CUDA prefetcher. Ref: https://github.com/NVIDIA/apex/issues/304# It may consums more GPU memory. Args: loader: Dataloader. opt (dict): Options.
| 82 | |
| 83 | |
| 84 | class CUDAPrefetcher(): |
| 85 | """CUDA prefetcher. |
| 86 | |
| 87 | Ref: |
| 88 | https://github.com/NVIDIA/apex/issues/304# |
| 89 | |
| 90 | It may consums more GPU memory. |
| 91 | |
| 92 | Args: |
| 93 | loader: Dataloader. |
| 94 | opt (dict): Options. |
| 95 | """ |
| 96 | |
| 97 | def __init__(self, loader, opt): |
| 98 | self.ori_loader = loader |
| 99 | self.loader = iter(loader) |
| 100 | self.opt = opt |
| 101 | self.stream = torch.cuda.Stream() |
| 102 | self.device = torch.device('cuda' if opt['num_gpu'] != 0 else 'cpu') |
| 103 | self.preload() |
| 104 | |
| 105 | def preload(self): |
| 106 | try: |
| 107 | self.batch = next(self.loader) # self.batch is a dict |
| 108 | except StopIteration: |
| 109 | self.batch = None |
| 110 | return None |
| 111 | # put tensors to gpu |
| 112 | with torch.cuda.stream(self.stream): |
| 113 | for k, v in self.batch.items(): |
| 114 | if torch.is_tensor(v): |
| 115 | self.batch[k] = self.batch[k].to(device=self.device, non_blocking=True) |
| 116 | |
| 117 | def next(self): |
| 118 | torch.cuda.current_stream().wait_stream(self.stream) |
| 119 | batch = self.batch |
| 120 | self.preload() |
| 121 | return batch |
| 122 | |
| 123 | def reset(self): |
| 124 | self.loader = iter(self.ori_loader) |
| 125 | self.preload() |