(self, args, logger=None)
| 80 | self.scheduler = scheduler |
| 81 | |
| 82 | def train(self, args, logger=None): |
| 83 | |
| 84 | ngpus_per_node = torch.cuda.device_count() |
| 85 | |
| 86 | # EMA Init |
| 87 | self.model.train() |
| 88 | self.ema = EMA(self.model, self.ema_m) |
| 89 | self.ema.register() |
| 90 | if args.resume == True: |
| 91 | self.ema.load(self.ema_model) |
| 92 | |
| 93 | # p(y) based on the labeled examples seen during training |
| 94 | dist_file_name = r"./data_statistics/" + args.dataset + '_' + str(args.num_labels) + '.json' |
| 95 | if args.dataset.upper() == 'IMAGENET': |
| 96 | p_target = None |
| 97 | else: |
| 98 | with open(dist_file_name, 'r') as f: |
| 99 | p_target = json.loads(f.read()) |
| 100 | p_target = torch.tensor(p_target['distribution']) |
| 101 | p_target = p_target.cuda(args.gpu) |
| 102 | # print('p_target:', p_target) |
| 103 | |
| 104 | p_model = None |
| 105 | |
| 106 | # for gpu profiling |
| 107 | start_batch = torch.cuda.Event(enable_timing=True) |
| 108 | end_batch = torch.cuda.Event(enable_timing=True) |
| 109 | start_run = torch.cuda.Event(enable_timing=True) |
| 110 | end_run = torch.cuda.Event(enable_timing=True) |
| 111 | |
| 112 | start_batch.record() |
| 113 | best_eval_acc, best_it = 0.0, 0 |
| 114 | |
| 115 | scaler = GradScaler() |
| 116 | amp_cm = autocast if args.amp else contextlib.nullcontext |
| 117 | |
| 118 | # eval for once to verify if the checkpoint is loaded correctly |
| 119 | if args.resume == True: |
| 120 | eval_dict = self.evaluate(args=args) |
| 121 | print(eval_dict) |
| 122 | |
| 123 | selected_label = torch.ones((len(self.ulb_dset),), dtype=torch.long, ) * -1 |
| 124 | selected_label = selected_label.cuda(args.gpu) |
| 125 | |
| 126 | classwise_acc = torch.zeros((args.num_classes,)).cuda(args.gpu) |
| 127 | |
| 128 | for (_, x_lb, y_lb), (x_ulb_idx, x_ulb_w, x_ulb_s) in zip(self.loader_dict['train_lb'], |
| 129 | self.loader_dict['train_ulb']): |
| 130 | # prevent the training iterations exceed args.num_train_iter |
| 131 | if self.it > args.num_train_iter: |
| 132 | break |
| 133 | |
| 134 | end_batch.record() |
| 135 | torch.cuda.synchronize() |
| 136 | start_run.record() |
| 137 | |
| 138 | num_lb = x_lb.shape[0] |
| 139 | num_ulb = x_ulb_w.shape[0] |
no test coverage detected