A sequential and parallel preprocessor

It would be nice to have a helper class that takes several pre-processors and applies them to the data one by one sequentially.

Recently, I have been working with madmom and I really like the idea of functions/classes API they provide + their processors API. We could have something similar throughout Bob generally I think.

mentioned in issue bob.learn.tensorflow#36 (closed)

Indeed, this would be great. In bob.bio.face, I currently have an implementation of specific preprocessor chains, e.g., face alignment and tan-triggs. Currently this is kind-of a mess, I define a preprocessor to be used as face detector inside the TanTriggs class: https://gitlab.idiap.ch/bob/bob.bio.face/blob/master/bob/bio/face/preprocessor/TanTriggs.py#L48

This gets particularly ugly, when you want to use this in the bob.bio.video preprocessor, where we get another level of indirection: https://gitlab.idiap.ch/bob/bob.bio.video/blob/master/bob/bio/video/preprocessor/Wrapper.py#L53 If you want to perform face detection with tan-triggs, the configuration would look like:

preprocessor = bob.bio.video.preprocessor.Wrapper(
  preprocessor = bob.bio.face.preprocessor.TanTriggs(
    face_cropper = bob.bio.face.preprocessor.FaceDetect(
      ...
    )
    ...
  )
  ...
)

While the first level of indirection is OK, the second level should be avoided at all costs, but this is what is currently implemented.

Hmm... thinking of it, even your approach would have a three-level indirection, using a (not yet implemented bob.bio.base.preprocessor.MultiPreprocessor):

preprocessor = bob.bio.video.preprocessor.Wrapper(
  preprocessor = bob.bio.base.preprocessor.MultiPreprocessor(
    preprocessors = [
      bob.bio.face.preprocessor.FaceDetect(
        ...
      ),
      bob.bio.face.preprocessor.TanTriggs(
        ...
      )
    ],
    ...
  )
  ...
)

Yes it can come handy in a lot of cases. I am thinking of these two preprocessors:

https://madmom.readthedocs.io/en/latest/modules/processors.html#madmom.processors.SequentialProcessor and https://madmom.readthedocs.io/en/latest/modules/processors.html#madmom.processors.ParallelProcessor

I can see the sequential preprocessor, but not the parallel one. Parallelization is done though gridtk and, therefore, I don't see a need for a different concept of parallelization.

a parallel processor is different in a sense that instead of passing the data into its processor one by one, it passes all at once and then outputs a list of them. For example:

>>> import numpy as np
>>> mean = np.mean
>>> multiply_by_2 = lambda x: x*2
>>> seq_proc = SeqProc([multiply_by_2, mean])
>>> parallel_proc = ParallelProc([multiply_by_2, mean])

>>> data = np.array([2, 4])
>>> seq_proc(data)
6
>>> parallel_proc(data)
[[4, 8], 3]

I see. What you are saying is that we could have a set of preprocessors for one image/original data sample. While this sounds reasonable, so far we have no extractor that would be able to integrate multiple preprocessed data samples into one feature vector. Hence, at the moment I don't see the need of a parallel processor, while it might be interesting in future.

mentioned in merge request !102 (merged)

closed via commit 8d7a645c

closed via merge request !102 (merged)

mentioned in commit 8d7a645c