Facenet: why use 512 dimension embeddings while not use 128 dimensions?

[billtiger]

in the last version,the bottlelecksize is 128,but i attention that you are using 512 now!
is that 512 dimension embeddings will have better performance?
@davidsandberg

[davidsandberg]

There seem to be cases where 128 dimensional embeddings are a little bit on the low side and increasing the dimensionality increases performance a little bit. But it's not any huge differences we are talking about.

[billtiger]

thank you for your reply! I have another question about the code:
what's the difference of the two operations:
(tf.cast(image, tf.float32) - 127.5)/128.0 and tf.image.per_image_standardization(image)) ?
because you have comment that --use_fixed_image_standardization.
@davidsandberg

[Alloshee]

so we are using the same inception architecture explained in the original paper but with the fully connected layer being 512 instead of 128 ??
is that right ?
@davidsandberg
@billtiger