Removed Tiling code
Browse filesAfter a bit longer testing, it's causing pretty big changes in colors. I need to check out why.
README.md
CHANGED
@@ -53,461 +53,3 @@ upscaled_image = vae(image).sample
|
|
53 |
# Save the reconstructed image
|
54 |
utils.save_image(upscaled_image, "test.png")
|
55 |
```
|
56 |
-
|
57 |
-
In case you want to run it on GPU and VRAM usage is too high, below you can find modified AsymmetricAutoencoderKL class with tiling support (and maybe slicing - it does not reduce VRAM usage for me, but it can be issue with ROCm on my platform). It's copy paste from AutoencoderKL with separated tile size for encode and decode.
|
58 |
-
|
59 |
-
```
|
60 |
-
class AsymmetricAutoencoderKL(ModelMixin, ConfigMixin):
|
61 |
-
r"""
|
62 |
-
Designing a Better Asymmetric VQGAN for StableDiffusion https://arxiv.org/abs/2306.04632 . A VAE model with KL loss
|
63 |
-
for encoding images into latents and decoding latent representations into images.
|
64 |
-
|
65 |
-
This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
|
66 |
-
for all models (such as downloading or saving).
|
67 |
-
|
68 |
-
Parameters:
|
69 |
-
in_channels (int, *optional*, defaults to 3): Number of channels in the input image.
|
70 |
-
out_channels (int, *optional*, defaults to 3): Number of channels in the output.
|
71 |
-
down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
|
72 |
-
Tuple of downsample block types.
|
73 |
-
down_block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`):
|
74 |
-
Tuple of down block output channels.
|
75 |
-
layers_per_down_block (`int`, *optional*, defaults to `1`):
|
76 |
-
Number layers for down block.
|
77 |
-
up_block_types (`Tuple[str]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
|
78 |
-
Tuple of upsample block types.
|
79 |
-
up_block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`):
|
80 |
-
Tuple of up block output channels.
|
81 |
-
layers_per_up_block (`int`, *optional*, defaults to `1`):
|
82 |
-
Number layers for up block.
|
83 |
-
act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
|
84 |
-
latent_channels (`int`, *optional*, defaults to 4): Number of channels in the latent space.
|
85 |
-
sample_size (`int`, *optional*, defaults to `32`): Sample input size.
|
86 |
-
norm_num_groups (`int`, *optional*, defaults to `32`):
|
87 |
-
Number of groups to use for the first normalization layer in ResNet blocks.
|
88 |
-
scaling_factor (`float`, *optional*, defaults to 0.18215):
|
89 |
-
The component-wise standard deviation of the trained latent space computed using the first batch of the
|
90 |
-
training set. This is used to scale the latent space to have unit variance when training the diffusion
|
91 |
-
model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
|
92 |
-
diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
|
93 |
-
/ scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
|
94 |
-
Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
|
95 |
-
"""
|
96 |
-
|
97 |
-
@register_to_config
|
98 |
-
def __init__(
|
99 |
-
self,
|
100 |
-
in_channels: int = 3,
|
101 |
-
out_channels: int = 3,
|
102 |
-
down_block_types: Tuple[str, ...] = ("DownEncoderBlock2D",),
|
103 |
-
down_block_out_channels: Tuple[int, ...] = (64,),
|
104 |
-
layers_per_down_block: int = 1,
|
105 |
-
up_block_types: Tuple[str, ...] = ("UpDecoderBlock2D",),
|
106 |
-
up_block_out_channels: Tuple[int, ...] = (64,),
|
107 |
-
layers_per_up_block: int = 1,
|
108 |
-
act_fn: str = "silu",
|
109 |
-
latent_channels: int = 4,
|
110 |
-
norm_num_groups: int = 32,
|
111 |
-
sample_size: int = 32,
|
112 |
-
scaling_factor: float = 0.18215,
|
113 |
-
use_quant_conv: bool = True,
|
114 |
-
use_post_quant_conv: bool = True,
|
115 |
-
) -> None:
|
116 |
-
super().__init__()
|
117 |
-
|
118 |
-
# pass init params to Encoder
|
119 |
-
self.encoder = Encoder(
|
120 |
-
in_channels=in_channels,
|
121 |
-
out_channels=latent_channels,
|
122 |
-
down_block_types=down_block_types,
|
123 |
-
block_out_channels=down_block_out_channels,
|
124 |
-
layers_per_block=layers_per_down_block,
|
125 |
-
act_fn=act_fn,
|
126 |
-
norm_num_groups=norm_num_groups,
|
127 |
-
double_z=True,
|
128 |
-
)
|
129 |
-
|
130 |
-
# pass init params to Decoder
|
131 |
-
self.decoder = MaskConditionDecoder(
|
132 |
-
in_channels=latent_channels,
|
133 |
-
out_channels=out_channels,
|
134 |
-
up_block_types=up_block_types,
|
135 |
-
block_out_channels=up_block_out_channels,
|
136 |
-
layers_per_block=layers_per_up_block,
|
137 |
-
act_fn=act_fn,
|
138 |
-
norm_num_groups=norm_num_groups,
|
139 |
-
)
|
140 |
-
|
141 |
-
self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1) if use_quant_conv else None
|
142 |
-
self.post_quant_conv = nn.Conv2d(latent_channels, latent_channels, 1) if use_post_quant_conv else None
|
143 |
-
|
144 |
-
self.use_slicing = False
|
145 |
-
self.use_tiling = False
|
146 |
-
|
147 |
-
# only relevant if vae tiling is enabled
|
148 |
-
self.tile_sample_min_size = self.config.sample_size
|
149 |
-
sample_size = (
|
150 |
-
self.config.sample_size[0]
|
151 |
-
if isinstance(self.config.sample_size, (list, tuple))
|
152 |
-
else self.config.sample_size
|
153 |
-
)
|
154 |
-
self.tile_latent_min_up_size = int(sample_size / (2 ** (len(self.config.up_block_out_channels) - 1)))
|
155 |
-
self.tile_latent_min_down_size = int(sample_size / (2 ** (len(self.config.down_block_out_channels) - 1)))
|
156 |
-
|
157 |
-
self.tile_overlap_factor = 0.25
|
158 |
-
|
159 |
-
self.register_to_config(block_out_channels=up_block_out_channels)
|
160 |
-
self.register_to_config(force_upcast=False)
|
161 |
-
|
162 |
-
def enable_tiling(self, use_tiling: bool = True):
|
163 |
-
r"""
|
164 |
-
Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
|
165 |
-
compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
|
166 |
-
processing larger images.
|
167 |
-
"""
|
168 |
-
self.use_tiling = use_tiling
|
169 |
-
|
170 |
-
def disable_tiling(self):
|
171 |
-
r"""
|
172 |
-
Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
|
173 |
-
decoding in one step.
|
174 |
-
"""
|
175 |
-
self.enable_tiling(False)
|
176 |
-
|
177 |
-
def enable_slicing(self):
|
178 |
-
r"""
|
179 |
-
Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
|
180 |
-
compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
|
181 |
-
"""
|
182 |
-
self.use_slicing = True
|
183 |
-
|
184 |
-
def disable_slicing(self):
|
185 |
-
r"""
|
186 |
-
Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
|
187 |
-
decoding in one step.
|
188 |
-
"""
|
189 |
-
self.use_slicing = False
|
190 |
-
|
191 |
-
def _encode(self, x: torch.Tensor) -> torch.Tensor:
|
192 |
-
batch_size, num_channels, height, width = x.shape
|
193 |
-
|
194 |
-
if self.use_tiling and (width > self.tile_sample_min_size or height > self.tile_sample_min_size):
|
195 |
-
return self._tiled_encode(x)
|
196 |
-
|
197 |
-
enc = self.encoder(x)
|
198 |
-
if self.quant_conv is not None:
|
199 |
-
enc = self.quant_conv(enc)
|
200 |
-
|
201 |
-
return enc
|
202 |
-
|
203 |
-
@apply_forward_hook
|
204 |
-
def encode(
|
205 |
-
self, x: torch.Tensor, return_dict: bool = True
|
206 |
-
) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]:
|
207 |
-
"""
|
208 |
-
Encode a batch of images into latents.
|
209 |
-
|
210 |
-
Args:
|
211 |
-
x (`torch.Tensor`): Input batch of images.
|
212 |
-
return_dict (`bool`, *optional*, defaults to `True`):
|
213 |
-
Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
|
214 |
-
|
215 |
-
Returns:
|
216 |
-
The latent representations of the encoded images. If `return_dict` is True, a
|
217 |
-
[`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned.
|
218 |
-
"""
|
219 |
-
if self.use_slicing and x.shape[0] > 1:
|
220 |
-
encoded_slices = [self._encode(x_slice) for x_slice in x.split(1)]
|
221 |
-
h = torch.cat(encoded_slices)
|
222 |
-
else:
|
223 |
-
h = self._encode(x)
|
224 |
-
|
225 |
-
posterior = DiagonalGaussianDistribution(h)
|
226 |
-
|
227 |
-
if not return_dict:
|
228 |
-
return (posterior,)
|
229 |
-
|
230 |
-
return AutoencoderKLOutput(latent_dist=posterior)
|
231 |
-
|
232 |
-
def _decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
|
233 |
-
if self.use_tiling and (z.shape[-1] > self.tile_latent_min_up_size or z.shape[-2] > self.tile_latent_min_up_size):
|
234 |
-
return self.tiled_decode(z, return_dict=return_dict)
|
235 |
-
|
236 |
-
if self.post_quant_conv is not None:
|
237 |
-
z = self.post_quant_conv(z)
|
238 |
-
|
239 |
-
dec = self.decoder(z)
|
240 |
-
|
241 |
-
if not return_dict:
|
242 |
-
return (dec,)
|
243 |
-
|
244 |
-
return DecoderOutput(sample=dec)
|
245 |
-
|
246 |
-
@apply_forward_hook
|
247 |
-
def decode(
|
248 |
-
self, z: torch.FloatTensor, return_dict: bool = True, generator=None
|
249 |
-
) -> Union[DecoderOutput, torch.FloatTensor]:
|
250 |
-
"""
|
251 |
-
Decode a batch of images.
|
252 |
-
|
253 |
-
Args:
|
254 |
-
z (`torch.Tensor`): Input batch of latent vectors.
|
255 |
-
return_dict (`bool`, *optional*, defaults to `True`):
|
256 |
-
Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
|
257 |
-
|
258 |
-
Returns:
|
259 |
-
[`~models.vae.DecoderOutput`] or `tuple`:
|
260 |
-
If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
|
261 |
-
returned.
|
262 |
-
|
263 |
-
"""
|
264 |
-
if self.use_slicing and z.shape[0] > 1:
|
265 |
-
decoded_slices = [self._decode(z_slice).sample for z_slice in z.split(1)]
|
266 |
-
decoded = torch.cat(decoded_slices)
|
267 |
-
else:
|
268 |
-
decoded = self._decode(z).sample
|
269 |
-
|
270 |
-
if not return_dict:
|
271 |
-
return (decoded,)
|
272 |
-
|
273 |
-
return DecoderOutput(sample=decoded)
|
274 |
-
|
275 |
-
def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
|
276 |
-
blend_extent = min(a.shape[2], b.shape[2], blend_extent)
|
277 |
-
for y in range(blend_extent):
|
278 |
-
b[:, :, y, :] = a[:, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, y, :] * (y / blend_extent)
|
279 |
-
return b
|
280 |
-
|
281 |
-
def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
|
282 |
-
blend_extent = min(a.shape[3], b.shape[3], blend_extent)
|
283 |
-
for x in range(blend_extent):
|
284 |
-
b[:, :, :, x] = a[:, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, x] * (x / blend_extent)
|
285 |
-
return b
|
286 |
-
|
287 |
-
def _tiled_encode(self, x: torch.Tensor) -> torch.Tensor:
|
288 |
-
r"""Encode a batch of images using a tiled encoder.
|
289 |
-
|
290 |
-
When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
|
291 |
-
steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is
|
292 |
-
different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the
|
293 |
-
tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the
|
294 |
-
output, but they should be much less noticeable.
|
295 |
-
|
296 |
-
Args:
|
297 |
-
x (`torch.Tensor`): Input batch of images.
|
298 |
-
|
299 |
-
Returns:
|
300 |
-
`torch.Tensor`:
|
301 |
-
The latent representation of the encoded videos.
|
302 |
-
"""
|
303 |
-
|
304 |
-
overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor))
|
305 |
-
blend_extent = int(self.tile_latent_min_down_size * self.tile_overlap_factor)
|
306 |
-
row_limit = self.tile_latent_min_down_size - blend_extent
|
307 |
-
|
308 |
-
# Split the image into 512x512 tiles and encode them separately.
|
309 |
-
rows = []
|
310 |
-
for i in range(0, x.shape[2], overlap_size):
|
311 |
-
row = []
|
312 |
-
for j in range(0, x.shape[3], overlap_size):
|
313 |
-
tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size]
|
314 |
-
tile = self.encoder(tile)
|
315 |
-
if self.config.use_quant_conv:
|
316 |
-
tile = self.quant_conv(tile)
|
317 |
-
row.append(tile)
|
318 |
-
rows.append(row)
|
319 |
-
result_rows = []
|
320 |
-
for i, row in enumerate(rows):
|
321 |
-
result_row = []
|
322 |
-
for j, tile in enumerate(row):
|
323 |
-
# blend the above tile and the left tile
|
324 |
-
# to the current tile and add the current tile to the result row
|
325 |
-
if i > 0:
|
326 |
-
tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
|
327 |
-
if j > 0:
|
328 |
-
tile = self.blend_h(row[j - 1], tile, blend_extent)
|
329 |
-
result_row.append(tile[:, :, :row_limit, :row_limit])
|
330 |
-
result_rows.append(torch.cat(result_row, dim=3))
|
331 |
-
|
332 |
-
enc = torch.cat(result_rows, dim=2)
|
333 |
-
return enc
|
334 |
-
|
335 |
-
def tiled_encode(self, x: torch.Tensor, return_dict: bool = True) -> AutoencoderKLOutput:
|
336 |
-
r"""Encode a batch of images using a tiled encoder.
|
337 |
-
|
338 |
-
When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
|
339 |
-
steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is
|
340 |
-
different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the
|
341 |
-
tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the
|
342 |
-
output, but they should be much less noticeable.
|
343 |
-
|
344 |
-
Args:
|
345 |
-
x (`torch.Tensor`): Input batch of images.
|
346 |
-
return_dict (`bool`, *optional*, defaults to `True`):
|
347 |
-
Whether or not to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
|
348 |
-
|
349 |
-
Returns:
|
350 |
-
[`~models.autoencoder_kl.AutoencoderKLOutput`] or `tuple`:
|
351 |
-
If return_dict is True, a [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain
|
352 |
-
`tuple` is returned.
|
353 |
-
"""
|
354 |
-
deprecation_message = (
|
355 |
-
"The tiled_encode implementation supporting the `return_dict` parameter is deprecated. In the future, the "
|
356 |
-
"implementation of this method will be replaced with that of `_tiled_encode` and you will no longer be able "
|
357 |
-
"to pass `return_dict`. You will also have to create a `DiagonalGaussianDistribution()` from the returned value."
|
358 |
-
)
|
359 |
-
deprecate("tiled_encode", "1.0.0", deprecation_message, standard_warn=False)
|
360 |
-
|
361 |
-
overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor))
|
362 |
-
blend_extent = int(self.tile_latent_min_up_size * self.tile_overlap_factor)
|
363 |
-
row_limit = self.tile_latent_min_up_size - blend_extent
|
364 |
-
|
365 |
-
# Split the image into 512x512 tiles and encode them separately.
|
366 |
-
rows = []
|
367 |
-
for i in range(0, x.shape[2], overlap_size):
|
368 |
-
row = []
|
369 |
-
for j in range(0, x.shape[3], overlap_size):
|
370 |
-
tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size]
|
371 |
-
tile = self.encoder(tile)
|
372 |
-
if self.config.use_quant_conv:
|
373 |
-
tile = self.quant_conv(tile)
|
374 |
-
row.append(tile)
|
375 |
-
rows.append(row)
|
376 |
-
result_rows = []
|
377 |
-
for i, row in enumerate(rows):
|
378 |
-
result_row = []
|
379 |
-
for j, tile in enumerate(row):
|
380 |
-
# blend the above tile and the left tile
|
381 |
-
# to the current tile and add the current tile to the result row
|
382 |
-
if i > 0:
|
383 |
-
tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
|
384 |
-
if j > 0:
|
385 |
-
tile = self.blend_h(row[j - 1], tile, blend_extent)
|
386 |
-
result_row.append(tile[:, :, :row_limit, :row_limit])
|
387 |
-
result_rows.append(torch.cat(result_row, dim=3))
|
388 |
-
|
389 |
-
moments = torch.cat(result_rows, dim=2)
|
390 |
-
posterior = DiagonalGaussianDistribution(moments)
|
391 |
-
|
392 |
-
if not return_dict:
|
393 |
-
return (posterior,)
|
394 |
-
|
395 |
-
return AutoencoderKLOutput(latent_dist=posterior)
|
396 |
-
|
397 |
-
def tiled_decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
|
398 |
-
r"""
|
399 |
-
Decode a batch of images using a tiled decoder.
|
400 |
-
|
401 |
-
Args:
|
402 |
-
z (`torch.Tensor`): Input batch of latent vectors.
|
403 |
-
return_dict (`bool`, *optional*, defaults to `True`):
|
404 |
-
Whether or not to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
|
405 |
-
|
406 |
-
Returns:
|
407 |
-
[`~models.vae.DecoderOutput`] or `tuple`:
|
408 |
-
If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
|
409 |
-
returned.
|
410 |
-
"""
|
411 |
-
overlap_size = int(self.tile_latent_min_up_size * (1 - self.tile_overlap_factor))
|
412 |
-
blend_extent = int(self.tile_sample_min_size * self.tile_overlap_factor)
|
413 |
-
row_limit = self.tile_sample_min_size - blend_extent
|
414 |
-
|
415 |
-
# Split z into overlapping 64x64 tiles and decode them separately.
|
416 |
-
# The tiles have an overlap to avoid seams between tiles.
|
417 |
-
rows = []
|
418 |
-
for i in range(0, z.shape[2], overlap_size):
|
419 |
-
row = []
|
420 |
-
for j in range(0, z.shape[3], overlap_size):
|
421 |
-
tile = z[:, :, i : i + self.tile_latent_min_up_size, j : j + self.tile_latent_min_up_size]
|
422 |
-
if self.config.use_post_quant_conv:
|
423 |
-
tile = self.post_quant_conv(tile)
|
424 |
-
decoded = self.decoder(tile)
|
425 |
-
row.append(decoded)
|
426 |
-
rows.append(row)
|
427 |
-
result_rows = []
|
428 |
-
for i, row in enumerate(rows):
|
429 |
-
result_row = []
|
430 |
-
for j, tile in enumerate(row):
|
431 |
-
# blend the above tile and the left tile
|
432 |
-
# to the current tile and add the current tile to the result row
|
433 |
-
if i > 0:
|
434 |
-
tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
|
435 |
-
if j > 0:
|
436 |
-
tile = self.blend_h(row[j - 1], tile, blend_extent)
|
437 |
-
result_row.append(tile[:, :, :row_limit, :row_limit])
|
438 |
-
result_rows.append(torch.cat(result_row, dim=3))
|
439 |
-
|
440 |
-
dec = torch.cat(result_rows, dim=2)
|
441 |
-
if not return_dict:
|
442 |
-
return (dec,)
|
443 |
-
|
444 |
-
return DecoderOutput(sample=dec)
|
445 |
-
|
446 |
-
def forward(
|
447 |
-
self,
|
448 |
-
sample: torch.Tensor,
|
449 |
-
sample_posterior: bool = False,
|
450 |
-
return_dict: bool = True,
|
451 |
-
generator: Optional[torch.Generator] = None,
|
452 |
-
) -> Union[DecoderOutput, torch.Tensor]:
|
453 |
-
r"""
|
454 |
-
Args:
|
455 |
-
sample (`torch.Tensor`): Input sample.
|
456 |
-
sample_posterior (`bool`, *optional*, defaults to `False`):
|
457 |
-
Whether to sample from the posterior.
|
458 |
-
return_dict (`bool`, *optional*, defaults to `True`):
|
459 |
-
Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
|
460 |
-
"""
|
461 |
-
x = sample
|
462 |
-
posterior = self.encode(x).latent_dist
|
463 |
-
if sample_posterior:
|
464 |
-
z = posterior.sample(generator=generator)
|
465 |
-
else:
|
466 |
-
z = posterior.mode()
|
467 |
-
dec = self.decode(z).sample
|
468 |
-
|
469 |
-
if not return_dict:
|
470 |
-
return (dec,)
|
471 |
-
|
472 |
-
return DecoderOutput(sample=dec)
|
473 |
-
|
474 |
-
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
|
475 |
-
def fuse_qkv_projections(self):
|
476 |
-
"""
|
477 |
-
Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
|
478 |
-
are fused. For cross-attention modules, key and value projection matrices are fused.
|
479 |
-
|
480 |
-
<Tip warning={true}>
|
481 |
-
|
482 |
-
This API is 🧪 experimental.
|
483 |
-
|
484 |
-
</Tip>
|
485 |
-
"""
|
486 |
-
self.original_attn_processors = None
|
487 |
-
|
488 |
-
for _, attn_processor in self.attn_processors.items():
|
489 |
-
if "Added" in str(attn_processor.__class__.__name__):
|
490 |
-
raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
|
491 |
-
|
492 |
-
self.original_attn_processors = self.attn_processors
|
493 |
-
|
494 |
-
for module in self.modules():
|
495 |
-
if isinstance(module, Attention):
|
496 |
-
module.fuse_projections(fuse=True)
|
497 |
-
|
498 |
-
self.set_attn_processor(FusedAttnProcessor2_0())
|
499 |
-
|
500 |
-
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
|
501 |
-
def unfuse_qkv_projections(self):
|
502 |
-
"""Disables the fused QKV projection if enabled.
|
503 |
-
|
504 |
-
<Tip warning={true}>
|
505 |
-
|
506 |
-
This API is 🧪 experimental.
|
507 |
-
|
508 |
-
</Tip>
|
509 |
-
|
510 |
-
"""
|
511 |
-
if self.original_attn_processors is not None:
|
512 |
-
self.set_attn_processor(self.original_attn_processors)
|
513 |
-
```
|
|
|
53 |
# Save the reconstructed image
|
54 |
utils.save_image(upscaled_image, "test.png")
|
55 |
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|