Spaces:
Running
on
Zero
Running
on
Zero
wondervictor
commited on
Update preprocessor.py
Browse files- preprocessor.py +104 -0
preprocessor.py
CHANGED
@@ -0,0 +1,104 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gc
|
2 |
+
import cv2
|
3 |
+
import numpy as np
|
4 |
+
import PIL.Image
|
5 |
+
import torch
|
6 |
+
from controlnet_aux import (
|
7 |
+
CannyDetector,
|
8 |
+
ContentShuffleDetector,
|
9 |
+
HEDdetector,
|
10 |
+
LineartAnimeDetector,
|
11 |
+
LineartDetector,
|
12 |
+
MidasDetector,
|
13 |
+
MLSDdetector,
|
14 |
+
NormalBaeDetector,
|
15 |
+
OpenposeDetector,
|
16 |
+
PidiNetDetector,
|
17 |
+
)
|
18 |
+
from controlnet_aux.util import HWC3
|
19 |
+
from transformers import pipeline
|
20 |
+
# from cv_utils import resize_image
|
21 |
+
# from depth_estimator import DepthEstimator
|
22 |
+
|
23 |
+
|
24 |
+
class DepthEstimator:
|
25 |
+
def __init__(self):
|
26 |
+
self.model = pipeline("condition/ckpts/dpt_large")
|
27 |
+
|
28 |
+
def __call__(self, image: np.ndarray, **kwargs) -> PIL.Image.Image:
|
29 |
+
detect_resolution = kwargs.pop("detect_resolution", 512)
|
30 |
+
image_resolution = kwargs.pop("image_resolution", 512)
|
31 |
+
image = np.array(image)
|
32 |
+
image = HWC3(image)
|
33 |
+
image = resize_image(image, resolution=detect_resolution)
|
34 |
+
image = PIL.Image.fromarray(image)
|
35 |
+
image = self.model(image)
|
36 |
+
image = image["depth"]
|
37 |
+
image = np.array(image)
|
38 |
+
image = HWC3(image)
|
39 |
+
image = resize_image(image, resolution=image_resolution)
|
40 |
+
return PIL.Image.fromarray(image)
|
41 |
+
|
42 |
+
def resize_image(input_image, resolution, interpolation=None):
|
43 |
+
H, W, C = input_image.shape
|
44 |
+
H = float(H)
|
45 |
+
W = float(W)
|
46 |
+
k = float(resolution) / max(H, W)
|
47 |
+
H *= k
|
48 |
+
W *= k
|
49 |
+
H = int(np.round(H / 64.0)) * 64
|
50 |
+
W = int(np.round(W / 64.0)) * 64
|
51 |
+
if interpolation is None:
|
52 |
+
interpolation = cv2.INTER_LANCZOS4 if k > 1 else cv2.INTER_AREA
|
53 |
+
img = cv2.resize(input_image, (W, H), interpolation=interpolation)
|
54 |
+
return img
|
55 |
+
|
56 |
+
|
57 |
+
class Preprocessor:
|
58 |
+
MODEL_ID = "condition/ckpts"
|
59 |
+
|
60 |
+
def __init__(self):
|
61 |
+
self.model = None
|
62 |
+
self.name = ""
|
63 |
+
|
64 |
+
def load(self, name: str) -> None:
|
65 |
+
if name == self.name:
|
66 |
+
return
|
67 |
+
if name == "HED":
|
68 |
+
self.model = HEDdetector.from_pretrained(self.MODEL_ID)
|
69 |
+
# elif name == "Midas":
|
70 |
+
# self.model = MidasDetector.from_pretrained(self.MODEL_ID)
|
71 |
+
elif name == "Lineart":
|
72 |
+
self.model = LineartDetector.from_pretrained(self.MODEL_ID)
|
73 |
+
elif name == "Canny":
|
74 |
+
self.model = CannyDetector()
|
75 |
+
elif name == "Depth":
|
76 |
+
# self.model = DepthEstimator()
|
77 |
+
self.model = MidasDetector.from_pretrained(self.MODEL_ID)
|
78 |
+
else:
|
79 |
+
raise ValueError
|
80 |
+
torch.cuda.empty_cache()
|
81 |
+
gc.collect()
|
82 |
+
self.name = name
|
83 |
+
|
84 |
+
def __call__(self, image: PIL.Image.Image, **kwargs) -> PIL.Image.Image:
|
85 |
+
if self.name == "Canny":
|
86 |
+
if "detect_resolution" in kwargs:
|
87 |
+
detect_resolution = kwargs.pop("detect_resolution")
|
88 |
+
image = np.array(image)
|
89 |
+
image = HWC3(image)
|
90 |
+
image = resize_image(image, resolution=detect_resolution)
|
91 |
+
image = self.model(image, **kwargs)
|
92 |
+
return PIL.Image.fromarray(image)
|
93 |
+
elif self.name == "Midas":
|
94 |
+
detect_resolution = kwargs.pop("detect_resolution", 512)
|
95 |
+
image_resolution = kwargs.pop("image_resolution", 512)
|
96 |
+
image = np.array(image)
|
97 |
+
image = HWC3(image)
|
98 |
+
image = resize_image(image, resolution=detect_resolution)
|
99 |
+
image = self.model(image, **kwargs)
|
100 |
+
image = HWC3(image)
|
101 |
+
image = resize_image(image, resolution=image_resolution)
|
102 |
+
return PIL.Image.fromarray(image)
|
103 |
+
else:
|
104 |
+
return self.model(image, **kwargs)
|