Commit
·
e4d55ce
1
Parent(s):
6c599c3
try adding custom handler
Browse files- .vscode/settings.json +1 -0
- handler.py +81 -0
- requirements.txt +1 -0
.vscode/settings.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{}
|
handler.py
ADDED
@@ -0,0 +1,81 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Dict, List, Any
|
2 |
+
import torch
|
3 |
+
import requests
|
4 |
+
from PIL import Image
|
5 |
+
from io import BytesIO
|
6 |
+
from diffusers import StableDiffusionPipeline, StableDiffusionImg2ImgPipeline, DDIMScheduler
|
7 |
+
|
8 |
+
# set device
|
9 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
10 |
+
|
11 |
+
if device.type != 'cuda':
|
12 |
+
raise ValueError("need to run on GPU")
|
13 |
+
|
14 |
+
model_id = "stabilityai/stable-diffusion-2-1-base"
|
15 |
+
|
16 |
+
class EndpointHandler():
|
17 |
+
def __init__(self, path=""):
|
18 |
+
# load the optimized model
|
19 |
+
self.textPipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
|
20 |
+
self.textPipe.scheduler = DDIMScheduler.from_config(self.textPipe.scheduler.config)
|
21 |
+
self.textPipe = self.textPipe.to(device)
|
22 |
+
|
23 |
+
# create an img2img model
|
24 |
+
self.imgPipe = StableDiffusionImg2ImgPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
|
25 |
+
self.imgPipe.scheduler = DDIMScheduler.from_config(self.imgPipe.scheduler.config)
|
26 |
+
self.imgPipe = self.imgPipe.to(device)
|
27 |
+
|
28 |
+
def __call__(self, data: Any) -> List[List[Dict[str, float]]]:
|
29 |
+
"""
|
30 |
+
Args:
|
31 |
+
data (:obj:):
|
32 |
+
includes the input data and the parameters for the inference.
|
33 |
+
Return:
|
34 |
+
A :obj:`dict`:. base64 encoded image
|
35 |
+
"""
|
36 |
+
prompt = data.pop("inputs", data)
|
37 |
+
url = data.pop("src", data)
|
38 |
+
response = requests.get(url)
|
39 |
+
init_image = Image.open(BytesIO(response.content)).convert("RGB")
|
40 |
+
init_image.thumbnail((512, 512))
|
41 |
+
|
42 |
+
params = data.pop("parameters", data)
|
43 |
+
|
44 |
+
# hyperparamters
|
45 |
+
num_inference_steps = params.pop("num_inference_steps", 25)
|
46 |
+
guidance_scale = params.pop("guidance_scale", 7.5)
|
47 |
+
negative_prompt = params.pop("negative_prompt", None)
|
48 |
+
height = params.pop("height", None)
|
49 |
+
width = params.pop("width", None)
|
50 |
+
manual_seed = params.pop("manual_seed", -1)
|
51 |
+
|
52 |
+
out = None
|
53 |
+
|
54 |
+
if data.get("url"):
|
55 |
+
generator = torch.Generator(device='cuda')
|
56 |
+
generator.manual_seed(manual_seed)
|
57 |
+
# run img2img pipeline
|
58 |
+
out = self.imgPipe(prompt,
|
59 |
+
image=init_image,
|
60 |
+
num_inference_steps=num_inference_steps,
|
61 |
+
guidance_scale=guidance_scale,
|
62 |
+
num_images_per_prompt=1,
|
63 |
+
negative_prompt=negative_prompt,
|
64 |
+
height=height,
|
65 |
+
width=width
|
66 |
+
)
|
67 |
+
else:
|
68 |
+
# run text pipeline
|
69 |
+
out = self.textPipe(prompt,
|
70 |
+
image=init_image,
|
71 |
+
num_inference_steps=num_inference_steps,
|
72 |
+
guidance_scale=guidance_scale,
|
73 |
+
num_images_per_prompt=1,
|
74 |
+
negative_prompt=negative_prompt,
|
75 |
+
height=height,
|
76 |
+
width=width
|
77 |
+
)
|
78 |
+
|
79 |
+
|
80 |
+
# return first generated PIL image
|
81 |
+
return out.images[0]
|
requirements.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
diffusers==0.10.2
|