These are chat archives for beniz/deepdetect

3rd
Oct 2016
dgtlmoon
@dgtlmoon
Oct 03 2016 19:42
can you suggest some other image dataset i can train on to test out the image similarity mechanism? i think my images have too much noise, even when retrained with 8000 images
Emmanuel Benazera
@beniz
Oct 03 2016 20:14
use the pre trained GoogLeNet and test on anything, like you own pictures. I remember I used this one
Esteban Siravegna
@Esiravegna
Oct 03 2016 22:24
@dgtlmoon have you considered using an autoencoder? Perhaps you could use one, train the network on the images you want to learn, and then use half of it (the embedding) to train another net (or using ANNOY or the like) in order to compute similarity?
btw. does somebody have any doc I can check on where to start in order to add a resnet with continous output as opposed to a classifier?