Few-shot instead of one-shot in demo inference

[open-mmlab]

Few-shot instead of one-shot in demo inference

[linyq17]

To adapt few-shot inference, forwarding support images and labels is enough. The model.before_forward_support() will handle them according to their labels. Also, the labels should in the range(0, number of classes).

[rlleshi]

Hi @linyq17

It's actually not working like that. I am forwarding a couple of support images and labels using process_support_images() but only the last forwarded images are being used to build the support features and labels.

In other words, if forwarding two different sets of support images, set-1, and set-2 (in this order), the results of the inference will be the selfsame as if forwarding only the latter set-2.

As already mentioned above, this seems to me to be the case because in before_forward_support() for meta_baseline_head we are resetting the support features.

[linyq17]

It seems the process_support_images was called multiple times. Each time this function will reset the head and the code is in here. If the support images need to be passed multiple times, this line of code needs to be modified.

[rlleshi]

Exactly. I had modified only that and I got different results. But wanted to confirm once. Thanks.

before_forward_query() (for meta baseline) on the other hand need not be modified.