search

szagoruyko / cvpr15deepcompare

Code and models for "Learning to Compare Image Patches via Convolutional Neural Networks"
http://imagine.enpc.fr/~zagoruys/deepcompare.html
Other
468 stars 169 forks source link

Choice of Criterion #16

Open Supersak80 opened 8 years ago

Supersak80 commented 8 years ago

@szagoruyko Does it make any difference whether you use the nn.MarginCriterion() or nn.HingeCriterion()? In the paper, you mentioned the use of a hinge loss with an L2 regularization term on the weights. Is that what is being used in the code?

szagoruyko commented 8 years ago

@Supersak80 what is nn.HingeCriterion? The code directly corresponds to the paper, we used nn.MarginCriterion with weight decay.

Supersak80 commented 8 years ago

@szagoruyko sorry, nn.HingeEmbeddingCriterion.

"Creates a criterion that measures the loss given an input x which is a 1-dimensional vector and a label y (1 or -1). This is usually used for measuring whether two inputs are similar or dissimilar, e.g. using the L1 pairwise distance, and is typically used for learning nonlinear embeddings or semi-supervised learning."