How to train a model with different task datasets at the same time?

[kamiLight]

Now i have two datasets with different tasks, dataset-A (detection task) and dataset-B (segmentation task), these two datasets has different image, i.e. one image only has detection label or segmentation label. The model has a shared backbone, and two head for detection task and segmentation task respectively. I want to train this model in multiple gpu, in first step, all gpu get the samples from dataset-A and forward/backward, in second step, all gpu get the samples from dataset-B and forward/backward, and so on until finish the training process, how can i achieve this? Thanks!