vkgs-sim / tests /test_ground.py
kimjunhan's picture
Add files using upload-large-folder tool
4c75661 verified
Raw History Blame Contribute Delete
2.18 kB
"""์ง€๋ฉด ๋ชจ๋ธ (nusim.recon.ground) โ€” GPU ์—†์ด CPU ํ…์„œ๋กœ."""
import math
import torch
from nusim.recon.ground import GROUND_THICKNESS, flatten, ground_loss, init_flat
def test_ground_loss_flat_vs_bumpy():
g = torch.Generator().manual_seed(0)
xy = torch.rand(20000, 2, generator=g) * 20
m = torch.ones(20000, dtype=torch.bool)
flat = torch.cat([xy, torch.zeros(20000, 1)], 1)
bumpy = torch.cat([xy, torch.randn(20000, 1, generator=g) * 0.1], 1)
assert float(ground_loss(flat, m)) < 1e-9
assert abs(float(ground_loss(bumpy, m)) - 0.01) < 0.002 # ๋ถ„์‚ฐ = 0.1ยฒ
def test_ground_loss_slope_is_small_per_patch():
# 5% ๊ฒฝ์‚ฌ: 1 m ์นธ ์•ˆ ๋†’์ด ์ฐจ 5 cm โ†’ ๋ถ„์‚ฐ โ‰ˆ 0.05ยฒ/12 โ€” ํ‰ํƒ„ํ•˜์ง€ ์•Š๋‹ค๊ณ  ํฌ๊ฒŒ ๋ฒŒ์ฃผ์ง€ ์•Š๋Š”๋‹ค
xy = torch.rand(20000, 2) * 20
z = 0.05 * xy[:, :1]
v = float(ground_loss(torch.cat([xy, z], 1), torch.ones(20000, dtype=torch.bool)))
assert v < 0.05 ** 2 / 12 * 1.5
def test_ground_loss_ignores_non_ground_and_has_grad():
means = torch.cat([torch.rand(1000, 2) * 5, torch.randn(1000, 1)], 1).requires_grad_(True)
mask = torch.zeros(1000, dtype=torch.bool)
assert float(ground_loss(means, mask)) == 0.0
mask[:500] = True
ground_loss(means, mask).backward()
assert means.grad[500:].abs().sum() == 0 and means.grad[:500, 2].abs().sum() > 0
def test_flatten_only_touches_ground():
p = {"quats": torch.nn.Parameter(torch.randn(4, 4)), "scales": torch.nn.Parameter(torch.randn(4, 3))}
before_q, before_s = p["quats"].data.clone(), p["scales"].data.clone()
mask = torch.tensor([True, False, True, False])
init_flat(p, mask)
assert torch.equal(p["quats"].data[0], torch.tensor([1.0, 0, 0, 0]))
assert math.isclose(float(p["scales"].data[2, 2]), math.log(GROUND_THICKNESS), rel_tol=1e-6)
assert float(p["scales"].data[0, 0]) == float(p["scales"].data[0, 1])
assert torch.equal(p["quats"].data[~mask], before_q[~mask])
assert torch.equal(p["scales"].data[~mask], before_s[~mask])
p["quats"].data[0] = torch.tensor([0.0, 1, 0, 0])
flatten(p, mask)
assert torch.equal(p["quats"].data[0], torch.tensor([1.0, 0, 0, 0]))