Upload pipeline.py with huggingface_hub
Browse files- pipeline.py +90 -0
pipeline.py
ADDED
@@ -0,0 +1,90 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import List
|
2 |
+
|
3 |
+
import torch
|
4 |
+
from diffusers import StableDiffusionModelEditingPipeline as SDTIME
|
5 |
+
from diffusers.models import AutoencoderKL, UNet2DConditionModel
|
6 |
+
from diffusers.pipelines.deprecated.stable_diffusion_variants.pipeline_stable_diffusion_model_editing import (
|
7 |
+
AUGS_CONST,
|
8 |
+
)
|
9 |
+
from diffusers.pipelines.stable_diffusion.safety_checker import (
|
10 |
+
StableDiffusionSafetyChecker,
|
11 |
+
)
|
12 |
+
from diffusers.schedulers.scheduling_utils import SchedulerMixin
|
13 |
+
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
|
14 |
+
|
15 |
+
|
16 |
+
class StableDiffusionModelEditingPipeline(SDTIME):
|
17 |
+
def __init__(
|
18 |
+
self,
|
19 |
+
vae: AutoencoderKL,
|
20 |
+
text_encoder: CLIPTextModel,
|
21 |
+
tokenizer: CLIPTokenizer,
|
22 |
+
unet: UNet2DConditionModel,
|
23 |
+
scheduler: SchedulerMixin,
|
24 |
+
safety_checker: StableDiffusionSafetyChecker,
|
25 |
+
feature_extractor: CLIPImageProcessor,
|
26 |
+
requires_safety_checker: bool = True,
|
27 |
+
with_to_k: bool = True,
|
28 |
+
with_augs: List[str] = AUGS_CONST,
|
29 |
+
) -> None:
|
30 |
+
super().__init__(
|
31 |
+
vae,
|
32 |
+
text_encoder,
|
33 |
+
tokenizer,
|
34 |
+
unet,
|
35 |
+
scheduler,
|
36 |
+
safety_checker,
|
37 |
+
feature_extractor,
|
38 |
+
requires_safety_checker,
|
39 |
+
with_to_k,
|
40 |
+
with_augs,
|
41 |
+
)
|
42 |
+
|
43 |
+
# get cross-attention layers
|
44 |
+
ca_layers = []
|
45 |
+
|
46 |
+
def append_ca(net_):
|
47 |
+
# In diffusers v1.15.0 and later, `CrossAttention` has been changed to `Attention`
|
48 |
+
# Refer to the pipeline in the fork:
|
49 |
+
# https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_model_editing.py#L135
|
50 |
+
if net_.__class__.__name__ == "Attention":
|
51 |
+
ca_layers.append(net_)
|
52 |
+
elif hasattr(net_, "children"):
|
53 |
+
for net__ in net_.children():
|
54 |
+
append_ca(net__)
|
55 |
+
|
56 |
+
# recursively find all cross-attention layers in unet
|
57 |
+
for net in self.unet.named_children():
|
58 |
+
if "down" in net[0]:
|
59 |
+
append_ca(net[1])
|
60 |
+
elif "up" in net[0]:
|
61 |
+
append_ca(net[1])
|
62 |
+
elif "mid" in net[0]:
|
63 |
+
append_ca(net[1])
|
64 |
+
|
65 |
+
# get projection matrices
|
66 |
+
self.ca_clip_layers = [l for l in ca_layers if l.to_v.in_features == 768]
|
67 |
+
assert len(self.ca_clip_layers) > 0
|
68 |
+
self.projection_matrices = [l.to_v for l in self.ca_clip_layers]
|
69 |
+
assert len(self.projection_matrices) > 0
|
70 |
+
|
71 |
+
if self.with_to_k:
|
72 |
+
projection_matrices = [l.to_k for l in self.ca_clip_layers]
|
73 |
+
self.projection_matrices = self.projection_matrices + projection_matrices
|
74 |
+
assert len(self.projection_matrices) > 0
|
75 |
+
|
76 |
+
@torch.no_grad()
|
77 |
+
def edit_model(
|
78 |
+
self,
|
79 |
+
source_prompt: str,
|
80 |
+
destination_prompt: str,
|
81 |
+
lamb: float = 0.1,
|
82 |
+
**kwargs,
|
83 |
+
) -> None:
|
84 |
+
# `restart_params` creates a copy of the object when restoring the original weights,
|
85 |
+
# which can lead to problems such as the device not being set correctly
|
86 |
+
# when exiting the pipeline. For these reasons, `restart_params` is set to `False`.
|
87 |
+
# If you want to restore the original weights, it is recommended to reload the pipeline.
|
88 |
+
super().edit_model(
|
89 |
+
source_prompt, destination_prompt, lamb, restart_params=False
|
90 |
+
)
|