Why training features?

[saiaman]

Hello great code and article. I've got a small question : you train the triplet with features and not image data : is this a good way ?
In fact i'm asking myself : the main goal is to get better features from the triplet so why do you use a ImageNetFeatureExtractor ?

[saideeptalari]

Good question. Actually we want to find the similarities between images and ImageNet classifiers extract decent features at the top layers on basic images. For the purpose of the blog post, I just used the pretrained model to get to the subject straight away. But, if you have a very complex dataset such as fashion, you need to train the whole model to optimise for extracting better features with triplet as a loss. It's a bit tricky to train though. Infact, it's good to train a classifier on the dataset first and let it extract accurate features and then use that as a feature extractor. Either ways, the desired accuracy can be achieved.

…

On Thu 23 May, 2019, 2:25 PM saiaman ***@***.*** wrote: Hello great code and article. I've got a small question : you train the triplet with features and not image data : is this a good way ? In fact i'm asking myself : the main goal is to get better features from the triplet so why do you use a ImageNetFeatureExtractor ? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#1?email_source=notifications&email_token=ACZKAPJ2R37Z5V4G5LXETMDPWZLZLA5CNFSM4HO3LSAKYY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4GVMX43Q>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACZKAPLXOVFKQBFUGG2ZSITPWZLZLANCNFSM4HO3LSAA> .