we try to train a multitask model with the mask for the images which does not correspond to a specific task, but with two totally different datasets, we have a batch which contains no images for the second task, so the implement we before done in the following line :
Scenario:
problem of entrain a model for multitasking by launching the following command:
python tools/train.py configs/multi_task.py --work-dir ../tmp/work_dir_multi
Context :
we try to train a multitask model with the mask for the images which does not correspond to a specific task, but with two totally different datasets, we have a batch which contains no images for the second task, so the implement we before done in the following line :
https://github.com/open-mmlab/mmclassification/blob/c7ec630c37dd2c7c997405a70b9fafa6a54e5f95/mmcls/models/heads/multi_task_head.py#L26
do not allow us backward calculation of the loss