| 66 | |
| 67 | |
| 68 | class CallBackLogging(object): |
| 69 | def __init__(self, frequent, total_step, batch_size, start_step=0,writer=None): |
| 70 | self.frequent: int = frequent |
| 71 | self.rank: int = distributed.get_rank() |
| 72 | self.world_size: int = distributed.get_world_size() |
| 73 | self.time_start = time.time() |
| 74 | self.total_step: int = total_step |
| 75 | self.start_step: int = start_step |
| 76 | self.batch_size: int = batch_size |
| 77 | self.writer = writer |
| 78 | |
| 79 | self.init = False |
| 80 | self.tic = 0 |
| 81 | |
| 82 | def __call__(self, |
| 83 | global_step: int, |
| 84 | loss: AverageMeter, |
| 85 | epoch: int, |
| 86 | fp16: bool, |
| 87 | learning_rate: float, |
| 88 | grad_scaler: torch.cuda.amp.GradScaler): |
| 89 | if self.rank == 0 and global_step > 0 and global_step % self.frequent == 0: |
| 90 | if self.init: |
| 91 | try: |
| 92 | speed: float = self.frequent * self.batch_size / (time.time() - self.tic) |
| 93 | speed_total = speed * self.world_size |
| 94 | except ZeroDivisionError: |
| 95 | speed_total = float('inf') |
| 96 | |
| 97 | #time_now = (time.time() - self.time_start) / 3600 |
| 98 | #time_total = time_now / ((global_step + 1) / self.total_step) |
| 99 | #time_for_end = time_total - time_now |
| 100 | time_now = time.time() |
| 101 | time_sec = int(time_now - self.time_start) |
| 102 | time_sec_avg = time_sec / (global_step - self.start_step + 1) |
| 103 | eta_sec = time_sec_avg * (self.total_step - global_step - 1) |
| 104 | time_for_end = eta_sec/3600 |
| 105 | if self.writer is not None: |
| 106 | self.writer.add_scalar('time_for_end', time_for_end, global_step) |
| 107 | self.writer.add_scalar('learning_rate', learning_rate, global_step) |
| 108 | self.writer.add_scalar('loss', loss.avg, global_step) |
| 109 | if fp16: |
| 110 | msg = "Speed %.2f samples/sec Loss %.4f LearningRate %.6f Epoch: %d Global Step: %d " \ |
| 111 | "Fp16 Grad Scale: %2.f Required: %1.f hours" % ( |
| 112 | speed_total, loss.avg, learning_rate, epoch, global_step, |
| 113 | grad_scaler.get_scale(), time_for_end |
| 114 | ) |
| 115 | else: |
| 116 | msg = "Speed %.2f samples/sec Loss %.4f LearningRate %.6f Epoch: %d Global Step: %d " \ |
| 117 | "Required: %1.f hours" % ( |
| 118 | speed_total, loss.avg, learning_rate, epoch, global_step, time_for_end |
| 119 | ) |
| 120 | logging.info(msg) |
| 121 | loss.reset() |
| 122 | self.tic = time.time() |
| 123 | else: |
| 124 | self.init = True |
| 125 | self.tic = time.time() |