(self, args, logger=None)
| 81 | self.scheduler = scheduler |
| 82 | |
| 83 | def warmup(self, args, logger=None): |
| 84 | ngpus_per_node = torch.cuda.device_count() |
| 85 | |
| 86 | self.model.train() |
| 87 | |
| 88 | # for gpu profiling |
| 89 | start_batch = torch.cuda.Event(enable_timing=True) |
| 90 | end_batch = torch.cuda.Event(enable_timing=True) |
| 91 | start_run = torch.cuda.Event(enable_timing=True) |
| 92 | end_run = torch.cuda.Event(enable_timing=True) |
| 93 | |
| 94 | warmup_it = 0 |
| 95 | start_batch.record() |
| 96 | |
| 97 | scaler = GradScaler() |
| 98 | amp_cm = autocast if args.amp else contextlib.nullcontext |
| 99 | |
| 100 | for _, x_lb, y_lb in self.loader_dict['train_lb']: |
| 101 | |
| 102 | # prevent the training iterations exceed args.num_train_iter |
| 103 | if warmup_it > 2048: |
| 104 | break |
| 105 | |
| 106 | end_batch.record() |
| 107 | torch.cuda.synchronize() |
| 108 | start_run.record() |
| 109 | |
| 110 | x_lb = x_lb.cuda(args.gpu) |
| 111 | y_lb = y_lb.cuda(args.gpu) |
| 112 | |
| 113 | num_lb = x_lb.shape[0] |
| 114 | |
| 115 | # inference and calculate sup/unsup losses |
| 116 | with amp_cm(): |
| 117 | |
| 118 | logits_x_lb = self.model(x_lb) |
| 119 | sup_loss = ce_loss(logits_x_lb, y_lb, use_hard_labels=True, reduction='mean') |
| 120 | |
| 121 | total_loss = sup_loss |
| 122 | |
| 123 | # parameter updates |
| 124 | if args.amp: |
| 125 | scaler.scale(total_loss).backward() |
| 126 | if (args.clip > 0): |
| 127 | torch.nn.utils.clip_grad_norm_(self.model.parameters(), args.clip) |
| 128 | scaler.step(self.optimizer) |
| 129 | scaler.update() |
| 130 | else: |
| 131 | total_loss.backward() |
| 132 | if (args.clip > 0): |
| 133 | torch.nn.utils.clip_grad_norm_(self.model.parameters(), args.clip) |
| 134 | self.optimizer.step() |
| 135 | |
| 136 | self.model.zero_grad() |
| 137 | |
| 138 | end_run.record() |
| 139 | torch.cuda.synchronize() |
| 140 |
no test coverage detected