Restoring pre-trained model error about facenet

[davidsandberg]

Hi!
Sorry for the delayed response...
Which revision of the facenet repository are you running?
From the error message it looks like a tensor name between the stored model and the creation of the network differs.
I just tried with the latest and that one seems to work fine.

[yobibyte]

Hi, David!

Thank you, it works now. I would also like to ask some questions about the implementation details. Do you have time? Should I do it here in the comments? Or email/skype/facebook is better?

[davidsandberg]

Hi,
Great that it works!
Feel free to ask questions in this thread.

[yobibyte]

The main question is about training loop here.

• On line 147 you start iterating over the epochs and for every epoch you call train() function.
• On line 175 you start iterating over the batches
• but for some reason from line 202 you iterate over batches again.

Why do you need to iterate over batches inside every batch iteration?

[davidsandberg]

This is a little confusing in the code and has to do with the fact that the facenet paper uses a batch size of 40_45 examples. But this requires quite a bit of GPU memory and to still be able to train using 40_45 examples the batch is split into fragments and processed sequentially in the code.

[yobibyte]

Got it. Thanks. Have you tried smaller batches? Does it dramatically affect the performance?

[davidsandberg]

I haven't experimented with different batch sizes so I don't have a feel for how it affects performance.

[yobibyte]

Okay. Thanks for all the comments.