forked from entmike/disco-diffusion-1
-
Notifications
You must be signed in to change notification settings - Fork 0
/
disco_xform_utils.py
181 lines (156 loc) · 7.11 KB
/
disco_xform_utils.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
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
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
from cv2 import imread
import torch, torchvision
from loguru import logger
try:
from pytorch3d import renderer
except:
logger.warning("""Cannot import pytorch3d renderer module. 3D animations won't work.""")
from midas import utils as midas_utils
from PIL import Image
import numpy as np
import math
from adabins.infer import InferenceHelper
MAX_ADABINS_AREA = 500000
MIN_ADABINS_AREA = 448 * 448
@torch.no_grad()
def transform_image_3d(
img_filepath,
midas_model,
midas_transform,
device,
rot_mat=torch.eye(3).unsqueeze(0),
translate=(0.0, 0.0, -0.04),
near=2000,
far=20000,
fov_deg=60,
padding_mode="border",
sampling_mode="bicubic",
midas_weight=0.3,
spherical=False,
):
img_pil = Image.open(open(img_filepath, "rb")).convert("RGB")
w, h = img_pil.size
image_tensor = torchvision.transforms.functional.to_tensor(img_pil).to(device)
use_adabins = midas_weight < 1.0
if use_adabins:
# AdaBins
"""
predictions using nyu dataset
"""
logger.debug("Running AdaBins depth estimation implementation...")
infer_helper = InferenceHelper(pretrained_path_base="models", dataset="nyu", device=device)
image_pil_area = w * h
if image_pil_area > MAX_ADABINS_AREA:
scale = math.sqrt(MAX_ADABINS_AREA) / math.sqrt(image_pil_area)
depth_input = img_pil.resize((int(w * scale), int(h * scale)), Image.LANCZOS) # LANCZOS is supposed to be good for downsampling.
elif image_pil_area < MIN_ADABINS_AREA:
scale = math.sqrt(MIN_ADABINS_AREA) / math.sqrt(image_pil_area)
depth_input = img_pil.resize((int(w * scale), int(h * scale)), Image.BICUBIC)
else:
depth_input = img_pil
try:
_, adabins_depth = infer_helper.predict_pil(depth_input)
if image_pil_area != MAX_ADABINS_AREA:
adabins_depth = (
torchvision.transforms.functional.resize(
torch.from_numpy(adabins_depth),
image_tensor.shape[-2:],
interpolation=torchvision.transforms.functional.InterpolationMode.BICUBIC,
)
.squeeze()
.to(device)
)
else:
adabins_depth = torch.from_numpy(adabins_depth).squeeze().to(device)
adabins_depth_np = adabins_depth.cpu().numpy()
except:
pass
torch.cuda.empty_cache()
# MiDaS
img_midas = midas_utils.read_image(img_filepath)
img_midas_input = midas_transform({"image": img_midas})["image"]
midas_optimize = True
# MiDaS depth estimation implementation
logger.debug("Running MiDaS depth estimation implementation...")
sample = torch.from_numpy(img_midas_input).float().to(device).unsqueeze(0)
if midas_optimize == True and device == torch.device("cuda"):
sample = sample.to(memory_format=torch.channels_last)
sample = sample.half()
prediction_torch = midas_model.forward(sample)
prediction_torch = torch.nn.functional.interpolate(
prediction_torch.unsqueeze(1),
size=img_midas.shape[:2],
mode="bicubic",
align_corners=False,
).squeeze()
prediction_np = prediction_torch.clone().cpu().numpy()
logger.debug("Finished depth estimation.")
torch.cuda.empty_cache()
# MiDaS makes the near values greater, and the far values lesser. Let's reverse that and try to align with AdaBins a bit better.
prediction_np = np.subtract(50.0, prediction_np)
prediction_np = prediction_np / 19.0
if use_adabins:
adabins_weight = 1.0 - midas_weight
depth_map = prediction_np * midas_weight + adabins_depth_np * adabins_weight
else:
depth_map = prediction_np
depth_map = np.expand_dims(depth_map, axis=0)
depth_tensor = torch.from_numpy(depth_map).squeeze().to(device)
pixel_aspect = 1.0 # really.. the aspect of an individual pixel! (so usually 1.0)
persp_cam_old = renderer.FoVPerspectiveCameras(near, far, pixel_aspect, fov=fov_deg, degrees=True, device=device)
persp_cam_new = renderer.FoVPerspectiveCameras(
near,
far,
pixel_aspect,
fov=fov_deg,
degrees=True,
R=rot_mat,
T=torch.tensor([translate]),
device=device,
)
# range of [-1,1] is important to torch grid_sample's padding handling
y, x = torch.meshgrid(
torch.linspace(-1.0, 1.0, h, dtype=torch.float32, device=device),
torch.linspace(-1.0, 1.0, w, dtype=torch.float32, device=device),
)
z = torch.as_tensor(depth_tensor, dtype=torch.float32, device=device)
xyz_old_world = torch.stack((x.flatten(), y.flatten(), z.flatten()), dim=1)
# Transform the points using pytorch3d. With current functionality, this is overkill and prevents it from working on Windows.
# If you want it to run on Windows (without pytorch3d), then the transforms (and/or perspective if that's separate) can be done pretty easily without it.
xyz_old_cam_xy = persp_cam_old.get_full_projection_transform().transform_points(xyz_old_world)[:, 0:2]
xyz_new_cam_xy = persp_cam_new.get_full_projection_transform().transform_points(xyz_old_world)[:, 0:2]
offset_xy = xyz_new_cam_xy - xyz_old_cam_xy
# affine_grid theta param expects a batch of 2D mats. Each is 2x3 to do rotation+translation.
identity_2d_batch = torch.tensor([[1.0, 0.0, 0.0], [0.0, 1.0, 0.0]], device=device).unsqueeze(0)
# coords_2d will have shape (N,H,W,2).. which is also what grid_sample needs.
coords_2d = torch.nn.functional.affine_grid(identity_2d_batch, [1, 1, h, w], align_corners=False)
offset_coords_2d = coords_2d - torch.reshape(offset_xy, (h, w, 2)).unsqueeze(0)
if spherical:
spherical_grid = get_spherical_projection(h, w, torch.tensor([0, 0], device=device), -0.4, device=device) # align_corners=False
stage_image = torch.nn.functional.grid_sample(
image_tensor.add(1 / 512 - 0.0001).unsqueeze(0),
offset_coords_2d,
mode=sampling_mode,
padding_mode=padding_mode,
align_corners=True,
)
new_image = torch.nn.functional.grid_sample(stage_image, spherical_grid, align_corners=True) # , mode=sampling_mode, padding_mode=padding_mode, align_corners=False)
else:
new_image = torch.nn.functional.grid_sample(
image_tensor.add(1 / 512 - 0.0001).unsqueeze(0),
offset_coords_2d,
mode=sampling_mode,
padding_mode=padding_mode,
align_corners=False,
)
img_pil = torchvision.transforms.ToPILImage()(new_image.squeeze().clamp(0, 1.0))
torch.cuda.empty_cache()
return img_pil
def get_spherical_projection(H, W, center, magnitude, device):
xx, yy = torch.linspace(-1, 1, W, dtype=torch.float32, device=device), torch.linspace(-1, 1, H, dtype=torch.float32, device=device)
gridy, gridx = torch.meshgrid(yy, xx)
grid = torch.stack([gridx, gridy], dim=-1)
d = center - grid
d_sum = torch.sqrt((d**2).sum(axis=-1))
grid += d * d_sum.unsqueeze(-1) * magnitude
return grid.unsqueeze(0)