Skip to first unread message
Filipe Trocado Ferreira
Nov 30, 2016, 12:29:20 PM11/30/16
to Caffe Users
What is the best way to train a network to solve a classification and regression task simultaneously and deal with missing labels in some training data?
Miquel Martí
Nov 30, 2016, 10:21:33 PM11/30/16
to Caffe Users
Check this paper UberNet : Training a ‘Universal’ Convolutional Neural Network for Low-, Mid-, and High-Level Vision using Diverse Datasets and Limited Memory and demo at http://cvn.ecp.fr/ubernet/ .
El dijous, 1 desembre de 2016 2:29:20 UTC+9, Filipe Trocado Ferreira va escriure:
Specific for convnets but hope it helps.
Best
El dijous, 1 desembre de 2016 2:29:20 UTC+9, Filipe Trocado Ferreira va escriure:
Filipe Trocado Ferreira
Dec 1, 2016, 6:04:39 PM12/1/16
to Caffe Users
Thanks for the response. I'm aware of the concept. My particular question is how to deal with missing labels. For example, how can I ignore some labels in Euclidean Loss?
Miquel Martí
Dec 6, 2016, 1:04:50 AM12/6/16
to Caffe Users
I think you will have to modify the Euclidean Loss layer yourself to give zero loss when there is no label available. I don't think it is possible to do it with the current implementation of the loss layers. You might also want to implement the asynchronous weight update idea of the paper to avoid problems when few or no samples for a given task were included in a mini-batch.
El divendres, 2 desembre de 2016 8:04:39 UTC+9, Filipe Trocado Ferreira va escriure:
El divendres, 2 desembre de 2016 8:04:39 UTC+9, Filipe Trocado Ferreira va escriure:
Reply all
Reply to author
Forward
0 new messages