Training on custom dataset

[jgrzeszczyk]

Hi, I would like to ask what do i need to change in network files in order to train on custom coco format dataset. I've changed every instance of num_classses, num_thing_classes, num_stuff_classes and modified the config accordingly. Training epoch runs correctly, but I am facing the following during validation

File "/K-Net/knet/det/kernel_update_head.py", line 372, in _get_target_single
    mask_targets[pos_inds, ...] = pos_mask_targets
RuntimeError: shape mismatch: value tensor of shape [54, 168, 216] cannot be broadcast to indexing result of shape [54, 84, 108]

[ZwwWayne]

It seems that the sizes of pos_mask_targets and mask_targets are different for some reason. However, during validation, the code should not run into _get_target_single.