:param sample: :param batch_idx: :param optimizer_idx: :return: {'loss': torch.Tensor, 'progress_bar': dict, 'tb_log': dict}
(self, sample, batch_idx, optimizer_idx=-1)
| 103 | raise NotImplementedError |
| 104 | |
| 105 | def training_step(self, sample, batch_idx, optimizer_idx=-1): |
| 106 | """ |
| 107 | |
| 108 | :param sample: |
| 109 | :param batch_idx: |
| 110 | :param optimizer_idx: |
| 111 | :return: {'loss': torch.Tensor, 'progress_bar': dict, 'tb_log': dict} |
| 112 | """ |
| 113 | # perform the main training step in a specific task |
| 114 | loss_ret = self._training_step(sample, batch_idx, optimizer_idx) |
| 115 | if loss_ret is None: |
| 116 | return {'loss': None} |
| 117 | total_loss, log_outputs = loss_ret |
| 118 | log_outputs = tensors_to_scalars(log_outputs) |
| 119 | |
| 120 | # add to epoch meter |
| 121 | for k, v in log_outputs.items(): |
| 122 | if '/' in k: |
| 123 | k_split = k.split("/") |
| 124 | assert len(k_split) == 2, "we only support one `/` in tag_name, i.e., `<tag>/<sub_tag>`" |
| 125 | k = k.replace("/", "_") |
| 126 | if k not in self.epoch_training_losses_meter: |
| 127 | self.epoch_training_losses_meter[k] = AvgrageMeter() |
| 128 | if not np.isnan(v): |
| 129 | self.epoch_training_losses_meter[k].update(v) |
| 130 | |
| 131 | if optimizer_idx >= 0: |
| 132 | for params_group_i in range(len(self.trainer.optimizers[optimizer_idx].param_groups)): |
| 133 | log_outputs[f'lr/optimizer{optimizer_idx}_params_group{params_group_i}'] = self.trainer.optimizers[optimizer_idx].param_groups[params_group_i]['lr'] |
| 134 | |
| 135 | # add to progress bar |
| 136 | progress_bar_log = {} |
| 137 | for k, v in log_outputs.items(): |
| 138 | if '/' in k: |
| 139 | k_split = k.split("/") |
| 140 | assert len(k_split) == 2, "we only support one `/` in tag_name, i.e., `<tag>/<sub_tag>`" |
| 141 | k = k.replace("/", "_") |
| 142 | assert k not in progress_bar_log, f"we got duplicate tags in log_outputs, check this `{k}`" |
| 143 | progress_bar_log[k] = v |
| 144 | |
| 145 | # add to progress bar |
| 146 | tb_log = {} |
| 147 | for k, v in log_outputs.items(): |
| 148 | if '/' in k: |
| 149 | tb_log[k] = v |
| 150 | else: |
| 151 | tb_log[f'tr/{k}'] = v |
| 152 | |
| 153 | if not isinstance(total_loss, torch.Tensor): |
| 154 | return {'loss': None} |
| 155 | self.epoch_training_losses_meter['total_loss'].update(total_loss.item()) |
| 156 | |
| 157 | return { |
| 158 | 'loss': total_loss, |
| 159 | 'progress_bar': progress_bar_log, |
| 160 | 'tb_log': tb_log |
| 161 | } |
| 162 |
no test coverage detected