v0.3.0
Browse files- pipeline.py +76 -0
pipeline.py
ADDED
@@ -0,0 +1,76 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from typing import List, Optional, Tuple, Union
|
3 |
+
|
4 |
+
import torch
|
5 |
+
from diffusers import DiffusionPipeline, ImagePipelineOutput, UNet2DModel
|
6 |
+
from diffusers.utils import randn_tensor
|
7 |
+
|
8 |
+
|
9 |
+
class ConsistencyPipeline(DiffusionPipeline):
|
10 |
+
unet: UNet2DModel
|
11 |
+
|
12 |
+
def __init__(
|
13 |
+
self,
|
14 |
+
unet: UNet2DModel,
|
15 |
+
) -> None:
|
16 |
+
super().__init__()
|
17 |
+
self.register_modules(unet=unet)
|
18 |
+
|
19 |
+
@torch.no_grad()
|
20 |
+
def __call__(
|
21 |
+
self,
|
22 |
+
steps: int = 1,
|
23 |
+
generator: Optional[
|
24 |
+
Union[torch.Generator, List[torch.Generator]]
|
25 |
+
] = None,
|
26 |
+
time_min: float = 0.002,
|
27 |
+
time_max: float = 80.0,
|
28 |
+
data_std: float = 0.5,
|
29 |
+
output_type: Optional[str] = "pil",
|
30 |
+
return_dict: bool = True,
|
31 |
+
**kwargs,
|
32 |
+
) -> Union[Tuple, ImagePipelineOutput]:
|
33 |
+
img_size = self.unet.config.sample_size
|
34 |
+
shape = (1, 3, img_size, img_size)
|
35 |
+
|
36 |
+
model = self.unet
|
37 |
+
|
38 |
+
time: float = time_max
|
39 |
+
|
40 |
+
sample = randn_tensor(shape, generator=generator) * time
|
41 |
+
|
42 |
+
for step in self.progress_bar(range(steps)):
|
43 |
+
if step > 0:
|
44 |
+
time = self.search_previous_time(time)
|
45 |
+
sigma = math.sqrt(time**2 - time_min**2 + 1e-6)
|
46 |
+
sample = sample + sigma * randn_tensor(
|
47 |
+
sample.shape, device=sample.device, generator=generator
|
48 |
+
)
|
49 |
+
|
50 |
+
out = model(
|
51 |
+
sample, torch.tensor([time], device=sample.device)
|
52 |
+
).sample
|
53 |
+
|
54 |
+
skip_coef = data_std**2 / (
|
55 |
+
(time - time_min) ** 2 + data_std**2
|
56 |
+
)
|
57 |
+
out_coef = data_std * time / (time**2 + data_std**2) ** (0.5)
|
58 |
+
|
59 |
+
sample = (sample * skip_coef + out * out_coef).clamp(-1.0, 1.0)
|
60 |
+
|
61 |
+
sample = (sample / 2 + 0.5).clamp(0, 1)
|
62 |
+
image = sample.cpu().permute(0, 2, 3, 1).numpy()
|
63 |
+
|
64 |
+
if output_type == "pil":
|
65 |
+
image = self.numpy_to_pil(image)
|
66 |
+
|
67 |
+
if not return_dict:
|
68 |
+
return (image,)
|
69 |
+
|
70 |
+
return ImagePipelineOutput(images=image)
|
71 |
+
|
72 |
+
# TODO: Implement greedy search on FID
|
73 |
+
def search_previous_time(
|
74 |
+
self, time, time_min: float = 0.002, time_max: float = 80.0
|
75 |
+
):
|
76 |
+
return (2 * time + time_min) / 3
|