MultiTask empty label problem

[peng-peng0913]

I have been doing MultiTask recently, but I have encountered a problem.
My data.list

image name | label (task_1) | label (task_2) | label (task_3) |

xxx1.jpg 0 2 1
xxx2.jpg -1 1 0

If the label is - 1, it means that the image is not labeled. Direct training will report an error. How can I handle the - 1 label to match the input.

[mzr1996]

Hello, here are two methods:

1. Format your annotation file in the MultiTaskDataset required format.
   https://github.com/open-mmlab/mmclassification/blob/654a337d3e93364328044d110b25441d907f4bd6/mmcls/datasets/multi_task.py#L32-L58
2. Implement your own dataset class to handle your annotation file, you can find tutorial here, and the returned info should be like

[{'img_path': xxx1.jpg,
  'gt_label': {
      'task1': 0,
      'task2': 2,
      'task3': 1,
     }
 },
 {'img_path': xxx2.jpg,
  'gt_label': {
      'task2': 1,
      'task3': 0,
     }
 }
 ...
]

[peng-peng0913]

I've been using mmcls version 0.24 and writing my own MultiTask. I now see that the new version of mmcls supports MultiTask. I will update my mmcls version. Thank you.