Track a series of values and provide access to smoothed values over a window or the global series average.
| 147 | |
| 148 | |
| 149 | class SmoothedValue(object): |
| 150 | """Track a series of values and provide access to smoothed values over a |
| 151 | window or the global series average. |
| 152 | """ |
| 153 | |
| 154 | def __init__(self, window_size=1000, fmt=None): |
| 155 | if fmt is None: |
| 156 | fmt = "{median:.4f} ({avg:.4f})" |
| 157 | self.deque = deque(maxlen=window_size) |
| 158 | self.total = 0.0 |
| 159 | self.count = 0 |
| 160 | self.fmt = fmt |
| 161 | |
| 162 | def update(self, value, n=1): |
| 163 | self.deque.append(value) |
| 164 | self.count += n |
| 165 | self.total += value * n |
| 166 | |
| 167 | def synchronize_between_processes(self): |
| 168 | """ |
| 169 | Warning: does not synchronize the deque! |
| 170 | """ |
| 171 | if not is_dist_avail_and_initialized(): |
| 172 | return |
| 173 | t = torch.tensor([self.count, self.total], dtype=torch.float64, device='cuda') |
| 174 | dist.barrier() |
| 175 | dist.all_reduce(t) |
| 176 | t = t.tolist() |
| 177 | self.count = int(t[0]) |
| 178 | self.total = t[1] |
| 179 | |
| 180 | @property |
| 181 | def median(self): |
| 182 | d = torch.tensor(list(self.deque)) |
| 183 | return d.median().item() |
| 184 | |
| 185 | @property |
| 186 | def avg(self): |
| 187 | d = torch.tensor(list(self.deque), dtype=torch.float32) |
| 188 | return d.mean().item() |
| 189 | |
| 190 | @property |
| 191 | def global_avg(self): |
| 192 | return self.total / self.count |
| 193 | |
| 194 | @property |
| 195 | def max(self): |
| 196 | return max(self.deque) |
| 197 | |
| 198 | @property |
| 199 | def value(self): |
| 200 | return self.deque[-1] |
| 201 | |
| 202 | def __str__(self): |
| 203 | return self.fmt.format( |
| 204 | median=self.median, |
| 205 | avg=self.avg, |
| 206 | global_avg=self.global_avg, |