_hw_tiling = False
        self.dims = dims
        self.z_sample_size = 1

        self.decoder_params = inspect.signature(self.decoder.forward).parameters

        # only relevant if vae tiling is enabled
        self.set_tiling_params(sample_size=sample_size, overlap_factor=0.25)

    def set_tiling_params(self, sample_size: int = 512, overlap_factor: float = 0.25):
        self.tile_sample_min_size = sample_size
        num_blocks = len(self.encoder.down_blocks)
        self.tile_latent_min_size = int(sample_size / (2 ** (num_blocks - 1)))
        self.tile_overlap_factor = overlap_factor

    def enable_z_tiling(self, z_sample_size: int = 8):
        r"""
        Enable tiling during VAE decoding.

        When this option is enabled, the VAE will split the input tensor in tiles to compute decoding in several
        steps. This is useful to save some memory and allow larger batch sizes.
        """
        self.use_z_tiling = z_sample_size > 1
        self.z_sample_size = z_sample_size
        assert (
            z_sample_size % 8 == 0 or z_sample_size == 1
        ), f"z_sample_size must be a multiple of 8 or 1. Got {z_sample_size}."

    def disable_z_tiling(self):
        r"""
        Disable tiling during VAE decoding. If `use_tiling` was previously invoked, this method will go back to computing
        decoding in one step.
        """
        self.use_z_tiling = False

    def enable_hw_tiling(self):
        r"""
        Enable tiling during VAE decoding along the height and width dimension.
        """
        self.use_hw_tiling = True

    def disable_hw_tiling(self):
        r"""
        Disable tiling during VAE decoding along the height and width dimension.
        """
        self.use_hw_tiling = False

    def _hw_tiled_encode(self, x: torch.FloatTensor, return_dict: bool = True):
        overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor))
        blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor)
        row_limit = self.tile_latent_min_size - blend_extent

        # Split the image into 512x512 tiles and encode them separately.
        rows = []
        for i in range(0, x.shape[3], overlap_size):
            row = []
            for j in range(0, x.shape[4], overlap_size):
                tile = x[
                    :,
                    :,
                    :,
                    i : i + self.tile_sample_min_size,
                    j : j + self.tile_sample_min_size,
                ]
                tile = self.encoder(tile)
                tile = self.quant_conv(tile)
                row.append(tile)
            rows.append(row)
        result_rows = []
        for i, row in enumerate(rows):
            result_row = []
            for j, tile in enumerate(row):
                # blend the above tile and the left tile
                # to the current tile and add the current tile to the result row
                if i > 0:
                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
                if j > 0:
                    tile = self.blend_h(row[j - 1], tile, blend_extent)
                result_row.append(tile[:, :, :, :row_limit, :row_limit])
            result_rows.append(torch.cat(result_row, dim=4))

        moments = torch.cat(result_rows, dim=3)
        return moments

    def blend_z(
        self, a: torch.Tensor, b: torch.Tensor, blend_extent: int
    ) -> torch.Tensor:
        blend_extent = min(a.shape[2], b.shape[2], blend_extent)
        for z in range(blend_extent):
            b[:, :, z, :, :] = a[:, :, -blend_extent + z, :, :] * (
                1 - z / blend_extent
            ) + b[:, :, z, :, :] * (z / blend_extent)
        return b

    def blend_v(
        self, a: torch.Tensor, b: torch.Tensor, blend_extent: int
    ) -> torch.Tensor:
        blend_extent = min(a.shape[3], b.shape[3], blend_extent)
        for y in range(blend_extent):
            b[:, :, :, y, :] = a[:, :, :, -blend_extent + y, :] * (
                1 - y / blend_extent
            ) + b[:, :, :, y, :] * (y / blend_extent)
        return b

    def blend_h(
        self, a: torch.Tensor, b: torch.Tensor, blend_extent: int
    ) -> torch.Tensor:
        blend_extent = min(a.shape[4], b.shape[4], blend_extent)
        for x in range(blend_extent):
            b[:, :, :, :, x] = a[:, :, :, :, -blend_extent + x] * (
                1 - x / blend_extent
            ) + b[:, :, :, :, x] * (x / blend_extent)
        return b

    def _hw_tiled_decode(self, z: torch.FloatTensor, target_shape):
        overlap_size = int(self.tile_latent_min_size * (1 - self.tile_overlap_factor))
        blend_extent = int(self.tile_sample_min_size * self.tile_overlap_factor)
        row_limit = self.tile_sample_min_size - blend_extent
        tile_target_shape = (
            *target_shape[:3],
            self.tile_sample_min_size,
            self.tile_sample_min_size,
        )
        # Split z into overlapping 64x64 tiles and decode them separately.
        # The tiles have an overlap to avoid seams between tiles.
        rows = []
        for i in range(0, z.shape[3], overlap_size):
            row = []
            for j in range(0, z.shape[4], overlap_size):
                tile = z[
                    :,
                    :,
                    :,
                    i : i + self.tile_latent_min_size,
                    j : j + self.tile_latent_min_size,
                ]
                tile = self.post_quant_conv(tile)
                decoded = self.decoder(tile, target_shape=tile_target_shape)
                row.append(decoded)
            rows.append(row)
        result_rows = []
        for i, row in enumerate(rows):
            result_row = []
            for j, tile in enumerate(row):
                # blend the above tile and the left tile
                # to the current tile and add the current tile to the result row
                if i > 0:
                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
                if j > 0:
                    tile = self.blend_h(row[j - 1], tile, blend_extent)
                result_row.append(tile[:, :, :, :row_limit, :row_limit])
            result_rows.append(torch.cat(result_row, dim=4))

        dec = torch.cat(result_rows, dim=3)
        return dec

    def encode(
        self, z: torch.FloatTensor, return_dict: bool = True
    ) -> Union[DecoderOutput, torch.FloatTensor]:
        if self.use_z_tiling and z.shape[2] > self.z_sample_size > 1:
            num_splits = z.shape[2] // self.z_sample_size
            sizes = [self.z_sample_size] * num_splits
            sizes = (
                sizes + [z.shape[2] - sum(sizes)]
                if z.shape[2] - sum(sizes) > 0
                else sizes
            )
            tiles = z.split(sizes, dim=2)
            moments_tiles = [
                (
                    self._hw_tiled_encode(z_tile, return_dict)
                    if self.use_hw_tiling
                    else self._encode(z_tile)
                )
                for z_tile in tiles
            ]
            moments = torch.cat(moments_tiles, dim=2)

        else:
            moments = (
                self._hw_tiled_encode(z, return_dict)
                if self.use_hw_tiling
                else self._encode(z)
            )

        posterior = DiagonalGaussianDistribution(moments)
        if not return_dict:
            return (posterior,)

        return AutoencoderKLOutput(latent_dist=posterior)

    def _encode(self, x: torch.FloatTensor) -> AutoencoderKLOutput:
        h = self.encoder(x)
        moments = self.quant_conv(h)
        return moments

    def _decode(
        self,
        z: torch.FloatTensor,
        target_shape=None,
        timesteps: Optional[torch.Tensor] = None,
    ) -> Union[DecoderOutput, torch.FloatTensor]:
        z = self.post_quant_conv(z)
        if "timesteps" in self.decoder_params:
            dec = self.decoder(z, target_shape=target_shape, timesteps=timesteps)
        else:
            dec = self.decoder(z, target_shape=target_shape)
        return dec

    def decode(
        self,
        z: torch.FloatTensor,
        return_dict: bool = True,
        target_shape=None,
        timesteps: Optional[torch.Tensor] = None,
    ) -> Union[DecoderOutput, torch.FloatTensor]:
        assert target_shape is not None, "target_shape must be provided for decoding"
        if self.use_z_tiling and z.shape[2] > self.z_sample_size > 1:
            reduction_factor = int(
                self.encoder.patch_size_t
                * 2
                ** (
                    len(self.encoder.down_blocks)
                    - 1
                    - math.sqrt(self.encoder.patch_size)
                )
            )
            split_size = self.z_sample_size // reduction_factor
            num_splits = z.shape[2] // split_size

            # copy target shape, and divide frame dimension (=2) by the context size
            target_shape_split = list(target_shape)
            target_shape_split[2] = target_shape[2] // num_splits

            decoded_tiles = [
                (
                    self._hw_tiled_decode(z_tile, target_shape_split)
                    if self.use_hw_tiling
                    else self._decode(z_tile, target_shape=target_shape_split)
                )
                for z_tile in torch.tensor_split(z, num_splits, dim=2)
            ]
            decoded = torch.cat(decoded_tiles, dim=2)
        else:
            decoded = (
                self._hw_tiled_decode(z, target_shape)
                if self.use_hw_tiling
                else self._decode(z, target_shape=target_shape, timesteps=timesteps)
            )

        if not return_dict:
            return (decoded,)

        return DecoderOutput(sample=decoded)

    def forward(
        self,
        sample: torch.FloatTensor,
        sample_posterior: bool = False,
        return_dict: bool = True,
        generator: Optional[torch.Generator] = None,
    ) -> Union[DecoderOutput, torch.FloatTensor]:
        r"""
        Args:
            sample (`torch.FloatTensor`): Input sample.
            sample_posterior (`bool`, *optional*, defaults to `False`):
                Whether to sample from the posterior.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether to return a [`DecoderOutput`] instead of a plain tuple.
            generator (`torch.Generator`, *optional*):
                Generator used to sample from the posterior.
        """
        x = sample
        posterior = self.encode(x).latent_dist
        if sample_posterior:
            z = posterior.sample(generator=generator)
        else:
            z = posterior.mode()
        dec = self.decode(z, target_shape=sample.shape).sample

        if not return_dict:
            return (dec,)

        return DecoderOutput(sample=dec)