Spaces:
Paused
Paused
File size: 5,546 Bytes
7f3c2df |
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 |
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact [email protected]
#
import torch
import sys
from datetime import datetime
import numpy as np
import random
import os
import cv2
def inverse_sigmoid(x):
return torch.log(x/(1-x))
def PILtoTorch(pil_image, resolution):
resized_image_PIL = pil_image.resize(resolution)
resized_image = torch.from_numpy(np.array(resized_image_PIL)) / 255.0
if len(resized_image.shape) == 3:
return resized_image.permute(2, 0, 1)
else:
return resized_image.unsqueeze(dim=-1).permute(2, 0, 1)
def PIL2toTorch(pil_image, resolution):
resized_image_PIL = pil_image.resize(resolution)
resized_image = torch.from_numpy(np.array(resized_image_PIL)) / 255.0 * (2.0 ** 16 - 1.0)
return resized_image
def decode_op(optical_png):
# use 'PIL Image.Open' to READ
"Convert from .png (h, w, 3-rgb) -> (h,w,2)(flow_x, flow_y) .. float32 array"
optical_png = optical_png[..., [2, 1, 0]] # bgr -> rgb
h, w, _c = optical_png.shape
assert optical_png.dtype == np.uint16 and _c == 3
"invalid flow flag: b == 0 for sky or other invalid flow"
invalid_points = np.where(optical_png[..., 2] == 0)
out_flow = torch.empty((h, w, 2))
decoded = 2.0 / (2**16 - 1.0) * optical_png.astype('f4') - 1
out_flow[..., 0] = torch.tensor(decoded[:, :, 0] * (w - 1)) # (pixel) delta_x : R
out_flow[..., 1] = torch.tensor(decoded[:, :, 1] * (h - 1)) # delta_y : G
out_flow[invalid_points[0], invalid_points[1], :] = 0 # B=0 for invalid flow
return out_flow
def get_expon_lr_func(
lr_init, lr_final, lr_delay_steps=0, lr_delay_mult=1.0, max_steps=1000000
):
"""
Copied from Plenoxels
Continuous learning rate decay function. Adapted from JaxNeRF
The returned rate is lr_init when step=0 and lr_final when step=max_steps, and
is log-linearly interpolated elsewhere (equivalent to exponential decay).
If lr_delay_steps>0 then the learning rate will be scaled by some smooth
function of lr_delay_mult, such that the initial learning rate is
lr_init*lr_delay_mult at the beginning of optimization but will be eased back
to the normal learning rate when steps>lr_delay_steps.
:param conf: config subtree 'lr' or similar
:param max_steps: int, the number of steps during optimization.
:return HoF which takes step as input
"""
def helper(step):
if step < 0 or (lr_init == 0.0 and lr_final == 0.0):
# Disable this parameter
return 0.0
if lr_delay_steps > 0:
# A kind of reverse cosine decay.
delay_rate = lr_delay_mult + (1 - lr_delay_mult) * np.sin(
0.5 * np.pi * np.clip(step / lr_delay_steps, 0, 1)
)
else:
delay_rate = 1.0
t = np.clip(step / max_steps, 0, 1)
log_lerp = np.exp(np.log(lr_init) * (1 - t) + np.log(lr_final) * t)
return delay_rate * log_lerp
return helper
def strip_lowerdiag(L):
uncertainty = torch.zeros((L.shape[0], 6), dtype=torch.float, device="cuda")
uncertainty[:, 0] = L[:, 0, 0]
uncertainty[:, 1] = L[:, 0, 1]
uncertainty[:, 2] = L[:, 0, 2]
uncertainty[:, 3] = L[:, 1, 1]
uncertainty[:, 4] = L[:, 1, 2]
uncertainty[:, 5] = L[:, 2, 2]
return uncertainty
def strip_symmetric(sym):
return strip_lowerdiag(sym)
def build_rotation(r):
norm = torch.sqrt(r[:,0]*r[:,0] + r[:,1]*r[:,1] + r[:,2]*r[:,2] + r[:,3]*r[:,3])
q = r / norm[:, None]
R = torch.zeros((q.size(0), 3, 3), device='cuda')
r = q[:, 0]
x = q[:, 1]
y = q[:, 2]
z = q[:, 3]
R[:, 0, 0] = 1 - 2 * (y*y + z*z)
R[:, 0, 1] = 2 * (x*y - r*z)
R[:, 0, 2] = 2 * (x*z + r*y)
R[:, 1, 0] = 2 * (x*y + r*z)
R[:, 1, 1] = 1 - 2 * (x*x + z*z)
R[:, 1, 2] = 2 * (y*z - r*x)
R[:, 2, 0] = 2 * (x*z - r*y)
R[:, 2, 1] = 2 * (y*z + r*x)
R[:, 2, 2] = 1 - 2 * (x*x + y*y)
return R
def build_scaling_rotation(s, r):
L = torch.zeros((s.shape[0], 3, 3), dtype=torch.float, device="cuda")
R = build_rotation(r)
L[:,0,0] = s[:,0]
L[:,1,1] = s[:,1]
L[:,2,2] = s[:,2]
L = R @ L
return L
DEFAULT_RANDOM_SEED = 0
def seedBasic(seed=DEFAULT_RANDOM_SEED):
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
np.random.seed(seed)
def seedTorch(seed=DEFAULT_RANDOM_SEED):
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# basic + tensorflow + torch
def seedEverything(seed=DEFAULT_RANDOM_SEED):
seedBasic(seed)
seedTorch(seed)
def safe_state(silent):
old_f = sys.stdout
class F:
def __init__(self, silent):
self.silent = silent
def write(self, x):
if not self.silent:
if x.endswith("\n"):
old_f.write(x.replace("\n", " [{}]\n".format(str(datetime.now().strftime("%d/%m %H:%M:%S")))))
else:
old_f.write(x)
def flush(self):
old_f.flush()
sys.stdout = F(silent)
random.seed(DEFAULT_RANDOM_SEED)
np.random.seed(DEFAULT_RANDOM_SEED)
torch.manual_seed(DEFAULT_RANDOM_SEED)
torch.cuda.set_device(torch.device("cuda:0"))
# sys.stdout = old_f
|