Test Facenet

@lcolbois but tests are passing.

Looks like the Tensorflow-related tests are not executed :

 bob.bio.face.test.test_embeddings.test_facenet_sanderberg ... SKIP: Skipping test since `tensorflow` is not available: No module named 'tensorflow.contrib'
bob.bio.face.test.test_embeddings.test_idiap_inceptionv1_casia ... SKIP: Skipping test since `tensorflow` is not available: No module named 'tensorflow.contrib'
bob.bio.face.test.test_embeddings.test_idiap_inceptionv1_msceleb ... SKIP: Skipping test since `tensorflow` is not available: No module named 'tensorflow.contrib'
bob.bio.face.test.test_embeddings.test_idiap_inceptionv2_casia ... SKIP: Skipping test since `tensorflow` is not available: cannot import name 'InceptionResnetv2_Casia_CenterLoss_2018' from 'bob.bio.face.embeddings' (/scratch/builds/bob/bob.bio.face/miniconda/conda-bld/bob.bio.face_1607000812207/_test_env_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_placehold_pla/lib/python3.7/site-packages/bob/bio/face/embeddings/__init__.py)
bob.bio.face.test.test_embeddings.test_idiap_inceptionv2_msceleb ... SKIP: Skipping test since `tensorflow` is not available: No module named 'tensorflow.contrib'

we need to remove these tensorflow.contrib imports.

Thanks, Laurent.

Weights are exactly the same from TF2 imported model and TF1. I dumped both and compared them one by one

/[MY_USER]/gitlab/bob/bob.nightlies/TF1.hdf5
/[MY_USER]/tpereira/gitlab/bob/bob.nightlies/TF2.hdf5

/[MY_USER]/gitlab/bob/bob.nightlies/compare_tf1_tf2.py

Normalization is also the sample.

I don't know what can be wrong

Any insights?

added 2 commits

62986b36 - Added reference embeddings for all TF networks
6cca593f - Added TF1/TF2 embedding comparison tests for all Inception networks. Fixed...

Compare with previous version

Hi, I added an embedding comparison tests for all network. Running the tests shows that Inception-v2 models works correctly, however Inception-v1 models have an embedding discrepancy w.r.t the TF1 version.

If the weights are the same, and the input normalization is the same that would suggest there is a difference in the architecture itself (i.e. the operation). I am not super familiar with the details of the InceptionResnet architecture, but if I brainstorm possible sources of error I can think of that would not impact the weights:

Wrong non-linear activation somewhere
Wrong pooling operation somewhere (e.g. mean pooling instead of max pooling)
Wrong handling of dropout / batch normalization layers at inference time

Wrong non-linear activation somewhere

Wrong pooling operation somewhere (e.g. mean pooling instead of max pooling)

Wrong handling of dropout / batch normalization layers at inference time

That's what I'm inspecting. So far everything matches. This one will be tough

PS: Dropout is not used at inference time. so this we can discard

PS: Dropout is not used at inference time. so this we can discard

Yes, I was wondering if by any chance it wrongly remained activated at inference time. But I am pretty sure this should not be the case as we are using Keras' predict_on_batch method

It's probably an activation or some residual scaling ....

If that can be helpful : looks like the intermediate outputs are the same in both models up to layer 23 ('Mixed_7a') of the Keras implementation (including the output of this layer).

After that the models diverge

Found a discrepancy ! In the TF1 implementation:

                # 5 x Inception-Resnet-C
                name = "block8_BN"
                trainable = is_trainable(name, trainable_variables, mode=mode)
                with slim.arg_scope(
                    [slim.batch_norm],
                        is_training=(mode == tf.estimator.ModeKeys.TRAIN),
                        trainable=trainable):

                    name = "block8"
                    trainable = is_trainable(name, trainable_variables, mode=mode)
                    net = slim.repeat(
                        net,
                        5,
                        block8,
                        scale=0.20,
                        trainable_variables=trainable)
                    end_points[name] = net

there are 5 block8 with scale 0.2.

In the TF2 implementation:

    # 5x block8 (Inception-ResNet-C block): 8 x 8 x 2080
    for block_idx in range(1, 5):
        layers.append(
            InceptionResnetBlock(
                n_channels=1792,
                scale=0.2,
                block_type="block8",
                block_idx=block_idx,
                name=f"block8_{block_idx}",
            )
        )

    layers.append(
        InceptionResnetBlock(
            n_channels=1792,
            scale=1.0,
            activation=None,
            block_type="block8",
            block_idx=5,
            name="block8_5",
        )
    )

there are 4 block8 with scale 0.2 and one with scale 1.0

Edit : Well it looks weirder than that actually, looks like one block8 is missing from the TF2 implementation. But we would see that is the number of weights no ?

I love you man.

That's it.

I'll push a fix and correct the checkpoints

There you go.

Facenet passes now with your test @lcolbois

Tests are passing with all InceptionResnetv1 models :-)

merged

mentioned in commit c566524f

but the tests are still skipped, what is the issue?

bob.bio.face.test.test_embeddings.test_facenet_sanderberg ... bob.pipelines.wrappers@2020-12-05 20:00:39,759 -- DEBUG: Sample|FaceNetSanderberg_20170.transform
 ok
 bob.bio.face.test.test_embeddings.test_idiap_inceptionv1_casia ... bob.pipelines.wrappers@2020-12-05 20:01:59,892 -- DEBUG: Sample|InceptionResnetv1_Casia.transform
 ok
 bob.bio.face.test.test_embeddings.test_idiap_inceptionv1_msceleb ... bob.pipelines.wrappers@2020-12-05 20:03:12,824 -- DEBUG: Sample|InceptionResnetv1_MsCel.transform
 ok
 bob.bio.face.test.test_embeddings.test_idiap_inceptionv2_casia ... bob.pipelines.wrappers@2020-12-05 20:05:50,973 -- DEBUG: Sample|InceptionResnetv2_Casia.transform
 ok
 bob.bio.face.test.test_embeddings.test_idiap_inceptionv2_msceleb ... bob.pipelines.wrappers@2020-12-05 20:08:47,854 -- DEBUG: Sample|InceptionResnetv2_MsCel.transform
 ok
 bob.bio.face.test.test_embeddings.test_idiap_inceptionv2_msceleb_memory_demanding ... bob.pipelines.wrappers@2020-12-05 20:11:51,421 -- DEBUG: Sample|InceptionResnetv2_MsCel.transform
 ok

Solved with !81 (merged)

Test Facenet

Merge request reports

Activity