janesjanes / sketchy

MIT License
161 stars 47 forks source link

Weight parameters of different models #6

Open pkmital opened 6 years ago

pkmital commented 6 years ago

I was interested in exploring sketchy's Triplet Loss + Classification Loss network, though having a little trouble dissecting the different losses in the network, and what weighting you would have used for each of these for the model described in your paper. If I try fine tuning with your model and using the sketchy db w/ 125 categories and removing the invalid data from your shared annotations lists, using what I think are image lists that you might have used, the Euclidean loss starts to blow up. Is this expected behavior, or did you weight this loss higher than 0 to ensure it did not blow up?

Is this network solver/prototxt exactly the same as what you used for the reported Triplet Loss + Classification Loss network in your paper: https://github.com/janesjanes/sketchy/blob/master/training/Triplet_googlenet_train_test.prototxt

It looks like the softmax loss for the sketch was set to 10, negative image to 0, and positive to 10. And for triplet loss, you had three separate ones set to 0.3, 0.3 and 1 (not sure I understand what each of these three separate losses are?), and for contrastive loss, all are set to 0. Finally, there is also a euclidean loss which is set to 0. Any advice/guidance would be much appreciated!