DDGS/example.py at main · splatsdotcom/DDGS · GitHub

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
import torch
import numpy as np
import math
from PIL import Image
import os
import time

import ddgs

# ------------------------------------------- #

def look_at(eye, target, up):
	f = (target - eye)
	f = f / torch.norm(f)
	u = up / torch.norm(up)
	s = torch.cross(f, u, dim=0)
	s = s / torch.norm(s)
	u = torch.cross(s, f, dim=0)
	m = torch.eye(4, dtype=torch.float32)

	m[0, :3] = s
	m[1, :3] = u
	m[2, :3] = -f
	m[0, 3] = -torch.dot(s, eye)
	m[1, 3] = -torch.dot(u, eye)
	m[2, 3] = torch.dot(f, eye)

	return m

def perspective(fovy, aspect, znear, zfar):
	tanHalfFov = math.tan(fovy / 2)
	m = torch.zeros((4, 4), dtype=torch.float32)
	m[0, 0] = 1 / (aspect * tanHalfFov)
	m[1, 1] = 1 / tanHalfFov
	m[2, 2] = -(zfar + znear) / (zfar - znear)
	m[2, 3] = -(2 * zfar * znear) / (zfar - znear)
	m[3, 2] = -1.0
	return m

# ------------------------------------------- #


def main():

	# torch setup:
	# ---------------
	torch.set_default_device('cuda')
	torch.manual_seed(0)

	# intrinsics:
	# ---------------
	width, height = 1920, 1080
	aspect = width / height
	fovy = math.radians(60)
	proj = perspective(fovy, aspect, 0.1, 100.0)
	focalX = width / (2 * math.tan(fovy / 2))
	focalY = focalX

	# GT data:
	# ---------------
	gtMeans = torch.tensor([
		[0.3,  0.0, 0.0],
		[-0.3, 0.1, 0.1],
		[0.0, -0.2, 0.0]
	], dtype=torch.float32, device='cuda')

	gtScales = torch.tensor([
		[0.2, 0.05, 0.05],
		[0.08, 0.3, 0.08],
		[0.1, 0.1, 0.25]
	], dtype=torch.float32, device='cuda')

	gtRotations = torch.tensor([
		[0.0, 0.0, 0.0, 1.0],
		[0.0, 0.0, 0.0, 1.0],
		[0.0, 0.0, 0.0, 1.0]
	], dtype=torch.float32, device='cuda')

	gtOpacities = torch.tensor([
		[1.0],
		[1.0],
		[1.0]
	], dtype=torch.float32, device='cuda')

	gtHarmonics = torch.tensor([
		[ [1.0, 0.0, 0.0] ],
		[ [0.0, 1.0, 0.0] ], # currently harmonics is just treated as RGB color
		[ [0.0, 0.0, 1.0] ]
	], dtype=torch.float32, device='cuda')

	# generate GT views + renderers:
	# ---------------
	numViews = 10
	views = []

	os.makedirs("renders", exist_ok=True)

	for i in range(numViews):
		angle = 2 * math.pi * i / numViews
		eye = torch.tensor([2.0 * math.sin(angle), 0.5, 2.0 * math.cos(angle)], device='cuda')
		target = torch.tensor([0.0, 0.0, 0.0], device='cuda')
		up = torch.tensor([0.0, 1.0, 0.0], device='cuda')
		view = look_at(eye, target, up)

		settings = ddgs.Settings(
			width=width, height=height,
			view=view, proj=proj,
			focalX=focalX, focalY=focalY,
			debug=False
		)
		renderer = ddgs.Renderer(settings)

		with torch.no_grad():
			img = renderer(gtMeans, gtScales, gtRotations, gtOpacities, gtHarmonics).color

		views.append((renderer, img))

		imgNp = (img.detach().cpu().clamp(0, 1).numpy() * 255).astype(np.uint8)
		Image.fromarray(imgNp).save(f"renders/gt_view_{i}.png")

	print("Saved ground truth renders...")

	# generate perturbed scene:
	# ---------------
	means = (gtMeans + 0.25 * torch.randn_like(gtMeans)).clone().detach().requires_grad_(True)
	scales = (gtScales + 0.1 * torch.randn_like(gtScales)).clone().detach().requires_grad_(True)
	rotations = (gtRotations + 0.5 * torch.randn_like(gtRotations)).clone().detach().requires_grad_(True)
	opacities = torch.rand_like(gtOpacities).requires_grad_(True)
	harmonics = (torch.rand_like(gtHarmonics) * 0.5 + gtHarmonics * 0.5).detach().clone().requires_grad_(True)

	optimizer = torch.optim.Adam([means, scales, rotations, opacities, harmonics], lr=1e-2)

	# train:
	# ---------------
	start = time.time()

	for step in range(500):
		optimizer.zero_grad()
		totalLoss = 0.0

		for (renderer, gtImg) in views:
			predImg = renderer(means, scales, rotations, opacities, harmonics).color
			loss = torch.nn.functional.mse_loss(predImg, gtImg)
			totalLoss += loss

		totalLoss.backward()
		optimizer.step()

		if step % 10 == 0:
			print(f"[Step {step:03d}] Loss = {totalLoss.item():.6f}")

			# save current render (takes too long):
			# with torch.no_grad():
			# 	renderer, _ = views[0]
			# 	img = renderer(means, scales, rotations, opacities, harmonics)
			# 	imgNp = (img.detach().cpu().clamp(0, 1).numpy() * 255).astype(np.uint8)
			# 	Image.fromarray(imgNp).save(f"renders/train_step_{step:03d}.png")

	print("Training complete...")
	print(f"Took {time.time() - start}s")

if __name__ == "__main__":
	main()