A lot of new features

• 905ffc94 small changes: nit, bug fix, small features
• 214fad7a Add scripts for training keras models using keras API
• 13ae3919 Add two functions to allow dynamic weighting of samples per batch
• c77e386e Add several extractors which are usefull at inference time
• 0b0055bf Add a generic Generator class alternative to BioGenerator
• 18e34e61 Add virtual adversarial training loss
• fcc10424 Fix the euclidean function so that its gradients don't become nan. Also moves the bytes_to_human function
• 5ad23cfd Add a hook to add some tensors to the summary
• 9f3b752b Add bob tf predict command
• 09edcdf7 Fix the image augmentation function
• 9921e122 Add dataset_to_tfrecord and dataset_from_tfrecord
• c3c9e9a1 Make bob tf cache command useful
• 316eca76 Add a new MLP architecture
• 6809b360 Fixes in inception architectures
• 0ab7c214 setup.py required changes for multiple features

Sorry for not merging these changes earlier The best I can do is to separate them into logical commits I think it would be easier to review each commit separately.

@pkorshunov if you could review this too that would be great.

I skipped my changes to estimators since they are not ready.

I can take care of them by next week. Tomorrow I will probably spend the day running the FSPR grading

After scrolling through the changes, everything is looking good, as far as I could see. But I will try to run my stuff with these changes to battle-test them.

• Am I correct to say that in tfrecords.py you don't have fixed_batch_size option with `batch_and_drop_remainder' applied to the database? Just to make sure.
• I noticed you have some training with adversarial learning. Do you have some papers on this?

I noticed you have some training with adversarial learning. Do you have some papers on this?

See here: https://github.com/takerum/vat_tf

Am I correct to say that in tfrecords.py you don't have fixed_batch_size option with `batch_and_drop_remainder' applied to the database? Just to make sure.

Can you point me to the code? I don't understand.

Can you point me to the code? I don't understand.

Something like that: https://gitlab.idiap.ch/bob/bob.learn.tensorflow/blob/predict/bob/learn/tensorflow/dataset/tfrecords.py#L254

As discussed offline, I don't use these functions. The only thing I did was to fix the augmentation code.

Yes, cool, I will run the code tomorrow and will let you know

I'm starting to look into it

I'm good with this one, thanks

I can't see tests for this one? Does make sense to have it?

Btw, are you using this in your work @amohammadi ?

no I am not using this. I just added it since I wrote it at some point. I think it's good to keep untested. It will be useful for TF 2 migration.

There are no tests for this one and for this we should do it?

Is there any reference in the literature for this @amohammadi ?

Thanks

I learned this from @ageorge I don't know if he found it from somewhere.

There are tests but they are doctests as you can see above.

With this commit we should remove some code from bob.ip.tensorflow_extractor to avoid code repetition.

so you are going to make bob.ip... depend on bob.learn...?

No corresponding tests for this one

No test cases for this one. Did the guy that you get the code provide some?

thanks

I'm good with this one.

thanks