WIP: TF2

Here we'll port some models to TF2 and add some training script examples

marked as a Work In Progress

global batch size is only relevant when you do multi-gpu

changed this line in version 2 of the diff

these reduction overrides should be here only when you want to average the loss using global batch size yourself which you have not done here. So please remove these reductions from the two losses as well.

Thanks man, I was looking for this (this makes the model diverges).

I'll push a fix

changed this line in version 2 of the diff

I am wondering if this is the correct package to implement functionality from deep networks. Actually, I am currently proposing a Master's project to build a generic package bob.bio.dnn that includes interfaces for several deep learning frameworks as well as pre-trained networks for feature extraction, and face detectors. @sebastien.marcel can tell you more about this. I know that some packages for feature extraction with CNNs already exist, but it would be nice to have a package that provides generic implementations for many frameworks.

Maybe that new package might be a good place the have scripts to train some networks, too.

Hi Manuel,

We have bob.bio.face_ongoing (https://gitlab.idiap.ch/bob/bob.bio.face_ongoing) and bob.bio.htface (https://gitlab.idiap.ch/bob/bob.bio.htface) that contains some scripts to train CNNs for respectively face and Heterogenous face recognition. Then we have the bob.ip.tensorflow_extractor that contains some feature extractors based on TensorFlow.

Furthermore, we have bob.learn.tensorflow (https://gitlab.idiap.ch/bob/bob.learn.tensorflow) and bob.learn.pytorch (https://gitlab.idiap.ch/bob/bob.learn.pytorch) that contains generic functionalities (some CNN backbones, specific ways to preprocess data using tensors, some specific losses,...) for respectively, TF and PyTorch.

As you can see, we have a lot of packages to maintain and a lot of stuff for a newcomer to look at (and here I'm mentioning only FR). I would like to keep all FR stuff in one place, so newcomers can i- Arrive, ii-) git clone, iii- Create a conda env and iv- work right away. This will also reduce our load of packages to maintain and would make the job a newcomer less stressful.

With this new iteration of our development, which I would be happy to brief you via Zoom if you would like, we are trying to reduce the number of packages. Hence, adding these scripts here it's just a move in this direction.

Thanks

Thanks, Tiago, I am aware of all of these packages. The idea of the proposed package is to consolidate all the efforts for face processing with DNNs in one package. For example, there are pre-trained face recognition network that are available in MxNet: https://github.com/deepinsight/insightface for which there exists no package in Bob yet -- and as you mentioned it might not be a good idea to add another package particularly for MxNet. There is a generic interface of OpenCV, which is able to use pre-trained networks from several types of frameworks (including the outdated Caffe framework, for which we still have some networks), which would be a good starting point for a generic interface.

But maybe you are right, bob.bio.face might actually be the right package to put all these things into since all of the methods are related to face recognition (at least as of now).

W.r.t. the Zoom call, I will send you an email.

For example, there are pre-trained face recognition network that are available in MxNet: https://github.com/deepinsight/insightface

We incorporated this in bob.bio.face too. It's in a branch

added 1 commit

• 196f184e - Update some problems in CenterLoss

Compare with previous version

We also have a tutorial to convert mxnext weights to tensorflow in https://gitlab.idiap.ch/biometric/mxnet-to-tensorflow but I guess you cannot see it.

To add to this conversation, I think FR is deep learning now so it doesn't make sense to have bob.bio.face and not have deep learning code in here. Although, the dependencies might become too big at some point if we start depending on several deep learning frameworks. So I think the best option is to keep these dependencies optional.

@tiago.pereira You have pushed a random empty file in here, pls remove that.

This value is based on our MS-CELEB tfrecords.

Batch size should be multiples of 8 when applying mixed_precision

pls comment why this layer is needed.

I have also been working with converting these networks, but this is cumbersome and might not work in all circumstances for all source and target frameworks and models. I was using mmdnn (https://github.com/microsoft/MMdnn) for the conversion. A generic interface like the one from OpenCV.DNN (https://github.com/opencv/opencv/wiki/Deep-Learning-in-OpenCV) might be a good starting point for any network from any framework, including ONNX.

I totally agree that we should keep all dependencies optional. AFAIR we have some infrastructure available to make tests optional, too, depending on the availability of frameworks.

I don't think it's worth it to implement a generic interface like OpenCV.DNN in Bob. Many of these frameworks will be deprecated anyway and sometimes the functionality is too different that you cannot provide a generic interface. Keras already tried to do this but gave up.

changed target branch from dask-pipelines to master

dealing with that in another MR.

closing this one

closed