Commit
·
a9ef0f5
1
Parent(s):
847ce27
test tiling
Browse files- handler.py +37 -123
handler.py
CHANGED
@@ -2,154 +2,68 @@ from typing import Dict, List, Any
|
|
2 |
import base64
|
3 |
from PIL import Image
|
4 |
from io import BytesIO
|
5 |
-
from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
|
6 |
-
#from diffusers import StableDiffusionControlNetPipeline, ControlNetModel, StableDiffusionSafetyChecker
|
7 |
-
# import Safety Checker
|
8 |
from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
|
9 |
-
from diffusers
|
10 |
|
11 |
import torch
|
12 |
|
13 |
|
14 |
-
import numpy as np
|
15 |
-
import cv2
|
16 |
-
import controlnet_hinter
|
17 |
|
18 |
-
# set device
|
19 |
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
20 |
if device.type != 'cuda':
|
21 |
raise ValueError("need to run on GPU")
|
22 |
# set mixed precision dtype
|
23 |
dtype = torch.bfloat16 if torch.cuda.get_device_capability()[0] == 8 else torch.float16
|
24 |
-
|
25 |
-
# controlnet mapping for controlnet id and control hinter
|
26 |
-
CONTROLNET_MAPPING = {
|
27 |
-
"canny_edge": {
|
28 |
-
"model_id": "lllyasviel/sd-controlnet-canny",
|
29 |
-
"hinter": controlnet_hinter.hint_canny
|
30 |
-
},
|
31 |
-
"pose": {
|
32 |
-
"model_id": "lllyasviel/sd-controlnet-openpose",
|
33 |
-
"hinter": controlnet_hinter.hint_openpose
|
34 |
-
},
|
35 |
-
"depth": {
|
36 |
-
"model_id": "lllyasviel/sd-controlnet-depth",
|
37 |
-
"hinter": controlnet_hinter.hint_depth
|
38 |
-
},
|
39 |
-
"scribble": {
|
40 |
-
"model_id": "lllyasviel/sd-controlnet-scribble",
|
41 |
-
"hinter": controlnet_hinter.hint_scribble,
|
42 |
-
},
|
43 |
-
"segmentation": {
|
44 |
-
"model_id": "lllyasviel/sd-controlnet-seg",
|
45 |
-
"hinter": controlnet_hinter.hint_segmentation,
|
46 |
-
},
|
47 |
-
"normal": {
|
48 |
-
"model_id": "lllyasviel/sd-controlnet-normal",
|
49 |
-
"hinter": controlnet_hinter.hint_normal,
|
50 |
-
},
|
51 |
-
"hed": {
|
52 |
-
"model_id": "lllyasviel/sd-controlnet-hed",
|
53 |
-
"hinter": controlnet_hinter.hint_hed,
|
54 |
-
},
|
55 |
-
"hough": {
|
56 |
-
"model_id": "lllyasviel/sd-controlnet-mlsd",
|
57 |
-
"hinter": controlnet_hinter.hint_hough,
|
58 |
-
}
|
59 |
-
}
|
60 |
-
|
61 |
|
62 |
class EndpointHandler():
|
63 |
-
|
64 |
-
|
65 |
-
self.control_type = "depth"
|
66 |
-
self.controlnet = ControlNetModel.from_pretrained(CONTROLNET_MAPPING[self.control_type]["model_id"],torch_dtype=dtype).to(device)
|
67 |
|
68 |
-
#processor = AutoProcessor.from_pretrained("CompVis/stable-diffusion-safety-checker")
|
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 |
-
# Check if a new controlnet is provided
|
99 |
-
if controlnet_type is not None and controlnet_type != self.control_type:
|
100 |
-
print(f"changing controlnet from {self.control_type} to {controlnet_type} using {CONTROLNET_MAPPING[controlnet_type]['model_id']} model")
|
101 |
-
self.control_type = controlnet_type
|
102 |
-
self.controlnet = ControlNetModel.from_pretrained(CONTROLNET_MAPPING[self.control_type]["model_id"],
|
103 |
-
torch_dtype=dtype).to(device)
|
104 |
-
self.pipe.controlnet = self.controlnet
|
105 |
-
|
106 |
-
|
107 |
-
targets = [self.pipe_without_controlnet.vae, self.pipe_without_controlnet.unet]
|
108 |
-
for target in targets:
|
109 |
-
for module in target.modules():
|
110 |
-
if isinstance(module, torch.nn.Conv2d):
|
111 |
-
module.padding_mode = "circular"
|
112 |
-
|
113 |
-
# hyperparamters
|
114 |
-
num_inference_steps = data.pop("num_inference_steps", 30)
|
115 |
-
guidance_scale = data.pop("guidance_scale", 7.4)
|
116 |
-
negative_prompt = data.pop("negative_prompt", None)
|
117 |
-
height = data.pop("height", None)
|
118 |
-
width = data.pop("width", None)
|
119 |
-
controlnet_conditioning_scale = data.pop("controlnet_conditioning_scale", 1.0)
|
120 |
-
|
121 |
-
test_var = data.pop("test_var", "DEFAULT")
|
122 |
-
tiling = data.pop("tiling", True)
|
123 |
-
|
124 |
-
print(f"prompt: {prompt}")
|
125 |
-
print(f"prompt: {test_var}")
|
126 |
-
|
127 |
-
# process image
|
128 |
-
image = self.decode_base64_image(image)
|
129 |
-
#control_image = CONTROLNET_MAPPING[self.control_type]["hinter"](image)
|
130 |
|
131 |
# run inference pipeline
|
132 |
-
|
133 |
prompt=prompt,
|
134 |
negative_prompt=negative_prompt,
|
135 |
-
#image=control_image,
|
136 |
-
#image=image,
|
137 |
num_inference_steps=num_inference_steps,
|
138 |
guidance_scale=guidance_scale,
|
139 |
num_images_per_prompt=1,
|
140 |
height=height,
|
141 |
width=width,
|
142 |
-
#controlnet_conditioning_scale=controlnet_conditioning_scale,
|
143 |
generator=self.generator
|
144 |
)
|
145 |
|
146 |
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
# helper to decode input image
|
151 |
-
def decode_base64_image(self, image_string):
|
152 |
-
base64_image = base64.b64decode(image_string)
|
153 |
-
buffer = BytesIO(base64_image)
|
154 |
-
image = Image.open(buffer)
|
155 |
-
return image
|
|
|
2 |
import base64
|
3 |
from PIL import Image
|
4 |
from io import BytesIO
|
|
|
|
|
|
|
5 |
from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
|
6 |
+
from diffusers import StableDiffusionPipeline
|
7 |
|
8 |
import torch
|
9 |
|
10 |
|
11 |
+
# import numpy as np
|
12 |
+
# import cv2
|
|
|
13 |
|
14 |
+
# # set device
|
15 |
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
16 |
if device.type != 'cuda':
|
17 |
raise ValueError("need to run on GPU")
|
18 |
# set mixed precision dtype
|
19 |
dtype = torch.bfloat16 if torch.cuda.get_device_capability()[0] == 8 else torch.float16
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
20 |
|
21 |
class EndpointHandler():
|
22 |
+
def __init__(self, path=""):
|
23 |
+
self.stable_diffusion_id = "Lykon/dreamshaper-8"
|
|
|
|
|
24 |
|
|
|
25 |
|
26 |
+
|
27 |
+
def patch_conv(cls):
|
28 |
+
init = cls.__init__
|
29 |
+
def __init__(self, *args, **kwargs):
|
30 |
+
return init(self, *args, **kwargs, padding_mode='circular')
|
31 |
+
cls.__init__ = __init__
|
32 |
+
|
33 |
+
patch_conv(torch.nn.Conv2d)
|
34 |
+
|
35 |
+
|
36 |
+
|
37 |
+
|
38 |
+
self.pipe = StableDiffusionPipeline.from_pretrained(self.stable_diffusion_id,torch_dtype=dtype,safety_checker=StableDiffusionSafetyChecker.from_pretrained("CompVis/stable-diffusion-safety-checker", torch_dtype=dtype)).to(device.type)
|
39 |
+
|
40 |
+
self.generator = torch.Generator(device=device.type).manual_seed(3)
|
41 |
+
|
42 |
+
def __call__(self, data: Any) -> List[List[Dict[str, float]]]:
|
43 |
+
# """
|
44 |
+
# :param data: A dictionary contains `inputs` and optional `image` field.
|
45 |
+
# :return: A dictionary with `image` field contains image in base64.
|
46 |
+
# """
|
47 |
+
prompt = data.pop("inputs", None)
|
48 |
+
num_inference_steps = data.pop("num_inference_steps", 30)
|
49 |
+
guidance_scale = data.pop("guidance_scale", 7.4)
|
50 |
+
negative_prompt = data.pop("negative_prompt", None)
|
51 |
+
height = data.pop("height", None)
|
52 |
+
width = data.pop("width", None)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
53 |
|
54 |
# run inference pipeline
|
55 |
+
out = self.pipe(
|
56 |
prompt=prompt,
|
57 |
negative_prompt=negative_prompt,
|
|
|
|
|
58 |
num_inference_steps=num_inference_steps,
|
59 |
guidance_scale=guidance_scale,
|
60 |
num_images_per_prompt=1,
|
61 |
height=height,
|
62 |
width=width,
|
|
|
63 |
generator=self.generator
|
64 |
)
|
65 |
|
66 |
|
67 |
+
# return first generate PIL image
|
68 |
+
return out.images[0]
|
69 |
+
|
|
|
|
|
|
|
|
|
|
|
|