Change targets of torchvision dataset

vision
1

Hello there.
I have some data in data folder and I read them like this
path = os.path.join(os.getcwd(),'../../kitti/dataset/sequences/',seq)
data_in_path = torchvision.datasets.ImageFolder(path)
the problem is that in my task the labels are not classification targets. each data has its own target which is a float tensor by shape (6,1).
how can I change the classification targets in the dataset i have loaded?
I also tried writing custom dataset and reading the image samples. but they were just the path of images. Is it possible for the dataloader to read images from their addresses?
Thanks

2

I think the best approach would still be to write a custom Dataset and read the images from their path by using e.g. PIL.Image.open. In fact, the ImageFolder does the same as seen here.
I don’t know how the targets are created, but I assume you could either create or load them using the same index as used for the image paths.

3

Thanks for answering Peter.
Yes. The indexes are the same.
I will try the solution you mentioned and write here about it.