A question about train skills

[wzp8023391]

Hi, nice to see the interesting work. I implement it in pytorch to train my private dataset (only 8 types, about 20000 images overall), I found that no matter I changed the support images types, the segmentation result kept unchagned, does it overfit?

[wzp8023391]

Another question is if the number of each training dataset is different, for example, A dataset have 1000 images, B dataset have 2000 images, when I train the universeg, should I resample B dataset as 1000 images, or train model directly? Or in each training batch, it have all kind of dataset? For example, a traininig batch have A, B, C, D dataset?

[VictorButoi]

Yeah imbalance between the number of datapoints for each task (dataset) is something that we also had to contend with!

If you are trying to emulate how we trained our models, then you should, in the get_item function of your dataloader, sample a task and then images belonging to that task. In this way, a training batch can indeed have a sampled (query, Support) pair from multiple tasks.