Batch-wise processing or image-by-image processing? (DINO V1)

Question

I've been trying to recreate the Dino V1 traning set up for a personal project. For which I've take majority of the code from this repo: https://github.com/facebookresearch/dino[dinov1 link]1

And rn I'm almost done with it except for one part in the main_dino.py file there is a function called train_one_epoch whereby in line 318 they have given:

teacher_output= teacher (images[:2]) # only the 2 global views pass through the teacher

Now I know how pytorch tensor indexing/slicing works. Hence, if images are a batch of images of a structure:

(batch size, num crops, c, h, w)

1. How would doing images[:2] get you the global crops of all the images in a given batch?
2. Are they processing images in batch here or is the "images" list here just a list containing multiple crops from a SINGLE input image?

Answer 1

Prior to the call of train_one_epoch() there was another modifiaction to the models, both the student and teacher models are wrapped with MultiCropWrapper class. Just take a look at the class' docstring as follows:

class MultiCropWrapper(nn.Module):
    """
    Perform forward pass separately on each resolution input.
    The inputs corresponding to a single resolution are clubbed and single
    forward is run on the same resolution inputs. Hence we do several
    forward passes = number of different resolutions used. We then
    concatenate all the output features and run the head forward on these
    concatenated features.
    """

So this MultiCropWrapper class handles the forward passes, and it is also mentioned that it does several forward passes for different resolutions.