Iterating through ImageFolder for sample, target

I am trying to use the ImageFolder class to read a bunch of images which are arranged this way:




Then I do something lilke:

ds = torchvision.datasets.ImageFolder( root=path_to, transform=p )

I can see the list of classes through:

list_of_classes=list(map(int, list(ds.classes)) )

Now reading through the docs, I iterate through the sample and targets lke so:

for idx, (sample, target) in enumerate(ds):
    print(sample, target)

For some reason, I notice that the target is showing the indices of ds.classes rather than the target value itself.

Is this the intended behaviour? I was expecting it to simply show me the classes (0,1,2) in this case.

Assuming dataset.classes returns the folder names, the corresponding indices should be the class indices (i.e. the target values), shouldn’t it?

@ptrblck: thank you for your reply but I am not sure I understand what you have said :frowning:

So, dataset.classes does return folder names, but when I loop through ds , the target is showing me indices of dataset.classes rather than the folder name :frowning:

To the NN I would pass the sample and the target(which here is the folder name), not the indices, right?

I am confused as to why this is… I obviously do not understand something :frowning:

So, when I do:

for idx, (sample, target) in enumerate(train_dataset):
    print(sample, list_of_classes[target] )

it seems to pick up the correct labels(aka target/folder_names). Is this the correct way to do it?

That is expected and the target should contain the class indices, not the names.
So e.g. for three folders your target should contain values in [0, 1, 2] :wink:
These target values are used to index the output of the model during the loss calculation, so they are used as numerical values instead of the folder names.

@ptrblck: thank you for this clarification. I think I got confused because my folder_names were also integers. It does make sense, since we compute the probability of classes at the end and thus indexing makes sense.

Thank you again. This was VERY helpful.