[proposal] Add an optional arg for specifying classes to load in ImageFolder

[toto-maru]

🚀 Feature

Add "classes" option totorchvision.datasets.ImageFolder.__init__() method, which specify the classes I want to use under "root", not every classes (folders) under "root".

Motivation

I'd like to specify classes to use under "root" path, but current ImageFolder loads every classes (dirs) under "root" path.
In Keras' ImageDataGenerator.flow_from_directory, I can specify classes to use under "root" by passing "classes" optional argument. I'd like pytorch ImageFolder to have the same argument as Keras' ImageDataGenerator.flow_from_directory (I've opened same issue which is already closed, at wrong repo before.).

Pitch

Add a new optional argument "classes" which takes list of classes I need, on
torchvision.datasets.ImageFolder.__init__() method, and passed to super class DatasetFolder then do selective class import in DatasetFolder._find_classes() method.
If argument "classes" is not passed, ImageFolder loads all classes under "root" (same as current behavior).

[vincentqb]

I can imagine that there would be a lot of different types of exceptions to take care of. I would err on the side of not having exceptions. The folders are fairly easy to restructure so as to fit the ImageFolder format. Thoughts?

[toto-maru]

Thank you for your reply. I was thinking just add codes like below to current class(folder) scan

if `class_list_arg_is_provided` and `provided_class_list_is_valid`:
    classes = [d.name for d in os.scandir(dir) if d.is_dir() and d.name in `class_list_provided`]
else:
    classes = [d.name for d in os.scandir(dir) if d.is_dir()]

and add a simple method verifying provided_class_list is valid by checking

1. the provided_class_list is of type list or tuple.
2. each element in provided_class_list is of type str.

Am I missing something and oversimplifying?

[fmassa]

Hi,

I agree with @vincentqb . There are multiple possible ways of providing a filter function for the class names, and I'm not sure that specifying it as a list is the most generic way. One could for example pass instead a filter function, which could support the case you mentioned, but others as well (for example, filtering via a regex).

This is a common-enough request that has been discussed in #145 for example, and we addressed it by letting _find_classes be a method of FolderDataset in #527. This gives full flexibility for the user, with just a bit of extra code. For example:

class MyDataset(DatasetFolder):
    def _find_classes(self, dir):
        # apply your custom rules here

dataset = MyDataset(...)

As such, I'm closing the issue but let me know if you disagree.