softmargin_triplet_loss about cosine_metric_learning HOT 7 CLOSED

If it's tensorboard, the faded line is the actual loss curve and the bold is the smoothed one. Change the smoothing to 0 and you will see.

from cosine_metric_learning.

In short, the triplets are generated on the GPU from the current batch. In each iteration, a fixed number of images are sampled for a number of random identities. The rest happens on GPU (therefore also the label array is needed). Sorry for the late response.

from cosine_metric_learning.

the label size is m , not n, why

from cosine_metric_learning.

Thats an error in the documentation. It should be N; has been fixed now.

from cosine_metric_learning.

thanks, how about the feature and labels example ?
feature [ [1,1] [2,2] [3,3] [4,4] ] >> (4,2) , the [1,1] and [2,2] are the same person
so the label will be the [ 1 1 2 2 ] >> (4,) right ?

from cosine_metric_learning.

this is loss image , do you know the mean of two line ?
I know the one is loss , but another , I don't know

from cosine_metric_learning.

one batch of images contains anchor, positive, negative samples, according to the labels vectors.

from cosine_metric_learning.