Facenet: how to inference an image?

[IvyGongoogle]

new to face recognition. Now I just want to inference an image, but I do not know run which py script in this repository.
Looking forward to your reply.
Thank you in advance.

[davidsandberg]

The script compare.py will calculate the embeddings for a few images and that is one way to run inference.

[RakshakTalwar]

@IvyGongoogle you can try the batch_represent.py script that I recently contributed.

It's under the tmp/ directory

[IvyGongoogle]

@davidsandberg @RakshakTalwar Thank you very much!

[AKSHAYUBHAT]

@RakshakTalwar @davidsandberg I am trying to do inference on set of images using your script to incorporate facenet into Deep Video Analytics. However it require pre-detected / cropped faces. I saw that there is an "align module" that uses MTCNN ported to tensorflow. Is it ready to use? And is there any test script demonstrating how to use it?
Thanks,

[RakshakTalwar]

@AKSHAYUBHAT you would use the following script to detect and align using MTCNN, then you can run it through batch-represent

https://github.com/davidsandberg/facenet/blob/master/src/align/align_dataset_mtcnn.py

[IvyGongoogle]

Now I use tf0.9, but it shows error:
Segmentation fault (core dumped)
it seems to be a hidden error.
Do I have to use tf1.0?

[IvyGongoogle]

@davidsandberg @RakshakTalwar

[RakshakTalwar]

@IvyGongoogle I was able to get it to work with tf0.12 as well, not sure if it'll work with 0.9

[AKSHAYUBHAT]

It works correctly with tf 1.0. One issue is that there is no way to set GPU memory percentage, I modified the script while incorporating. You can do something like this

        gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.15)
        sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options, log_device_placement=False))
        with sess.as_default():

[pandearti]

Hi @davidsandberg @RakshakTalwar

• which script has the infer function that will identify the class of given input image?
• I tried compare.py, it only outputs the distances between given input images. Does not identify the class (person name in this case)
• even if I use batch_generate.py, I will eventually need to call infer function for classification.

[IvyGongoogle]

@davidsandberg @RakshakTalwar I also need the script as @pandearti said above.

[davidsandberg]

I have added a script 'classifier.py' that can train a small classifier on the facenet features. There is also a wiki page showing how to use it. Hopefully that will help.

[IvyGongoogle]

@davidsandberg thank you very much!