Facenet: Pre-trained model returns 512 vector, I need 128

[nora-illanes]

Hello everybody,

I'm using either 20180408-102900 or 20180402-114759 for .pb pre-trained model in a very small set of images. It works well using classify.py (I'm only interested in the embedding). However, I really need the 128 element vector, since the final dataset will contain 10*10^3 images.

How can I make the embedding coming from the LFW-trained-model return such a vector?

Really need this help.

[seth814]

Why do you feel like you need to reduce dimensions? I would recommend you just use the full 512 vector. I recently trained on the full 3.3 million images from vggface2 and it went okay. There shouldn't be any reason why you want to reduce these from a classification perspective. Of course you might run out of RAM while loading all the data, so you should batch data during training using a neural network like a simple hidden layer into softmax using fit_generator from keras or something.

If you want to try it, just perform PCA:

https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.PCA.html

This will usually create more general vectors and you will probably lose a lot of accuracy I would guess. Unless you are looking for features that measure similarity then this would be a good idea.

[nora-illanes]

Why do you feel like you .....There shouldn't be any reason why you want to reduce these from a classification perspective.

I need to create a library of embedding vectors to save it to a gigantic file.

If you want to try it, just perform PCA:
.... Unless you are looking for features that measure similarity then this would be a good idea.

Hmmm, I am looking for features of similarities. Thank you for your reply. The numbers in the vector are of huge dimensions, but I guess it is like that.

Thank you again!

[ryecomp]

How about training to get your own model ?
you can specify --embedding=128 when calling train_*.py script

[nora-illanes]

Thanks for your reply!

How about training to get your own model ?
you can specify --embedding=128 when calling train_*.py script

I will try this. Why are you running train_? I have only one script from facenet repository, and that is train.py.

[ryecomp]

you can find train_tripletloss and train_softmax in src folder

[neverUseThisName]

Maybe you can apply PCA to 512-d vectors and reduce them to 128-d.

[seth814]

I would agree with @ryecomp here. As long as reducing dims doesn't affect performance. @nora-illanes siamese nets can be trained with a triplet loss to reference euclidean distance later or softmax activation in the attempt of classification.

[WangLuyao627]

Hello, Have you managed to turn 512 dimensions into 128 dimensions？How do you do?