Shiroi-max
commited on
Commit
•
1688bb3
1
Parent(s):
0782bce
Upload pipeline.py
Browse files- pipeline.py +66 -0
pipeline.py
ADDED
@@ -0,0 +1,66 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import List, Optional, Tuple, Union
|
2 |
+
from diffusers import DiffusionPipeline, ImagePipelineOutput, randn_tensor
|
3 |
+
|
4 |
+
import torch
|
5 |
+
|
6 |
+
|
7 |
+
class DDPMConditionalPipeline(DiffusionPipeline):
|
8 |
+
model_cpu_offload_seq = "unet"
|
9 |
+
|
10 |
+
def __init__(self, unet, scheduler):
|
11 |
+
super().__init__()
|
12 |
+
self.register_modules(unet=unet, scheduler=scheduler)
|
13 |
+
|
14 |
+
@torch.no_grad()
|
15 |
+
def __call__(
|
16 |
+
self,
|
17 |
+
label,
|
18 |
+
batch_size: int = 1,
|
19 |
+
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
20 |
+
num_inference_steps: int = 1000,
|
21 |
+
output_type: Optional[str] = "pil",
|
22 |
+
return_dict: bool = True,
|
23 |
+
) -> Union[ImagePipelineOutput, Tuple]:
|
24 |
+
# Sample gaussian noise to begin loop
|
25 |
+
if isinstance(self.unet.config.sample_size, int):
|
26 |
+
image_shape = (
|
27 |
+
batch_size,
|
28 |
+
self.unet.config.in_channels,
|
29 |
+
self.unet.config.sample_size,
|
30 |
+
self.unet.config.sample_size,
|
31 |
+
)
|
32 |
+
else:
|
33 |
+
image_shape = (
|
34 |
+
batch_size,
|
35 |
+
self.unet.config.in_channels,
|
36 |
+
*self.unet.config.sample_size,
|
37 |
+
)
|
38 |
+
|
39 |
+
if self.device.type == "mps":
|
40 |
+
# randn does not work reproducibly on mps
|
41 |
+
image = randn_tensor(image_shape, generator=generator)
|
42 |
+
image = image.to(self.device)
|
43 |
+
else:
|
44 |
+
image = randn_tensor(image_shape, generator=generator, device=self.device)
|
45 |
+
|
46 |
+
# set step values
|
47 |
+
self.scheduler.set_timesteps(num_inference_steps)
|
48 |
+
|
49 |
+
for t in self.progress_bar(self.scheduler.timesteps):
|
50 |
+
# 1. predict noise model_output
|
51 |
+
model_output = self.unet(image, t, label).sample
|
52 |
+
|
53 |
+
# 2. compute previous image: x_t -> x_t-1
|
54 |
+
image = self.scheduler.step(
|
55 |
+
model_output, t, image, generator=generator
|
56 |
+
).prev_sample
|
57 |
+
|
58 |
+
image = (image / 2 + 0.5).clamp(0, 1)
|
59 |
+
image = image.cpu().permute(0, 2, 3, 1).numpy()
|
60 |
+
if output_type == "pil":
|
61 |
+
image = self.numpy_to_pil(image)
|
62 |
+
|
63 |
+
if not return_dict:
|
64 |
+
return (image,)
|
65 |
+
|
66 |
+
return ImagePipelineOutput(images=image)
|