Open InnovArul opened 6 years ago
@InnovArul In the prototxt file (https://github.com/jasjeetIM/Mask-RCNN/blob/master/models/mask_rcnn_prototxts/resnet_50layers_conv/train_val.prototxt#L3741), the model ignore the labels which equal -1 when computing the loss.
Can anyone give a clarification on the intuition of using no_loss_mask = -1? https://github.com/jasjeetIM/Mask-RCNN/blob/master/functions/mask_rcnn/mask_get_minibatch.m#L131
As far as I can tell, cross-entropy loss expects target = [0, 1].