File size: 6,333 Bytes
69b6acd | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 | import os
import cv2
import time
import random
import numpy as np
import torch
from torchmetrics import MeanMetric as TorchMeanMetric
from nerv.utils.tensor import to_numpy
def datetime2str(form='%Y-%m-%d_%H-%M-%S'):
datetime = time.strftime(form, time.localtime())
return datetime
def assert_array_shape(xyz, shapes=()):
"""Check array shape.
Args:
xyz (np.ndarray): array
shape (tuple of tuple of ints, optional): possible target shapes,
-1 means arbitrary. Defaults to ((-1, 3)).
"""
if not shapes:
raise ValueError('"shapes" cannot be empty')
if isinstance(shapes[0], int):
shapes = (shapes, )
flags = {x: True for x in range(len(shapes))}
for idx, shape in enumerate(shapes):
if len(xyz.shape) != len(shape):
flags[idx] = False
for dim, num in enumerate(shape):
if num == -1:
continue
elif xyz.shape[dim] != num:
flags[idx] = False
if sum(flags.values()) == 0: # None of the possible shape works
raise ValueError(
f'Input array {xyz.shape} is not in target shapes {shapes}!')
def array_equal(a, b):
"""Compare if two arrays are the same.
Args:
a/b: can be np.ndarray or torch.Tensor.
"""
if a.shape != b.shape:
return False
try:
assert (a == b).all()
return True
except:
return False
def array_in_list(array, lst):
"""Judge whether an array is in a list."""
for v in lst:
if array_equal(array, v):
return True
return False
def set_seed(seed=1, deterministic=False):
"""Set the random seed of the environment for reproducibility.
Args:
seed (int): the random seed to set.
deterministic (bool, optional): whether to use deterministic torch
backend. Default: False.
"""
print('Using random seed:', seed)
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
if deterministic:
print('Using deterministic pytorch backends')
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
class AverageMeter(object):
"""Computes and stores the average and current value."""
def __init__(self, device=torch.device('cpu')):
self.reset()
self.device = device
def reset(self):
self.val = 0.
self.avg = 0.
self.sum = 0.
self.count = 0.
def update(self, val, n=1):
self.val = val
self.sum += val * n
self.count += n
self.avg = self.sum / self.count
def compute(self):
return torch.tensor(self.avg).to(self.device)
def to(self, device):
self.device = device
return self
class MeanMetric:
"""Compatibility wrapper around `torchmetrics.MeanMetric`.
Provides a small API expected by the codebase: `update(value, weight=1.0)`,
`compute()` and `to(device)`. Works with torchmetrics 1.x and 2.x.
"""
def __init__(self, device, nan_strategy="warn", compute_on_step=True, dist_sync_on_step=False, process_group=None, dist_sync_fn=None):
self.device = device
# create underlying torchmetrics MeanMetric and move to device
self._metric = TorchMeanMetric()
try:
self._metric.to(device)
except Exception:
pass
def update(self, value, weight=1.0):
# torchmetrics MeanMetric.update accepts (value, weight) in recent versions,
# older variations may accept only value — try both.
try:
return self._metric.update(value, weight)
except TypeError:
return self._metric.update(value)
def compute(self):
out = self._metric.compute()
if isinstance(out, torch.Tensor):
return out.to(self.device)
try:
return torch.tensor(out).to(self.device)
except Exception:
return out
def to(self, device):
try:
self._metric.to(device)
except Exception:
pass
self.device = device
return self
def _convert4save_img(array):
"""Convert a image array to be saveable."""
# extend channel axis
if len(array.shape) == 2:
array = np.stack([array] * 3, axis=-1)
# [C, H, W] --> [H, W, C], where C == 3
if array.shape[0] == 3:
array = array.transpose(1, 2, 0)
return np.ascontiguousarray(array)
def _convert4save_video(array):
"""Convert a video array to be saveable."""
# extend channel axis
if len(array.shape) == 3:
array = np.stack([array] * 3, axis=-1)
# [T, C, H, W] --> [T, H, W, C], where C == 3
if array.shape[1] == 3:
array = array.transpose(0, 2, 3, 1)
return np.ascontiguousarray(array)
def convert4save(array, is_video=False):
"""Check the dtype and value range of input array for save.
Need to convert to [(T), H, W, C] with np.uint8 value range [0, 255].
Args:
array (np.ndarray or torch.Tensor): array to be converted.
is_video (bool, optional): whether the array is a video or image.
Default: False (means array is an image).
Returns:
np.ndarray: the converted array ready for save (image or video).
"""
array = to_numpy(array)
if 'int' in str(array.dtype):
assert 0 <= array.min() <= array.max() <= 255
elif 'float' in str(array.dtype):
assert 0. <= array.min() <= array.max() <= 1.
array = np.round(array * 255.).astype(np.uint8)
if is_video:
return _convert4save_video(array)
else:
return _convert4save_img(array)
def save_video(video, save_path, fps=30, codec='mp4v'):
"""video: np.ndarray of shape [M, H, W, 3]"""
os.makedirs(os.path.dirname(save_path), exist_ok=True)
video = convert4save(video, is_video=True)
video = video[..., [2, 1, 0]]
H, W = video.shape[-3:-1]
assert save_path.split('.')[-1] == 'mp4' # save as mp4 file
# opencv has opposite dimension definition as numpy
size = [W, H]
out = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*codec), fps, size)
for i in range(video.shape[0]):
out.write(video[i])
out.release()
|