The usage of a global config file is not clear

The only thing I see is:

  PATH                  A configuration file containing one or more of
                        "database", "preprocessor", "extractor", "algorithm"
                        and/or "grid" (default: None)

which is far less than what you can actually do with the config file.

For example I am writing a config file like this:

#!/usr/bin/env python

import bob.bio.face

verbose = 2
grid = 'gpu'
env = ['PATH', 'LD_LIBRARY_PATH']

database = 'replay-img-licit'


# Detects the face amd crops it without eye detection
CROPPED_IMAGE_HEIGHT = 224
CROPPED_IMAGE_WIDTH = 224
preprocessor = bob.bio.face.preprocessor.FaceDetect(
  face_cropper=bob.bio.face.preprocessor.FaceCrop(
      cropped_image_size=(CROPPED_IMAGE_HEIGHT, CROPPED_IMAGE_WIDTH)),
  use_flandmark=False
)

extractor = 'vgg_features'

algorithm = 'distance-cosine'

but writing it was not possible without reading the code.

especially the fact that you should use env to replace --environment.

This should be covered on issue #36 (closed). Why did we close that?

@amohammadi: These names correspond to the internally attribute values used by argparse. I just kept them. And yes, of course, a bit of doc would help to figure those out. Not sure how to generate a comprehensive list though. Possible?

Not sure how to generate a comprehensive list though. Possible?

What do you mean?

Is there a script/hack we can use that will dump all possible options and variable names together?

Hmm... we might write a script that fills in all possible command line options into a configuration file, and comments them out, such as:

# preprocessor = ...

# extractor = ...

# algorithm = ...

For options with a default, these default values might be written there, too:

# verbose = 0

# grid = None

# groups = ['dev']

Writing such a script should not be that difficult, however, providing documentation for each entry might be more cumbersome, and getting them in a reasonable order might be difficult, too.

created branch 45-the-usage-of-a-global-config-file-is-not-clear

mentioned in merge request !65 (merged)

Another issue that I see with the global configuration file is that variables that do not correspond to any option are silently ignored.

This might be good in the example of Amir above, i.e., where CROPPED_IMAGE_HEIGHT and CROPPED_IMAGE_WIDTH are defined globally and used inside a preprocessor configuration.

On the other hand, if options are misspelled (e.g., using verbosity instead of verbose), these options are just ignored without raising any warning. Maybe, we can come up with a test that checks if every variable inside the configuration file was actually used as a command line option, and emit a warning if not. We might skip all-uppercase variable names, in this case.

I have a brief idea on how to implement that:

1. When parsing over the command line options using: https://gitlab.idiap.ch/bob/bob.bio.base/blob/master/bob/bio/base/tools/command_line.py#L156 we can simply set all variable inside config to None (after reading them out, of course)
2. Any variable in config that is not None and that is not all uppercase will be reported as unknown (using logger.warn, for example)

Do you think this is worth investigating?

Too late, it is implemented, see !65 (merged) ;-)

Do you think this is worth investigating?

Extremely important! Thank you for fixing this.

closed via commit 0c672d5e

mentioned in commit 0c672d5e

mentioned in issue #65 (closed)