(self, args, logger=None)
| 82 | self.scheduler = scheduler |
| 83 | |
| 84 | def train(self, args, logger=None): |
| 85 | |
| 86 | ngpus_per_node = torch.cuda.device_count() |
| 87 | |
| 88 | # EMA Init |
| 89 | self.model.train() |
| 90 | self.ema = EMA(self.model, self.ema_m) |
| 91 | self.ema.register() |
| 92 | if args.resume == True: |
| 93 | self.ema.load(self.ema_model) |
| 94 | |
| 95 | # for gpu profiling |
| 96 | start_batch = torch.cuda.Event(enable_timing=True) |
| 97 | end_batch = torch.cuda.Event(enable_timing=True) |
| 98 | start_run = torch.cuda.Event(enable_timing=True) |
| 99 | end_run = torch.cuda.Event(enable_timing=True) |
| 100 | |
| 101 | start_batch.record() |
| 102 | best_eval_acc, best_it = 0.0, 0 |
| 103 | |
| 104 | scaler = GradScaler() |
| 105 | amp_cm = autocast if args.amp else contextlib.nullcontext |
| 106 | |
| 107 | # eval for once to verify if the checkpoint is loaded correctly |
| 108 | if args.resume == True: |
| 109 | eval_dict = self.evaluate(args=args) |
| 110 | print(eval_dict) |
| 111 | |
| 112 | self.lb_prob_t = torch.ones((args.num_classes)).cuda(args.gpu) / args.num_classes |
| 113 | self.ulb_prob_t = torch.ones((args.num_classes)).cuda(args.gpu) / args.num_classes |
| 114 | self.prob_max_mu_t = 1.0 / args.num_classes |
| 115 | self.prob_max_var_t = 1.0 |
| 116 | |
| 117 | for (_, x_lb, y_lb), (x_ulb_idx, x_ulb_w, x_ulb_s) in zip(self.loader_dict['train_lb'], |
| 118 | self.loader_dict['train_ulb']): |
| 119 | |
| 120 | # prevent the training iterations exceed args.num_train_iter |
| 121 | if self.it > args.num_train_iter: |
| 122 | break |
| 123 | |
| 124 | end_batch.record() |
| 125 | torch.cuda.synchronize() |
| 126 | start_run.record() |
| 127 | |
| 128 | num_lb = x_lb.shape[0] |
| 129 | num_ulb = x_ulb_w.shape[0] |
| 130 | assert num_ulb == x_ulb_s.shape[0] |
| 131 | |
| 132 | x_lb, x_ulb_w, x_ulb_s = x_lb.cuda(args.gpu), x_ulb_w.cuda(args.gpu), x_ulb_s.cuda(args.gpu) |
| 133 | y_lb = y_lb.cuda(args.gpu) |
| 134 | |
| 135 | inputs = torch.cat((x_lb, x_ulb_w, x_ulb_s)) |
| 136 | |
| 137 | # inference and calculate sup/unsup losses |
| 138 | with amp_cm(): |
| 139 | logits = self.model(inputs) |
| 140 | logits_x_lb = logits[:num_lb] |
| 141 | logits_x_ulb_w, logits_x_ulb_s = logits[num_lb:].chunk(2) |
no test coverage detected