Track a series of values and provide access to smoothed values over a window or the global series average.
| 202 | |
| 203 | |
| 204 | class SmoothedValue(object): |
| 205 | """Track a series of values and provide access to smoothed values over a |
| 206 | window or the global series average. |
| 207 | """ |
| 208 | |
| 209 | def __init__(self, window_size=20, fmt=None): |
| 210 | if fmt is None: |
| 211 | fmt = "{median:.6f} ({global_avg:.6f})" |
| 212 | self.deque = deque(maxlen=window_size) |
| 213 | self.total = 0.0 |
| 214 | self.count = 0 |
| 215 | self.fmt = fmt |
| 216 | |
| 217 | def update(self, value, n=1): |
| 218 | self.deque.append(value) |
| 219 | self.count += n |
| 220 | self.total += value * n |
| 221 | |
| 222 | def synchronize_between_processes(self): |
| 223 | """ |
| 224 | Warning: does not synchronize the deque! |
| 225 | """ |
| 226 | if not is_dist_avail_and_initialized(): |
| 227 | return |
| 228 | t = torch.tensor([self.count, self.total], dtype=torch.float64, device='cuda') |
| 229 | dist.barrier() |
| 230 | dist.all_reduce(t) |
| 231 | t = t.tolist() |
| 232 | self.count = int(t[0]) |
| 233 | self.total = t[1] |
| 234 | |
| 235 | @property |
| 236 | def median(self): |
| 237 | d = torch.tensor(list(self.deque)) |
| 238 | return d.median().item() |
| 239 | |
| 240 | @property |
| 241 | def avg(self): |
| 242 | d = torch.tensor(list(self.deque), dtype=torch.float32) |
| 243 | return d.mean().item() |
| 244 | |
| 245 | @property |
| 246 | def global_avg(self): |
| 247 | return self.total / self.count |
| 248 | |
| 249 | @property |
| 250 | def max(self): |
| 251 | return max(self.deque) |
| 252 | |
| 253 | @property |
| 254 | def value(self): |
| 255 | return self.deque[-1] |
| 256 | |
| 257 | def __str__(self): |
| 258 | return self.fmt.format( |
| 259 | median=self.median, |
| 260 | avg=self.avg, |
| 261 | global_avg=self.global_avg, |