search

bernard24 / RIS

Implementation of the approach described in the paper "Recurrent Instance Segmentation" https://arxiv.org/abs/1511.08250.
MIT License
27 stars 17 forks source link

is it normal for a slow convergence? #14

Open kuaitoukid opened 7 years ago

kuaitoukid commented 7 years ago

I am trying to apply this method to detect other objects, however, even there is only one training sample without augmentation operation, this framework is hard to convergence? Is it normal? I follow the training method for people segmentation, and only train the 8-2 and lstm model.

isn4 commented 7 years ago

@bernard24 I've noticed slow convergence as well with the plants segmentation. I've got about 2000 training images and when training for 100,000 iterations it's taking roughly 48 hours to get through only 60,000 iterations.

bernard24 commented 7 years ago

Not sure if this is your case, but as far as I have seen, convergence is significantly harder to achieve if the instances are very different among them (e.g. one experiment in the paper is on segmenting people, but segmenting simultaneously people and other objects becomes harder). In any case, starting training with a low number of iterations for the ConvLSTM (e.g. 2) and increasing it only when it achieves convergence, seems to help.

@isn4 do you mean that the time per iteration is too slow?